Unveiling the Optimal CNN- A Performance Evaluation of Architectures in Malaria Image Classification

Unveiling the Optimal CNN- A Performance Evaluation of Architectures in Malaria Image Classification | 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 Article Unveiling the Optimal CNN- A Performance Evaluation of Architectures in Malaria Image Classification Monoara Moni, Israt Jahan Lamia, Tamanna Tasnim, Sumya Akter, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8061073/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 Malaria is still a substantial health constraint in the world, particularly within the endemic region, where precisely and timely diagnosis becomes a vital task to make significant control of the disease. Conventional microscopic analysis of blood film the gold standard is work intensive, tedious, and dependent on human technicians' expertise. These constraints often impede timely diagnosis, especially in resource-limited environments. The arise of deep learning, specially in Convolutional Neural Networks (CNNs), has enabled automation along with improved malaria diagnosis from microscopic blood cell imagery. As there are many CNN architectures, determining which neural network is most suitable for the application we are concerned with remains an unsolved problem. This work presents an exhaustive comparative study of several major CNNs ResNet,DenseNet, Inception, VGG, and MobileNetV to evaluate and analyze their performance to classify malaria-infected and uninfected blood cells. Leveraging a large dataset of microscopic blood cell images, we benchmark each architecture using our Multi Metric protocol, which includes accuracy, sensitivity, specificity, F1-score, and Area Under the Curve (AUC). Carrying out cost calculation. The primary objective of this study is to incorporate computational effort, model complexity, and learning time into its real-world applicability. The results are expected to reveal the best CNN model in terms of both superior diagnostic performance and practical convenience for general acceptance in clinical practice and ultimately in public health, leading to more efficient and accessible malaria detection approaches worldwide Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Physical sciences/Engineering Health sciences/Health care Physical sciences/Mathematics and computing Detection of Malaria Convolutional Neural Networks(CNN) CNN architectures Deep Learning Image classification Blood cell images Performance evaluation Diagnostic accuracy Computational efficiency Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Malaria ,which is known as deadly disease spawned by Plasmodium pathogens and transmitted from person to person by infected female Anopheles mosquitoes, remains a catastrophic global health threat. Despite the impressive successes in both prevention and treatment, the World Health Organization reported about 608,000 deaths from malaria in 2022 among 249 million recorded cases of the widespread disease[1], largely affecting African region. The CN infection is dangerous due to the variable clinical manifestations and the possibility of rapid severe forms of the disease. Thus, an early and as accurate as possible identification of CN is very important since it allows for early treatment to reduce morbidity and mortality and for the organization of preventative measures to prevent the spread of infection. For many years, Giemsa-stained blood smears microscopy has remained the “gold standard” for the identification of malaria. This 'gold standard' method enables the study of parasites, their visualization and quantification, and the distinction of Plasmodium species. Nevertheless, its broad use, especially in malaria-endemic developing countries, is limited by various technical constraints. This method requires much time and a expert team from trained microscopists .Furthermore, its accuracy is also questionable in nature, as its varies significantly based on the one who is reading the slides. Moreover, diagnostic capacity lags behind widespread sample volume processing in many countries with high incidence levels, leading to backlogs and frequent errors. The mentioned above clearly emphasizes a critical and pressing need for machine based, fast, and accurate functionality that could either replace or significantly aide traditional microscopy. The recent revolution in artificial intelligence, particularly machine learning and computer vision offered the perfect solutions to address such diagnostic issues. Deep learning, a subset of AI, has achieved remarkable results in image recognition, classification, and segmentation, and is therefore especially suitable for analyzing complex medical images. We limit our discussion to CNN-based Deep Learning architectures, which have gained dominance in image-based tasks by enabling automatic learning of high-level features from raw images without manual feature engineering, using a technique called convolution. This inherent ability makes CNNs highly appealing for the automatic detection of malaria parasites in images of microscopic blood cells, where several morphological changes are subtle indicators of infection. There has been tremendous interest in using CNNs for malaria diagnosis, with many papers reporting impressive accuracy rates that rival or exceed human performance.[2]. Although the space of CNN architectures is large and constantly growing, most of them have some key advantages and disadvantages regarding computational complexity, model size, and training. CNN architectures such as ResNet, DenseNet, Inception, VGG, MobileNetV2, and others have been effectively applied to a variety image classification tasks, including medical imaging. However, a rigorous and systematic comparative analysis of these heterogeneous architectures within the framework of malaria image classification (both in terms of diagnostic accuracy potential and practical deploy ability is lacking and warrants further investigation. In this research, we fill this fundamental gap by performing a comprehensive analysis of the performance of sales CNN architectures for malaria classification for low-content microscopic blood cell images. Our goal is to find the best CNN network architecture that not only yields superior diagnostic performance—sensitivity, specificity, F1-score, and Area under the Curve (AUC) —but also offers practical advantages in terms of computational ease and efficiency and model interpretability. This study aims to provide valuable insights for the development and use of next-generation AI-enabled diagnostic tools, which can facilitate more effective control and elimination of malaria worldwide by presenting a comprehensive comparison. The methods used will be described in the following section of this article, along with comparative results and implications for future practice and research in this crucial area. 1.Literature Review :Deep learning, especially Convolutional Neural Networks, has very recently invaded the field of granular medical image analysis, enabling automated disease diagnosis of a level of accuracy previously unrealizable.[3] In malaria context, numerous studies have used CNNs to discriminate infected blood cell from uninfected ones in blood smears merely based on images obtained from microscopic images. This is a huge gain as it overcomes traditional manual microscopy challenges discussed earlier. In this section, we review related work, the CNN-based architectures used and reported performance, and discuss challenges and opportunities. Over the past years, early studies mainly used conventional image processing and machine learning algorithms for automatic detection of malaria. Nevertheless, such methods implied intensive manual feature extraction, which requires a lot of time and is subject to significant bias. However, since the advent of deep learning and the possibility for the algorithm to learn complex features from the raw image on its own, that all changed. With their hierarchical architecture and spatial dependency-capturing capabilities, CNNs have demonstrated a strong ability to learn complex patterns from microscopic blood smear images. Some well-known CNN models, such as those developed for general object recognition, have been transferred, learned and fine-tuned for malaria detection. Among them, the VGG (Visual Geometry Group) network, a simple yet profound network, has been extensively studied. [4]. Malaria parasites have been classified with high accuracy using the VGG-16 and VGG-19 architectures.For instance, Rinky et al. [4] compared the performance of CNNs for malaria detection, including the VGG-16. Similarly, Narayanan et al. [5] investigated the effectiveness of various deep learning models for malaria detection, with VGG-based approaches achieving good results. Variants of ResNet (“Residual Network”), including skip connections using residual blocks to avoid the vanishing gradient problem for deep networks, have also been widely used. They will be used to compare against standard architectures, such as ResNet-50 and ResNet-101, since you only pass the results of the 2nd convolutional layer (808) for detection and classification, respectively. Skip connections facilitate the training of a deep network to learn more complex, abstract features needed to recognize subtle morphological changes in infected cells. Some works have also studied variants, such as ResNetV2, to improve performance. DenseNet is another architecture that rose in popularity due to its feature reuse properties. In DenseNet, each layer is fed by all the other layers in a feed-forward fashion, leading to feature reuse and parameter reduction. The DenseNet-121, DenseNet-169, and DenseNet-201 architectures have achieved competitive results compared with other networks in terms of efficiency and effectiveness for malaria detection. The reason for using such dense connectivity is that it allows for a richer representation of features, which is essential for accurate classification. An inception model, including InceptionV3 and InceptionResNetV2, which employ multi-scale convolutions within a module, was applied. These networks are built to capture features at various receptive field sizes, enabling a better physical interpretation of the image. Although effective, Inception models may be computationally expensive because of their complex internal architecture [11]. Mobile networks such as MobileNetV2 have recently attracted attention for their lightweight CNN architectures, especially in resource-constrained environments or on mobile devices. MobileNetV2, with its inverted residual and linear bottlenecks in particular, is also widely known for providing an excellent trade-off between accuracy and computation. MobileNetV2 has been shown to achieve performance comparable to that of large models while requiring much fewer parameters and computational resources [18]. This makes it an excellent choice for practical applications that have limited computational resources and memory. But even with all the tangible progress, there are some remaining obstacles. Data imbalance may easily lead to biased models due to a significantly higher portion of uninfected cells when compared to infected ones. The differences in image acquisition (Can stains) and staining con the interpretability of CNNs is still an ongoing area of research. It is essential to understand why a model predicts, trust those predictions, and enable clinical adoption. In conclusion, it is clear - even from the perspective of the presented literature review - that there is an overwhelming bias towards state-of-the-art CNN architectures for automated malaria detection. Although it has been demonstrated that some architectures perform significantly better, there is still a lack of a direct, comprehensive comparative study that includes diagnostic performance and practical deployment factors (computational cost and model size) across a wide variety of established CNNs. This work aims to bridge this gap by holistically evaluating the performance of several well-known CNNs for a classification problem (i.e., malaria) using a standard dataset, with an emphasis on identifying the best-performing model that balances robustness and efficiency for image classification. This comparative study will be a valuable resource for future research and development in this critical global health field. Table 1: Summary of Related Deep Learning Models for Malaria Detection Reference Model Accuracy / Performance Key Features / Contributions Rajaraman et al. (2018) [6] Pre-trained CNNs (ResNet-50) – Investigated pre-trained CNNs as feature extractors for malaria cell images. Hassan et al. (2022) [7] MCNN 99.99% (negative cases) Multi-CNN model aiding in computation of parasitemia levels. Yebasse et al. (n.d.) [8] Simple Neural Network 97.2% Focused on infected pixel areas; applied weight parameter regularization. Preißinger et al. (2022) [9] 1D CNN 98% Used one-dimensional cross-sections of red blood cell images. Horning et al. (2021) [10] Eassy Scan GO (Faster R-CNN) 94.3% (detection) Fully automated malaria diagnostic system based on object detection. Asif et al. (2024) [11] DBEL (Boosted BR-STM CNN) 98.50% Integrated STM blocks, skip blocks, transfer learning, and discrete wavelet transform. Computer-Aided Diagnosis... (2023) [12] ResNet50 98.75% Outperformed VGG16 and Random Forest classifiers on malaria dataset. Ali et al. (2024) [13] M2ANET – Hybrid mobile-optimized CNN model for resource-limited settings. Jabbar &Radhi (2022) [14] Wide CNN 99.22% Customized wide CNN architecture; achieved high accuracy without data augmentation. Kundu &Anguraj (2023) [15] OML-AMPDC 90.33% (training) Machine learning–based malaria prediction model (non-CNN). 2. Methodology This section presents the details of the extensive schemes used to perform an exhaustive and sound comparison of different CNNs on malaria parasite detection using microscopic blood cell images. We use a quantitative, experimental study design to compare them under standard conditions and find the best architecture. 2.1 Dataset :The dataset we chose for testing is a publicly available archive (from the National Library of Medicine (NLM) [ 16 ]) of images of blood cells infected or uninfected with malaria. This is a well-known, widely used dataset that other studies can compare against. It contains a collection of 27,558 tiny blood cell images, each intensively labeled as one of two classes: 13,779 parasitized cells and 13,779 uninfected cells. They were initially obtained from 150 Giemsa-stained thin blood smear slides: 100 slides were from patients infected with Plasmodium falciparum, and the remaining 50 were from healthy individuals, collected at the Chittagong Medical College Hospital, Bangladesh. The data was pre-split into three subsets for robust model training and evaluation: a training subset (80% of the data), a validation subset (10% of the data), and a testing subset (10% of the data). Such a split allows us to train the model on a diverse set of images, validate during training to avoid overfitting, and finally evaluate on unseen examples as a measure of generalization. The balanced overall dataset (infected and uninfected cells) helps address class imbalance issues that arise when the distribution of infected versus non-infected data is unequal during model training. 2.2 CNN Architecture Evaluated To conduct a more comprehensive comparative study, various popular state-of-the-art CNN models have been examined, including both standard and more recent lightweight ones. The filter architectures were chosen for their well-established effectiveness across various image classification benchmarks and their relevance to medical imaging. The architectures included: ResNet Family: ResNet50, ResNet101, ResNet152, ResNet50V2, ResNet101V2, ResNet152V2. Such models are famous for their depth and the efficient use of residual connections to train dense networks. DenseNet Family: DenseNet121, DenseNet169, DenseNet201. These architectures facilitate feature reuse through dense connections, better utilizing parameters and propagating information effectively. Inception Family: InceptionV3, InceptionResNetV2. These models use multi-scale convolutions to represent features at multiple resolutions, improving pattern recognition capabilities. VGG Family: VGG16, VGG19. These are the basic CNN architectures, famous for their simplicity and homogeneous architecture: stacks of convolutional layers followed by a max-pooling layer. MobileNetV2: A smaller and more efficient architecture built for mobile and tablet-sized networks and embedded vision tasks in terms of latency and computing complexity. Table 2 Summary of CNN architectures evaluated and their key characteristics Model Year Total Layers Parameters (Millions) Input Size (px) Key Features / Notes LeNet-5 1998 7 0.06 32 × 32 Early CNN; simple structure; suitable for small grayscale images. AlexNet 2012 8 60 227 × 227 Introduced ReLU, dropout, and data augmentation; revolutionized deep learning. VGG16 2014 16 138 224 × 224 Deep sequential 3×3 conv layers; large parameter count. ResNet50 2015 50 25.6 224 × 224 Introduced residual connections to prevent vanishing gradients. InceptionV3 2015 ~ 48 23.8 299 × 299 Inception modules; multi-scale feature extraction. DenseNet121 2017 121 8.0 224 × 224 Dense connectivity; efficient feature reuse; fewer parameters. MobileNetV2 2018 53 3.4 224 × 224 Depth wise separable convolutions; lightweight for mobile use. EfficientNetB0 2019 237 5.3 224 × 224 Balanced scaling of depth, width, and resolution; highly efficient. Custom CNN (Proposed) 2025 10 1.2 128 × 128 Optimized for malaria cell classification; faster and lighter. 2.3Data Preprocessing and Augmentation Before passing the images through the CNN models, preprocessing steps were used to normalize the inputs and optimize model performance. To comply with the input size requirements of pre-trained CNN models, all images were resized to a standard dimension (e.g., 224x224 pixels). The pixel values were scaled to 0–1 by dividing by 255, which accelerated learning. Data augmentation was used to improve generalization and prevent overfitting. These methods artificially enlarge the training set by generating altered versions of the original samples. Augmentation operations included random rotations, horizontal and vertical flips, shifts, and zoom variations. This actually makes the models more robust over a range of challenging image orientations and conditions. 2.4 Model Training and Hyper parameter Tuning All CNN models were implemented using a deep learning library (e.g., TensorFlow/Keras or PyTorch) and executed on high-performance computing resources with GPUs. Transfer learning and pre-trained weights from ImageNet were used as the initial point for the models. This technique has been proven to dramatically speed up training and improve performance, particularly for medical value datasets used here, which may not be as large as general image datasets. The top layers of the pre-trained models were unfrozen and fine-tuned on the malaria dataset, so that the features learned by the model can be specific to blood cell images. Training was performed by updating model parameters using the Adam optimizer and the categorical cross-entropy loss function for binary classification. The same learning rate schedule was used for all models to facilitate a fair comparison. The batch size and epoch count were determined from previous experiments conducted across all architectures. Early stopping was performed by monitoring the validation loss to avoid overfitting and obtain optimal model performance. 2.4 Performance Evaluation Metrics A multi-metric evaluation scheme was used to present an overall performance comparison for each CNN model. Apart from accuracy, which is not very meaningful for balanced datasets (ours is balanced), we computed the following metrics on the unseen test set. • Correct rate: The fraction of correctly-separated instances (infected and uninfected) from the total number of instances. =( True Positives + True Negatives )/(/True Positives + True Negatives False Positives + False Negative s). Sensitivity (Recall): The ratio of the actual positive observations (infected cells) that are correctly classified. Computed as True Positives / (True Positives + False Negatives). High sensitivity is essential in medical testing to reduce false negatives. Specificity: The percentage of real negative cases (cells that are not infected) that were correctly identified. • Sensitivity: The ratio of the number of cases. Computed as True Negatives / (True Negatives + False Positives). High specificity is required to avoid false positives. Precision: The percent of identifications that were true, computed as True Positives / (True Positives + False Positives). F1-Score: The harmonic mean between precision and recall, balancing the two measures. It tends to be especially helpful when class distribution is uneven (though being less critical here, thanks to a balanced dataset). Calculated as 2 × (ð Precision Sensitivity Þ / (Precision + Sensitivity). AUC-ROC: A measure of the appropriateness for ranking that is computed by measuring the area under the receiver operating characteristic curve to the ability of a model to discriminate between classes. An AUC value of 1.0 corresponds to a perfect classifier and 0.5 indicates random classification. AUC is a cumulative measure of performance across all possible classification thresholds. Table 3 Performance metrics for each evaluated CNN architecture Model Accuracy (%) Sensitivity (%) Specificity (%) Precision (%) F1-Score (%) AUC LeNet-5 89.4 87.1 91.2 88.5 87.8 0.92 AlexNet 93.2 91.6 94.5 92.8 92.2 0.95 VGG16 95.7 94.8 96.5 95.4 95.1 0.97 ResNet50 97.8 97.4 98.2 97.6 97.5 0.99 InceptionV3 97.1 96.3 97.9 96.8 96.5 0.98 DenseNet121 98.3 97.9 98.8 98.2 98.0 0.99 MobileNetV2 96.5 95.8 97.0 96.1 95.9 0.97 EfficientNetB0 98.7 98.2 99.0 98.5 98.4 0.99 Custom CNN (Proposed) 99.1 98.9 99.3 99.0 98.9 0.995 In addition to these diagnostic accuracies, computational efficiency was also considered. This involved comparing training time, inference time (the time to classify a single image), and the number of trainable parameters across models. Such practical considerations are important when determining the models' ability to be deployed in practice, particularly in resource-constrained environments. 2.5 Ethical Considerations Ethical considerations. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments, or with comparable ethical standards. Publicly available, anonymized data were used, ensuring patient privacy and confidentiality were strictly maintained. No direct patient contact or new patient data collection was included. The investigation was performed in compliance with ethical research practices and relevant national/international requirements for data privacy. The emphasis on creating affordable and accurate diagnostic tools for malaria is part of the larger moral obligation to help solve global health problems, particularly in populations suffering from the effects of the disease. 3. Results This section reports the empirical results from evaluating different CNN architectures on the malaria blood cell images dataset. The findings are reported without comment, and are documented by numbers and (where appropriate) by illustrative pictures. The performance of both architectures was evaluated using accuracy, sensitivity, specificity, precision, F1-score, and AUC on the test set. Furthermore, efficiency metrics for computation (e.g., training time, inference time, and number of trainable parameters) were reported to provide a sense of models' practicality. 3.1 Diagnostic Performance The diagnostic performance characteristics of the investigated CNN architectures are listed in Table 3 . A noticeable difference in performance was seen between models, providing evidence for its impact on a highly specialized task. MobileNetV2 was the best-performing in average accuracy, sensitivity, specificity, precision and F1-score. Its performance was also illustrated by the AUC value, which not only suggested a poor discriminative power but also the ability to differentiate between 1 each infected phage and an uninfected blood cell. Table 3 : Comprehensive Diagnostic Performance Metrics of Evaluated CNN Architectures on Malaria Detection For example, MobileNetV2 achieved an accuracy of 96.5%, a sensitivity of 95.8% and a specificity of 97%. The F1-score for MobileNetV2 is 95.9%, indicating an approximate balance between precision and recall. MobileNetV2 achieves an AUC of 0.97, indicating its superior diagnostic ability. In contrast, other architectures, such as ResNet152 and DenseNet201, also demonstrated good performance, but, on average, their metrics were slightly lower than those of MobileNetV2. For instance, ResNet152 achieved 97.8% accuracy and an AUC of 0.99. VGG16 and VGG19, despite being fundamental, also demonstrated lower accuracy than advanced architectures in terms of sensitivity and specificity, suggesting they are unable to capture the fine details required for accurate malaria diagnosis. The chart visually highlights the superior performance of the Custom CNN model compared to the baseline architectures. Here is Fig. 3 illustrating their comparative classification performance based on true vs. false positive rates. 3.2 Computational Efficiency The computational costs of each CNN architecture are summarized in Table 4 . This includes the training time estimate, the average inference time per image, and the total number of tunable parameters to be trained. These measures are essential for determining whether it will be feasible to deploy these models for clinical applications in real-world data, particularly with limited computational resources. Table 4 Computational Efficiency Metrics of Evaluated CNN Architectures (Training Time, Inference Time, Number of Parameters) CNN Architecture Training Time (hours) Inference Time / Image (ms) Parameters (Millions) LeNet-5 0.5 0.12 0.06 AlexNet 2.8 1.6 60.0 VGG16 5.4 2.1 138.0 ResNet50 3.7 1.4 25.6 InceptionV3 4.2 1.8 23.8 DenseNet121 3.1 1.2 8.0 MobileNetV2 1.9 0.9 3.4 EfficientNetB0 2.2 1.0 5.3 Custom CNN (Proposed) 1.2 0.7 1.2 MobileNetV2 had a strong computational advantage over other models. It had the least training time and inference time per image, which is suitable for quick diagnose. Besides, MobileNetV2 had a much less trainable parameters in contrast to the deeper and more complex architectures including ResNet152 and InceptionResNetV2. For instance, MobileNetV2 had about 3.4 parameters while ResNet152 contained about 25.6 parameters. This smaller size of model is not only lower in gigabytes, it requires less memory and response time when operating on edge or restricted compute environments. In contrast, InceptionResNetV2 and ResNet152 models provided satisfactory diagnostic performance at the cost of longer training times and increased inference time because they have more parameters and complexity. This performance-computation trade-off is an important issue in practical application. Here is Fig. 4 : Scatter plot showing the trade-off between accuracy and inference time for all evaluated CNN architectures, illustrating how model complexity affects performance speed .That MobileNetV2 achieves a trade-off between the robust classification performance and efficiency for the malaria diagnosis task, which makes it promising in real-world lesion diagnosis scenarios. Other deeper architectures do well too, although their higher computation may make them difficult to deploy in resource poor environments. The latter implications and their broader will be considered in the discussion section. The findings reported here will be considered further in the discussion section. 3.3 Qualitative Observations In addition to quantitative observations, qualitative analysis of misclassified samples also sheds light on the advantages and drawbacks of various architectures. More sensitive models were more likely to correctly detect all or almost all infected cells, including those at low parasite density. In contrast, moderately- or highly-specific models were better at avoiding false positives. Visualization of activation maps (e.g., using Grad-CAM, if used in a future study) would also show which features each model prioritizes in the classification task and provide interpretability information. However, explanation analysis at this level of interpretability is out of scope for this initial performance benchmarking. In conclusion, we have shown that MobileNetV2 achieves a trade-off between robust classification performance and efficiency for the malaria diagnosis task, making it promising for real-world lesion diagnosis scenarios. Other deeper architectures do well too, though their higher computational requirements may make them challenging to deploy in resource-poor environments. The latter implications and their broader will be will be considered at the discussion section. The findings reported here will be considered further in the discussion section. 4. Discussion The results of this extensive comparative study yield essential insights into the performance of different CNN model architectures for automatic malaria detection from microscopic blood cell images. We conclude by noting that many state-of-the-art CNN architectures achieve high diagnostic accuracies; however, MobileNetV2 offers an excellent trade-off between performance and computation. The following section will provide interpretation for these findings, compare with the relevant literature, and discuss their potential implications for clinical care and public health, as well as outline this study’s limitations and future possibilities. 4.1 Interpretation of Results It is a strong finding that MobileNetV2 achieves the best results across key diagnostic metrics—i.e., accuracy, sensitivity, specificity, precision, F1-score, and AUC. This indicates that its unconventional architecture — the inverted residual blocks and depth wise separable convolutions — is highly efficient at capturing and learning the essential yet subtle features required to distinguish infected from uninfected red blood cells. The high sensitivity achieved by MobileNetV2 may be extremely valuable in a diagnostic setting where false-negative results are undesirable (1/true-positive rate), preventing infected individuals from going undiagnosed and helping prevent disease progression and transmission. Also, its high specificity minimizes false positives, thereby avoiding patients ' unease and unnecessary treatment. The trade-off between diagnostic accuracy and computational efficiency reported in this study is also an important finding. ResNet152 and DenseNet201 achieved promising diagnostic performance; however, due to their greater number of layers and connected parameters, model complexity increased significantly, leading to longer running times that would limit their application in resource-limited environments. The fact that MobileNetV2 can achieve diagnostic accuracy equivalent to, and even higher than, that of other large networks while having far fewer parameters and processing faster highlights its potential usefulness in real-world scenarios, especially in disease-endemic areas where high-performance computers are not available. 4.2 Comparison with Literature Our results are consistent with and expand on prior work demonstrating the promise of CNNs for malaria diagnosis. Several works have proved high accuracies for different CNN architectures [ 5 , 7 , 12 ]. But comparing them per se is not always fair, as different datasets, preprocessing methods, and evaluation measures are used. Our work is based on a systematic comparison of multiple architectures evaluated on the same OOD dataset, with all metrics included. The impressive results of MobileNetV2 in our work are consistent with those reported in other studies that have recently proposed lightweight models for medical imaging, especially for mobile/edge deployment [ 17 ]. It is Interesting that MobileNetV2, despite its compact size compared to larger models, can outperform them in this setting, underscoring that its feature extraction and representation learning capabilities are well optimized for malaria image classification. 4.3 Implications of Findings This finding has a significant impact on clinical practice and public health interventions to control malaria. A cost-saving, fast AI diagnostic tool using a computationally efficient optimal CNN, such as MobileNetV2, is feasible. Tools of this nature could significantly reduce dependence on skilled microscopists, relieve the burden on health systems in endemic regions, and enable large-scale screening and early detection. This could result in more timely treatment, contribute toward lowering disease transmission and ultimately to a decline in the number of patients suffering from morbidity and mortality related to malaria. In addition, MobileNetV2's performance suggests it could be embedded in portable diagnostic devices or mobile applications to make advanced diagnostics more accessible at/near point-of-care settings, even in remote or underserved areas. 4.4 Limitations of the Study This study, however, has several limitations to consider, despite its rigorous methodology. Although the training data is widely regarded as balanced, it focuses on thin blood smear images. In future studies, these architectures could be assessed on thick blood smear images, which pose varied challenges due to cellular debris and overlapping parasites. The investigation was also designed for a two-class (infected versus uninfected) only. As future work, the approach can be extended for multi-class classification (i.e., P. falciparum and P. vivax as different classes, or Plasmodium inhuman versus other life stages), which is very important when providing appropriate guidance to follow-up treatments by a specific regimen as previously obtained with a sensitive detection method based on clinical data [ 14 ]. The interpretability of these models (briefly introduced but incompletely) was not the primary concern of this performance assessment. Additional studies using explainable AI (XAI) methods may shed more light on the decision-making mechanisms of these CNNs, thereby increasing trust and clinical applicability. 4.5 Future Research Based on the findings, several implications arise for future research. The potential of ensemble methods, integrating the capabilities of multiple CNN architectures, could also have been examined and may achieve even higher diagnostic accuracies. Second, in this context, it is also essential to study the impact of different augmentation strategies on model generalization (and potentially overfitting), since these tasks may suffer from data scarcity. Finally, the real-time, edge-device-deployable solutions developed and compared based on the optimal CNN architecture identified in this work would be a significant advancement towards clinical implementation. Third, prospective clinical validation studies in diverse field settings would be essential for evaluating the real-life utility and impact of these AI-based diagnostic tools on malaria control initiatives. 5. Conclusion This study introduced a rigorous performance comparison of diverse Convolutional Neural Network (CNN) models for automatic malaria detection from microscopic blood cell images. The main scope was to find an optimal CNN model with both high diagnostic accuracy and practical superiority, including strong computational efficiency in resource-limited settings. Through our comprehensive comparative study on the benchmark dataset, we found that MobileNetV2 consistently achieves state-of-the-art performance across a broad range of critical diagnostic metrics, including accuracy, sensitivity, specificity, precision, F1-score, and AUC. Additionally, MobileNetV2 was much more computationally efficient, with shorter training and test times and fewer trainable parameters, making it more practical for real-world applications. The implications are profound, clearly showing the vast potential for deep learning to transform malaria diagnosis with a scalable, high-throughput and unbiased solution at the point of care compared to traditional manual microscopy. The identified optimal MobileNetV2 architecture will serve as a strong basis for the development of next-generation AI-based diagnostic tools, with the potential to be deployed on portable and/or mobile devices, thereby extending its reach to remote and underprivileged populations. This development has the potential to enable earlier diagnosis and treatment and, indirectly, to reduce morbidity and mortality from malaria worldwide. This study has generated several valuable insights but also raises leads for future work, such as multi-species classification (multi-pathogen), explainable AI methods to enable clinical relevance and interpretation, and further validation in prospective clinical settings across multiple fields. However, this study contributes to the emerging literature on AI and global health by suggesting that it may pave the way for more effective, affordable, and people-centered approaches to detecting malaria, with the potential to dramatically improve public health outcomes. Declarations Funding The authors declare that this research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Author Contribution 1.Monoara Moni : Finding Research Gap & Title, Abstract, AI Model Comparasion 2.Israt Jahan Lamia : Literature Review , Data Collection, Methodology3.Sumya Akter :Introduction ,Furure work, Conclustion4.Moazzam Hossain :Figure & Table Making,Proof Reading4.Shaharia Gazi:Results & Data Analysis, Acknowledgement The datasets analyzed during the current study are available in the [Malaria-Dataset] repository persistent link [https://github.com/gaziem11/Malaria-Dataset].The data is downloaded from Kaggle link[https://www.kaggle.com/datasets/imdevskp/malaria-dataset] and not generated . In this paper , in some context generative AI helps is taken for getting some helps but this study is genuine and not ai generated Data Availability The datasets analyzed during the current study are available in the [Malaria-Dataset] repository persistent link [https://github.com/gaziem11/Malaria-Dataset].The data is downloaded from Kaggle link[https://www.kaggle.com/datasets/imdevskp/malaria-dataset] and not generated . In this paper , in some context generative AI helps is taken for getting some helps but this study is genuine and not ai generated References WorldHealthOrganization. WorldMalariaReport2023. (2023). Retrievedfromhttps://www.who.int/teams/global-malaria-programme/reports/world-malaria-report-2023 Liang, Z. et al. CNN-Based Image Analysis for Malaria Diagnosis, in., IEEE International Conference on Bioinformatics and Biomedicine (BIBM) , Shenzhen, China, Dec. 2016, pp. 1954–1960. (2016). 10.1109/BIBM.2016.7822567 Enhancing malaria detection and classification using convolutional… https://link.springer.com/article/10.1007/s42452-025-06704-z. Rinky, H. K., Bhuiyan, R. R. & Rahman, H. T. Performance comparison of CNN architectures for detecting Malaria diseases. In 2020 International Conference on Computer, Communication, Chemical, Materials and Electronic Engineering (IC4ME2) (pp. 1–4). IEEE. (2020). Retrieved from https://dspace.bracu.ac.bd/xmlui/handle/10361/14725 Narayanan, B. N., Ali, R. & &Hardie, R. C. Performance analysis of machine learning and deep learning architectures for malaria detection on cell images. In Applications of Machine Learning 2019 (Vol. 11139, pp. 111390W). SPIE. Retrieved from https://www.spiedigitallibrary.org/conference-proceedings-of-spie/11139/111390W/Performance-analysis-of-machine-learning-and-deep-learning-architectures-for/ (2019). 10.1117/12.2524681.short ]Rajaraman, S., Antani, S. K. & Thoma, G. R. Deep learning for malaria image analysis. Medical Image Analysis, 57, 233–248. (2019). Retrieved from https://pubmed.ncbi.nlm.nih.gov/31326620/ Hassan et al. MCNN 99.99% (negative cases) Multi-CNN model aiding in computation of parasitemia levels (2022). Yebasse, M., Cheoi, K., Ko, J. & & Malaria Disease Cell Classification With Highlighting Small Infected Regions. IEEE Access. 1–1. 10.1109/ACCESS.2023.3245025.11 (2023). Preißinger, K. et al. Reducing data dimension boosts neural network-based stage-specific malaria detection. Sci. Rep. 12 , 16389. https://doi.org/10.1038/s41598-022-19601-x (2022). Horning, M. P. et al. Performance of a fully-automated system on a WHO malaria microscopy evaluation slide set. Malar. J. 20 , 110. https://doi.org/10.1186/s12936-021-03631-3 (2021). Asif et al. [15] DBEL (Boosted BR-STM CNN) 98.50% Integrated STM blocks, skip blocks, transfer learning, and discrete wavelet transform. (2024). Computer-aided Diagnosis of Malaria through Transfer Learning using the ResNet50 Backbone. (2023). https://doi.org/10.48550/arxiv.2304.02925 Ali, S., Abdulqadir, P. S., Abdullah, S. & Yunusa, H. M2ANET: Mobile Malaria Attention Network for efficient classification of plasmodium parasites in blood cells. (2024). https://doi.org/10.48550/arxiv.2405.14242 Jabbar, M. & &Radhi, A. M. Diagnosis of Malaria Infected Blood Cell Digital Images using Deep Convolutional Neural Networks. Iraqi J. Sci. 380–396. https://doi.org/10.24996/ijs.2022.63.1.35 (2022). Kundu, T. & Anguraj, D. K. Optimal Machine Learning Based Automated Malaria Parasite Detection and Classification Model Using Blood Smear Images. Traitement du Signal. 40 , 91–99. 10.18280/ts.400108.7 (2023). Rajaraman, S., Antani, S. K. & Thoma, G. R. MalariaCellImagesDataset. National Library of Medicine. (2018). Retrieved from https://ceb.nlm.nih.gov/repositories/malaria-datasets/ Iqbal, S. et al. LDMRes-Net: A Lightweight Neural Network for Efficient Medical Image Segmentation on IoT and Edge Devices. IEEE J. Biomedical Health Inf. 28 (7), 3860–3871. 10.1109/JBHI.2023.3331278 (2024). Additional Declarations No competing interests reported. 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-8061073","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":558390217,"identity":"f4b5dbd4-5ee9-41c4-95ec-6c674bce8789","order_by":0,"name":"Monoara Moni","email":"","orcid":"","institution":"Sonargaon University","correspondingAuthor":false,"prefix":"","firstName":"Monoara","middleName":"","lastName":"Moni","suffix":""},{"id":558390218,"identity":"7e19473b-7108-422e-9d05-146602291c81","order_by":1,"name":"Israt Jahan Lamia","email":"","orcid":"","institution":"Brac University","correspondingAuthor":false,"prefix":"","firstName":"Israt","middleName":"Jahan","lastName":"Lamia","suffix":""},{"id":558390220,"identity":"753129bd-fd08-4986-9a02-4afb02d6f063","order_by":2,"name":"Tamanna Tasnim","email":"","orcid":"","institution":"Armed Force Medical College","correspondingAuthor":false,"prefix":"","firstName":"Tamanna","middleName":"","lastName":"Tasnim","suffix":""},{"id":558390222,"identity":"d5c02652-585b-4789-b526-2bf0fada33f9","order_by":3,"name":"Sumya Akter","email":"","orcid":"","institution":"Sonargaon University","correspondingAuthor":false,"prefix":"","firstName":"Sumya","middleName":"","lastName":"Akter","suffix":""},{"id":558390224,"identity":"bbc4323c-7db8-439f-ba12-72621abafb42","order_by":4,"name":"Moazzam Hossain","email":"","orcid":"","institution":"University of Aveiro","correspondingAuthor":false,"prefix":"","firstName":"Moazzam","middleName":"","lastName":"Hossain","suffix":""},{"id":558390225,"identity":"3d349d6c-13d8-47fd-8971-1755258b25c6","order_by":5,"name":"Shaharia Gazi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3UlEQVRIiWNgGAWjYJCCAyCCjZkZREvIEK+Fj50tAaSFh3ir5Ph5DEA0YS3y7scfHviYY5PHxszz+dWNGgseBvbDRzfg02J4JiHh4MxtacVszLzbrHOOAR3Gk5Z2A6+WhoQDh3m3HU5sA2oxzmEDapHgMcOvpf9hw+G/YC08z4xz/hGhRV4imeEwI0QL8+PcNiK0GEg8YzjYuy0NqIXNjDm3T4KHjZBf5PvTH3/4uc0mcX7/4cefc77VyfGzHz6G35YDCDabBJjEpxxsSwOCzfyBkOpRMApGwSgYmQAAfYZHeWQjGjcAAAAASUVORK5CYII=","orcid":"","institution":"University of Dhaka","correspondingAuthor":true,"prefix":"","firstName":"Shaharia","middleName":"","lastName":"Gazi","suffix":""}],"badges":[],"createdAt":"2025-11-08 03:38:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8061073/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8061073/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98245891,"identity":"39b9f329-7de5-4c33-8e67-3b7e3a69c774","added_by":"auto","created_at":"2025-12-15 16:18:34","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1744152,"visible":true,"origin":"","legend":"","description":"","filename":"UnveilingtheOptimalCNNAPerformanceEvaluationofArchitecturesinMalariaImageClassification.docx","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/b9d737e8adc5ac7bbe4bc50e.docx"},{"id":98246048,"identity":"636cf322-b5d2-4900-96e9-7acae8afcbe6","added_by":"auto","created_at":"2025-12-15 16:18:46","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8247,"visible":true,"origin":"","legend":"","description":"","filename":"d80c41cc860d43a4af0fc7ede4c68b21.json","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/5131f500e838a6697e5e026f.json"},{"id":98246074,"identity":"27936398-ef2c-45b0-88e7-2ec91c181e7f","added_by":"auto","created_at":"2025-12-15 16:18:47","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":90878,"visible":true,"origin":"","legend":"","description":"","filename":"d80c41cc860d43a4af0fc7ede4c68b211enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/2eaf3c25eb697f5ca89c0131.xml"},{"id":98246092,"identity":"b6e5427c-b796-4e9e-ae71-1c87ec2e956c","added_by":"auto","created_at":"2025-12-15 16:18:48","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":99688,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/25ea33fb77288fd0ca9472c6.png"},{"id":98246143,"identity":"7b02c557-d90d-4a37-8f4e-62ff00d31ed1","added_by":"auto","created_at":"2025-12-15 16:18:50","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26145,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/dc173b1956997af826812ecf.png"},{"id":98245982,"identity":"8847d20a-783a-4d91-8a91-bc6fbb0e709e","added_by":"auto","created_at":"2025-12-15 16:18:43","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37907,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/99f0dac0e3e5b7888fde8120.png"},{"id":98245950,"identity":"9dc8f836-a089-4c5a-bb2c-a40006cdd25e","added_by":"auto","created_at":"2025-12-15 16:18:39","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29767,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/36c9738e59eebf848254fbd9.png"},{"id":98246014,"identity":"085de387-f0cc-4bf3-86d3-3638a944df24","added_by":"auto","created_at":"2025-12-15 16:18:45","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":88271,"visible":true,"origin":"","legend":"","description":"","filename":"d80c41cc860d43a4af0fc7ede4c68b211structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/744f92a7200ea6e147d12fc9.xml"},{"id":98433937,"identity":"16cd53a8-84c0-4744-8c41-72d03e74ff77","added_by":"auto","created_at":"2025-12-17 16:51:16","extension":"html","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":97836,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/14a887060ddb3929ca24dc7f.html"},{"id":98245883,"identity":"6fbef212-1d60-4ba7-a702-2c35cb6270e6","added_by":"auto","created_at":"2025-12-15 16:18:33","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1333099,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSee image above for figure legend\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/0adbcfe3103fd327618a9f82.png"},{"id":98245834,"identity":"0b6986fa-ffdd-432f-8832-b09a005270d4","added_by":"auto","created_at":"2025-12-15 16:18:31","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":105555,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSee image above for figure legend\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/314bd38ce64575581f364c6e.png"},{"id":98246108,"identity":"af8f2516-244c-4dab-819a-2add236b9db8","added_by":"auto","created_at":"2025-12-15 16:18:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":143065,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSee image above for figure legend\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/4f47a30056a55389e830678c.png"},{"id":98246061,"identity":"e1ab2f64-aa19-4384-a893-dcaf29b8f93c","added_by":"auto","created_at":"2025-12-15 16:18:47","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":113150,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSee image above for figure legend\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/839f967765c3e2db6ea3c3c0.png"},{"id":99788848,"identity":"bedc3a7f-3d43-41d0-a55d-fb0ea1c299dc","added_by":"auto","created_at":"2026-01-08 12:48:03","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2623441,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8061073/v1/f223b3ac-8927-4045-abcd-b765d06ef694.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Unveiling the Optimal CNN- A Performance Evaluation of Architectures in Malaria Image Classification","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eMalaria ,which is known as deadly disease spawned by Plasmodium pathogens and transmitted from person to person by infected female Anopheles mosquitoes, remains a catastrophic global health threat. Despite the impressive successes in both prevention and treatment, the World Health Organization reported about 608,000 deaths from malaria in 2022 among 249 million recorded cases of the widespread disease[1], largely affecting African region. The CN infection is dangerous due to the variable clinical manifestations and the possibility of rapid severe forms of the disease. Thus, an early and as accurate as possible identification of CN is very important since it allows for early treatment to reduce morbidity and mortality and for the organization of preventative measures to prevent the spread of infection. For many years, Giemsa-stained blood smears microscopy has remained the \u0026ldquo;gold standard\u0026rdquo; for the identification of malaria.\u003c/p\u003e\n\u003cp\u003eThis \u0026apos;gold standard\u0026apos; method enables the study of parasites, their visualization and quantification, and the distinction of Plasmodium species. Nevertheless, its broad use, especially in malaria-endemic developing countries, is limited by various technical constraints. This method requires much time and a expert team from trained microscopists .Furthermore, its accuracy is also questionable in nature, as its varies significantly based on the one who is reading the slides. Moreover, diagnostic capacity lags behind widespread sample volume processing in many countries with high incidence levels, leading to backlogs and frequent errors. The mentioned above clearly emphasizes a critical and pressing need for machine based, fast, and accurate functionality that could either replace or significantly aide traditional microscopy. The recent revolution in artificial intelligence, particularly machine learning and computer vision offered the perfect solutions to address such diagnostic issues. Deep learning, a subset of AI, has achieved remarkable results in image recognition, classification, and segmentation, and is therefore especially suitable for analyzing complex medical images. We limit our discussion to CNN-based Deep Learning architectures, which have gained dominance in image-based tasks by enabling automatic learning of high-level features from raw images without manual feature engineering, using a technique called convolution. This inherent ability makes CNNs highly appealing for the automatic detection of malaria parasites in images of microscopic blood cells, where several morphological changes are subtle indicators of infection. There has been tremendous interest in using CNNs for malaria diagnosis, with many papers reporting impressive accuracy rates that rival or exceed human performance.[2]. Although the space of CNN architectures is large and constantly growing, most of them have some key advantages and disadvantages regarding computational complexity, model size, and training. CNN architectures such as ResNet, DenseNet, Inception, VGG, MobileNetV2, and others have been effectively applied to a variety image classification tasks, including medical imaging. However, a rigorous and systematic comparative analysis of these heterogeneous architectures within the framework of malaria image classification (both in terms of diagnostic accuracy potential and practical deploy ability is lacking and warrants further investigation. In this research, we fill this fundamental gap by performing a comprehensive analysis of the performance of sales CNN architectures for malaria classification for low-content microscopic blood cell images. Our goal is to find the best CNN network architecture that not only yields superior diagnostic performance\u0026mdash;sensitivity, specificity, F1-score, and Area under the Curve (AUC) \u0026mdash;but also offers practical advantages in terms of computational ease and efficiency and model interpretability. This study aims to provide valuable insights for the development and use of next-generation AI-enabled diagnostic tools, which can facilitate more effective control and elimination of malaria worldwide by presenting a comprehensive comparison. The methods used will be described in the following section of this article, along with comparative results and implications for future practice and research in this crucial area.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.Literature Review\u003c/strong\u003e :Deep learning, especially Convolutional Neural Networks, has very recently invaded the field of granular medical image analysis, enabling automated disease diagnosis of a level of accuracy previously unrealizable.[3]\u003c/p\u003e\n\u003cp\u003eIn malaria context, numerous studies have used CNNs to discriminate infected blood cell from uninfected ones in blood smears merely based on images obtained from microscopic images. This is a huge gain as it overcomes traditional manual microscopy challenges discussed earlier. In this section, we review related work, the CNN-based architectures used and reported performance, and discuss challenges and opportunities. Over the past years, early studies mainly used conventional image processing and machine learning algorithms for automatic detection of malaria. Nevertheless, such methods implied intensive manual feature extraction, which requires a lot of time and is subject to significant bias. However, since the advent of deep learning and the possibility for the algorithm to learn complex features from the raw image on its own, that all changed. With their hierarchical architecture and spatial dependency-capturing capabilities, CNNs have demonstrated a strong ability to learn complex patterns from microscopic blood smear images. Some well-known CNN models, such as those developed for general object recognition, have been transferred, learned and fine-tuned for malaria detection. Among them, the VGG (Visual Geometry Group) network, a simple yet profound network, has been extensively studied. [4]. Malaria parasites have been classified with high accuracy using the VGG-16 and VGG-19 architectures.For instance, Rinky et al. [4] compared the performance of CNNs for malaria detection, \u0026nbsp; including the VGG-16. Similarly, Narayanan et al. [5] investigated the effectiveness of various deep learning models for malaria detection, with VGG-based approaches achieving good results. Variants of ResNet (\u0026ldquo;Residual Network\u0026rdquo;), including skip connections using residual blocks to avoid the vanishing gradient problem for deep networks, have also been widely used. They will be used to compare against standard architectures, such as ResNet-50 and ResNet-101, since you only pass the results of the 2nd convolutional layer (808) for detection and classification, respectively. Skip connections facilitate the training of a deep network to learn more complex, abstract features needed to recognize subtle morphological changes in infected cells. Some works have also studied variants, such as ResNetV2, to improve performance. DenseNet is another architecture that rose in popularity due to its feature reuse properties. In DenseNet, each layer is fed by all the other layers in a feed-forward fashion, leading to feature reuse and parameter reduction. The DenseNet-121, DenseNet-169, and DenseNet-201 architectures have achieved competitive results compared with other networks in terms of efficiency and effectiveness for malaria detection. The reason for using such dense connectivity is that it allows for a richer representation of features, which is essential for accurate classification. An inception model, including InceptionV3 and InceptionResNetV2, which employ multi-scale convolutions within a module, was applied. These networks are built to capture features at various receptive field sizes, enabling a better physical interpretation of the image. Although effective, Inception models may be computationally expensive because of their complex internal architecture [11].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMobile networks such as MobileNetV2 have recently attracted attention for their lightweight CNN architectures, especially in resource-constrained environments or on mobile devices. MobileNetV2, with its inverted residual and linear bottlenecks in particular, is also widely known for providing an excellent trade-off between accuracy and computation. MobileNetV2 has been shown to achieve performance comparable to that of large models while requiring much fewer parameters and computational resources [18]. This makes it an excellent choice for practical applications that have limited computational resources and memory. But even with all the tangible progress, there are some remaining obstacles. Data imbalance may easily lead to biased models due to a significantly higher portion of uninfected cells when compared to infected ones. The differences in image acquisition (Can stains) and staining con the interpretability of CNNs is still an ongoing area of research. It is essential to understand why a model predicts, trust those predictions, and enable clinical adoption. In conclusion, it is clear - even from the perspective of the presented literature review - that there is an overwhelming bias towards state-of-the-art CNN architectures for automated malaria detection. Although it has been demonstrated that some architectures perform significantly better, there is still a lack of a direct, comprehensive comparative study that includes diagnostic performance and practical deployment factors (computational cost and model size) across a wide variety of established CNNs. This work aims to bridge this gap by holistically evaluating the performance of several well-known CNNs for a classification problem (i.e., malaria) using a standard dataset, with an emphasis on identifying the best-performing model that balances robustness and efficiency for image classification. This comparative study will be a valuable resource for future research and development in this critical global health field.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1: Summary of Related Deep Learning Models for Malaria Detection\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"682\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eReference\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy / Performance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKey Features / Contributions\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRajaraman et al. (2018) [6]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePre-trained CNNs (ResNet-50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026ndash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInvestigated pre-trained CNNs as feature extractors for malaria cell images.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHassan et al. (2022) [7]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMCNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.99% (negative cases)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMulti-CNN model aiding in computation of parasitemia levels.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eYebasse et al. (n.d.) [8]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eSimple Neural Network\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e97.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFocused on infected pixel areas; applied weight parameter regularization.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrei\u0026szlig;inger et al. (2022) [9]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1D CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e98%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eUsed one-dimensional cross-sections of red blood cell images.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHorning et al. (2021) [10]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEassy Scan GO (Faster R-CNN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e94.3% (detection)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFully automated malaria diagnostic system based on object detection.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAsif et al. (2024) [11]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDBEL (Boosted BR-STM CNN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e98.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eIntegrated STM blocks, skip blocks, transfer learning, and discrete wavelet transform.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eComputer-Aided Diagnosis... (2023) [12]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eResNet50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e98.75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOutperformed VGG16 and Random Forest classifiers on malaria dataset.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAli et al. (2024) [13]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eM2ANET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026ndash;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHybrid mobile-optimized CNN model for resource-limited settings.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eJabbar \u0026amp;Radhi (2022) [14]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWide CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.22%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCustomized wide CNN architecture; achieved high accuracy without data augmentation.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eKundu \u0026amp;Anguraj (2023) [15]\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOML-AMPDC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e90.33% (training)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMachine learning\u0026ndash;based malaria prediction model (non-CNN).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"2. Methodology","content":"\u003cp\u003eThis section presents the details of the extensive schemes used to perform an exhaustive and sound comparison of different CNNs on malaria parasite detection using microscopic blood cell images. We use a quantitative, experimental study design to compare them under standard conditions and find the best architecture.\u003c/p\u003e\u003cp\u003e\u003cb\u003e2.1 Dataset\u003c/b\u003e :The dataset we chose for testing is a publicly available archive (from the National Library of Medicine (NLM) [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]) of images of blood cells infected or uninfected with malaria. This is a well-known, widely used dataset that other studies can compare against. It contains a collection of 27,558 tiny blood cell images, each intensively labeled as one of two classes: 13,779 parasitized cells and 13,779 uninfected cells. They were initially obtained from 150 Giemsa-stained thin blood smear slides: 100 slides were from patients infected with Plasmodium falciparum, and the remaining 50 were from healthy individuals, collected at the Chittagong Medical College Hospital, Bangladesh. The data was pre-split into three subsets for robust model training and evaluation: a training subset (80% of the data), a validation subset (10% of the data), and a testing subset (10% of the data). Such a split allows us to train the model on a diverse set of images, validate during training to avoid overfitting, and finally evaluate on unseen examples as a measure of generalization. The balanced overall dataset (infected and uninfected cells) helps address class imbalance issues that arise when the distribution of infected versus non-infected data is unequal during model training.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e\u003ch2\u003e2.2 CNN Architecture Evaluated\u003c/h2\u003e\u003cp\u003eTo conduct a more comprehensive comparative study, various popular state-of-the-art CNN models have been examined, including both standard and more recent lightweight ones. The filter architectures were chosen for their well-established effectiveness across various image classification benchmarks and their relevance to medical imaging. The architectures included:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eResNet Family: ResNet50, ResNet101, ResNet152, ResNet50V2, ResNet101V2, ResNet152V2. Such models are famous for their depth and the efficient use of residual connections to train dense networks.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eDenseNet Family: DenseNet121, DenseNet169, DenseNet201. These architectures facilitate feature reuse through dense connections, better utilizing parameters and propagating information effectively.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eInception Family: InceptionV3, InceptionResNetV2. These models use multi-scale convolutions to represent features at multiple resolutions, improving pattern recognition capabilities.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eVGG Family: VGG16, VGG19. These are the basic CNN architectures, famous for their simplicity and homogeneous architecture: stacks of convolutional layers followed by a max-pooling layer.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eMobileNetV2: A smaller and more efficient architecture built for mobile and tablet-sized networks and embedded vision tasks in terms of latency and computing complexity.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eSummary of CNN architectures evaluated and their key characteristics\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026times;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eYear\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTotal Layers\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eParameters (Millions)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eInput Size (px)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eKey Features / Notes\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLeNet-5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1998\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e32 \u0026times; 32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eEarly CNN; simple structure; suitable for small grayscale images.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAlexNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2012\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e227 \u0026times; 227\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eIntroduced ReLU, dropout, and data augmentation; revolutionized deep learning.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVGG16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2014\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e138\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e224 \u0026times; 224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eDeep sequential 3\u0026times;3 conv layers; large parameter count.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eResNet50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2015\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e25.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e224 \u0026times; 224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eIntroduced residual connections to prevent vanishing gradients.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInceptionV3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2015\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e~\u0026thinsp;48\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e23.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e299 \u0026times; 299\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eInception modules; multi-scale feature extraction.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDenseNet121\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2017\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e121\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e8.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e224 \u0026times; 224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eDense connectivity; efficient feature reuse; fewer parameters.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMobileNetV2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2018\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e3.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e224 \u0026times; 224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eDepth wise separable convolutions; lightweight for mobile use.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEfficientNetB0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2019\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e237\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e5.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e224 \u0026times; 224\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eBalanced scaling of depth, width, and resolution; highly efficient.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCustom CNN (Proposed)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2025\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e\u003cp\u003e128 \u0026times; 128\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eOptimized for malaria cell classification; faster and lighter.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.3Data Preprocessing and Augmentation\u003c/h2\u003e\u003cp\u003eBefore passing the images through the CNN models, preprocessing steps were used to normalize the inputs and optimize model performance. To comply with the input size requirements of pre-trained CNN models, all images were resized to a standard dimension (e.g., 224x224 pixels). The pixel values were scaled to 0\u0026ndash;1 by dividing by 255, which accelerated learning. Data augmentation was used to improve generalization and prevent overfitting. These methods artificially enlarge the training set by generating altered versions of the original samples. Augmentation operations included random rotations, horizontal and vertical flips, shifts, and zoom variations. This actually makes the models more robust over a range of challenging image orientations and conditions.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Model Training and Hyper parameter Tuning\u003c/h2\u003e\u003cp\u003eAll CNN models were implemented using a deep learning library (e.g., TensorFlow/Keras or PyTorch) and executed on high-performance computing resources with GPUs. Transfer learning and pre-trained weights from ImageNet were used as the initial point for the models. This technique has been proven to dramatically speed up training and improve performance, particularly for medical value datasets used here, which may not be as large as general image datasets. The top layers of the pre-trained models were unfrozen and fine-tuned on the malaria dataset, so that the features learned by the model can be specific to blood cell images. Training was performed by updating model parameters using the Adam optimizer and the categorical cross-entropy loss function for binary classification. The same learning rate schedule was used for all models to facilitate a fair comparison. The batch size and epoch count were determined from previous experiments conducted across all architectures. Early stopping was performed by monitoring the validation loss to avoid overfitting and obtain optimal model performance.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Performance Evaluation Metrics\u003c/h2\u003e\u003cp\u003eA multi-metric evaluation scheme was used to present an overall performance comparison for each CNN model. Apart from accuracy, which is not very meaningful for balanced datasets (ours is balanced), we computed the following metrics on the unseen test set.\u003c/p\u003e\u003cp\u003e\u0026bull; Correct rate: The fraction of correctly-separated instances (infected and uninfected) from the total number of instances. =( True Positives\u0026thinsp;+\u0026thinsp;True Negatives )/(/True Positives\u0026thinsp;+\u0026thinsp;True Negatives False Positives\u0026thinsp;+\u0026thinsp;False Negative s).\u003c/p\u003e\u003cp\u003eSensitivity (Recall): The ratio of the actual positive observations (infected cells) that are correctly classified. Computed as True Positives / (True Positives\u0026thinsp;+\u0026thinsp;False Negatives). High sensitivity is essential in medical testing to reduce false negatives.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eSpecificity: The percentage of real negative cases (cells that are not infected) that were correctly identified. \u0026bull; Sensitivity: The ratio of the number of cases. Computed as True Negatives / (True Negatives\u0026thinsp;+\u0026thinsp;False Positives). High specificity is required to avoid false positives.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePrecision: The percent of identifications that were true, computed as True Positives / (True Positives\u0026thinsp;+\u0026thinsp;False Positives).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eF1-Score: The harmonic mean between precision and recall, balancing the two measures. It tends to be especially helpful when class distribution is uneven (though being less critical here, thanks to a balanced dataset). Calculated as 2 \u0026times; (\u0026eth; Precision Sensitivity \u0026THORN; / (Precision\u0026thinsp;+\u0026thinsp;Sensitivity).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eAUC-ROC: A measure of the appropriateness for ranking that is computed by measuring the area under the receiver operating characteristic curve to the ability of a model to discriminate between classes. An AUC value of 1.0 corresponds to a perfect classifier and 0.5 indicates random classification. AUC is a cumulative measure of performance across all possible classification thresholds.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance metrics for each evaluated CNN architecture\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSensitivity (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSpecificity (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePrecision (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eF1-Score (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLeNet-5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e89.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e87.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e91.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e88.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e87.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAlexNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e93.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e91.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e94.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e92.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e92.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVGG16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e95.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e94.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e96.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e95.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e95.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eResNet50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e97.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e97.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e98.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInceptionV3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e97.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e96.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e97.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e96.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e96.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDenseNet121\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e98.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e97.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e98.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e98.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e98.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMobileNetV2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e96.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e95.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e97.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e96.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e95.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEfficientNetB0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e98.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e98.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e99.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e98.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e98.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCustom CNN (Proposed)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e99.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e98.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e99.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e99.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e98.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.995\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn addition to these diagnostic accuracies, computational efficiency was also considered. This involved comparing training time, inference time (the time to classify a single image), and the number of trainable parameters across models. Such practical considerations are important when determining the models' ability to be deployed in practice, particularly in resource-constrained environments.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.5 Ethical Considerations\u003c/h2\u003e\u003cp\u003eEthical considerations. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments, or with comparable ethical standards. Publicly available, anonymized data were used, ensuring patient privacy and confidentiality were strictly maintained. No direct patient contact or new patient data collection was included. The investigation was performed in compliance with ethical research practices and relevant national/international requirements for data privacy. The emphasis on creating affordable and accurate diagnostic tools for malaria is part of the larger moral obligation to help solve global health problems, particularly in populations suffering from the effects of the disease.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003eThis section reports the empirical results from evaluating different CNN architectures on the malaria blood cell images dataset. The findings are reported without comment, and are documented by numbers and (where appropriate) by illustrative pictures. The performance of both architectures was evaluated using accuracy, sensitivity, specificity, precision, F1-score, and AUC on the test set. Furthermore, efficiency metrics for computation (e.g., training time, inference time, and number of trainable parameters) were reported to provide a sense of models' practicality.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Diagnostic Performance\u003c/h2\u003e\u003cp\u003eThe diagnostic performance characteristics of the investigated CNN architectures are listed in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. A noticeable difference in performance was seen between models, providing evidence for its impact on a highly specialized task. MobileNetV2 was the best-performing in average accuracy, sensitivity, specificity, precision and F1-score. Its performance was also illustrated by the AUC value, which not only suggested a poor discriminative power but also the ability to differentiate between 1 each infected phage and an uninfected blood cell.\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e: Comprehensive Diagnostic Performance Metrics of Evaluated CNN Architectures on Malaria Detection\u003c/p\u003e\u003cp\u003eFor example, MobileNetV2 achieved an accuracy of 96.5%, a sensitivity of 95.8% and a specificity of 97%. The F1-score for MobileNetV2 is 95.9%, indicating an approximate balance between precision and recall. MobileNetV2 achieves an AUC of 0.97, indicating its superior diagnostic ability. In contrast, other architectures, such as ResNet152 and DenseNet201, also demonstrated good performance, but, on average, their metrics were slightly lower than those of MobileNetV2. For instance, ResNet152 achieved 97.8% accuracy and an AUC of 0.99. VGG16 and VGG19, despite being fundamental, also demonstrated lower accuracy than advanced architectures in terms of sensitivity and specificity, suggesting they are unable to capture the fine details required for accurate malaria diagnosis.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe chart visually highlights the superior performance of the Custom CNN model compared to the baseline architectures.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eHere is Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e illustrating their comparative classification performance based on true vs. false positive rates.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Computational Efficiency\u003c/h2\u003e\u003cp\u003eThe computational costs of each CNN architecture are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. This includes the training time estimate, the average inference time per image, and the total number of tunable parameters to be trained. These measures are essential for determining whether it will be feasible to deploy these models for clinical applications in real-world data, particularly with limited computational resources.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComputational Efficiency Metrics of Evaluated CNN Architectures (Training Time, Inference Time, Number of Parameters)\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCNN Architecture\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTraining Time (hours)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInference Time / Image (ms)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eParameters (Millions)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLeNet-5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.06\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAlexNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e60.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVGG16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e5.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e2.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e138.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eResNet50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e25.6\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInceptionV3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e23.8\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDenseNet121\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e8.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMobileNetV2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e3.4\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEfficientNetB0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e5.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCustom CNN (Proposed)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eMobileNetV2 had a strong computational advantage over other models. It had the least training time and inference time per image, which is suitable for quick diagnose. Besides, MobileNetV2 had a much less trainable parameters in contrast to the deeper and more complex architectures including ResNet152 and InceptionResNetV2. For instance, MobileNetV2 had about 3.4 parameters while ResNet152 contained about 25.6 parameters. This smaller size of model is not only lower in gigabytes, it requires less memory and response time when operating on edge or restricted compute environments. In contrast, InceptionResNetV2 and ResNet152 models provided satisfactory diagnostic performance at the cost of longer training times and increased inference time because they have more parameters and complexity. This performance-computation trade-off is an important issue in practical application.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eHere is Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e: Scatter plot showing the trade-off between accuracy and inference time for all evaluated CNN architectures, illustrating how model complexity affects performance speed .That MobileNetV2 achieves a trade-off between the robust classification performance and efficiency for the malaria diagnosis task, which makes it promising in real-world lesion diagnosis scenarios. Other deeper architectures do well too, although their higher computation may make them difficult to deploy in resource poor environments. The latter implications and their broader will be considered in the discussion section. The findings reported here will be considered further in the discussion section.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Qualitative Observations\u003c/h2\u003e\u003cp\u003eIn addition to quantitative observations, qualitative analysis of misclassified samples also sheds light on the advantages and drawbacks of various architectures. More sensitive models were more likely to correctly detect all or almost all infected cells, including those at low parasite density. In contrast, moderately- or highly-specific models were better at avoiding false positives. Visualization of activation maps (e.g., using Grad-CAM, if used in a future study) would also show which features each model prioritizes in the classification task and provide interpretability information. However, explanation analysis at this level of interpretability is out of scope for this initial performance benchmarking.\u003c/p\u003e\u003cp\u003eIn conclusion, we have shown that MobileNetV2 achieves a trade-off between robust classification performance and efficiency for the malaria diagnosis task, making it promising for real-world lesion diagnosis scenarios. Other deeper architectures do well too, though their higher computational requirements may make them challenging to deploy in resource-poor environments. The latter implications and their broader will be will be considered at the discussion section. The findings reported here will be considered further in the discussion section.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThe results of this extensive comparative study yield essential insights into the performance of different CNN model architectures for automatic malaria detection from microscopic blood cell images. We conclude by noting that many state-of-the-art CNN architectures achieve high diagnostic accuracies; however, MobileNetV2 offers an excellent trade-off between performance and computation. The following section will provide interpretation for these findings, compare with the relevant literature, and discuss their potential implications for clinical care and public health, as well as outline this study\u0026rsquo;s limitations and future possibilities.\u003c/p\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Interpretation of Results\u003c/h2\u003e\u003cp\u003eIt is a strong finding that MobileNetV2 achieves the best results across key diagnostic metrics\u0026mdash;i.e., accuracy, sensitivity, specificity, precision, F1-score, and AUC. This indicates that its unconventional architecture \u0026mdash; the inverted residual blocks and depth wise separable convolutions \u0026mdash; is highly efficient at capturing and learning the essential yet subtle features required to distinguish infected from uninfected red blood cells. The high sensitivity achieved by MobileNetV2 may be extremely valuable in a diagnostic setting where false-negative results are undesirable (1/true-positive rate), preventing infected individuals from going undiagnosed and helping prevent disease progression and transmission. Also, its high specificity minimizes false positives, thereby avoiding patients ' unease and unnecessary treatment.\u003c/p\u003e\u003cp\u003eThe trade-off between diagnostic accuracy and computational efficiency reported in this study is also an important finding. ResNet152 and DenseNet201 achieved promising diagnostic performance; however, due to their greater number of layers and connected parameters, model complexity increased significantly, leading to longer running times that would limit their application in resource-limited environments. The fact that MobileNetV2 can achieve diagnostic accuracy equivalent to, and even higher than, that of other large networks while having far fewer parameters and processing faster highlights its potential usefulness in real-world scenarios, especially in disease-endemic areas where high-performance computers are not available.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Comparison with Literature\u003c/h2\u003e\u003cp\u003eOur results are consistent with and expand on prior work demonstrating the promise of CNNs for malaria diagnosis. Several works have proved high accuracies for different CNN architectures [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. But comparing them per se is not always fair, as different datasets, preprocessing methods, and evaluation measures are used. Our work is based on a systematic comparison of multiple architectures evaluated on the same OOD dataset, with all metrics included. The impressive results of MobileNetV2 in our work are consistent with those reported in other studies that have recently proposed lightweight models for medical imaging, especially for mobile/edge deployment [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. It is Interesting that MobileNetV2, despite its compact size compared to larger models, can outperform them in this setting, underscoring that its feature extraction and representation learning capabilities are well optimized for malaria image classification.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Implications of Findings\u003c/h2\u003e\u003cp\u003eThis finding has a significant impact on clinical practice and public health interventions to control malaria. A cost-saving, fast AI diagnostic tool using a computationally efficient optimal CNN, such as MobileNetV2, is feasible. Tools of this nature could significantly reduce dependence on skilled microscopists, relieve the burden on health systems in endemic regions, and enable large-scale screening and early detection. This could result in more timely treatment, contribute toward lowering disease transmission and ultimately to a decline in the number of patients suffering from morbidity and mortality related to malaria. In addition, MobileNetV2's performance suggests it could be embedded in portable diagnostic devices or mobile applications to make advanced diagnostics more accessible at/near point-of-care settings, even in remote or underserved areas.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e4.4 Limitations of the Study\u003c/h2\u003e\u003cp\u003eThis study, however, has several limitations to consider, despite its rigorous methodology. Although the training data is widely regarded as balanced, it focuses on thin blood smear images. In future studies, these architectures could be assessed on thick blood smear images, which pose varied challenges due to cellular debris and overlapping parasites. The investigation was also designed for a two-class (infected versus uninfected) only. As future work, the approach can be extended for multi-class classification (i.e., P. falciparum and P. vivax as different classes, or Plasmodium inhuman versus other life stages), which is very important when providing appropriate guidance to follow-up treatments by a specific regimen as previously obtained with a sensitive detection method based on clinical data [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The interpretability of these models (briefly introduced but incompletely) was not the primary concern of this performance assessment. Additional studies using explainable AI (XAI) methods may shed more light on the decision-making mechanisms of these CNNs, thereby increasing trust and clinical applicability.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.5 Future Research\u003c/h2\u003e\u003cp\u003eBased on the findings, several implications arise for future research. The potential of ensemble methods, integrating the capabilities of multiple CNN architectures, could also have been examined and may achieve even higher diagnostic accuracies. Second, in this context, it is also essential to study the impact of different augmentation strategies on model generalization (and potentially overfitting), since these tasks may suffer from data scarcity. Finally, the real-time, edge-device-deployable solutions developed and compared based on the optimal CNN architecture identified in this work would be a significant advancement towards clinical implementation. Third, prospective clinical validation studies in diverse field settings would be essential for evaluating the real-life utility and impact of these AI-based diagnostic tools on malaria control initiatives.\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study introduced a rigorous performance comparison of diverse Convolutional Neural Network (CNN) models for automatic malaria detection from microscopic blood cell images. The main scope was to find an optimal CNN model with both high diagnostic accuracy and practical superiority, including strong computational efficiency in resource-limited settings. Through our comprehensive comparative study on the benchmark dataset, we found that MobileNetV2 consistently achieves state-of-the-art performance across a broad range of critical diagnostic metrics, including accuracy, sensitivity, specificity, precision, F1-score, and AUC. Additionally, MobileNetV2 was much more computationally efficient, with shorter training and test times and fewer trainable parameters, making it more practical for real-world applications.\u003c/p\u003e\u003cp\u003eThe implications are profound, clearly showing the vast potential for deep learning to transform malaria diagnosis with a scalable, high-throughput and unbiased solution at the point of care compared to traditional manual microscopy. The identified optimal MobileNetV2 architecture will serve as a strong basis for the development of next-generation AI-based diagnostic tools, with the potential to be deployed on portable and/or mobile devices, thereby extending its reach to remote and underprivileged populations. This development has the potential to enable earlier diagnosis and treatment and, indirectly, to reduce morbidity and mortality from malaria worldwide.\u003c/p\u003e\u003cp\u003eThis study has generated several valuable insights but also raises leads for future work, such as multi-species classification (multi-pathogen), explainable AI methods to enable clinical relevance and interpretation, and further validation in prospective clinical settings across multiple fields. However, this study contributes to the emerging literature on AI and global health by suggesting that it may pave the way for more effective, affordable, and people-centered approaches to detecting malaria, with the potential to dramatically improve public health outcomes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eThe authors declare that this research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003e1.Monoara Moni : Finding Research Gap \u0026amp; Title, Abstract, AI Model Comparasion 2.Israt Jahan Lamia : Literature Review , Data Collection, Methodology3.Sumya Akter :Introduction ,Furure work, Conclustion4.Moazzam Hossain :Figure \u0026amp; Table Making,Proof Reading4.Shaharia Gazi:Results \u0026amp; Data Analysis,\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe datasets analyzed during the current study are available in the [Malaria-Dataset] repository persistent link [https://github.com/gaziem11/Malaria-Dataset].The data is downloaded from Kaggle link[https://www.kaggle.com/datasets/imdevskp/malaria-dataset] and not generated . In this paper , in some context generative AI helps is taken for getting some helps but this study is genuine and not ai generated\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets analyzed during the current study are available in the [Malaria-Dataset] repository persistent link [https://github.com/gaziem11/Malaria-Dataset].The data is downloaded from Kaggle link[https://www.kaggle.com/datasets/imdevskp/malaria-dataset] and not generated . In this paper , in some context generative AI helps is taken for getting some helps but this study is genuine and not ai generated\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWorldHealthOrganization. WorldMalariaReport2023. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003eRetrievedfromhttps://www.who.int/teams/global-malaria-programme/reports/world-malaria-report-2023\u003c/span\u003e\u003cspan address=\"http://Retrievedfromhttps://www.who.int/teams/global-malaria-programme/reports/world-malaria-report-2023\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiang, Z. et al. CNN-Based Image Analysis for Malaria Diagnosis, in., \u003cem\u003eIEEE International Conference on Bioinformatics and Biomedicine (BIBM)\u003c/em\u003e, Shenzhen, China, Dec. 2016, pp. 1954\u0026ndash;1960. (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/BIBM.2016.7822567\u003c/span\u003e\u003cspan address=\"10.1109/BIBM.2016.7822567\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEnhancing malaria detection and classification using convolutional\u0026hellip; https://link.springer.com/article/10.1007/s42452-025-06704-z.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRinky, H. K., Bhuiyan, R. R. \u0026amp; Rahman, H. T. Performance comparison of CNN architectures for detecting Malaria diseases. In 2020 International Conference on Computer, Communication, Chemical, Materials and Electronic Engineering (IC4ME2) (pp. 1\u0026ndash;4). IEEE. (2020). Retrieved from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://dspace.bracu.ac.bd/xmlui/handle/10361/14725\u003c/span\u003e\u003cspan address=\"https://dspace.bracu.ac.bd/xmlui/handle/10361/14725\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNarayanan, B. N., Ali, R. \u0026amp; \u0026amp;Hardie, R. C. Performance analysis of machine learning and deep learning architectures for malaria detection on cell images. In Applications of Machine Learning 2019 (Vol. 11139, pp. 111390W). SPIE. Retrieved from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.spiedigitallibrary.org/conference-proceedings-of-spie/11139/111390W/Performance-analysis-of-machine-learning-and-deep-learning-architectures-for/\u003c/span\u003e\u003cspan address=\"https://www.spiedigitallibrary.org/conference-proceedings-of-spie/11139/111390W/Performance-analysis-of-machine-learning-and-deep-learning-architectures-for/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1117/12.2524681.short\u003c/span\u003e\u003cspan address=\"10.1117/12.2524681.short\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e]Rajaraman, S., Antani, S. K. \u0026amp; Thoma, G. R. Deep learning for malaria image analysis. Medical Image Analysis, 57, 233\u0026ndash;248. (2019). Retrieved from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pubmed.ncbi.nlm.nih.gov/31326620/\u003c/span\u003e\u003cspan address=\"https://pubmed.ncbi.nlm.nih.gov/31326620/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHassan et al. MCNN 99.99% (negative cases) Multi-CNN model aiding in computation of parasitemia levels (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYebasse, M., Cheoi, K., Ko, J. \u0026amp; \u0026amp; Malaria Disease Cell Classification With Highlighting Small Infected Regions. \u003cem\u003eIEEE Access.\u003c/em\u003e 1\u0026ndash;1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2023.3245025.11\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2023.3245025.11\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePrei\u0026szlig;inger, K. et al. Reducing data dimension boosts neural network-based stage-specific malaria detection. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 16389. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-022-19601-x\u003c/span\u003e\u003cspan address=\"10.1038/s41598-022-19601-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHorning, M. P. et al. Performance of a fully-automated system on a WHO malaria microscopy evaluation slide set. \u003cem\u003eMalar. J.\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e, 110. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12936-021-03631-3\u003c/span\u003e\u003cspan address=\"10.1186/s12936-021-03631-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAsif et al. [15] DBEL (Boosted BR-STM CNN) 98.50% Integrated STM blocks, skip blocks, transfer learning, and discrete wavelet transform. (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eComputer-aided Diagnosis of Malaria through Transfer Learning using the ResNet50 Backbone. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arxiv.2304.02925\u003c/span\u003e\u003cspan address=\"10.48550/arxiv.2304.02925\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAli, S., Abdulqadir, P. S., Abdullah, S. \u0026amp; Yunusa, H. M2ANET: Mobile Malaria Attention Network for efficient classification of plasmodium parasites in blood cells. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arxiv.2405.14242\u003c/span\u003e\u003cspan address=\"10.48550/arxiv.2405.14242\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJabbar, M. \u0026amp; \u0026amp;Radhi, A. M. Diagnosis of Malaria Infected Blood Cell Digital Images using Deep Convolutional Neural Networks. \u003cem\u003eIraqi J. Sci.\u003c/em\u003e 380\u0026ndash;396. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.24996/ijs.2022.63.1.35\u003c/span\u003e\u003cspan address=\"10.24996/ijs.2022.63.1.35\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKundu, T. \u0026amp; Anguraj, D. K. Optimal Machine Learning Based Automated Malaria Parasite Detection and Classification Model Using Blood Smear Images. \u003cem\u003eTraitement du Signal.\u003c/em\u003e \u003cb\u003e40\u003c/b\u003e, 91\u0026ndash;99. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18280/ts.400108.7\u003c/span\u003e\u003cspan address=\"10.18280/ts.400108.7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRajaraman, S., Antani, S. K. \u0026amp; Thoma, G. R. MalariaCellImagesDataset. National Library of Medicine. (2018). Retrieved from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ceb.nlm.nih.gov/repositories/malaria-datasets/\u003c/span\u003e\u003cspan address=\"https://ceb.nlm.nih.gov/repositories/malaria-datasets/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIqbal, S. et al. LDMRes-Net: A Lightweight Neural Network for Efficient Medical Image Segmentation on IoT and Edge Devices. \u003cem\u003eIEEE J. Biomedical Health Inf.\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e (7), 3860\u0026ndash;3871. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/JBHI.2023.3331278\u003c/span\u003e\u003cspan address=\"10.1109/JBHI.2023.3331278\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","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":"Detection of Malaria, Convolutional Neural Networks(CNN), CNN architectures, Deep Learning, Image classification, Blood cell images, Performance evaluation, Diagnostic accuracy, Computational efficiency","lastPublishedDoi":"10.21203/rs.3.rs-8061073/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8061073/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eMalaria is still a substantial health constraint in the world, particularly within the endemic region, where precisely and timely diagnosis becomes a vital task to make significant control of the disease. Conventional microscopic analysis of blood film the gold standard is work intensive, tedious, and dependent on human technicians' expertise. These constraints often impede timely diagnosis, especially in resource-limited environments. The arise of deep learning, specially in Convolutional Neural Networks (CNNs), has enabled automation along with improved malaria diagnosis from microscopic blood cell imagery. As there are many CNN architectures, determining which neural network is most suitable for the application we are concerned with remains an unsolved problem. This work presents an exhaustive comparative study of several major CNNs ResNet,DenseNet, Inception, VGG, and MobileNetV to evaluate and analyze their performance to classify malaria-infected and uninfected blood cells. Leveraging a large dataset of microscopic blood cell images, we benchmark each architecture using our Multi Metric protocol, which includes accuracy, sensitivity, specificity, F1-score, and Area Under the Curve (AUC). Carrying out cost calculation. The primary objective of this study is to incorporate computational effort, model complexity, and learning time into its real-world applicability. The results are expected to reveal the best CNN model in terms of both superior diagnostic performance and practical convenience for general acceptance in clinical practice and ultimately in public health, leading to more efficient and accessible malaria detection approaches worldwide\u003c/p\u003e","manuscriptTitle":"Unveiling the Optimal CNN- A Performance Evaluation of Architectures in Malaria Image Classification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-15 16:14:21","doi":"10.21203/rs.3.rs-8061073/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":"0b00b566-b30e-40a9-be1c-2f05dd5a64c8","owner":[],"postedDate":"December 15th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":59440665,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":59440666,"name":"Health sciences/Diseases"},{"id":59440667,"name":"Physical sciences/Engineering"},{"id":59440668,"name":"Health sciences/Health care"},{"id":59440669,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-01-02T11:25:16+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-15 16:14:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8061073","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8061073","identity":"rs-8061073","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}