Differential Diagnosis of Cognitive Disorders using Deep Learning Techniques based on Neuroimaging Data

Differential Diagnosis of Cognitive Disorders using Deep Learning Techniques based on Neuroimaging Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Differential Diagnosis of Cognitive Disorders using Deep Learning Techniques based on Neuroimaging Data Mehran Panahi, Ali Asghar Safaei, Foad Ghaderi, Fatemeh Panahi This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7998240/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background The worldwide increase in cognitive diseases has generated a greater demand for precise, non-invasive diagnostic tools with the ability of differentiating neurodegenerative diseases with similar clinical symptoms and subtle differences in brain anatomy. Alzheimer’s disease (AD), frontotemporal dementia (FTD), and Parkinson’s disease (PD) are typified by cognitive impairment, hence differential diagnosis becomes crucial. Methods This work utilizes three-dimensional convolutional neural networks (3D-CNNs) on 1,629 T1-weighted structural MRI images gathered from well-established datasets (ADNI, NIFD, and PPMI). Preprocessing involved format conversion, image registration, brain extraction, image cropping, volumetric downsampling, and intensity normalization. We built binary classification models for each disorder against healthy controls, and then a combined multiclass model for simultaneous discrimination between AD, FTD, PD, and controls. In order to prevent data leakage, subject-level data partitioning was performed instead of image-based splitting. Training was done without synthetic augmentation. With the application of Focal Loss and Maxout layers—innovations not utilized in previous research covered here—our strict preprocessing and architecture outperform many of the existing solutions in sensitivity and diagnostic accuracy. Results The binary models performed with very good accuracies: 98% (AD), 98% (FTD), and 93% (PD). The merged multiclass model achieved an overall accuracy of 95%, class-specific sensitivities of 100% (AD), 95% (FTD), and 95% (PD), and an AUC close to 1.00. Conclusions Our model’s strict preprocessing methods and novel architecture surpass numerous current methods, demonstrating the promise of 3D-CNN architectures as powerful clinical decision support systems for differential diagnosis of cognitive disorders. Artificial Intelligence and Machine Learning Deep Learning Alzheimer’s Disease Frontotemporal Dementia Parkinson’s Disease Magnetic Resonance Imaging (MRI) 3D Convolutional Neural Network Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 1. Introduction Cognitive impairment is one of the leading causes of disability in late life and is featured in the top most common syndromes of this phase of life. The global population affected by dementia is increasing rapidly, with projections estimating it will reach around 135 million by 2050 [1]. The World Health Organization (WHO) reports that 57 million people had dementia in 2021 and nearly 10 million new cases are found every year. Dementia is the seventh greatest cause of death worldwide and the greatest single cause of disability and dependency in older adults. It is a nonspecific term for a decline in mental capacities, including memory, cognition, attention, executive function (such as problem-solving and decision-making), language, and visuospatial perception. The most important subtypes of dementia include Alzheimer’s disease, vascular dementia, dementia with Lewy bodies, and frontotemporal dementia [2]. Additionally, patients with Parkinson’s disease are at an increased risk of developing dementia as the disease progresses [3]. This paper aims to develop deep learning-based models to diagnose cognitive and neurodegenerative diseases such as Alzheimer’s disease, frontotemporal dementia, and Parkinson’s disease using available MRI data. All the three diseases have been discussed in the following paragraphs. Alzheimer’s disease (AD) is the most prevalent form of dementia globally, accounting for approximately 60–70% of all dementia cases [4]. It is a progressive neurodegenerative disorder with no known treatment yet [5]. Neuroimaging results show that cerebral atrophy, decrease in hippocampal volume, ventriculomegaly, and cortical thinning are structural brain changes which can be detected years prior to clinical symptom onset and follow a temporal profile over the course of the illness [6]. The hippocampus is one of the earliest regions affected in AD, largely due to its critical role in memory and learning, making it particularly vulnerable to AD-related damage. Hippocampal atrophy will probably be the most crucial neuroimaging biomarker of the disease [7]. At the early period of AD, hippocampal volume reduction can be up to 20% versus controls; later it may exceed 30% [8]. AD will cause atrophy in certain areas of the brain, i.e., the temporal, parietal, and frontal lobes. Certain effects of this atrophy are dilatation of the ventricles of the brain [5]. It also features cortical thinning in AD, typically involving distant areas of the brain such as the entorhinal cortex and medial temporal lobe and progressing as the disease advances [9]. The structural changes in Alzheimer’s disease, as observed through MRI, are illustrated in Figure 1. Frontotemporal dementia (FTD) is characterized by the progressive degeneration of neurons in the frontal and temporal lobes of the cerebral cortex [11]. It is recognized as the most common cause of early-onset dementia, typically affecting individuals under the age of 65 [12]. To date, no disease-modifying treatment has been approved for FTD [11], making timely and accurate diagnosis of paramount clinical importance. In contrast to Alzheimer’s disease, which initially affects memory, the early manifestations of FTD primarily involve deficits in social interaction, language, and behavioral regulation [11]. Delayed or incorrect diagnosis often results in inappropriate therapeutic interventions and unfavorable outcomes [13]. While memory impairment is a hallmark early symptom of AD—typically resulting from early damage to the hippocampus [7]—this feature is usually absent in the initial stages of FTD. Nevertheless, FTD is frequently misdiagnosed as AD [14], due to overlapping clinical features such as cognitive decline, cerebral atrophy, and gradual deterioration in language, behavior, and personality [11–12]. Structural MRI scans commonly reveal atrophy in the anterior regions of the frontal and temporal lobes, a neuroanatomical pattern that aids in distinguishing FTD from AD [15]. Figure 2 illustrates the structural brain changes in MRI associated with FTD. Parkinson’s disease (PD) is a progressive neurodegenerative disorder that primarily affects the motor system. Its core clinical features include resting tremor, muscular rigidity, bradykinesia, and postural instability. As the disease progresses, non-motor symptoms such as cognitive impairment, sleep disturbances, and depression also become increasingly common among patients [17]. PD typically begins after the age of 60 [18], with a reported prevalence of approximately 0.5% to 1% in individuals aged 65 to 69, increasing to around 3% in older age groups [17]. Similar to other neurodegenerative disorders, structural brain changes in PD can be observed on MRI, particularly in the substantia nigra—a brain region critical for dopamine production. Degeneration in this area leads to decreased dopamine levels, which underlies many of the motor symptoms of the disease. The pattern of brain atrophy in PD is often diffuse and widespread [19]. Figure 3 present the distribution of brain atrophy in Parkinson’s disease. Up to 78% of individuals with PD may develop Parkinson’s disease dementia (PDD) during the course of the illness. It is hypothesized that the accumulation of alpha-synuclein protein contributes to a shared pathological basis between PD and dementia with Lewy bodies [3]. Speech-related abnormalities—such as reduced vocal intensity and monotonic speech—as well as executive dysfunction, often emerge in advanced stages. These symptoms are also observed in other forms of dementia, reinforcing the need for early cognitive monitoring and the use of differential diagnostic tools [18]. Each of these disorders requires distinct pharmacological treatments, rehabilitative approaches, and care strategies. However, the considerable overlap in clinical symptoms among them significantly complicates the diagnostic process. In some cases, this can lead to misdiagnosis, resulting not only in substantial costs for patients and healthcare systems but also delays in appropriate treatment or the administration of unsuitable interventions [20, 21]. Therefore, early, accurate, and differential diagnosis is of critical importance in the management of neurodegenerative disorders. Neuroimaging is generally considered to be a key tool in the diagnosis of cognitive disorders because it reveals structural and metabolic changes in the brain. Structural signs such as brain atrophy can be observed in these scans. However, some studies have shown that deep learning-based models are much more accurate in diagnosing these disorders through neuroimaging than human experts [22]. This may be because some patterns related to cognitive disorders are so subtle that they may be beyond the ability of human vision to perceive. Recent advances in artificial intelligence, especially deep learning, offer promising solutions to this challenge. Deep learning, especially in the field of computer vision, has the potential to recognize subtle patterns in neuroimaging and provide more accurate diagnoses and prognoses. The successes achieved in this field, and in particular models based on convolutional networks in the diagnosis of various diseases, indicate the high potential of this technology in improving diagnostic processes. These models have shown the ability to identify various diseases with high accuracy by using neuroimaging data such as MRI [23-27] The main objective of the present study was to develop and evaluate 3D convolutional neural network models based on MRI for accurate and differential diagnosis of cognitive impairments associated with Alzheimer’s disease, frontotemporal dementia, and Parkinson’s disease. In the first step, three binary classification models were designed to distinguish each disorder from a normal control group. In the second step, a unified multiclass model was developed for differential diagnosis among the three conditions. Overall, our goal was to minimize diagnostic errors and achieve a high level of accuracy and reliability to enable these models to serve as decision support tools in clinical settings. To do this, we used the entire brain volume in the form of 3D models instead of using 2D slices, which had their own preprocessing and computational challenges. To overcome the challenge of data leakage, the models were trained in a subject-based manner. In the differential diagnosis model of disorders, we in addition used maxout layers to achieve greater accuracy, which was not used in previous studies we reviewed. On the other hand, in the differential diagnosis model, to overcome the challenge of data imbalance, we used the focal loss function instead of artificial data augmentation, which was also not used in previous studies we reviewed. In the end, our aim was to develop models that provide diagnoses with high accuracy, sensitivity, and as well as high reliability as decision support tools. The rest of the paper is organized as follows: Relevant studies in the classification of cognitive disorders are introduced in the “Related Works” section. The “Methods and Materials” section describes the dataset, preprocessing techniques and model training. The “Results” section is dedicated to the research results based on the evaluation criteria. The “Discussion” section analyzes the research and evaluates the quality of the results. Finally, the “Conclusions and Future Work” section concludes the paper and suggests directions for future research. 2. Related Works Several previous studies have significantly contributed to the foundation of the present research, particularly in terms of methodological inspiration and focus on cognitive disorders that are similar to the conditions addressed in this study. These works collectively highlight the impressive accuracy of machine learning and deep learning models in diagnosing cognitive impairments using neuroimaging data. Table 1 provides an overview of the key investigations reviewed in this section. Etminani et al. demonstrated, in a study of 3D-CNN models, that these models possess an excellent capacity for diagnosing dementia with Lewy bodies (DLB), AD, and mild cognitive impairment (MCI) based on 18F-FDG PET data. The study included 757 cases. The suggested model attained an area under the ROC curve (AUC) of approximately 96% for DLB, 96.4% for AD, 71.4% for MCI, and 94.7% for the cognitively normal (CN), surpassing human experts according to several evaluation metrics [22]. Ramazan et al. also tried to propose a model that was able to detect AD and its prodromal phases. They employed a dataset of 138 f-MRI-based cases. Three ResNet-18-based convolutional models trained on a dataset of 138 subjects were applied in the study and attained an overall classification accuracy of over 97% [28]. In a similar study, Ahila et al. evaluated a deep learning model for AD diagnosis using 18F-FDG PET data. The dataset consisted of 635 CN and 220 AD, and the suggested model reached approximately 97% accuracy [29]. Akindele et al. also developed a hybrid deep learning framework for AD classification, consisting of 2D and 3D CNNs trained on MRI data. The model achieved 98.9%–99.99% accuracy with AUC of 100% for both settings. The dataset consisted of 2D and 3D MRI scans collected from 46 AD and 23 CN subjects. Multiple MRI scans were acquired for each participant, totaling 708 images from both CN and AD [30]. Ma et al. focused on an approach to discriminate between FTD, AD, and CN cases using deep learning and generative adversarial networks (GANs) with structural MRI data. In this study, MRI scans were divided into 87 regions of interest (ROIs) and divided into multiscale patches by k-means clustering. From each patch, statistical features such as cortical thickness, volume, and intensity were extracted. Artificial features were generated through GANs to address data scarcity. A multilayer perceptron (MLP) was then trained on both real and artificial features to classify individuals as AD, FTD, or CN. This framework achieved an overall accuracy of 88.28% on 1954 MRI images [31]. In a similar study for the classification of AD, FTD, and CN based on 18F-FDG PET scans, Rogeau et al. proposed a 3D CNN. The dataset consisted of 199 AD patients, 192 FTD patients, and 200 CN cases. Their proposed model achieved an accuracy of 89.8% and AUCs were 93.3% for AD, 95.3% for FTD, and 99.9% for CN. AUC besides was 93.3% for AD, 95.3% for FTD, and 99.9% for the CN [32]. For the diagnosis of PD, Esmailzadeh et al. applied a 3D CNN to a brain MRI dataset consisted of 3D scans including 452 PD patients and 204 CN individuals. Notably, the designed model yielded 100% classification accuracy in distinguishing PD from CN [33]. S. Chakraborty et al. also designed a 3D CNN model in the same type of research for diagnosing PD vs. CN from MRI data for 406 subjects with 203 PD patients and 203 CN subjects. The suggested model achieved 95.29% accuracy, average sensitivity of 0.943, and AUC score of 0.98 for both classes [34]. Dyrba et al. also trained another 3D CNN on 663 MRI scans to discriminate AD and MCI. The model, which was validated on three independent cohorts, had high accuracy in discriminating AD from CN (AUC ≥ 0.91) and moderate accuracy in discriminating MCI from CN (AUC ≈ 0.74). The network achieved an accuracy of 88.3% (confidence interval) in discriminating AD from CN cases [35]. Z. Hu et al. developed an MRI-based model combining 3D CNN and Attention mechanisms to classify AD, MCI, and CN. A total of 520 images were used: 130 AD, 260 MCI, and 130 CN. Four classification experiments were performed based on the suggested hybrid model: (1) AD vs. CN, (2) MCI vs. CN, (3) AD vs. MCI, and (4) multi-class classification of AD, MCI, and CN. The model achieved binary classification accuracies of 94.23%, 95.67%, and 97.75%, respectively, and for the multi-class task accuracy of 89.71% [36]. S. De Francesco et al. prioritized demographic, clinical, and MRI information to construct a model for the differential diagnosis of AD, FTD, DLB, and CN. The study involved 506 subjects: 110 AD, 135 FTD, 153 DLB, and 108 CN cases. MRI scans were analyzed to obtain volumetric and cortical thickness measurements and white matter lesion metrics. These imaging characteristics, along with demographic and clinical information, were fed into a support vector machine model named MUQUBIA for differential diagnosis. The most informative of all features were age, gender, dementia staging scale, and 19 neuroimaging parameters. This model yielded an overall AUC of 98% with 87.5% accuracy, 88% sensitivity, and 88% F1 score [37]. El-Assy et al. Lastly, utilizing two interconnected yet separate convolutional models proposed an architecture that achieved over 99% accuracy in distinguishing Alzheimer’s disease from its various prodromal stages. The dataset comprised 1,296 MRI scans categorized into five groups: healthy controls, early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), AD, and general MCI cases. The proposed dual-model architecture demonstrated exceptional performance in multi-stage classification of cognitive decline [38]. In Table 1 , the accuracy criteria for the models used in related works can be seen along with the type of disorders under investigation, the modality used and the number of subjects were summarized. Table 1 . overview of the key investigations reviewed in this paper. Study Accuracy (%) Disorders Modality Subjects # K. Etminani et al. [ 22 ] 78.9 (DLB, AD, MCI, CN) PET 757 F. Ramzan et al. [ 28 ] 97.92 (CN, SMC, EMCI, MCI, LMCI, AD) fMRI 138 A. Ahila et al. [ 29 ] 96.8 (AD, CN) PET 855 R. G. Akindele et al. [ 30 ] 99.9 (AD, CN) MRI 69 D. Ma et al. [ 31 ] 88.28 (AD, FTD, CN) MRI 1954 A. Rogeau et al. [ 32 ] 89.8 (AD, FTD, CN) PET 496 S. Esmaeilzadeh et al. [ 33 ] 100 (PD, CN) MRI 656 S. Chakraborty et al. [ 34 ] 95.29 (PD, CN) MRI 406 M. Dyrba et al. [ 35 ] AD vs CN: 75.5–88.3 MCI vs CN: 63.1–75.4 (AD, MCI, CN) MRI 663 Z. Hu et al. [ 36 ] AD vs NC: 94.23 MCI vs NC: 95.67 AD vs MCI: 97.75 AD vs MCI vs NC: 89.71 (AD, MCI, CN) MRI 520 S. De Francesco et al. [ 37 ] 87.5 (AD, FTD, DLB, CN) MRI 506 A. M. El-Assy et al. [ 38 ] 99.30 (CN, EMCI, LMCI, MCI, AD) MRI 1296 AD, Alzheimer’s disease; FTD, Frontotemporal dementia; DLB, dementia with Lewy bodies; PD, Parkinson’s disease; MCI, mild cognitive impairment; CN, cognitively normal; EMCI, early mild cognitive impairment; LMCI, late mild cognitive impairment 3. Materials and Methods 3.1. Dataset In the current study, a dataset of subject-based T1-weighted MRI images was utilized for diagnosing AD, FTD, and PD from the CN. This allows for both differentiation of each disorder from the control group and differential diagnosis between the disorders. The sample size available in the dataset here employed is 1,629 images of the classes, of which 20% were used for testing and 80% for training. The distribution of data for all the classes is presented in Figure 4 . The CN groups are shown in blue in the figure, while yellow indicates the disorders. The data utilized were from trusted databases within the neuroimaging discipline. These include the ADNI 1 , NIFD 2 , and PPMI 3 databases, which are among the largest and most reliable datasets for AD, FTD, and PD, respectively. 3.2. Data Preprocessing For preparing brain imaging data for deep learning models, specifically convolutional neural networks (CNNs), the pipeline started by converting medical imaging files from DICOM to NIfTI format with the help of mricron software. As the NIfTI file format is compact and simplified, it is more suitable for handling large-scale MRI datasets and hence efficient storage and processing. The second step was image registration, where all MRI scans are spatially registered to a common coordinate frame. This alignment ensures that corresponding anatomical points across different subjects match in spatial coordinates, which is essential for conducting comparative and group-level analyses. In this study, registration was performed using the FSL software suite by aligning each image to the MNI152-T1-1mm.nii.gz template—a widely adopted standardized brain atlas generated by averaging T1-weighted scans from 152 healthy individuals [39]. By mapping all patient images into this common space, variations in anatomical scale, orientation, and positioning across individuals are minimized, thereby improving the reliability of voxel-wise analysis and model training. This standardization is particularly important in neuroimaging studies where structural variability may obscure disease-specific features. The use of robust registration pipelines, as surveyed in [40], enhances the reproducibility and interpretability of deep learning models by reducing non-pathological variance. Next, brain extraction was performed using the Brain Extraction Tool (BET) in FSL [41]. This step removes non-brain elements such as the skull, scalp, and surrounding tissues, isolating only the brain region. Brain extraction is particularly crucial for training CNNs for several reasons. First, eliminating extraneous tissue reduces background noise, resulting in cleaner input data. Second, it sharpens the model’s focus on relevant neurological features while filtering out distractions. Third, simplifying the input structure improves model efficiency and training speed. Fourth, reducing irrelevant variability helps minimize overfitting and enhances the model’s ability to generalize. Finally, removing non-brain data decreases the overall computational load, optimizing performance on large datasets. Figure 5 provides a visual example of the brain extraction process. The subsequent step in the preprocessing pipeline involved image cropping , a crucial operation aimed at reducing data volume and eliminating redundant regions beyond the brain. By trimming empty peripheral space surrounding the brain, this step ensures that the input data is spatially focused on anatomically relevant structures. Cropping not only minimizes file size and memory usage but also accelerates processing and model training. More significantly, it enhances model performance by removing background noise and irrelevant data from non-brain tissues. The cropping was followed by down-sampling to decrease the spatial resolution of the MRI scans further. This was a necessary step for computation and model optimization. All images were resampled into a 1.5 mm voxel in the three spatial dimensions (x, y, z). This resolution was chosen for two main reasons: first, increasing voxel size reduces the total number of voxels and thus data volume; second, it maintains reasonable image quality, balancing compression with retention of important neuroanatomical details. Following resampling, each scan was resized to (101, 103, 127) voxels to ensure dataset uniformity while preserving sufficient structural integrity for analysis. Intensity normalization was the last part of the MRI preprocessing procedure, an important process involving the minimization of intensity scale variation across images. Variation often occurs as a result of differing scanner settings, acquisition parameters, or physiological differences between subjects. To counteract this, voxel intensities were renormalized to a common range via Min-Max normalization, in which the minimum and maximum values of each image are mapped to 0 and 1, respectively. Intensity normalization, a widespread practice in neuroimaging research [42], decreases the model’s sensitivity to irrelevant intensity differences, increases class separability, and facilitates generalization to new unseen data. 3.3. Model Training This section outlines the training procedures for the models developed to detect AD, FTD, and PD. The training process begins with the development of binary classification models , each designed to distinguish a specific disorder from the cognitively normal (CN) group. Following that, a multiclass classification model is introduced. This comprehensive model aims to perform differential diagnosis across all three conditions, enabling the classification of subjects into one of the neurodegenerative categories based on MRI data. 3.3.1. Binary Models The AD detection model takes volumetric MRI data as input and employs three-dimensional convolutional layers to learn spatial features in all directions, allowing it to capture subtle, disease-specific structural patterns in the brain. Figure 6 shows the model architecture overview, along with the output shape and number of parameters of each layer. To accompany this, Figure 7 displays the hierarchical organization of the network, depicting the step-by-step development of information from raw input images to ultimate classification. The model’s input is a 3D MRI scan of dimensions (103, 127, 101, 1), where the final entry represents a single grayscale channel. The initial convolutional layer consists of 32 filters of size 3×3×3, activated by the ReLU function. This setup enables the model to learn simple features like edges and texture patterns, without keeping the number of trainable parameters high at this point. The ensuing feature maps are batch normalized to stabilize learning and speed up convergence. Then, 3D max-pooling with kernel size 2×2×2 is applied to downsample the spatial resolution. This pooling layer reduces each axis’s dimensions by half, reducing memory consumption and the risk of overfitting. The second convolutional layer doubles the filter count to 64, allowing the same kernel size. Doubling allows the model to learn more complex spatial features. Another 2×2×2 pooling layer down-samples the resulting feature maps, preserving the trade-off between feature abstraction and computational efficiency. The third convolutional block has 128 filters with again 3×3×3 kernels and ReLU activation. The network is learning abstract and high-level features of brain morphology at this stage. A final max-pooling layer further reduces the dimensions of feature maps before the fully connected layers. The output from the last pooling layer is input into a flattening layer, which transforms the 3D feature maps into a one-dimensional vector for dense layers. The first dense layer consists of 128 neurons with ReLU activation to enable the model to learn complex combinations of the features extracted. A Dropout layer is then added to avoid overfitting by randomly dropping out 30% of the neurons during training. Later dense layers have 64 and 32 neurons, respectively—both with ReLU activation. This progressive decrease in dimensionality aids in distillation and summarization of the feature representation without sacrificing non-linearity. Lastly, the output layer has one neuron with a Sigmoid activation function that outputs a probability score between 0 and 1. The score is indicative of the model’s confidence that the input scan is an Alzheimer’s case or a normal control. Training was limited to 100 epochs. Early stopping was used to prevent overfitting, a typical issue with deep learning. Training would stop automatically if validation loss would experience no improvement in a streak of 10 epochs. This way, it is ensured that the model is not learning from noise or spurious patterns anymore after substantial performance improvements have leveled off. Moreover, the best weights of the model (on the validation set) during training were saved, maintaining the optimum state and avoiding subsequent degradation in performance. A mini-batch size of 4 was employed to trade off between two significant considerations: memory limitations and stability of learning. Smaller batches allow for more frequent model weight updates, which can improve generalization and provide finer-grained convergence, particularly when dealing with high-dimensional 3D neuroimaging data. This alternative also prevented GPU memory overflow, which was an important factor during training on computationally demanding models. Due to the high computational requirements of training 3D CNNs—stemming from their volumetric input and extensive parameter space—the experiments were run on an NVIDIA A100 GPU with more than 80 GB of RAM. This setup provided the necessary memory capacity and processing power. Initial experiments also verified that smaller batch sizes performed better than larger ones, possibly because of more accurate gradient updates and lesser overfitting. Binary Cross-entropy was used as the loss function to quantify the divergence between predicted probabilities and true binary labels for this binary classification problem (AD vs. CN). The Adagrad optimizer was used to optimize the models. This option was especially appropriate for our architecture, which has more than 25 million parameters. Adagrad adaptively learns rates for each parameter from previous gradients, allowing more stable and faster convergence. One of the main strong points of Adagrad is that it eliminates steep oscillations in gradient updates. By accumulating squared gradients over time, it reduces the learning rate for frequently updated parameters, and increases it for less frequently updated ones. This adaptiveness not only stabilizes training but also reduces the necessity of carrying out extensive manual learning rate tuning—particularly useful in complicated 3D-based architectures. Although other widely used optimizers like Adam, RMSprop, and SGD were also experimented with, Adagrad was more stable, with a better final classification accuracy. Its adaptiveness to gradient volatility was especially helpful for the complexity and high dimensionality of our model. In Figure 7, the architecture of the model has been visualized. As it can be seen, the whole volume of the brain is entered into the model in 3D form as an input layer, and in each convolution layer, the 3D volumes of the next layers are created by using 3D filters. The model structure and methodology applied to the FTD detection model are the same as those used for AD detection. Nevertheless, the key difference lies in the composition of the control group. In this case, a merged set is utilized that includes the CN cases of the AD model along with control data for FTD. This integration strategy was pursued to enhance the heterogeneity of the control group as well as offset the limited number of FTD-specific control samples. While the AD model served as the reference design, the FTD detection model was intentionally simplified. The largest modification occurred in the Dense (fully connected) layers, where the number of neurons in each layer was reduced by half. This modification significantly decreased the total count of trainable parameters and lessened the computational burden. More importantly, the simplified network achieved improved detection performance for FTD. The simplified network was better able to capture discriminative features for FTD. Figure 8 summarizes the structure of the model, including output dimensions and parameter counts for each layer. Figure 9 provides a visual representation of the hierarchical architecture, illustrating the progression of input data through the network to the final classification output. As can be seen in this figure, the number of parameters compared to the Alzheimer’s disease diagnosis model is almost halved. The 3D CNN architecture employed for detecting PD was the same as the AD detection model. The PD model directly took the design rationale, layering, and overall architecture from the AD model, which was the main reference architecture. This modeling consistency provides a consistent methodological foundation within experiments and allows comparative analysis among the various diagnostic tasks. Aside from architectural consistency, the same training setups—loss function, optimizer, batch size, early stopping condition, and hardware configuration—were utilized as those of the earlier models. This methodological consistency was required both to maintain experimental rigor and to be able to reasonably compare model performances across different neurological disorders. 3.3.2. Differential Diagnosis: An Integrated Unified Multiclass-Diagnosis Model This section focuses on a single deep learning model for the differential diagnosis of three major neurodegenerative diseases—AD, FTD, and PD—and the CN group. The overall goal was to train a single model that could take 3D MRI data as input and classify a subject into one of four diagnostic groups. This multiclass architecture was specifically designed to address the challenges posed by overlapping clinical symptoms, which sometimes complicate traditional diagnostic processes. By consolidating all classes into a single predictive system, the model not only aims to improve diagnostic efficiency, but also reduces the risk of misclassification that can arise in binary models. This unified strategy lays the groundwork for more robust, scalable, and clinically practical tools for distinguishing between related neurodegenerative conditions in real-world settings. The test training sets of this multiclass model were formed by merging the data that used in the three aforementioned binary classification models. All samples were combined in one dataset. The Differential Diagnosis model has some basic differences from the binary classification models. Rather than one output neuron with the Sigmoid activation function, this architecture has four output neurons with the Softmax activation function, enabling the model to predict probabilities of all classes simultaneously in a forward pass. Furthermore, the network architecture substitutes standard Dense layers with Maxout layers, which had greater representational power for features. Unlike regular activation functions that calculate one value per unit, Maxout layers calculate several parallel channels (four in our case per unit) and return the highest value among them. This adaptivity enabled the network to learn more complex features. However, it caused computational cost: Maxout layers drastically increase the number of parameters, which can affect both model complexity and memory usage. In order to manage class imbalance, a prevalent problem in multiclass medical imaging problems, the model uses the Focal Loss function. The loss function was particularly designed to diminish the contribution of dominant classes and direct learning towards more difficult, minority examples. The α (alpha) coefficients were set to be inversely related to class frequencies so that minority classes would have higher contributions to the loss function. The γ (gamma) was finely adjusted to focus on mislabeled or confusing samples. This Focal Loss variant brought tremendous performance gains. Not only did it suppress overfitting and improve generalization, but it also helped the model achieve balanced performance across all four classes, particularly underrepresented classes. Simultaneous adjustment of α and γ was instrumental in offering both high accuracy and strong class-wise distinction. Equation 1 presents the general form of the Focal Loss function as applied in this study [43]. Along with its architectural and loss function innovations, the multiclass model also benefited from the Exponential Linear Unit (ELU) for the activation function for convolutional layers. ELU was used instead of ReLU as it can learn to handle negative activations and also produce smoother and more continuous gradients. This also had the secondary effect of improved training stability and convergence properties, especially in the more challenging multiclass scenario. ELU was used right after batch normalization and convolutional layers to provide efficient non-linear feature map transformation. The optimizer, batch size, learning schedule, and early stopping were kept identical to the binary models for methodological consistency across experiments. The model architecture summary shows a dramatic rise in complexity. In particular, there are over 100 million parameters to train, most of which are in the Maxout layers. While this is a large computation, we believe it is proportionate to the high dimensionality of the input (3D MRI volumes) and the need for accurately discrimination between four diagnostic categories. Figure 11 provides a visual representation of the model’s hierarchical structure, demonstrating the progression of data from raw input to final multiclass output. As it can be seen, in this model, the ELU activation function and Maxout layers was used, and for this reason, the number of parameters is much more than the previous models. In this visualization, like the architecture of previous models, the entire brain volume was used to train the model, but to differentiate among disorders, the last layer in the dense layers had 4 outputs that belong to each of the disorders. 4. Results This section is dedicated to evaluating and analyzing the results obtained from the diagnostic models, including separate models for AD, FTD, and PD. Evaluations are presented using confusion matrices and quantitative indicators such as accuracy, precision, recall (sensitivity), which reveal how well the models perform in classifying new data. Following that, the performance of the unified multiclass model, designed for differential diagnosis among all disorders, is examined, highlighting its accuracy in distinguishing each class from the others. 4.1. Binary Classification Models As illustrated in Figure 12 , the confusion matrix for the Alzheimer’s disease (AD) detection model shows exceptional classification accuracy. All 77 AD cases were correctly identified, while only 4 out of 89 control samples were misclassified. This yielded an overall accuracy of 98% and an F1-score close to 1 , reflecting both high sensitivity and precision. Of particular note is the model’s perfect recall (1.00) for the AD class, meaning no Alzheimer’s cases were missed. Additionally, the precision for the control group reached 1.00 , indicating zero false positives. These results demonstrate the model’s strong ability to cleanly separate the AD and CN classes. The detailed performance metrics are presented in Table 2 . Table 2. Classification performance of the AD diagnosis model Class Precision Recall F1-Score Specificity Support AD 0.95 1.00 0.97 0.96 77 CN 1.00 0.96 0.98 1.00 89 Overall Accuracy 0.98 Macro Avg 0.98 0.98 0.98 0.98 166 Weighted Avg 0.98 0.98 0.98 0.98 166 Regarding the FTD classification model, as shown in the confusion matrix in Figure 13, the model demonstrated strong performance in distinguishing FTD from the CN group. This was despite a small number of misclassifications (3 cases in the FTD class). Out of 37 FTD cases, 34 were correctly identified, and all control group cases were classified correctly. An overall accuracy of 98% indicates excellent model performance. The precision of 1.00 means that all instances predicted as FTD were indeed FTD cases—none of the control samples were misclassified as FTD. On the other hand, the model achieved a recall of 92% for the FTD class, reflecting a relatively good sensitivity. The model also showed excellent performance in identifying control samples, with a 100% recall, indicating it correctly detected all normal control cases. Table 3 summarizes the classification performance of the FTD detection model based on various evaluation metrics. Table 3. Classification performance of the FTD model Class Precision Recall F1-Score Specificity Support FTD 1.00 0.92 0.96 1.00 37 CN 0.97 1.00 0.99 0.92 112 Overall Accuracy 0.98 Macro Avg 0.99 0.96 0.97 0.96 149 Weighted Avg 0.98 0.98 0.98 0.94 149 The confusion matrix in Figure 14 demonstrates that PD diagnosis model also performed well, albeit with slightly more misclassifications. Out of 62 PD samples , 60 were correctly identified , while 2 were misclassified as controls . Among 36 CN samples, 5 were misclassified as PD , resulting in an overall accuracy of 93% . The model maintained a high recall of 97% for the PD class, indicating strong sensitivity and few false negatives. However, the slightly lower specificity in the CN class (86%) suggests room for improvement in reducing false positives. Full performance metrics are provided in Table 4 . Table 4. Classification performance of the PD model Class Precision Recall F1-Score Specificity Support PD 0.92 0.97 0.94 0.86 62 CN 0.94 0.86 0.90 0.97 36 Overall Accuracy 0.93 Macro Avg 0.93 0.91 0.92 0.91 98 Weighted Avg 0.93 0.93 0.93 0.90 98 4.2. Differential Diagnosis: The Integrated Unified Multiclass-Diagnosis Model From the training curves of the combined model in Figure 15 , it is observed that the model converged after approximately 50 epochs. The smallest difference between training and validation accuracy indicates the absence of overfitting. The graphs indicate that despite a large number of parameters of the model (over 100 million), it converged using techniques such as Focal Loss, early stopping, and a very small batch size. The accuracy of training data improved step by step over the course of training, to nearly 99%, demonstrating that the model effectively acquired the training set characteristics while consistently reducing error. The validation accuracy trend was similar and reached to 95%. Some sharp spikes of validation accuracy in certain epochs are to be anticipated with Focal Loss since changes in the distribution of error—especially underrepresented classes—can lead to dramatic shifts in gradient behavior. Training loss decreased steadily to a minimal value, and validation loss also showed a decreasing trend. Absence of sudden loss spikes can be interpreted to mean training stability. Although the disparity between training and validation loss at times seemed large, it dropped near the final epochs, indicating greater model stability. As illustrated by Figure 16 , the differential diagnosis model’s confusion matrix demonstrates high performance in detecting all three neurodegenerative diseases, with high accuracy in classifying normal control. Notably, all 77 cases of AD were accurately distinguished, with perfect recall (1.00) for this class. 2 out of 37 FTD cases and 3 out of 62 PD instances were misclassified, a total of 171 correct predictions among 176 cases of diseases. The model achieved an overall classification accuracy of 95%, marking its stability in processing a multiclass neuroimaging task. One of the strongest findings was the model’s consistently high specificity in all classes, indicating a low rate of false positives—a primary requirement in clinical application where false diagnosis can lead to inappropriate or delayed treatment. Even when there was intrinsic class imbalance, particularly between the CN and disorder classes, use of Focal Loss played a crucial role in maintaining balance in the learning process. It allowed the model to focus more on challenging-to-classify and minority samples, in essence preventing the network from becoming biased towards majority classes. Further, the model possessed high precision for the CN and FTD classes, which signifies its ability to detect subtle differences in structural brain characteristics. This is especially important in distinguishing overlapping clinical characteristics that can be common in neurodegenerative diseases. Collectively, these results show that the proposed multiclass model is a safe and generalizable differential diagnosis scheme with great promise for practical application in real-world clinical decision support systems. A complete summary of classification metrics is provided in Table 5 . Table 5. Classification performance of the unified model Class Precision Recall F1-Score Specificity Support CN 0.97 0.93 0.95 0.98 148 AD 0.95 1.00 0.97 0.98 77 FTD 0.97 0.95 0.96 1.00 37 PD 0.91 0.95 0.93 0.98 62 Overall Accuracy 0.95 Macro Avg 0.95 0.96 0.95 0.98 324 Weighted Avg 0.95 0.95 0.95 0.98 324 To facilitate a complete visualization of the model's diagnostic performance, a combined ROC curve was plotted for all four classes, as illustrated in Figure 17. In this case, a one-vs-rest strategy was adopted, where each class was considered the positive class against all others. For example, when testing the detection capability of the model for the CN class, all samples that were not CN were considered negative, and CN samples positive. For each class, the true positive rate (TPR) and false positive rate (FPR) were calculated at various decision thresholds to construct the respective ROC curves. The area under the curve (AUC) was then computed to quantify the model’s ability to discriminate each class from the rest. As opposed to accuracy, which is threshold-dependent and possibly misleading in imbalanced datasets, AUC estimates classifier performance in a threshold-independent fashion, and is especially useful in multiclass medical diagnosis. This is based on prominent work in machine learning, which has shown that AUC estimates classification performance more reliably and informatively than accuracy alone [44]. The AUC values close to 1 for each class confirm the model’s very high ability for accurate differential diagnose. 5. Discussion One of the prime priorities in current research was the avoidance of data leakage. This can be a serious problem in neuroimaging research because it may bias model assessment and lead to artificially inflated accuracy. To avoid this, all data splitting was at the subject level in the present study, rather than the slice or image level. This ensured that no MRI scans of a single subject were in both test and train sets—something that is a typical error in longitudinal investigations wherein multiple scans are collected from a single subject at different time points. In deep learning research, data leakage is a concept that refers to the unwanted use of test data features in the model training process. It is most classically evident in data split at the level of the image, where multiple images of a single subject appear in train and test datasets. The result is a biased impression of model performance since the network can learn to remember the details of single subjects rather than generalized features of the disorder. This is well supported by current systematic reviews, one of which estimated that nearly half of Alzheimer’s diagnosis based on deep learning researchs are contaminated with data leakage, heavily compromising their clinical utility [45]. In our initial experiments, models were trained on longitudinal data without strict separation by subjects. Even with simpler models and no special layers like Maxout or Focal Loss, they still achieved near-perfect classifications in spite of this. Our hypothesis was that the model was more sensitive to the overall structure of individuals’ brains, rather than learning disease-specific patterns, because in the longitudinal data, the brain structure of an individual is repeated across multiple scans. This outcome had two consequences. On the positive side, it represented the model’s amazing capacity in picking up unique anatomical features of a single brain—suggesting a level of accuracy perhaps beyond human capability in detecting such inter-subject differentials. On the other hand, it also suggested the model learned non-generalizable disease-related features. In fact, the model had learned subject-specific anatomical features, not just disease-related ones. The results showed that the accuracy of the models decreased significantly compared to the previous case using longitudinal data. This decrease confirmed the assumption that data leakage in the previous method had caused false accuracy. Finally, in this study, to ensure that data leakage was prevented, we decided to conduct all experiments during training and testing. Specifically, we avoided placing different scans from the same individual in the training and testing sets. We also took great care to avoid the use of synthetic data augmentation, so models learned from actual patient scans rather than replicated or imitated patterns. Having mitigated data leakage, class imbalance presented the next major challenge, further compounded by relatively lower FTD data availability compared to AD or PD. To address this, we employed the Focal Loss function, which allows the model to learn from hard-to-classify and underrepresented classes by automatically balancing the contribution of each sample to the loss function. The α and γ hyperparameters were optimally tuned to balance the contribution of the smaller and larger classes. We also employed Maxout layers in the model. These contributed extra architectural complexity and computational overhead—bringing the number of model parameters in the multiclass model to in excess of 100 million—though enhanced the model’s ability to extract meaningful patterns from complex and sparse data. Another contribution of this work is the strict preprocessing pipeline, in particular the addition both of the brain extraction and image registration. Image registration established a shared spatial reference in the MRI scans and minimized the effects of position inconstancy and orientation variation of the head across subjects. Otherwise, models would be prone to learning irrelevant spatial noise instead of pathology. Brain extraction served further to ensure that models were trained only on intracranial tissue and not noise from non-brain tissue like the scalp and the skull. These preprocessing steps are sometimes omitted in other studies or are often not applied together. However, their implementation can be essential in improving model focus and reducing bias—especially in high-dimensional 3D data where each voxel carries a weight in the learning process. When we compare our results to similar research, several unique features of our methodology’s are apparent. Some previous works employed PET scans, which—while effective—are more intrusive, requiring radioactive tracers to be injected [46]. In contrast, our technique uses only T1-weighted MRI, a safer, more general modality. In addition, some previous works truncated 3D MRI volumes to 2D slices in order to increase their number of examples. While this technique inflates the number of examples, it probably causes heavy leakage of data—especially where multiple slices per subject are found to appear in both the train and test folds. In addition, 2D slices cannot model the brain as a volume, and may fail to pick up on important structural details in unsampled regions. In contrast, our 3D CNN model takes the entire brain volume as input, and is therefore able to process all neuroanatomical changes. While computationally less efficient, this results in richer and better spatially-coherent models of pathology. In addition, we avoided using synthetic augmentation techniques that might introduce bias or lead to learning artificially enhanced features. Instead, we opted for real, subject-level MRI data, taking care not to let any single subject contribute multiple times to the model. In particular, while previous works tried to address class imbalance by data augmentation, or by using synthetically image generation, our Focal Loss-based model—a more elegant solution that involves no data manipulation—has no equivalent in previous works that were compared in the present study. Although under tighter restrictions—imbalance in the data, no data augmentation, subject-based design, and 3D volumetric analysis—our model performed comparably, and in most cases, even better in terms of sensitivity and generalizability. In summary, design decisions for this study were made to make it useful for clinical applications. From strict data manipulation procedures to advanced loss function and preprocessing methods, every aspect was chosen in a careful effort to ensure that the model’s diagnostic accuracy was not merely statistically impressive but truly clinically significant, reliable, and practical. Table 6 is the comparison of this study with works of relevance. Table 6 summarizes the comparison of the present study with related works. Table 6. Comparative table of the present study with related works Study Accuracy (%) Sensitivity (%) AUC (%) Modality Subjects # Key Distinctions Present Study (Unified Model) 95 AD=100 FTD=95 PD=95 99–100 MRI 1629 - Brain extraction and image registration applied - Strict subject-level partitioning to prevent data leakage - Use of Focal Loss - Use of Maxout layers - Full 3D brain volume analysis K. Etminani et al. [ 22 ] 78.9 AD=91 MCI=17 DLB=86 71.4–96.4 PET 757 - Requires radioactive tracer injection (PET) - Significantly lower accuracy - lower sensitivity in AD F. Ramzan et al. [ 28 ] 97.92 AD=98.01 MCI=97.81 LMCI=97.43 EMCI=97.38 SMC=100 99.96–99.97 fMRI/2D 138 → 850,080 2D images - 3D scans converted to a large number of 2D slices - No subject-based split or slice duplication control - No registration or brain extraction applied - lower sensitivity in AD A. Ahila et al. [ 29 ] 96.8 AD=94 95 PET 855 - PET required - 3D scans converted to 2D - No brain extraction - Lower accuracy - lower sensitivity in AD R. G. Akindele et al. [ 30 ] 99.9 AD=100 100 MRI 69 subjects (708 scans) - No registration or brain extraction - Performed augmentation - 3D images reconstructed from 2D slices D. Ma et al. [ 31 ] 88.28 AD=84.66 FTD=77.82 95 MRI 1954 - Lower accuracy - Patch-based feature extraction - Synthetic data generated using GAN - Significantly lower accuracy - lower sensitivity in AD and FTD A. Rogeau et al. [ 32 ] 89.8 AD=75 FTD=95 93.3–99.9 PET 496 (591 scans) - Lower accuracy - PET-based - Unequal scan distribution across subjects - lower sensitivity in AD S. Esmaeilzadeh et al. [ 33 ] 100 PD=100 100 MRI 656 - No image registration - Applied data augmentation S. Chakraborty et al. [ 34 ] 95.29 PD=94.3 98 MRI 406 - No brain extraction - Lower sensitivity - Balanced dataset selection M. Dyrba et al. [ 35 ] AD vs CN: 75.5–88.3 MCI vs CN: 63.1–75.4 – AD vs CN: 82.8–97.8MCI vs CN: 66.7–84 MRI 663 - Lower accuracy - Only gray matter used - Data augmentation applied - Significantly lower accuracy Z. Hu et al. [ 36 ] AD vs NC: 94.23 MCI vs NC: 95.67 AD vs MCI: 97.75 AD vs MCI vs NC: 89.71 – – MRI 520 - Lower accuracy - Only gray matter used - Sensitivity and AUC not reported S. De Francesco et al. [ 37 ] 87.5 AD=73 FTD=89 DLB=95 92.09–99.94 MRI 506 - Lower accuracy - lower sensitivity in FTD A. M. El-Assy et al. [ 38 ] 99.30 AD=100 EMCI=89 LMCI=100 MCI=95 97.37–100 MRI/2D 1296 - No registration or brain extraction - Class balancing via augmentation - 2D slices extracted from 3D volumes (AD, Alzheimer’s disease; FTD, Frontotemporal dementia; DLB, dementia with Lewy bodies; PD, Parkinson’s disease; MCI, mild cognitive impairment; CN, cognitively normal; EMCI, early mild cognitive impairment; LMCI, late mild cognitive impairment) 6. Conclusion and Future Works The current study intended to examine the capabilities of convolutional neural networks in differential diagnosis of neurodegenerative diseases such as Alzheimer’s disease, frontotemporal dementia, and Parkinson’s disease based on MRI data. Considering the complex structure of the diseases and challenges in their diagnosis, the current study intended to provide a model supported by deep learning approaches that could be of high accuracy and credibility in differential diagnosis of the mentioned diseases and differentiating from one another. The results of this study indicated that the proposed models can be employed as a complementary decision-support tool in the diagnosis of neurodegenerative diseases. These models were capable to differentiate Alzheimer’s disease, frontotemporal dementia, and Parkinson’s disease from the control group with 98%, 98%, and 93% accuracy, respectively. The multiclass model in achieved 95% accuracy in differential diagnosis. Comparison with prior works revealed that the proposed model performs superior to most prior 2D and 3D approaches and in the meantime, avoided leakage of data. Ultimately, according to the results, it is possible to conclude that deep learning is much potent in analysis of medical imaging data and is capable of playing an effective role in enhancing neurological diseases’ diagnosis methods, notably Alzheimer’s disease, frontotemporal dementia, and Parkinson’s disease. The current study demonstrated that 3D convolutional neural networks are capable of analyzing information from the images with precise accuracy and differentiating several diseases with good quality. This enables doctors to make optimum diagnostic decisions and place patients on a suitable treatment path. Based on the findings of this study and the challenge of differential diagnosis of neurodegenerative diseases, one of the most critical research and optimization directions suggested in future work is determining the stage of disease development. The models presented in this study focused on differential diagnosis between neurodegenerative diseases, but future work can separate patients at different stages of the disease based on investigating the extent of brain atrophy and shrinking in critical structures like the hippocampus and cerebral cortex. The advantages of disease staging are that AI-based models are able to measure the progression of the disease in different patients based on examining the structural pattern of the brain and help manage treatment. Incorporation of the feature of determining the stage of the disease opens the door to compare cohorts of patients and investigate the influence of different treatments on disease development. Another suggestion towards model upgrading is the inclusion of dementia with Lewy bodies data. As mentioned before, this form of dementia pathologically and physiologically shares some similarities with Parkinson’s disease [3]. Adding this disorder to the model, therefore, is capable of bridging the prevailing gap in disease differentiation and making the system an integrated model of differential diagnosis of neurodegenerative cognitive diseases. Although data concerning this disorder were not captured in this study due to limitations in resources and data availability, in future research, the incorporation of data from this disorder is capable of upgrading the model to be more precise in disease distinction and ready to be employed in real clinical practice. The proposed models here are constructed only on MRI data but in practice, cognitive tests such as Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) are also crucial for assessing neurodegenerative disease patients [47]. Including this data with MRI images will increase the accuracy and validity, especially in the differential diagnosis model. Declarations Funding Declaration: No funding was received for this work. Competing Interests Declaration: The authors declare no competing interests. References Prince M, Wimo A, Guerchet M, Gemma-Claire A, Wu YT, Prina M. World Alzheimer Report 2015: The Global Impact of Dementia - An analysis of prevalence, incidence, cost and trends [Research report]. Alzheimer’s Dis Int [Internet]. 2015. Available from: https://hal.archives-ouvertes.fr/hal-03495438 Dementia: Key facts. WHO. [Internet]. 2025. Available from: https://www.who.int/news-room/fact-sheets/detail/dementia Gomperts SN. Lewy body dementias: Dementia with lewy bodies and Parkinson disease dementia. Contin Lifelong Learn Neurol. 2016;22(2, Dementia):435–63. Ferri CP, Prince M, Brayne C, Brodaty H, Fratiglioni L, Ganguli M, et al. Global prevalence of dementia: A Delphi consensus study. Lancet. 2005;366(9503):2112–7. Deture MA, Dickson DW. The neuropathological diagnosis of Alzheimer’s disease. Mol Neurodegener. 2019;14(1). Coupé P, Manjón JV, Lanuza E, Catheline G. Lifespan Changes of the Human Brain In Alzheimer’s Disease. Sci Rep. 2019;9(1). Rao YL, Ganaraja B, Murlimanju B V., Joy T, Krishnamurthy A, Agrawal A. Hippocampus and its involvement in Alzheimer’s disease: a review. 3 Biotech. 2022;12(2). Johnson KA, Fox NC, Sperling RA, Klunk WE. Brain imaging in Alzheimer disease. Cold Spring Harb Perspect Med. 2012;2(4). Dickerson BC, Wolk DA. MRI cortical thickness biomarker predicts AD-like CSF and cognitive decline in normal adults. Neurology. 2012;78(2):84–90. van Oostveen WM, de Lange ECM. Imaging techniques in alzheimer’s disease: A review of applications in early diagnosis and longitudinal monitoring. Int J Mol Sci. 2021;22(4):1–34. Finger EC. Frontotemporal dementias. Contin Lifelong Learn Neurol. 2016;22(2, Dementia):464–89. Rabinovici GD, Miller BL. Frontotemporal lobar degeneration: Epidemiology, pathophysiology, diagnosis and management. CNS Drugs. 2010;24(5):375–98. Bertoux M, O’Callaghan C, Flanagan E, Hornberger M. Frontotemporal Dementia (FTD). In: Pachana NA, editor. Encyclopedia of Geropsychology [Internet]. Singapore: Springer Singapore; 2015. p. 1–17. Available from: https://doi.org/10.1007/978-981-287-080-3_311-1 Lindau M, Almkvist O, Kushi J, Boone K, Johansson SE, Wahlund LO, et al. First symptoms - Frontotemporal dementia versus Alzheimer’s disease. Dement Geriatr Cogn Disord. 2000;11(5):286–93. Bocti C, Rockel C, Roy P, Gao FQ, Black SE. Topographical patterns of lobar atrophy in frontotemporal dementia and Alzheimer’s disease. Dement Geriatr Cogn Disord. 2006;21(5–6):364–72. What is frontotemporal degeneration?. Mayo Clinic [Internet].2024. Available from: https://newsnetwork.mayoclinic.org/discussion/what-is-frontotemporal-degeneration/ Kouli A, Torsney KM, Kuan WL. Parkinson’s Disease: Etiology, Neuropathology, and Pathogenesis. Park Dis Pathog Clin Asp. 2018;3–26. Parkinson’s Disease. National Institute of Neurological Disorders and Stroke [Internet].Available from: https://www.ninds.nih.gov/health-information/disorders/parkinsons-disease Zeighami Y, Ulla M, Iturria-Medina Y, Dadar M, Zhang Y, Larcher KMH, et al. Network structure of brain atrophy in de novo parkinson’s disease. Elife. 2015;4(September 2015). Akram AS, Grezenko H, Singh P, Ahmed M, Hassan BD, Hagenahalli Anand V, et al. Advancing the Frontier: Neuroimaging Techniques in the Early Detection and Management of Neurodegenerative Diseases. Cureus. 2024; Mathur S, Gawas C, Ahmad IZ, Wani M, Tabassum H. Neurodegenerative disorders: Assessing the impact of natural vs drug-induced treatment options. Aging Med. 2023;6(1):82–97. Etminani K, Soliman A, Davidsson A, Chang JR, Martínez-Sanchis B, Byttner S, et al. A 3D deep learning model to predict the diagnosis of dementia with Lewy bodies, Alzheimer’s disease, and mild cognitive impairment using brain 18F-FDG PET. Eur J Nucl Med Mol Imaging. 2022;49(2):563–84. Aggarwal R, Sounderajah V, Martin G, Ting DSW, Karthikesalingam A, King D, et al. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis. npj Digit Med. 2021;4(1). Chan HP, Samala RK, Hadjiiski LM, Zhou C. Deep Learning in Medical Image Analysis. Adv Exp Med Biol. 2020;1213:3–21. Kourounis G, Elmahmudi AA, Thomson B, Hunter J, Ugail H, Wilson C. Computer image analysis with artificial intelligence: a practical introduction to convolutional neural networks for medical professionals. Postgrad Med J. 2023;99(1178):1287–94. Mäenpää SM, Korja M. Diagnostic test accuracy of externally validated convolutional neural network (CNN) artificial intelligence (AI) models for emergency head CT scans – A systematic review. Int J Med Inform. 2024;189. Chakraborty C, Bhattacharya M, Pal S, Lee SS. From machine learning to deep learning: Advances of the recent data-driven paradigm shift in medicine and healthcare. Curr Res Biotechnol. 2024;7. Ramzan F, Khan MUG, Rehmat A, Iqbal S, Saba T, Rehman A, et al. A Deep Learning Approach for Automated Diagnosis and Multi-Class Classification of Alzheimer’s Disease Stages Using Resting-State fMRI and Residual Neural Networks. J Med Syst. 2020;44(2). A A, M P, Hamdi M, Bourouis S, Rastislav K, Mohmed F. Evaluation of Neuro Images for the Diagnosis of Alzheimer’s Disease Using Deep Learning Neural Network. Front public Heal. 2022;10:834032. Akindele RG, Adebayo S, Kanda PS, Yu M. AlzhiNet: Traversing from 2DCNN to 3DCNN, Towards Early Detection and Diagnosis of Alzheimer’s Disease. 2024; Available from: http://arxiv.org/abs/2410.02714 Ma D, Lu D, Popuri K, Wang L, Beg MF. Differential Diagnosis of Frontotemporal Dementia, Alzheimer’s Disease, and Normal Aging Using a Multi-Scale Multi-Type Feature Generative Adversarial Deep Neural Network on Structural Magnetic Resonance Images. Front Neurosci. 2020;14. Rogeau A, Hives F, Bordier C, Lahousse H, Roca V, Lebouvier T, et al. A 3D convolutional neural network to classify subjects as Alzheimer’s disease, frontotemporal dementia or healthy controls using brain 18F-FDG PET. Neuroimage. 2024;288. Esmaeilzadeh S, Yang Y, Adeli E. End-to-End Parkinson Disease Diagnosis using Brain MR-Images by 3D-CNN. 2018; Available from: http://arxiv.org/abs/1806.05233 Chakraborty S, Aich S, Kim HC. Detection of Parkinson’s disease from 3T T1 weighted MRI scans using 3D convolutional neural network. Diagnostics. 2020;10(6). Dyrba M, Hanzig M, Altenstein S, Bader S, Ballarini T, Brosseron F, et al. Improving 3D convolutional neural network comprehensibility via interactive visualization of relevance maps: evaluation in Alzheimer’s disease. Alzheimer’s Res Ther. 2021;13(1). Hu Z, Wang Y, Xiao L. Alzheimer’s disease diagnosis by 3D-SEConvNeXt. J Big Data. 2025;12(1). De Francesco S, Crema C, Archetti D, Muscio C, Reid RI, Nigri A, et al. Differential diagnosis of neurodegenerative dementias with the explainable MRI based machine learning algorithm MUQUBIA. Sci Rep. 2023;13(1). El-Assy AM, Amer HM, Ibrahim HM, Mohamed MA. A novel CNN architecture for accurate early detection and classification of Alzheimer’s disease using MRI data. Sci Rep. 2024;14(1). Mandal PK, Mahajan R, Dinov ID. Structural brain atlases: Design, rationale, and applications in normal and pathological cohorts. J Alzheimer’s Dis. 2012;31(SUPPL. 3). Viergever MA, Maintz JBA, Klein S, Murphy K, Staring M, Pluim JPW. A survey of medical image registration – under review. Med Image Anal. 2016;33:140–4. Jenkinson M, Beckmann CF, Behrens TEJ, Woolrich MW, Smith SM. FSL. Neuroimage [Internet]. 2012;62(2):782–90. Available from: https://www.sciencedirect.com/science/article/pii/S1053811911010603 Nyúl LG, Udupa JK, Zhang X. New variants of a method of MRI scale standardization. IEEE Trans Med Imaging. 2000;19(2):143–50. Lin TY, Goyal P, Girshick R, He K, Dollar P. Focal Loss for Dense Object Detection. IEEE Trans Pattern Anal Mach Intell. 2020;42(2):318–27. Huang J, Ling CX. Using AUC and accuracy in evaluating learning algorithms. IEEE Trans Knowl Data Eng. 2005;17(3):299–310. Wen J, Thibeau-Sutre E, Diaz-Melo M, Samper-González J, Routier A, Bottani S, et al. Convolutional neural networks for classification of Alzheimer’s disease: Overview and reproducible evaluation. Med Image Anal. 2020;63. Al-Sharify ZT, Al-Sharify TA, Al-Sharify NT, Naser HY. A critical review on medical imaging techniques (CT and PET scans) in the medical field. IOP Conf Ser Mater Sci Eng. 2020;870(1). Gonzalez Kelso I, Tadi P. Cognitive Assessment. In: StatPearls [Internet]. [Updated 2022 Nov 7]. Treasure Island (FL): StatPearls Publishing; 2022 Nov 7 [cited 2025 Jan]. Available from: https://www.ncbi.nlm.nih.gov/books/NBK556049/ Footnotes Alzheimer's Disease Neuroimaging Initiative ( http://adni.loni.usc.edu/ ) Neuroimaging in Frontotemporal Dementia ( http://memory.ucsf.edu/research/studies/nifd ) Parkinson's Progression Markers Initiative ( http://www.ppmi-info.org/ ) Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7998240","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":538043444,"identity":"6bb386bb-f833-42c0-9010-42f8fcec02af","order_by":0,"name":"Mehran Panahi","email":"data:image/png;base64,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","orcid":"","institution":"Tarbiat Modares University, Department of Data Science, Faculty of Interdisciplinary Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Mehran","middleName":"","lastName":"Panahi","suffix":""},{"id":538045881,"identity":"c0f161f9-11ef-405c-a4f4-0d031fc3f515","order_by":1,"name":"Ali Asghar Safaei","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+UlEQVRIie3PsUrDQBzH8b8cXJeLrhdOkldICDiJz/ILBbeCkFEhmS6L+ASKr9AidD4I6GLp2qmksxnSpXQo4k0iQq64OdwXDo4/fPjzJ/L5/m2XRHxUEYnvSXKMXFsizF8JSfwkjs5qw/o9rqLT8CNvu7t1FFds09PNepDId/DwHuOMq8lL+vhaZInhmaSkGF5jwCnYs1yryVwJjnxKdGFvwaCIly3bHlCWOlxY8onyuRrtnCRZgZRAAy6DuQo0QEa4t6SrjVbneEu1sLc8PSCdNqKQcJBoOW62HW7juF7M2m4H+6lnfX8YJkQn1a8Bs88FfD6fz3e8L+6rTJkY9kfYAAAAAElFTkSuQmCC","orcid":"","institution":"Tarbiat Modares University, Department of Data Science, Faculty of Interdisciplinary Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Ali","middleName":"Asghar","lastName":"Safaei","suffix":""},{"id":538049065,"identity":"b43d4e9b-382e-4930-9012-5025916e1a54","order_by":2,"name":"Foad Ghaderi","email":"","orcid":"","institution":"Tarbiat Modares University,, Electrical and Computer Engineering Department","correspondingAuthor":false,"prefix":"","firstName":"Foad","middleName":"","lastName":"Ghaderi","suffix":""},{"id":538049066,"identity":"61bf90dd-1514-4519-99cb-59cd8e6ddb2b","order_by":3,"name":"Fatemeh Panahi","email":"","orcid":"","institution":"University of Ottawa","correspondingAuthor":false,"prefix":"","firstName":"Fatemeh","middleName":"","lastName":"Panahi","suffix":""}],"badges":[],"createdAt":"2025-10-31 12:01:50","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7998240/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7998240/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95221105,"identity":"47bedfd3-d0ce-4fdf-9fc3-a28b92c4f13d","added_by":"auto","created_at":"2025-11-05 16:18:17","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2303857,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/5b19e78223c27ebd62954ceb.docx"},{"id":95006122,"identity":"1a6c1eeb-8039-48cb-bf07-22be6ced1191","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7998240.json","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/57616876abfc444f1313e44e.json"},{"id":95221229,"identity":"4a24b75c-924e-42b0-ba34-fa498ea80d62","added_by":"auto","created_at":"2025-11-05 16:18:39","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":168239,"visible":true,"origin":"","legend":"","description":"","filename":"rs79982400enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/8c5f9acebaca95d83b18a8b5.xml"},{"id":95006138,"identity":"58d3ff71-ddbf-4c87-bf33-a89fe9c27892","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":280073,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/40a6c0ed9b748ac511997c0b.png"},{"id":95221252,"identity":"2ad704dd-8619-45df-b289-55ed205fff92","added_by":"auto","created_at":"2025-11-05 16:18:43","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69468,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/340a6dc9d2c95e47f11b68db.png"},{"id":95221049,"identity":"e0a4b575-3859-462f-b009-35593128a5f4","added_by":"auto","created_at":"2025-11-05 16:18:07","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":254551,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/e0f125d01a3c8bfd29f1c9fb.png"},{"id":95221420,"identity":"fbb10193-db89-48a6-a346-882769dcfcc1","added_by":"auto","created_at":"2025-11-05 16:18:59","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17157,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/f09a5b6606ba94cf77a89b6a.png"},{"id":95221093,"identity":"4a7d8303-aacf-4368-ae4c-7023f7ace9eb","added_by":"auto","created_at":"2025-11-05 16:18:14","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15983,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/920de461f16b03c5c56fb681.png"},{"id":95006136,"identity":"de082aa3-be9b-43b5-8ffc-3005d3ca6b21","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15512,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/230f7477a9cb229bcf9c78f1.png"},{"id":95006131,"identity":"52d66d0d-5fc9-4796-a20f-31cfd8427a88","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":112646,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/82ac0956db00c93dee39617f.png"},{"id":95221072,"identity":"9223c088-f86a-46aa-bdc4-0a2412426e6e","added_by":"auto","created_at":"2025-11-05 16:18:11","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20948,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/af6f1705443c52ac67fd6ba8.png"},{"id":95006150,"identity":"34c837bf-6b84-473e-9d0c-f09e59370c14","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43368,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/542a08cc0fa985a6dccb22df.png"},{"id":95221382,"identity":"ac817609-e9b0-4d78-bccd-fd2513536af9","added_by":"auto","created_at":"2025-11-05 16:18:53","extension":"jpeg","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26350,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/19ab61e0bcfe8b8cfc02e63c.jpeg"},{"id":95221425,"identity":"7d058013-3556-4ab1-bd43-251f5555db38","added_by":"auto","created_at":"2025-11-05 16:18:59","extension":"jpeg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":104721,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/d70e439b3e7b8e5f640e37e8.jpeg"},{"id":95006147,"identity":"996416a3-1b3e-4d7a-8fa8-b466d70fddaa","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61387,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/c346dc5c648487c04dfdd7bc.png"},{"id":95221346,"identity":"d61bf1dd-9175-4cb4-802b-64d3112eacc1","added_by":"auto","created_at":"2025-11-05 16:18:50","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":139224,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/2f329253245500b03060f2e0.jpeg"},{"id":95006154,"identity":"fbd3ed74-e7f3-4e1a-84be-f12696140520","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71374,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/0f92ab68507d1301bb9673a1.png"},{"id":95006175,"identity":"cd70eacf-63bb-44e3-9679-48509b05b016","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":245682,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/e66d9962d9d4092236e1b141.png"},{"id":95221522,"identity":"7ea75993-e44c-4566-82c0-c037d337ee63","added_by":"auto","created_at":"2025-11-05 16:19:14","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71640,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/e17b60caa3d11fdf901b4ca4.png"},{"id":95006162,"identity":"a433f697-b2ac-45d2-8d20-4a16cad1de32","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":245414,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/d36dfad015df9e008ae007f5.png"},{"id":95220924,"identity":"beeac316-de40-4fdf-bd4c-42d8ad2f7014","added_by":"auto","created_at":"2025-11-05 16:17:10","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":38336,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/1d6d0429c2d2ad6abc2a642e.png"},{"id":95221503,"identity":"48b2521b-85f1-4408-adf4-cb17d23b8458","added_by":"auto","created_at":"2025-11-05 16:19:12","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22493,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/232f7079d0e95e5f2c71e5a5.png"},{"id":95006156,"identity":"3e1e3eaa-25fa-45f9-8de4-b3cfba3d68dd","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":59568,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/c70cbafb19d4ba16549184a7.png"},{"id":95006160,"identity":"3090f8de-81e5-4a31-ba92-0bde73f34dfa","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5509,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/34e8505043e4b720b9819d23.png"},{"id":95221253,"identity":"e0d27d24-b037-4749-8015-03231fc0c381","added_by":"auto","created_at":"2025-11-05 16:18:43","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5846,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/8b4a1f65fa3064351b4a490d.png"},{"id":95006171,"identity":"20d3792e-90d1-4f26-9367-a91ea4acea43","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5107,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/1dc67a89d3d9efa7c4f9fdbc.png"},{"id":95006161,"identity":"98517f33-4c00-43e6-812f-347ed1e0e189","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35698,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/f29acbfe5872445387b3ec25.png"},{"id":95221540,"identity":"3448366d-83ff-4d11-ac2a-8b77c52f8417","added_by":"auto","created_at":"2025-11-05 16:19:16","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6507,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/158ba3d59bd03199d570e51a.png"},{"id":95006178,"identity":"f7100355-0a9b-4051-a11b-58a23fba71e4","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10652,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/39ca45b7fdc731e2c541bb67.png"},{"id":95006164,"identity":"5b8fba70-e472-42e6-bd6a-ace925079115","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":58511,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/253c0ce1f31226c042496cc9.png"},{"id":95006172,"identity":"43fb1a3a-dbbf-498c-ae88-0dd2c37a3efb","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":51825,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/228ae6752a965487ac7aeba7.png"},{"id":95006163,"identity":"ae40909e-2cb3-49e4-b9ef-4e77d51139ba","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13799,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/060002eccee1e3874168501e.png"},{"id":95006165,"identity":"e9587473-7109-4705-bae7-16e5a6f3a581","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49881,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/0ecb1cf0c2b8ea39881a8613.png"},{"id":95006166,"identity":"07808409-4ad3-43f5-8f71-bfb86e612e72","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28800,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/ff83a5fdf04a2815f367b41b.png"},{"id":95006168,"identity":"af492b3d-a52c-447d-ade3-78c59bd53f7c","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56632,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/199efe25d591617e22b96af1.png"},{"id":95221356,"identity":"cd98bbe0-4db5-41af-ae1e-42dc0dff2f92","added_by":"auto","created_at":"2025-11-05 16:18:51","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28914,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/fc0538e9707644ba47b54601.png"},{"id":95006173,"identity":"66cc2599-9192-4ffc-b0b3-80f166676c7e","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56572,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/58fc135a0c0e336f7486f13d.png"},{"id":95220911,"identity":"b9e2df4d-098c-4c02-a028-34a1519b94a7","added_by":"auto","created_at":"2025-11-05 16:17:02","extension":"xml","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":166031,"visible":true,"origin":"","legend":"","description":"","filename":"rs79982400structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/eb7f33501b6a35ab976fbe05.xml"},{"id":95220857,"identity":"16bcaad6-0f1e-4e8e-bbfe-1c2da3848ca9","added_by":"auto","created_at":"2025-11-05 16:16:09","extension":"html","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":176959,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/41dac47cb2506632f804b9b9.html"},{"id":95006123,"identity":"cc38d36f-34f9-4be3-8611-08b67d42f5a4","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":261544,"visible":true,"origin":"","legend":"\u003cp\u003eStructural changes in the brain in AD\u003c/p\u003e\n\u003cp\u003e)AD causes atrophy of various parts of the brain, especially the hippocampus, as well as enlargement of the ventricles. This figure compares the brain of a healthy person (left) and a person suffering from Alzheimer’s (right) [10](\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/c25dcdc4b04fae0c82ae96b7.png"},{"id":95221385,"identity":"44f1171a-8c5d-434f-b8e8-ab54c18f337b","added_by":"auto","created_at":"2025-11-05 16:18:53","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":34882,"visible":true,"origin":"","legend":"\u003cp\u003eStructural changes in the brain in FTD.\u003c/p\u003e\n\u003cp\u003e(This figure shows the brain of a cognitively normal person (left) compared to a person suffering from FTD (right). As can be seen, atrophy is visible in the frontal part of the brain affected by this disorder [16].)\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/661f608ee5615defbf631274.jpeg"},{"id":95006124,"identity":"1fb3fad5-adb7-4b55-8fbe-d8e601cea267","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":101286,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of atrophy in Parkinson’s disease in different brain regions [19]\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/2dcd0c775a0782abbbce3e3e.jpeg"},{"id":95221095,"identity":"021556c5-f52a-4957-a668-d9140b4f13a0","added_by":"auto","created_at":"2025-11-05 16:18:14","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":62129,"visible":true,"origin":"","legend":"\u003cp\u003eData distribution across each class.\u003c/p\u003e\n\u003cp\u003e(Yellow indicates the number of subjects for each disorder and blue indicates the number of subjects belonging to the control group for each disorder)\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/a0119043d65588a0732f78af.png"},{"id":95006127,"identity":"c3b8183c-a1f2-4bb5-9320-ec4ed8077103","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":79780,"visible":true,"origin":"","legend":"\u003cp\u003eMRI on which brain extraction was performed using FSL software\u003c/p\u003e\n\u003cp\u003e(MRI on which the brain extraction operation was performed using FSL software. The figure shows the MRI before the operation (left) and after the operation, when excess tissue has been removed (right).)\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/71ec6fca9be0fca26bcc6315.jpeg"},{"id":95006125,"identity":"d7e89b28-e199-49cb-8261-f989a10fd0c9","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":70326,"visible":true,"origin":"","legend":"\u003cp\u003eSummary of the Alzheimer’s disease diagnostic model\u003c/p\u003e\n\u003cp\u003e(different layers of the model, the output dimensions of each layer, and the number of parameters of each layer)\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/20ef5308b60bfda1d158af75.png"},{"id":95006132,"identity":"850e92a7-2b1b-4e42-9270-180bd88612a3","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":272704,"visible":true,"origin":"","legend":"\u003cp\u003e3D_CNN architecture of the Alzheimer’s disease diagnosis model\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/a4d3d1a686b214a3455bd9ea.png"},{"id":95221140,"identity":"8cc1d071-2283-43e9-9b30-0bda1c21bedc","added_by":"auto","created_at":"2025-11-05 16:18:26","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":70416,"visible":true,"origin":"","legend":"\u003cp\u003eSummary of the Frontotemporal dementia diagnostic model\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/0e0a4bee752cd3487e9e5a30.png"},{"id":95006143,"identity":"f544bec0-fe4d-4517-afbb-277b836ab3d7","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":272331,"visible":true,"origin":"","legend":"\u003cp\u003e3D_CNN architecture of the Frontotemporal dementia diagnostic model.\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/503937431d03641a2d4c7913.png"},{"id":95220890,"identity":"ebabd9e1-0814-487f-9310-a67b2bf70abf","added_by":"auto","created_at":"2025-11-05 16:16:46","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":68877,"visible":true,"origin":"","legend":"\u003cp\u003eSummary of the Differential Diagnosis Model\u003c/p\u003e","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/0e2f510421d5a82bea43475d.png"},{"id":95220844,"identity":"806bb22c-b36d-4673-b272-73e271c4a0e6","added_by":"auto","created_at":"2025-11-05 16:15:57","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":289675,"visible":true,"origin":"","legend":"\u003cp\u003e3D_CNN architecture of the Differential Diagnosis Model\u003c/p\u003e","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/72415e354397bf06f6b80585.png"},{"id":95221060,"identity":"a70cf227-c63a-439a-b099-9da48ee18bb1","added_by":"auto","created_at":"2025-11-05 16:18:08","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":18550,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix for the AD classification model\u003c/p\u003e","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/2fd3ef3a13f3aa3fc7419805.png"},{"id":95006142,"identity":"dd98a357-4493-4215-b988-e88166e64bca","added_by":"auto","created_at":"2025-11-03 09:37:48","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":17229,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix for the FTD classification model\u003c/p\u003e","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/cf253d1e7769f24a8d3b0769.png"},{"id":95006159,"identity":"6d56ad1a-6805-4697-a23e-2c87eb438708","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":17029,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix for the PD classification model\u003c/p\u003e","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/0b85f517e593819d10158c2d.png"},{"id":95006149,"identity":"d86c6f6a-0d41-42fb-93d4-7e7046239670","added_by":"auto","created_at":"2025-11-03 09:37:49","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":116342,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy and loss curves of the unified multiclass model\u003c/p\u003e","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/2662d9334904b8537ce5113d.png"},{"id":95221395,"identity":"2283344a-e73f-4d83-9c69-4231ebcdabe6","added_by":"auto","created_at":"2025-11-05 16:18:55","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":20948,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix of the unified multiclass model\u003c/p\u003e","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/c5fad7a53735d03bf6b0b2c9.png"},{"id":95221504,"identity":"c7f56a5e-8192-4eb5-8e35-bdce6512d194","added_by":"auto","created_at":"2025-11-05 16:19:12","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":43368,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic curves and AUC value for each class in the unified multiclass model\u003c/p\u003e","description":"","filename":"floatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/111b9a70fc83bcfd8e139908.png"},{"id":95312036,"identity":"b83424c5-6516-4c1a-906d-0f29b946eafe","added_by":"auto","created_at":"2025-11-06 15:45:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3153074,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7998240/v1/c9e0e59f-fb0a-4b1a-902f-dc432ab2521c.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eDifferential Diagnosis of Cognitive Disorders using Deep Learning Techniques based on Neuroimaging Data\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eCognitive impairment is one of the leading causes of disability in late life and is featured in the top most common syndromes of this phase of life. The global population affected by dementia is increasing rapidly, with projections estimating it will reach around 135 million by 2050 [1]. The World Health Organization (WHO) reports that 57 million people had dementia in 2021 and nearly 10 million new cases are found every year. Dementia is the seventh greatest cause of death worldwide and the greatest single cause of disability and dependency in older adults. It is a nonspecific term for a decline in mental capacities, including memory, cognition, attention, executive function (such as problem-solving and decision-making), language, and visuospatial perception. The most important subtypes of dementia include Alzheimer\u0026rsquo;s disease, vascular dementia, dementia with Lewy bodies, and frontotemporal dementia [2]. Additionally, patients with Parkinson\u0026rsquo;s disease are at an increased risk of developing dementia as the disease progresses [3]. This paper aims to develop deep learning-based models to diagnose cognitive and neurodegenerative diseases such as Alzheimer\u0026rsquo;s disease, frontotemporal dementia, and Parkinson\u0026rsquo;s disease using available MRI data. All the three diseases have been discussed in the following paragraphs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAlzheimer\u0026rsquo;s disease (AD)\u0026nbsp;\u003c/strong\u003eis the most prevalent form of dementia globally, accounting for approximately 60\u0026ndash;70% of all dementia cases [4]. It is a progressive neurodegenerative disorder with no known treatment yet [5]. Neuroimaging results show that cerebral atrophy, decrease in hippocampal volume, ventriculomegaly, and cortical thinning are structural brain changes which can be detected years prior to clinical symptom onset and follow a temporal profile over the course of the illness [6].\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThe hippocampus is one of the earliest regions affected in AD, largely due to its critical role in memory and learning, making it particularly vulnerable to AD-related damage. Hippocampal atrophy will probably be the most crucial neuroimaging biomarker of the disease [7]. At the early period of AD, hippocampal volume reduction can be up to 20% versus controls; later it may exceed 30% [8]. AD will cause atrophy in certain areas of the brain, i.e., the temporal, parietal, and frontal lobes. Certain effects of this atrophy are dilatation of the ventricles of the brain [5]. It also features cortical thinning in AD, typically involving distant areas of the brain such as the entorhinal cortex and medial temporal lobe and progressing as the disease advances [9]. The structural changes in Alzheimer\u0026rsquo;s disease, as observed through MRI, are illustrated in \u003cstrong\u003e\u003cem\u003eFigure 1.\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFrontotemporal dementia (FTD)\u0026nbsp;\u003c/strong\u003eis characterized by the progressive degeneration of neurons in the frontal and temporal lobes of the cerebral cortex [11]. It is recognized as the most common cause of early-onset dementia, typically affecting individuals under the age of 65 [12]. To date, no disease-modifying treatment has been approved for FTD [11], making timely and accurate diagnosis of paramount clinical importance. In contrast to Alzheimer\u0026rsquo;s disease, which initially affects memory, the early manifestations of FTD primarily involve deficits in social interaction, language, and behavioral regulation [11]. Delayed or incorrect diagnosis often results in inappropriate therapeutic interventions and unfavorable outcomes [13].\u003c/p\u003e\n\u003cp\u003eWhile memory impairment is a hallmark early symptom of AD\u0026mdash;typically resulting from early damage to the hippocampus [7]\u0026mdash;this feature is usually absent in the initial stages of FTD. Nevertheless, FTD is frequently misdiagnosed as AD [14], due to overlapping clinical features such as cognitive decline, cerebral atrophy, and gradual deterioration in language, behavior, and personality [11\u0026ndash;12]. Structural MRI scans commonly reveal atrophy in the anterior regions of the frontal and temporal lobes, a neuroanatomical pattern that aids in distinguishing FTD from AD [15]. \u003cstrong\u003e\u003cem\u003eFigure 2\u003c/em\u003e\u003c/strong\u003e illustrates the structural brain changes in MRI associated with FTD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eParkinson\u0026rsquo;s disease (PD)\u0026nbsp;\u003c/strong\u003eis a progressive neurodegenerative disorder that primarily affects the motor system. Its core clinical features include resting tremor, muscular rigidity, bradykinesia, and postural instability. As the disease progresses, non-motor symptoms such as cognitive impairment, sleep disturbances, and depression also become increasingly common among patients [17]. PD typically begins after the age of 60 [18], with a reported prevalence of approximately 0.5% to 1% in individuals aged 65 to 69, increasing to around 3% in older age groups [17]. Similar to other neurodegenerative disorders, structural brain changes in PD can be observed on MRI, particularly in the substantia nigra\u0026mdash;a brain region critical for dopamine production. Degeneration in this area leads to decreased dopamine levels, which underlies many of the motor symptoms of the disease. The pattern of brain atrophy in PD is often diffuse and widespread [19]. \u003cstrong\u003e\u003cem\u003eFigure 3\u003c/em\u003e\u003c/strong\u003e present the distribution of brain atrophy in Parkinson\u0026rsquo;s disease. Up to 78% of individuals with PD may develop Parkinson\u0026rsquo;s disease dementia (PDD) during the course of the illness. It is hypothesized that the accumulation of alpha-synuclein protein contributes to a shared pathological basis between PD and dementia with Lewy bodies [3]. Speech-related abnormalities\u0026mdash;such as reduced vocal intensity and monotonic speech\u0026mdash;as well as executive dysfunction, often emerge in advanced stages. These symptoms are also observed in other forms of dementia, reinforcing the need for early cognitive monitoring and the use of differential diagnostic tools [18].\u003c/p\u003e\n\u003cp\u003eEach of these disorders requires distinct pharmacological treatments, rehabilitative approaches, and care strategies. However, the considerable overlap in clinical symptoms among them significantly complicates the diagnostic process. In some cases, this can lead to misdiagnosis, resulting not only in substantial costs for patients and healthcare systems but also delays in appropriate treatment or the administration of unsuitable interventions [20, 21]. Therefore, early, accurate, and differential diagnosis is of critical importance in the management of neurodegenerative disorders. Neuroimaging is generally considered to be a key tool in the diagnosis of cognitive disorders because it reveals structural and metabolic changes in the brain. Structural signs such as brain atrophy can be observed in these scans. However, some studies have shown that deep learning-based models are much more accurate in diagnosing these disorders through neuroimaging than human experts [22]. This may be because some patterns related to cognitive disorders are so subtle that they may be beyond the ability of human vision to perceive.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eRecent advances in artificial intelligence, especially deep learning, offer promising solutions to this challenge. Deep learning, especially in the field of computer vision, has the potential to recognize subtle patterns in neuroimaging and provide more accurate diagnoses and prognoses. The successes achieved in this field, and in particular models based on convolutional networks in the diagnosis of various diseases, indicate the high potential of this technology in improving diagnostic processes. These models have shown the ability to identify various diseases with high accuracy by using neuroimaging data such as MRI [23-27]\u003c/p\u003e\n\u003cp\u003eThe main objective of the present study was to develop and evaluate 3D convolutional neural network models based on MRI for accurate and differential diagnosis of cognitive impairments associated with Alzheimer\u0026rsquo;s disease, frontotemporal dementia, and Parkinson\u0026rsquo;s disease. In the first step, three binary classification models were designed to distinguish each disorder from a normal control group. In the second step, a unified multiclass model was developed for differential diagnosis among the three conditions. Overall, our goal was to minimize diagnostic errors and achieve a high level of accuracy and reliability to enable these models to serve as decision support tools in clinical settings. To do this, we used the entire brain volume in the form of 3D models instead of using 2D slices, which had their own preprocessing and computational challenges. To overcome the challenge of data leakage, the models were trained in a subject-based manner. In the differential diagnosis model of disorders, we in addition used maxout layers to achieve greater accuracy, which was not used in previous studies we reviewed. On the other hand, in the differential diagnosis model, to overcome the challenge of data imbalance, we used the focal loss function instead of artificial data augmentation, which was also not used in previous studies we reviewed.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eIn the end, our aim was to develop models that provide diagnoses with high accuracy, sensitivity, and as well as high reliability as decision support tools.\u003c/p\u003e\n\u003cp\u003eThe rest of the paper is organized as follows: Relevant studies in the classification of cognitive disorders are introduced in the \u0026ldquo;Related Works\u0026rdquo; section. The \u0026ldquo;Methods and Materials\u0026rdquo; section describes the dataset, preprocessing techniques and model training. The \u0026ldquo;Results\u0026rdquo; section is dedicated to the research results based on the evaluation criteria. The \u0026ldquo;Discussion\u0026rdquo; section analyzes the research and evaluates the quality of the results. Finally, the \u0026ldquo;Conclusions and Future Work\u0026rdquo; section concludes the paper and suggests directions for future research.\u003c/p\u003e"},{"header":"2. Related Works","content":"\u003cp\u003eSeveral previous studies have significantly contributed to the foundation of the present research, particularly in terms of methodological inspiration and focus on cognitive disorders that are similar to the conditions addressed in this study. These works collectively highlight the impressive accuracy of machine learning and deep learning models in diagnosing cognitive impairments using neuroimaging data. Table 1 provides an overview of the key investigations reviewed in this section.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eEtminani et al. demonstrated, in a study of 3D-CNN models, that these models possess an excellent capacity for diagnosing dementia with Lewy bodies (DLB), AD, and mild cognitive impairment (MCI) based on 18F-FDG PET data. The study included 757 cases. The suggested model attained an area under the ROC curve (AUC) of approximately 96% for DLB, 96.4% for AD, 71.4% for MCI, and 94.7% for the cognitively normal (CN), surpassing human experts according to several evaluation metrics [22].\u003c/p\u003e\n\u003cp\u003eRamazan et al. also tried to propose a model that was able to detect AD and its prodromal phases. They employed a dataset of 138 f-MRI-based cases. Three ResNet-18-based convolutional models trained on a dataset of 138 subjects were applied in the study and attained an overall classification accuracy of over 97% [28].\u003c/p\u003e\n\u003cp\u003eIn a similar study, Ahila et al. evaluated a deep learning model for AD diagnosis using 18F-FDG PET data. The dataset consisted of 635 CN and 220 AD, and the suggested model reached approximately 97% accuracy [29].\u003c/p\u003e\n\u003cp\u003eAkindele et al. also developed a hybrid deep learning framework for AD classification, consisting of 2D and 3D CNNs trained on MRI data. The model achieved 98.9%\u0026ndash;99.99% accuracy with AUC of 100% for both settings. The dataset consisted of 2D and 3D MRI scans collected from 46 AD and 23 CN subjects. Multiple MRI scans were acquired for each participant, totaling 708 images from both CN and AD [30].\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003eMa et al. focused on an approach to discriminate between FTD, AD, and CN cases using deep learning and generative adversarial networks (GANs) with structural MRI data. In this study, MRI scans were divided into 87 regions of interest (ROIs) and divided into multiscale patches by k-means clustering. From each patch, statistical features such as cortical thickness, volume, and intensity were extracted. Artificial features were generated through GANs to address data scarcity. A multilayer perceptron (MLP) was then trained on both real and artificial features to classify individuals as AD, FTD, or CN. This framework achieved an overall accuracy of 88.28% on 1954 MRI images [31].\u003c/p\u003e\n\u003cp\u003eIn a similar study for the classification of AD, FTD, and CN based on 18F-FDG PET scans, Rogeau et al. proposed a 3D CNN. The dataset consisted of 199 AD patients, 192 FTD patients, and 200 CN cases. Their proposed model achieved an accuracy of 89.8% and AUCs were 93.3% for AD, 95.3% for FTD, and 99.9% for CN. AUC besides was 93.3% for AD, 95.3% for FTD, and 99.9% for the CN [32].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor the diagnosis of PD, Esmailzadeh et al. applied a 3D CNN to a brain MRI dataset consisted of 3D scans including 452 PD patients and 204 CN individuals. Notably, the designed model yielded 100% classification accuracy in distinguishing PD from CN [33].\u003c/p\u003e\n\u003cp\u003eS. Chakraborty et al. also designed a 3D CNN model in the same type of research for diagnosing PD vs. CN from MRI data for 406 subjects with 203 PD patients and 203 CN subjects. The suggested model achieved 95.29% accuracy, average sensitivity of 0.943, and AUC score of 0.98 for both classes [34].\u003c/p\u003e\n\u003cp\u003eDyrba et al. also trained another 3D CNN on 663 MRI scans to discriminate AD and MCI. The model, which was validated on three independent cohorts, had high accuracy in discriminating AD from CN (AUC \u0026ge; 0.91) and moderate accuracy in discriminating MCI from CN (AUC \u0026asymp; 0.74). The network achieved an accuracy of 88.3% (confidence interval) in discriminating AD from CN cases [35].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eZ. Hu et al. developed an MRI-based model\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003ecombining 3D CNN and Attention mechanisms to classify AD, MCI, and CN. A total of 520 images were used: 130 AD, 260 MCI, and 130 CN. Four classification experiments were performed based on the suggested hybrid model: (1) AD vs. CN, (2) MCI vs. CN, (3) AD vs. MCI, and (4) multi-class classification of AD, MCI, and CN. The model achieved binary classification accuracies of 94.23%, 95.67%, and 97.75%, respectively, and for the multi-class task accuracy of 89.71% [36].\u003c/p\u003e\n\u003cp\u003eS. De Francesco et al. prioritized demographic, clinical, and MRI information to construct a model for the differential diagnosis of AD, FTD, DLB, and CN. The study involved 506 subjects: 110 AD, 135 FTD, 153 DLB, and 108 CN cases. MRI scans were analyzed to obtain volumetric and cortical thickness measurements and white matter lesion metrics. These imaging characteristics, along with demographic and clinical information, were fed into a support vector machine model named MUQUBIA for differential diagnosis. The most informative of all features were age, gender, dementia staging scale, and 19 neuroimaging parameters. This model yielded an overall AUC of 98% with 87.5% accuracy, 88% sensitivity, and 88% F1 score [37].\u003c/p\u003e\n\u003cp\u003eEl-Assy et al. Lastly, utilizing two interconnected yet separate convolutional models proposed an architecture that achieved over 99% accuracy in distinguishing Alzheimer\u0026rsquo;s disease from its various prodromal stages. The dataset comprised 1,296 MRI scans categorized into five groups: healthy controls, early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), AD, and general MCI cases. The proposed dual-model architecture demonstrated exceptional performance in multi-stage classification of cognitive decline [38].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn \u003cstrong\u003e\u003cem\u003eTable 1\u003c/em\u003e\u003c/strong\u003e, the accuracy criteria for the models used in related works can be seen along with the type of disorders under investigation, the modality used and the number of subjects were summarized.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eTable 1\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e.\u003c/strong\u003e overview of the key investigations reviewed in this paper.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"642\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eStudy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDisorders\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModality\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSubjects #\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eK. Etminani et al. [\u003c/strong\u003e22\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e78.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e\u0026nbsp;(DLB, AD, MCI, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003ePET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e757\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF. Ramzan et al. [\u003c/strong\u003e28\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e97.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(CN, SMC, EMCI, MCI, LMCI, AD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003efMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e138\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eA. Ahila et al. [\u003c/strong\u003e29\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e96.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(AD, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003ePET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e855\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eR. G. Akindele et al. [\u003c/strong\u003e30\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e99.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(AD, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e69\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eD. Ma et al. [\u003c/strong\u003e31\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e88.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e\u0026nbsp;(AD, FTD, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e1954\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eA. Rogeau et al. [\u003c/strong\u003e32\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e89.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(AD, FTD, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003ePET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e496\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eS. Esmaeilzadeh et al. [\u003c/strong\u003e33\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(PD, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e656\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eS. Chakraborty et al. [\u003c/strong\u003e34\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e95.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(PD, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e406\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eM. Dyrba et al. [\u003c/strong\u003e35\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003eAD vs CN:\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;75.5\u0026ndash;88.3\u003c/p\u003e\n \u003cp\u003eMCI vs CN:\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;63.1\u0026ndash;75.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(AD, MCI, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e663\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eZ. Hu et al. [\u003c/strong\u003e36\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003eAD vs NC: 94.23\u003c/p\u003e\n \u003cp\u003eMCI vs NC: 95.67\u003c/p\u003e\n \u003cp\u003eAD vs MCI: 97.75\u003c/p\u003e\n \u003cp\u003eAD vs MCI vs NC: 89.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(AD, MCI, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e520\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eS. De Francesco et al. [\u003c/strong\u003e37\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e87.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e\u0026nbsp; (AD, FTD, DLB, CN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e506\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.1433%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eA. M. El-Assy et al. [\u003c/strong\u003e38\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7819%;\"\u003e\n \u003cp\u003e99.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.5794%;\"\u003e\n \u003cp\u003e(CN, EMCI, LMCI, MCI, AD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.215%;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10.2804%;\"\u003e\n \u003cp\u003e1296\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAD, Alzheimer\u0026rsquo;s disease; FTD, Frontotemporal dementia; DLB, dementia with Lewy bodies; PD, Parkinson\u0026rsquo;s disease; MCI, mild cognitive impairment; CN, cognitively normal; EMCI, early mild cognitive impairment; LMCI, late mild cognitive impairment\u0026nbsp;\u003c/p\u003e"},{"header":"3. Materials and Methods","content":"\u003cp\u003e\u003cstrong\u003e3.1. Dataset\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the current study, a dataset of subject-based T1-weighted MRI images was utilized for diagnosing AD, FTD, and PD from the CN. This allows for both differentiation of each disorder from the control group and differential diagnosis between the disorders. The sample size available in the dataset here employed is 1,629 images of the classes, of which 20% were used for testing and 80% for training. The distribution of data for all the classes is presented in \u003cstrong\u003e\u003cem\u003eFigure 4\u003c/em\u003e\u003c/strong\u003e. The CN groups are\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eshown\u0026nbsp;in blue in the figure, while yellow indicates the disorders. The data utilized were from trusted databases within the neuroimaging discipline. These include the ADNI\u003ca href=\"#_ftn1\" name=\"_ftnref1\" title=\"\"\u003e\u003c/a\u003e\u003csup\u003e1\u003c/sup\u003e, NIFD\u003ca href=\"#_ftn2\" name=\"_ftnref2\" title=\"\"\u003e\u003c/a\u003e\u003csup\u003e2\u003c/sup\u003e, and PPMI\u003ca href=\"#_ftn3\" name=\"_ftnref3\" title=\"\"\u003e\u003c/a\u003e\u003csup\u003e3\u003c/sup\u003e databases, which are among the largest and most reliable datasets for AD, FTD, and PD, respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2. Data Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor preparing brain imaging data for deep learning models, specifically convolutional neural networks (CNNs), the pipeline started by converting medical imaging files from DICOM to NIfTI format with the help of mricron software. As the NIfTI file format is compact and simplified, it is more suitable for handling large-scale MRI datasets and hence efficient storage and processing.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eThe second step was image registration, where all MRI scans are spatially registered to a common coordinate frame. This alignment ensures that corresponding anatomical points across different subjects match in spatial coordinates, which is essential for conducting comparative and group-level analyses. In this study, registration was performed using the FSL software suite by aligning each image to the \u003cstrong\u003eMNI152-T1-1mm.nii.gz\u003c/strong\u003e template\u0026mdash;a widely adopted standardized brain atlas generated by averaging T1-weighted scans from 152 healthy individuals [39]. By mapping all patient images into this common space, variations in anatomical scale, orientation, and positioning across individuals are minimized, thereby improving the reliability of voxel-wise analysis and model training. This standardization is particularly important in neuroimaging studies where structural variability may obscure disease-specific features. The use of robust registration pipelines, as surveyed in [40], enhances the reproducibility and interpretability of deep learning models by reducing non-pathological variance. Next, \u003cstrong\u003ebrain extraction\u003c/strong\u003e was performed using the Brain Extraction Tool (BET) in FSL [41]. This step removes non-brain elements such as the skull, scalp, and surrounding tissues, isolating only the brain region. Brain extraction is particularly crucial for training CNNs for several reasons. First, eliminating extraneous tissue reduces background noise, resulting in cleaner input data. Second, it sharpens the model\u0026rsquo;s focus on relevant neurological features while filtering out distractions. Third, simplifying the input structure improves model efficiency and training speed. Fourth, reducing irrelevant variability helps minimize overfitting and enhances the model\u0026rsquo;s ability to generalize. Finally, removing non-brain data decreases the overall computational load, optimizing performance on large datasets.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eFigure 5 provides a visual example of the brain extraction process.\u003c/p\u003e\n\u003cp\u003eThe subsequent step in the preprocessing pipeline involved \u003cstrong\u003eimage cropping\u003c/strong\u003e, a crucial operation aimed at reducing data volume and eliminating redundant regions beyond the brain. By trimming empty peripheral space surrounding the brain, this step ensures that the input data is spatially focused on anatomically relevant structures. Cropping not only minimizes file size and memory usage but also accelerates processing and model training. More significantly, it enhances model performance by removing background noise and irrelevant data from non-brain tissues. The cropping was followed by down-sampling to decrease the spatial resolution of the MRI scans further. This was a necessary step for computation and model optimization. All images were resampled into a 1.5 mm voxel in the three spatial dimensions (x, y, z). This resolution was chosen for two main reasons: first, increasing voxel size reduces the total number of voxels and thus data volume; second, it maintains reasonable image quality, balancing compression with retention of important neuroanatomical details. Following resampling, each scan was resized to (101, 103, 127) voxels to ensure dataset uniformity while preserving sufficient structural integrity for analysis. Intensity normalization was the last part of the MRI preprocessing procedure, an important process involving the minimization of intensity scale variation across images. Variation often occurs as a result of differing scanner settings, acquisition parameters, or physiological differences between subjects. To counteract this, voxel intensities were renormalized to a common range via Min-Max normalization, in which the minimum and maximum values of each image are mapped to 0 and 1, respectively. Intensity normalization, a widespread practice in neuroimaging research [42], decreases the model\u0026rsquo;s sensitivity to irrelevant intensity differences, increases class separability, and facilitates generalization to new unseen data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3. Model Training\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis section outlines the training procedures for the models developed to detect AD, FTD, and PD. The training process begins with the development of \u003cstrong\u003ebinary classification models\u003c/strong\u003e, each designed to distinguish a specific disorder from the cognitively normal (CN) group. Following that, a \u003cstrong\u003emulticlass classification model\u003c/strong\u003e is introduced. This comprehensive model aims to perform \u003cstrong\u003edifferential diagnosis\u003c/strong\u003e across all three conditions, enabling the classification of subjects into one of the neurodegenerative categories based on MRI data.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003cstrong\u003e3.3.1. Binary Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe AD detection model takes volumetric MRI data as input and employs three-dimensional convolutional layers to learn spatial features in all directions, allowing it to capture subtle, disease-specific structural patterns in the brain. Figure 6 shows the model architecture overview, along with the output shape and number of parameters of each layer. To accompany this, Figure 7 displays the hierarchical organization of the network, depicting the step-by-step development of information from raw input images to ultimate classification. The model\u0026rsquo;s input is a 3D MRI scan of dimensions (103, 127, 101, 1), where the final entry represents a single grayscale channel. The initial convolutional layer consists of 32 filters of size 3\u0026times;3\u0026times;3, activated by the ReLU function. This setup enables the model to learn simple features like edges and texture patterns, without keeping the number of trainable parameters high at this point. The ensuing feature maps are batch normalized to stabilize learning and speed up convergence. Then, 3D max-pooling with kernel size 2\u0026times;2\u0026times;2 is applied to downsample the spatial resolution. This pooling layer reduces each axis\u0026rsquo;s dimensions by half, reducing memory consumption and the risk of overfitting. The second convolutional layer doubles the filter count to 64, allowing the same kernel size. Doubling allows the model to learn more complex spatial features. Another 2\u0026times;2\u0026times;2 pooling layer down-samples the resulting feature maps, preserving the trade-off between feature abstraction and computational efficiency. The third convolutional block has 128 filters with again 3\u0026times;3\u0026times;3 kernels and ReLU activation. The network is learning abstract and high-level features of brain morphology at this stage. A final max-pooling layer further reduces the dimensions of feature maps before the fully connected layers. The output from the last pooling layer is input into a flattening layer, which transforms the 3D feature maps into a one-dimensional vector for dense layers. The first dense layer consists of 128 neurons with ReLU activation to enable the model to learn complex combinations of the features extracted. A Dropout layer is then added to avoid overfitting by randomly dropping out 30% of the neurons during training. Later dense layers have 64 and 32 neurons, respectively\u0026mdash;both with ReLU activation. This progressive decrease in dimensionality aids in distillation and summarization of the feature representation without sacrificing non-linearity. Lastly, the output layer has one neuron with a Sigmoid activation function that outputs a probability score between 0 and 1. The score is indicative of the model\u0026rsquo;s confidence that the input scan is an Alzheimer\u0026rsquo;s case or a normal control.\u003c/p\u003e\n\u003cp\u003eTraining was limited to 100 epochs. Early stopping was used to prevent overfitting, a typical issue with deep learning. Training would stop automatically if validation loss would experience no improvement in a streak of 10 epochs. This way, it is ensured that the model is not learning from noise or spurious patterns anymore after substantial performance improvements have leveled off. Moreover, the best weights of the model (on the validation set) during training were saved, maintaining the optimum state and avoiding subsequent degradation in performance. A mini-batch size of 4 was employed to trade off between two significant considerations: memory limitations and stability of learning. Smaller batches allow for more frequent model weight updates, which can improve generalization and provide finer-grained convergence, particularly when dealing with high-dimensional 3D neuroimaging data. This alternative also prevented GPU memory overflow, which was an important factor during training on computationally demanding models. Due to the high computational requirements of training 3D CNNs\u0026mdash;stemming from their volumetric input and extensive parameter space\u0026mdash;the experiments were run on an NVIDIA A100 GPU with more than 80 GB of RAM. This setup provided the necessary memory capacity and processing power. Initial experiments also verified that smaller batch sizes performed better than larger ones, possibly because of more accurate gradient updates and lesser overfitting. Binary Cross-entropy was used as the loss function to quantify the divergence between predicted probabilities and true binary labels for this binary classification problem (AD vs. CN). The Adagrad optimizer was used to optimize the models. This option was especially appropriate for our architecture, which has more than 25 million parameters. Adagrad adaptively learns rates for each parameter from previous gradients, allowing more stable and faster convergence. One of the main strong points of Adagrad is that it eliminates steep oscillations in gradient updates. By accumulating squared gradients over time, it reduces the learning rate for frequently updated parameters, and increases it for less frequently updated ones. This adaptiveness not only stabilizes training but also reduces the necessity of carrying out extensive manual learning rate tuning\u0026mdash;particularly useful in complicated 3D-based architectures. Although other widely used optimizers like Adam, RMSprop, and SGD were also experimented with, Adagrad was more stable, with a better final classification accuracy. Its adaptiveness to gradient volatility was especially helpful for the complexity and high dimensionality of our model.\u003c/p\u003e\n\u003cp\u003eIn Figure 7, the architecture of the model has been visualized. As it can be seen, the whole volume of the brain is entered into the model in 3D form as an input layer, and in each convolution layer, the 3D volumes of the next layers are created by using 3D filters.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003eThe model structure and methodology applied to the FTD detection model are the same as those used for AD detection. Nevertheless, the key difference lies in the composition of the control group. In this case, a merged set is utilized that includes the CN cases of the AD model along with control data for FTD. This integration strategy was pursued to enhance the heterogeneity of the control group as well as offset the limited number of FTD-specific control samples. While the AD model served as the reference design, the FTD detection model was intentionally simplified. The largest modification occurred in the Dense (fully connected) layers, where the number of neurons in each layer was reduced by half. This modification significantly decreased the total count of trainable parameters and lessened the computational burden. More importantly, the simplified network achieved improved detection performance for FTD. The simplified network was better able to capture discriminative features for FTD. \u003cstrong\u003eFigure\u003cem\u003e\u0026nbsp;\u003c/em\u003e8\u003c/strong\u003e summarizes the structure of the model, including output dimensions and parameter counts for each layer. \u003cstrong\u003eFigure 9\u003c/strong\u003e provides a visual representation of the hierarchical architecture, illustrating the progression of input data through the network to the final classification output.\u003c/p\u003e\n\u003cp\u003eAs can be seen in this figure, the number of parameters compared to the Alzheimer\u0026rsquo;s disease diagnosis model is almost halved.\u003c/p\u003e\n\u003cp\u003eThe 3D CNN architecture employed for detecting PD was the same as the AD detection model. The PD model directly took the design rationale, layering, and overall architecture from the AD model, which was the main reference architecture. This modeling consistency provides a consistent methodological foundation within experiments and allows comparative analysis among the various diagnostic tasks. Aside from architectural consistency, the same training setups\u0026mdash;loss function, optimizer, batch size, early stopping condition, and hardware configuration\u0026mdash;were utilized as those of the earlier models. This methodological consistency was required both to maintain experimental rigor and to be able to reasonably compare model performances across different neurological disorders.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3.2. Differential Diagnosis: An Integrated Unified Multiclass-Diagnosis Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis section focuses on a single deep learning model for the differential diagnosis of three major neurodegenerative diseases\u0026mdash;AD, FTD, and PD\u0026mdash;and the CN group. The overall goal was to train a single model that could take 3D MRI data as input and classify a subject into one of four diagnostic groups.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eThis multiclass architecture was specifically designed to address the challenges posed by overlapping clinical symptoms, which sometimes complicate traditional diagnostic processes. By consolidating all classes into a single predictive system, the model not only aims to improve diagnostic efficiency, but also \u003cstrong\u003ereduces the risk of misclassification\u003c/strong\u003e that can arise in binary models. This unified strategy lays the groundwork for more robust, scalable, and clinically practical tools for distinguishing between related neurodegenerative conditions in real-world settings.\u003c/p\u003e\n\u003cp\u003eThe test training sets of this multiclass model were formed by merging the data that used in the three aforementioned binary classification models. All samples were combined in one dataset. The Differential Diagnosis model has some basic differences from the binary classification models. Rather than one output neuron with the Sigmoid activation function, this architecture has four output neurons with the Softmax activation function, enabling the model to predict probabilities of all classes simultaneously in a forward pass. Furthermore, the network architecture substitutes standard Dense layers with Maxout layers, which had greater representational power for features. Unlike regular activation functions that calculate one value per unit, Maxout layers calculate several parallel channels (four in our case per unit) and return the highest value among them. This adaptivity enabled the network to learn more complex features. However, it caused computational cost: Maxout layers drastically increase the number of parameters, which can affect both model complexity and memory usage. In order to manage class imbalance, a prevalent problem in multiclass medical imaging problems, the model uses the Focal Loss function. The loss function was particularly designed to diminish the contribution of dominant classes and direct learning towards more difficult, minority examples. The \u0026alpha; (alpha) coefficients were set to be inversely related to class frequencies so that minority classes would have higher contributions to the loss function. The \u0026gamma; (gamma) was finely adjusted to focus on mislabeled or confusing samples. This Focal Loss variant brought tremendous performance gains. Not only did it suppress overfitting and improve generalization, but it also helped the model achieve balanced performance across all four classes, particularly underrepresented classes. Simultaneous adjustment of \u0026alpha; and \u0026gamma; was instrumental in offering both high accuracy and strong class-wise distinction. \u003cstrong\u003e\u003cem\u003eEquation 1\u003c/em\u003e\u003c/strong\u003e presents the general form of the Focal Loss function as applied in this study [43].\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1762157155.png\" width=\"692\" height=\"52\"\u003e\u003c/p\u003e\n\u003cp\u003eAlong with its architectural and loss function innovations, the multiclass model also benefited from the Exponential Linear Unit (ELU) for the activation function for convolutional layers. ELU was used instead of ReLU as it can learn to handle negative activations and also produce smoother and more continuous gradients. This also had the secondary effect of improved training stability and convergence properties, especially in the more challenging multiclass scenario. ELU was used right after batch normalization and convolutional layers to provide efficient non-linear feature map transformation. The optimizer, batch size, learning schedule, and early stopping were kept identical to the binary models for methodological consistency across experiments. The model architecture summary shows a dramatic rise in complexity. In particular, there are over 100 million parameters to train, most of which are in the Maxout layers. While this is a large computation, we believe it is proportionate to the high dimensionality of the input (3D MRI volumes) and the need for accurately discrimination between four diagnostic categories. Figure 11 provides a visual representation of the model\u0026rsquo;s hierarchical structure, demonstrating the progression of data from raw input to final multiclass output.\u003c/p\u003e\n\u003cp\u003eAs it can be seen, in this model, the ELU activation function and Maxout layers was used, and for this reason, the number of parameters is much more than the previous models.\u003c/p\u003e\n\u003cp dir=\"\"\u003eIn this visualization, like the architecture of previous models, the entire brain volume was used to train the model, but to differentiate among disorders, the last layer in the dense layers had 4 outputs that belong to each of the disorders.\u003c/p\u003e"},{"header":"4. Results","content":"\u003cp\u003eThis section is dedicated to evaluating and analyzing the results obtained from the diagnostic models, including separate models for AD, FTD, and PD. Evaluations are presented using confusion matrices and quantitative indicators such as accuracy, precision, recall (sensitivity), which reveal how well the models perform in classifying new data. Following that, the performance of the unified multiclass model, designed for differential diagnosis among all disorders, is examined, highlighting its accuracy in distinguishing each class from the others.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1. Binary Classification Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAs illustrated in \u003cstrong\u003e\u003cem\u003eFigure 12\u003c/em\u003e\u003c/strong\u003e, the confusion matrix for the Alzheimer\u0026rsquo;s disease (AD) detection model shows exceptional classification accuracy. All 77 AD cases were correctly identified, while only \u003cstrong\u003e4 out of 89\u003c/strong\u003e control samples were misclassified. This yielded an \u003cstrong\u003eoverall accuracy of 98%\u003c/strong\u003e and an \u003cstrong\u003eF1-score close to 1\u003c/strong\u003e, reflecting both high sensitivity and precision.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eOf particular note is the model\u0026rsquo;s \u003cstrong\u003eperfect recall (1.00)\u003c/strong\u003e for the AD class, meaning no Alzheimer\u0026rsquo;s cases were missed. Additionally, the \u003cstrong\u003eprecision for the control group reached 1.00\u003c/strong\u003e, indicating zero false positives. These results demonstrate the model\u0026rsquo;s strong ability to cleanly separate the AD and CN classes. The detailed performance metrics are presented in \u003cstrong\u003e\u003cem\u003eTable 2\u003c/em\u003e\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eTable 2.\u003c/em\u003e\u0026nbsp;\u003c/strong\u003eClassification performance of the AD diagnosis model\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eClass\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSupport\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e77\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e89\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eOverall Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.98\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMacro Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e166\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eWeighted Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e166\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eRegarding the FTD classification model, as shown in the confusion matrix in Figure 13, the model demonstrated strong performance in distinguishing FTD from the CN group. This was despite a small number of misclassifications (3 cases in the FTD class). Out of 37 FTD cases, 34 were correctly identified, and all control group cases were classified correctly. An overall accuracy of 98% indicates excellent model performance. The precision of 1.00 means that all instances predicted as FTD were indeed FTD cases\u0026mdash;none of the control samples were misclassified as FTD. On the other hand, the model achieved a recall of 92% for the FTD class, reflecting a relatively good sensitivity. The model also showed excellent performance in identifying control samples, with a 100% recall, indicating it correctly detected all normal control cases. \u003cstrong\u003e\u003cem\u003eTable 3\u003c/em\u003e\u003c/strong\u003e summarizes the classification performance of the FTD detection model based on various evaluation metrics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eTable 3.\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eClassification performance of the FTD model\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eClass\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSupport\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFTD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e37\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e112\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eOverall Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.98\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMacro Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e149\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eWeighted Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e149\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe confusion matrix in \u003cstrong\u003eFigure 14\u003c/strong\u003e demonstrates that PD diagnosis model also performed well, albeit with slightly more misclassifications. Out of 62 PD samples\u003cstrong\u003e,\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e60 were correctly identified\u003c/strong\u003e, while \u003cstrong\u003e2 were misclassified as controls\u003c/strong\u003e. Among 36 CN samples, \u003cstrong\u003e5 were misclassified as PD\u003c/strong\u003e, resulting in an \u003cstrong\u003eoverall accuracy of 93%\u003c/strong\u003e.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eThe model maintained a \u003cstrong\u003ehigh recall of 97%\u003c/strong\u003e for the PD class, indicating strong sensitivity and few false negatives. However, the slightly lower specificity in the CN class (86%) suggests room for improvement in reducing false positives. Full performance metrics are provided in \u003cstrong\u003eTable 4\u003c/strong\u003e\u003cem\u003e.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eTable 4.\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eClassification performance of the PD model\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eClass\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSupport\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e62\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eOverall Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.93\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMacro Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eWeighted Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003e4.2. Differential Diagnosis: The Integrated Unified Multiclass-Diagnosis Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFrom the training curves of the combined model in \u003cstrong\u003e\u003cem\u003eFigure 15\u003c/em\u003e\u003c/strong\u003e, it is observed that the model converged after approximately 50 epochs. The smallest difference between training and validation accuracy indicates the absence of overfitting. The graphs indicate that despite a large number of parameters of the model (over 100 million), it converged using techniques such as Focal Loss, early stopping, and a very small batch size. The accuracy of training data improved step by step over the course of training, to nearly 99%, demonstrating that the model effectively acquired the training set characteristics while consistently reducing error. The validation accuracy trend was similar and reached to 95%. Some sharp spikes of validation accuracy in certain epochs are to be anticipated with Focal Loss since changes in the distribution of error\u0026mdash;especially underrepresented classes\u0026mdash;can lead to dramatic shifts in gradient behavior. Training loss decreased steadily to a minimal value, and validation loss also showed a decreasing trend. Absence of sudden loss spikes can be interpreted to mean training stability. Although the disparity between training and validation loss at times seemed large, it dropped near the final epochs, indicating greater model stability.\u003c/p\u003e\n\u003cp\u003eAs illustrated by \u003cstrong\u003e\u003cem\u003eFigure 16\u003c/em\u003e\u003c/strong\u003e, the differential diagnosis model\u0026rsquo;s confusion matrix demonstrates high performance in detecting all three neurodegenerative diseases, with high accuracy in classifying normal control. Notably, all 77 cases of AD were accurately distinguished, with perfect recall (1.00) for this class. 2 out of 37 FTD cases and 3 out of 62 PD instances were misclassified, a total of 171 correct predictions among 176 cases of diseases. The model achieved an overall classification accuracy of 95%, marking its stability in processing a multiclass neuroimaging task. One of the strongest findings was the model\u0026rsquo;s consistently high specificity in all classes, indicating a low rate of false positives\u0026mdash;a primary requirement in clinical application where false diagnosis can lead to inappropriate or delayed treatment. Even when there was intrinsic class imbalance, particularly between the CN and disorder classes, use of Focal Loss played a crucial role in maintaining balance in the learning process. It allowed the model to focus more on challenging-to-classify and minority samples, in essence preventing the network from becoming biased towards majority classes. Further, the model possessed high precision for the CN and FTD classes, which signifies its ability to detect subtle differences in structural brain characteristics. This is especially important in distinguishing overlapping clinical characteristics that can be common in neurodegenerative diseases. Collectively, these results show that the proposed multiclass model is a safe and generalizable differential diagnosis scheme with great promise for practical application in real-world clinical decision support systems. A complete summary of classification metrics is provided in \u003cstrong\u003e\u003cem\u003eTable 5\u003c/em\u003e\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eTable 5.\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eClassification performance of the unified model\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eClass\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSupport\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e148\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e77\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFTD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e37\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e62\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eOverall Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.95\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMacro Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e324\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eWeighted Avg\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e324\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTo facilitate a complete visualization of the model\u0026apos;s diagnostic performance, a combined ROC curve was plotted for all four classes, as illustrated in Figure 17. In this case, a one-vs-rest strategy was adopted, where each class was considered the positive class against all others. For example, when testing the detection capability of the model for the CN class, all samples that were not CN were considered negative, and CN samples positive.\u003cspan dir=\"RTL\"\u003e\u0026nbsp;\u003c/span\u003eFor each class, the true positive rate (TPR) and false positive rate (FPR) were calculated at various decision thresholds to construct the respective ROC curves. The area under the curve (AUC) was then computed to quantify the model\u0026rsquo;s ability to discriminate each class from the rest. As opposed to accuracy, which is threshold-dependent and possibly misleading in imbalanced datasets, AUC estimates classifier performance in a threshold-independent fashion, and is especially useful in multiclass medical diagnosis. This is based on prominent work in machine learning, which has shown that AUC estimates classification performance more reliably and informatively than accuracy alone [44]. The AUC values close to 1 for each class confirm the model\u0026rsquo;s very high ability for accurate differential diagnose.\u003c/p\u003e"},{"header":"5. Discussion","content":"\u003cp\u003eOne of the prime priorities in current research was the avoidance of data leakage. This can be a serious problem in neuroimaging research because it may bias model assessment and lead to artificially inflated accuracy. To avoid this, all data splitting was at the subject level in the present study, rather than the slice or image level. This ensured that no MRI scans of a single subject were in both test and train sets\u0026mdash;something that is a typical error in longitudinal investigations wherein multiple scans are collected from a single subject at different time points. In deep learning research, data leakage is a concept that refers to the unwanted use of test data features in the model training process. It is most classically evident in data split at the level of the image, where multiple images of a single subject appear in train and test datasets. The result is a biased impression of model performance since the network can learn to remember the details of single subjects rather than generalized features of the disorder. This is well supported by current systematic reviews, one of which estimated that nearly half of Alzheimer\u0026rsquo;s diagnosis based on deep learning researchs are contaminated with data leakage, heavily compromising their clinical utility [45].\u003c/p\u003e\n\u003cp\u003eIn our initial experiments, models were trained on longitudinal data without strict separation by subjects. Even with simpler models and no special layers like Maxout or Focal Loss, they still achieved near-perfect classifications in spite of this. Our hypothesis was that the model was more sensitive to the overall structure of individuals\u0026rsquo; brains, rather than learning disease-specific patterns, because in the longitudinal data, the brain structure of an individual is repeated across multiple scans. This outcome had two consequences. On the positive side, it represented the model\u0026rsquo;s amazing capacity in picking up unique anatomical features of a single brain\u0026mdash;suggesting a level of accuracy perhaps beyond human capability in detecting such inter-subject differentials. On the other hand, it also suggested the model learned non-generalizable disease-related features. In fact, the model had learned subject-specific anatomical features, not just disease-related ones. The results showed that the accuracy of the models decreased significantly compared to the previous case using longitudinal data. This decrease confirmed the assumption that data leakage in the previous method had caused false accuracy. Finally, in this study, to ensure that data leakage was prevented, we decided to conduct all experiments during training and testing.\u0026nbsp;Specifically, we avoided placing different scans from the same individual in the training and testing sets. We also took great care to avoid the use of synthetic data augmentation, so models learned from actual patient scans rather than replicated or imitated patterns.\u003c/p\u003e\n\u003cp\u003eHaving mitigated data leakage, class imbalance presented the next major challenge, further compounded by relatively lower FTD data availability compared to AD or PD. To address this, we employed the Focal Loss function, which allows the model to learn from hard-to-classify and underrepresented classes by automatically balancing the contribution of each sample to the loss function. The \u0026alpha; and \u0026gamma; hyperparameters were optimally tuned to balance the contribution of the smaller and larger classes. We also employed Maxout layers in the model. These contributed extra architectural complexity and computational overhead\u0026mdash;bringing the number of model parameters in the multiclass model to in excess of 100 million\u0026mdash;though enhanced the model\u0026rsquo;s ability to extract meaningful patterns from complex and sparse data.\u003c/p\u003e\n\u003cp\u003eAnother contribution of this work is the strict preprocessing pipeline, in particular the addition both of the brain extraction and image registration. Image registration established a shared spatial reference in the MRI scans and minimized the effects of position inconstancy and orientation variation of the head across subjects. Otherwise, models would be prone to learning irrelevant spatial noise instead of pathology. Brain extraction served further to ensure that models were trained only on intracranial tissue and not noise from non-brain tissue like the scalp and the skull. These preprocessing steps are sometimes omitted in other studies or are often not applied together. However, their implementation can be essential in improving model focus and reducing bias\u0026mdash;especially in high-dimensional 3D data where each voxel carries a weight in the learning process.\u003c/p\u003e\n\u003cp\u003eWhen we compare our results to similar research, several unique features of our methodology\u0026rsquo;s are apparent. Some previous works employed PET scans, which\u0026mdash;while effective\u0026mdash;are more intrusive, requiring radioactive tracers to be injected [46]. In contrast, our technique uses only T1-weighted MRI, a safer, more general modality. In addition, some previous works truncated 3D MRI volumes to 2D slices in order to increase their number of examples. While this technique inflates the number of examples, it probably causes heavy leakage of data\u0026mdash;especially where multiple slices per subject are found to appear in both the train and test folds. In addition, 2D slices cannot model the brain as a volume, and may fail to pick up on important structural details in unsampled regions. In contrast, our 3D CNN model takes the entire brain volume as input, and is therefore able to process all neuroanatomical changes. While computationally less efficient, this results in richer and better spatially-coherent models of pathology. In addition, we avoided using synthetic augmentation techniques that might introduce bias or lead to learning artificially enhanced features. Instead, we opted for real, subject-level MRI data, taking care not to let any single subject contribute multiple times to the model. In particular, while previous works tried to address class imbalance by data augmentation, or by using synthetically image generation, our Focal Loss-based model\u0026mdash;a more elegant solution that involves no data manipulation\u0026mdash;has no equivalent in previous works that were compared in the present study. Although under tighter restrictions\u0026mdash;imbalance in the data, no data augmentation, subject-based design, and 3D volumetric analysis\u0026mdash;our model performed comparably, and in most cases, even better in terms of sensitivity and generalizability. In summary, design decisions for this study were made to make it useful for clinical applications. From strict data manipulation procedures to advanced loss function and preprocessing methods, every aspect was chosen in a careful effort to ensure that the model\u0026rsquo;s diagnostic accuracy was not merely statistically impressive but truly clinically significant, reliable, and practical. Table 6 is the comparison of this study with works of relevance. \u003cstrong\u003e\u003cem\u003eTable 6\u003c/em\u003e\u003c/strong\u003e summarizes the comparison of the present study with related works.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eTable 6.\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eComparative table of the present study with related works\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"642\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eStudy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;(%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModality\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSubjects\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e#\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eKey Distinctions\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePresent Study (Unified Model)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=100\u003c/p\u003e\n \u003cp\u003eFTD=95\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;PD=95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e99\u0026ndash;100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e1629\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- Brain extraction and image registration applied\u003c/p\u003e\n \u003cp\u003e- Strict subject-level partitioning to prevent data leakage\u003c/p\u003e\n \u003cp\u003e- Use of Focal Loss\u003c/p\u003e\n \u003cp\u003e- Use of Maxout layers\u003c/p\u003e\n \u003cp\u003e- Full 3D brain volume analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eK. Etminani et al. [\u003c/strong\u003e22\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e78.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=91\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eMCI=17\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;DLB=86\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e71.4\u0026ndash;96.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003ePET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e757\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- Requires radioactive tracer injection (PET)\u003c/p\u003e\n \u003cp\u003e- Significantly lower accuracy\u003cbr\u003e\u0026nbsp;- lower sensitivity in AD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF. Ramzan et al. [\u003c/strong\u003e28\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e97.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=98.01 MCI=97.81 LMCI=97.43 EMCI=97.38 SMC=100\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e99.96\u0026ndash;99.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003efMRI/2D\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e138 \u0026rarr; 850,080 2D images\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- 3D scans converted to a large number of 2D slices\u003c/p\u003e\n \u003cp\u003e- No subject-based split or slice duplication control\u003c/p\u003e\n \u003cp\u003e- No registration or brain extraction applied\u003cbr\u003e\u0026nbsp;- lower sensitivity in AD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eA. Ahila et al. [\u003c/strong\u003e29\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e96.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003ePET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e855\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- PET required\u003c/p\u003e\n \u003cp\u003e- 3D scans converted to 2D\u003c/p\u003e\n \u003cp\u003e- No brain extraction\u003c/p\u003e\n \u003cp\u003e- Lower accuracy\u003cbr\u003e\u0026nbsp;- lower sensitivity in AD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eR. G. Akindele et al. [\u003c/strong\u003e30\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e99.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e69 subjects (708 scans)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- No registration or brain extraction\u003c/p\u003e\n \u003cp\u003e- Performed augmentation\u003c/p\u003e\n \u003cp\u003e- 3D images reconstructed from 2D slices\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eD. Ma et al. [\u003c/strong\u003e31\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e88.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=84.66 FTD=77.82\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e1954\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- Lower accuracy\u003c/p\u003e\n \u003cp\u003e- Patch-based feature extraction\u003c/p\u003e\n \u003cp\u003e- Synthetic data generated using GAN\u003cbr\u003e\u0026nbsp;- Significantly lower accuracy\u003cbr\u003e\u0026nbsp;- lower sensitivity in AD and FTD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eA. Rogeau et al. [\u003c/strong\u003e32\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e89.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=75\u003c/p\u003e\n \u003cp\u003eFTD=95\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e93.3\u0026ndash;99.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003ePET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e496 (591 scans)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- Lower accuracy\u003c/p\u003e\n \u003cp\u003e- PET-based\u003c/p\u003e\n \u003cp\u003e- Unequal scan distribution across subjects\u003cbr\u003e\u0026nbsp;- lower sensitivity in AD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eS. Esmaeilzadeh et al. [\u003c/strong\u003e33\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003ePD=100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e656\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- No image registration\u003c/p\u003e\n \u003cp\u003e- Applied data augmentation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eS. Chakraborty et al. [\u003c/strong\u003e34\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e95.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003ePD=94.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e406\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- No brain extraction\u003c/p\u003e\n \u003cp\u003e- Lower sensitivity\u003c/p\u003e\n \u003cp\u003e- Balanced dataset selection\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eM. Dyrba et al. [\u003c/strong\u003e35\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eAD vs CN: 75.5\u0026ndash;88.3\u003c/p\u003e\n \u003cp\u003eMCI vs CN: 63.1\u0026ndash;75.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026ndash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003eAD vs CN: 82.8\u0026ndash;97.8MCI vs CN: 66.7\u0026ndash;84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e663\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- Lower accuracy\u003c/p\u003e\n \u003cp\u003e- Only gray matter used\u003c/p\u003e\n \u003cp\u003e- Data augmentation applied\u003c/p\u003e\n \u003cp\u003e- Significantly lower accuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eZ. Hu et al. [\u003c/strong\u003e36\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eAD vs NC: 94.23\u003c/p\u003e\n \u003cp\u003eMCI vs NC: 95.67\u003c/p\u003e\n \u003cp\u003eAD vs MCI: 97.75\u003c/p\u003e\n \u003cp\u003eAD vs MCI vs NC: 89.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026ndash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e\u0026ndash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e520\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- Lower accuracy\u003c/p\u003e\n \u003cp\u003e- Only gray matter used\u003c/p\u003e\n \u003cp\u003e- Sensitivity and AUC not reported\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eS. De Francesco et al. [\u003c/strong\u003e37\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e87.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=73 FTD=89 DLB=95\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e92.09\u0026ndash;99.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e506\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- Lower accuracy\u0026nbsp;\u003cspan dir=\"RTL\"\u003e\u003cbr\u003e\u0026nbsp;\u003c/span\u003e- lower sensitivity in FTD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eA. M. El-Assy et al. [\u003c/strong\u003e38\u003cstrong\u003e]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e99.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003eAD=100\u003c/p\u003e\n \u003cp\u003eEMCI=89 LMCI=100 MCI=95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003e97.37\u0026ndash;100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eMRI/2D\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e1296\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e- No registration or brain extraction\u003c/p\u003e\n \u003cp\u003e- Class balancing via augmentation\u003c/p\u003e\n \u003cp\u003e- 2D slices extracted from 3D volumes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e(AD, Alzheimer\u0026rsquo;s disease; FTD, Frontotemporal dementia; DLB, dementia with Lewy bodies; PD, Parkinson\u0026rsquo;s disease; MCI, mild cognitive impairment; CN, cognitively normal; EMCI, early mild cognitive impairment; LMCI, late mild cognitive impairment)\u003c/p\u003e"},{"header":"6. Conclusion and Future Works","content":"\u003cp\u003eThe current study intended to examine the capabilities of convolutional neural networks in differential diagnosis of neurodegenerative diseases such as Alzheimer\u0026rsquo;s disease, frontotemporal dementia, and Parkinson\u0026rsquo;s disease based on MRI data. Considering the complex structure of the diseases and challenges in their diagnosis, the current study intended to provide a model supported by deep learning approaches that could be of high accuracy and credibility in differential diagnosis of the mentioned diseases and differentiating from one another. The results of this study indicated that the proposed models can be employed as a complementary decision-support tool in the diagnosis of neurodegenerative diseases. These models were capable to differentiate Alzheimer\u0026rsquo;s disease, frontotemporal dementia, and Parkinson\u0026rsquo;s disease from the control group with 98%, 98%, and 93% accuracy, respectively. The multiclass model in achieved 95% accuracy in differential diagnosis. Comparison with prior works revealed that the proposed model performs superior to most prior 2D and 3D approaches and in the meantime, avoided leakage of data. Ultimately, according to the results, it is possible to conclude that deep learning is much potent in analysis of medical imaging data and is capable of playing an effective role in enhancing neurological diseases\u0026rsquo; diagnosis methods, notably Alzheimer\u0026rsquo;s disease, frontotemporal dementia, and Parkinson\u0026rsquo;s disease. The current study demonstrated that 3D convolutional neural networks are capable of analyzing information from the images with precise accuracy and differentiating several diseases with good quality. This enables doctors to make optimum diagnostic decisions and place patients on a suitable treatment path.\u003c/p\u003e\n\u003cp\u003eBased on the findings of this study and the challenge of differential diagnosis of neurodegenerative diseases, one of the most critical research and optimization directions suggested in future work is determining the stage of disease development. The models presented in this study focused on differential diagnosis between neurodegenerative diseases, but future work can separate patients at different stages of the disease based on investigating the extent of brain atrophy and shrinking in critical structures like the hippocampus and cerebral cortex. The advantages of disease staging are that AI-based models are able to measure the progression of the disease in different patients based on examining the structural pattern of the brain and help manage treatment. Incorporation of the feature of determining the stage of the disease opens the door to compare cohorts of patients and investigate the influence of different treatments on disease development.\u003c/p\u003e\n\u003cp\u003eAnother suggestion towards model upgrading is the inclusion of dementia with Lewy bodies data. As mentioned before, this form of dementia pathologically and physiologically shares some similarities with Parkinson\u0026rsquo;s disease [3]. Adding this disorder to the model, therefore, is capable of bridging the prevailing gap in disease differentiation and making the system an integrated model of differential diagnosis of neurodegenerative cognitive diseases. Although data concerning this disorder were not captured in this study due to limitations in resources and data availability, in future research, the incorporation of data from this disorder is capable of upgrading the model to be more precise in disease distinction and ready to be employed in real clinical practice.\u003c/p\u003e\n\u003cp\u003eThe proposed models here are constructed only on MRI data but in practice, cognitive tests such as Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) are also crucial for assessing neurodegenerative disease patients [47]. Including this data with MRI images will increase the accuracy and validity, especially in the differential diagnosis model.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding Declaration:\u0026nbsp;\u003c/strong\u003eNo funding was received for this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests Declaration:\u0026nbsp;\u003c/strong\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePrince M, Wimo A, Guerchet M, Gemma-Claire A, Wu YT, Prina M. World Alzheimer Report 2015: The Global Impact of Dementia - An analysis of prevalence, incidence, cost and trends [Research report]. Alzheimer\u0026rsquo;s Dis Int [Internet]. 2015. Available from: https://hal.archives-ouvertes.fr/hal-03495438\u003c/li\u003e\n\u003cli\u003eDementia: Key facts. WHO. [Internet]. 2025. Available from: https://www.who.int/news-room/fact-sheets/detail/dementia\u003c/li\u003e\n\u003cli\u003eGomperts SN. Lewy body dementias: Dementia with lewy bodies and Parkinson disease dementia. Contin Lifelong Learn Neurol. 2016;22(2, Dementia):435\u0026ndash;63. \u003c/li\u003e\n\u003cli\u003eFerri CP, Prince M, Brayne C, Brodaty H, Fratiglioni L, Ganguli M, et al. Global prevalence of dementia: A Delphi consensus study. Lancet. 2005;366(9503):2112\u0026ndash;7. \u003c/li\u003e\n\u003cli\u003eDeture MA, Dickson DW. The neuropathological diagnosis of Alzheimer\u0026rsquo;s disease. Mol Neurodegener. 2019;14(1). \u003c/li\u003e\n\u003cli\u003eCoup\u0026eacute; P, Manj\u0026oacute;n JV, Lanuza E, Catheline G. Lifespan Changes of the Human Brain In Alzheimer\u0026rsquo;s Disease. Sci Rep. 2019;9(1). \u003c/li\u003e\n\u003cli\u003eRao YL, Ganaraja B, Murlimanju B V., Joy T, Krishnamurthy A, Agrawal A. Hippocampus and its involvement in Alzheimer\u0026rsquo;s disease: a review. 3 Biotech. 2022;12(2). \u003c/li\u003e\n\u003cli\u003eJohnson KA, Fox NC, Sperling RA, Klunk WE. Brain imaging in Alzheimer disease. Cold Spring Harb Perspect Med. 2012;2(4). \u003c/li\u003e\n\u003cli\u003eDickerson BC, Wolk DA. MRI cortical thickness biomarker predicts AD-like CSF and cognitive decline in normal adults. Neurology. 2012;78(2):84\u0026ndash;90. \u003c/li\u003e\n\u003cli\u003evan Oostveen WM, de Lange ECM. Imaging techniques in alzheimer\u0026rsquo;s disease: A review of applications in early diagnosis and longitudinal monitoring. Int J Mol Sci. 2021;22(4):1\u0026ndash;34. \u003c/li\u003e\n\u003cli\u003eFinger EC. Frontotemporal dementias. Contin Lifelong Learn Neurol. 2016;22(2, Dementia):464\u0026ndash;89. \u003c/li\u003e\n\u003cli\u003eRabinovici GD, Miller BL. Frontotemporal lobar degeneration: Epidemiology, pathophysiology, diagnosis and management. CNS Drugs. 2010;24(5):375\u0026ndash;98. \u003c/li\u003e\n\u003cli\u003eBertoux M, O\u0026rsquo;Callaghan C, Flanagan E, Hornberger M. Frontotemporal Dementia (FTD). In: Pachana NA, editor. Encyclopedia of Geropsychology [Internet]. Singapore: Springer Singapore; 2015. p. 1\u0026ndash;17. Available from: https://doi.org/10.1007/978-981-287-080-3_311-1\u003c/li\u003e\n\u003cli\u003eLindau M, Almkvist O, Kushi J, Boone K, Johansson SE, Wahlund LO, et al. First symptoms - Frontotemporal dementia versus Alzheimer\u0026rsquo;s disease. Dement Geriatr Cogn Disord. 2000;11(5):286\u0026ndash;93. \u003c/li\u003e\n\u003cli\u003eBocti C, Rockel C, Roy P, Gao FQ, Black SE. Topographical patterns of lobar atrophy in frontotemporal dementia and Alzheimer\u0026rsquo;s disease. Dement Geriatr Cogn Disord. 2006;21(5\u0026ndash;6):364\u0026ndash;72. \u003c/li\u003e\n\u003cli\u003eWhat is frontotemporal degeneration?. Mayo Clinic [Internet].2024. Available from: https://newsnetwork.mayoclinic.org/discussion/what-is-frontotemporal-degeneration/\u003c/li\u003e\n\u003cli\u003eKouli A, Torsney KM, Kuan WL. Parkinson\u0026rsquo;s Disease: Etiology, Neuropathology, and Pathogenesis. Park Dis Pathog Clin Asp. 2018;3\u0026ndash;26. \u003c/li\u003e\n\u003cli\u003eParkinson\u0026rsquo;s Disease. National Institute of Neurological Disorders and Stroke [Internet].Available from: https://www.ninds.nih.gov/health-information/disorders/parkinsons-disease\u003c/li\u003e\n\u003cli\u003eZeighami Y, Ulla M, Iturria-Medina Y, Dadar M, Zhang Y, Larcher KMH, et al. Network structure of brain atrophy in de novo parkinson\u0026rsquo;s disease. Elife. 2015;4(September 2015). \u003c/li\u003e\n\u003cli\u003eAkram AS, Grezenko H, Singh P, Ahmed M, Hassan BD, Hagenahalli Anand V, et al. Advancing the Frontier: Neuroimaging Techniques in the Early Detection and Management of Neurodegenerative Diseases. Cureus. 2024; \u003c/li\u003e\n\u003cli\u003eMathur S, Gawas C, Ahmad IZ, Wani M, Tabassum H. Neurodegenerative disorders: Assessing the impact of natural vs drug-induced treatment options. Aging Med. 2023;6(1):82\u0026ndash;97. \u003c/li\u003e\n\u003cli\u003eEtminani K, Soliman A, Davidsson A, Chang JR, Mart\u0026iacute;nez-Sanchis B, Byttner S, et al. A 3D deep learning model to predict the diagnosis of dementia with Lewy bodies, Alzheimer\u0026rsquo;s disease, and mild cognitive impairment using brain 18F-FDG PET. Eur J Nucl Med Mol Imaging. 2022;49(2):563\u0026ndash;84. \u003c/li\u003e\n\u003cli\u003eAggarwal R, Sounderajah V, Martin G, Ting DSW, Karthikesalingam A, King D, et al. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis. npj Digit Med. 2021;4(1). \u003c/li\u003e\n\u003cli\u003eChan HP, Samala RK, Hadjiiski LM, Zhou C. Deep Learning in Medical Image Analysis. Adv Exp Med Biol. 2020;1213:3\u0026ndash;21. \u003c/li\u003e\n\u003cli\u003eKourounis G, Elmahmudi AA, Thomson B, Hunter J, Ugail H, Wilson C. Computer image analysis with artificial intelligence: a practical introduction to convolutional neural networks for medical professionals. Postgrad Med J. 2023;99(1178):1287\u0026ndash;94. \u003c/li\u003e\n\u003cli\u003eM\u0026auml;enp\u0026auml;\u0026auml; SM, Korja M. Diagnostic test accuracy of externally validated convolutional neural network (CNN) artificial intelligence (AI) models for emergency head CT scans \u0026ndash; A systematic review. Int J Med Inform. 2024;189. \u003c/li\u003e\n\u003cli\u003eChakraborty C, Bhattacharya M, Pal S, Lee SS. From machine learning to deep learning: Advances of the recent data-driven paradigm shift in medicine and healthcare. Curr Res Biotechnol. 2024;7. \u003c/li\u003e\n\u003cli\u003eRamzan F, Khan MUG, Rehmat A, Iqbal S, Saba T, Rehman A, et al. A Deep Learning Approach for Automated Diagnosis and Multi-Class Classification of Alzheimer\u0026rsquo;s Disease Stages Using Resting-State fMRI and Residual Neural Networks. J Med Syst. 2020;44(2). \u003c/li\u003e\n\u003cli\u003eA A, M P, Hamdi M, Bourouis S, Rastislav K, Mohmed F. Evaluation of Neuro Images for the Diagnosis of Alzheimer\u0026rsquo;s Disease Using Deep Learning Neural Network. Front public Heal. 2022;10:834032. \u003c/li\u003e\n\u003cli\u003eAkindele RG, Adebayo S, Kanda PS, Yu M. AlzhiNet: Traversing from 2DCNN to 3DCNN, Towards Early Detection and Diagnosis of Alzheimer\u0026rsquo;s Disease. 2024; Available from: http://arxiv.org/abs/2410.02714\u003c/li\u003e\n\u003cli\u003eMa D, Lu D, Popuri K, Wang L, Beg MF. Differential Diagnosis of Frontotemporal Dementia, Alzheimer\u0026rsquo;s Disease, and Normal Aging Using a Multi-Scale Multi-Type Feature Generative Adversarial Deep Neural Network on Structural Magnetic Resonance Images. Front Neurosci. 2020;14. \u003c/li\u003e\n\u003cli\u003eRogeau A, Hives F, Bordier C, Lahousse H, Roca V, Lebouvier T, et al. A 3D convolutional neural network to classify subjects as Alzheimer\u0026rsquo;s disease, frontotemporal dementia or healthy controls using brain 18F-FDG PET. Neuroimage. 2024;288. \u003c/li\u003e\n\u003cli\u003eEsmaeilzadeh S, Yang Y, Adeli E. End-to-End Parkinson Disease Diagnosis using Brain MR-Images by 3D-CNN. 2018; Available from: http://arxiv.org/abs/1806.05233\u003c/li\u003e\n\u003cli\u003eChakraborty S, Aich S, Kim HC. Detection of Parkinson\u0026rsquo;s disease from 3T T1 weighted MRI scans using 3D convolutional neural network. Diagnostics. 2020;10(6). \u003c/li\u003e\n\u003cli\u003eDyrba M, Hanzig M, Altenstein S, Bader S, Ballarini T, Brosseron F, et al. Improving 3D convolutional neural network comprehensibility via interactive visualization of relevance maps: evaluation in Alzheimer\u0026rsquo;s disease. Alzheimer\u0026rsquo;s Res Ther. 2021;13(1). \u003c/li\u003e\n\u003cli\u003eHu Z, Wang Y, Xiao L. Alzheimer\u0026rsquo;s disease diagnosis by 3D-SEConvNeXt. J Big Data. 2025;12(1). \u003c/li\u003e\n\u003cli\u003eDe Francesco S, Crema C, Archetti D, Muscio C, Reid RI, Nigri A, et al. Differential diagnosis of neurodegenerative dementias with the explainable MRI based machine learning algorithm MUQUBIA. Sci Rep. 2023;13(1). \u003c/li\u003e\n\u003cli\u003eEl-Assy AM, Amer HM, Ibrahim HM, Mohamed MA. A novel CNN architecture for accurate early detection and classification of Alzheimer\u0026rsquo;s disease using MRI data. Sci Rep. 2024;14(1). \u003c/li\u003e\n\u003cli\u003eMandal PK, Mahajan R, Dinov ID. Structural brain atlases: Design, rationale, and applications in normal and pathological cohorts. J Alzheimer\u0026rsquo;s Dis. 2012;31(SUPPL. 3). \u003c/li\u003e\n\u003cli\u003eViergever MA, Maintz JBA, Klein S, Murphy K, Staring M, Pluim JPW. A survey of medical image registration \u0026ndash; under review. Med Image Anal. 2016;33:140\u0026ndash;4. \u003c/li\u003e\n\u003cli\u003eJenkinson M, Beckmann CF, Behrens TEJ, Woolrich MW, Smith SM. FSL. Neuroimage [Internet]. 2012;62(2):782\u0026ndash;90. Available from: https://www.sciencedirect.com/science/article/pii/S1053811911010603\u003c/li\u003e\n\u003cli\u003eNy\u0026uacute;l LG, Udupa JK, Zhang X. New variants of a method of MRI scale standardization. IEEE Trans Med Imaging. 2000;19(2):143\u0026ndash;50. \u003c/li\u003e\n\u003cli\u003eLin TY, Goyal P, Girshick R, He K, Dollar P. Focal Loss for Dense Object Detection. IEEE Trans Pattern Anal Mach Intell. 2020;42(2):318\u0026ndash;27. \u003c/li\u003e\n\u003cli\u003eHuang J, Ling CX. Using AUC and accuracy in evaluating learning algorithms. IEEE Trans Knowl Data Eng. 2005;17(3):299\u0026ndash;310. \u003c/li\u003e\n\u003cli\u003eWen J, Thibeau-Sutre E, Diaz-Melo M, Samper-Gonz\u0026aacute;lez J, Routier A, Bottani S, et al. Convolutional neural networks for classification of Alzheimer\u0026rsquo;s disease: Overview and reproducible evaluation. Med Image Anal. 2020;63. \u003c/li\u003e\n\u003cli\u003eAl-Sharify ZT, Al-Sharify TA, Al-Sharify NT, Naser HY. A critical review on medical imaging techniques (CT and PET scans) in the medical field. IOP Conf Ser Mater Sci Eng. 2020;870(1).\u003c/li\u003e\n\u003cli\u003eGonzalez Kelso I, Tadi P. Cognitive Assessment. In: StatPearls [Internet]. [Updated 2022 Nov 7]. Treasure Island (FL): StatPearls Publishing; 2022 Nov 7 [cited 2025 Jan]. Available from: https://www.ncbi.nlm.nih.gov/books/NBK556049/\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e Alzheimer's Disease Neuroimaging Initiative (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://adni.loni.usc.edu/\u003c/span\u003e\u003cspan address=\"http://adni.loni.usc.edu/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e Neuroimaging in Frontotemporal Dementia (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://memory.ucsf.edu/research/studies/nifd\u003c/span\u003e\u003cspan address=\"http://memory.ucsf.edu/research/studies/nifd\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e Parkinson's Progression Markers Initiative (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.ppmi-info.org/\u003c/span\u003e\u003cspan address=\"http://www.ppmi-info.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Tarbiat Modares University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Deep Learning, Alzheimer’s Disease, Frontotemporal Dementia, Parkinson’s Disease, Magnetic Resonance Imaging (MRI), 3D Convolutional Neural Network","lastPublishedDoi":"10.21203/rs.3.rs-7998240/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7998240/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eThe worldwide increase in cognitive diseases has generated a greater demand for precise, non-invasive diagnostic tools with the ability of differentiating neurodegenerative diseases with similar clinical symptoms and subtle differences in brain anatomy. Alzheimer\u0026rsquo;s disease (AD), frontotemporal dementia (FTD), and Parkinson\u0026rsquo;s disease (PD) are typified by cognitive impairment, hence differential diagnosis becomes crucial.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eThis work utilizes three-dimensional convolutional neural networks (3D-CNNs) on 1,629 T1-weighted structural MRI images gathered from well-established datasets (ADNI, NIFD, and PPMI). Preprocessing involved format conversion, image registration, brain extraction, image cropping, volumetric downsampling, and intensity normalization. We built binary classification models for each disorder against healthy controls, and then a combined multiclass model for simultaneous discrimination between AD, FTD, PD, and controls. In order to prevent data leakage, subject-level data partitioning was performed instead of image-based splitting. Training was done without synthetic augmentation. With the application of Focal Loss and Maxout layers\u0026mdash;innovations not utilized in previous research covered here\u0026mdash;our strict preprocessing and architecture outperform many of the existing solutions in sensitivity and diagnostic accuracy.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe binary models performed with very good accuracies: 98% (AD), 98% (FTD), and 93% (PD). The merged multiclass model achieved an overall accuracy of 95%, class-specific sensitivities of 100% (AD), 95% (FTD), and 95% (PD), and an AUC close to 1.00.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eOur model\u0026rsquo;s strict preprocessing methods and novel architecture surpass numerous current methods, demonstrating the promise of 3D-CNN architectures as powerful clinical decision support systems for differential diagnosis of cognitive disorders.\u003c/p\u003e","manuscriptTitle":"Differential Diagnosis of Cognitive Disorders using Deep Learning Techniques based on Neuroimaging Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-03 09:37:43","doi":"10.21203/rs.3.rs-7998240/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b3296327-d128-4d7a-b1f2-7539b8ef9a55","owner":[],"postedDate":"November 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":57319562,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-11-03T09:37:44+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-03 09:37:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7998240","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7998240","identity":"rs-7998240","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}