Tri-UNet: A Brain Age Prediction Method Based on Different Scale Features of Magnetic Resonance Imaging

Tri-UNet: A Brain Age Prediction Method Based on Different Scale Features of Magnetic Resonance Imaging | 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 Tri-UNet: A Brain Age Prediction Method Based on Different Scale Features of Magnetic Resonance Imaging Yu Pang, Yihuai Cai This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3820912/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 3 You are reading this latest preprint version Abstract In the process of human aging, significant age-related changes occur in brain tissue. To assist individuals in assessing the degree of brain aging, screening for disease risks, and further diagnosing age-related diseases, it is crucial to develop an accurate method for predicting brain age. This paper proposes a multi-scale feature fusion method called Tri-UNet based on the U-Net network structure, as well as a brain region information fusion method based on multi-channel input networks. These methods address the issue of insufficient image feature learning in brain neuroimaging data. They can effectively utilize features at different scales of MRI and fully leverage feature information from different regions of the brain. In the end, experiments were conducted on the Cam-CAN dataset, resulting in a minimum Mean Absolute Error (MAE) of 7.46. The results demonstrate that this method provides a new approach to feature learning at different scales in brain age prediction tasks, contributing to the advancement of the field and holding significance for practical applications in the context of elderly education. Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Health care Physical sciences/Engineering Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Introduction The structure of the human brain undergoes gradual changes throughout the entire lifespan 1 . Knickmeyer et al. 2 , in a longitudinal study of brain structure MRI in healthy individuals, found that the brain volume of newborns is typically only half that of adults, increasing to 90% of adult size by adulthood. Research by Good et al. 3 indicates a gradual decrease in gray matter volume throughout adulthood, with white matter volume following an inverted "U-shaped" curve, reaching its peak in middle age. Binte et al. 4 conducted a morphological analysis of brain structure using structural MRI data from 619 participants aged 19.98 to 83.62 years in the publicly available IXI dataset. The study revealed that as humans reach middle age, the brain undergoes aging with increasing age, accompanied by structural atrophy. The most significant manifestation of this aging process is a decline in cognitive abilities. Furthermore, brain aging increases the risk of brain diseases, especially age-related neurological disorders such as mild cognitive impairment, Alzheimer's disease, schizophrenia, epilepsy, traumatic brain injury. 5–10 The various diseases that accompany brain aging impose significant burdens on individuals, families, and society as a whole. While aging is inevitable, interventions in early life can be employed to prevent or reduce brain damage, potentially preventing common neurological diseases 11–13 . Therefore, accurately assessing the degree of individual brain aging to effectively identify the brain's health status has become a crucial research topic. With the rapid development of deep learning technology and its expanding applications, an increasing number of researchers are employing deep learning methods to predict brain age. Jiang et al. 14 obtained brain MRI data from a total of 1454 healthy subjects aged 18–90 from five publicly available datasets. The data were split into a training set with 1303 samples and a test set with 151 samples. The researchers then segmented the brain into seven different functional networks using the cortical parcellation template (CorticalParcellation_Yeo2011). Subsequently, three methods—3D CNN, Gaussian Process Regression (GPR), and Relevance Vector Regression (RVR)—were employed to train models. All three methods performed best on the Frontoparietal Network (FPN), Dorsal Attention Network (DAN), and Default Mode Network (DMN) brain networks. The 3D CNN method demonstrated average Mean Absolute Errors (MAE) of 5.55 years, 5.77 years, and 6.07 years for the three networks, outperforming the machine learning methods GPR and RVR. This work suggests that Convolutional Neural Networks (CNNs) hold significant potential for predicting brain age. Bintsi et al. 15 proposed a patch-based brain age prediction framework using an ensemble approach of 3D Convolutional Neural Network (CNN) and linear regression. The model was trained on the UK Biobank dataset, yielding a final average MAE of 2.46 years. By utilizing patches, the researchers identified brain regions with the greatest impact on brain age prediction, providing interpretable results and advancing anatomical research in the context of deep learning for brain age prediction. This study confirmed the hippocampus and regions such as the ventricles as most relevant to age prediction. Traditional brain age prediction methods based on anatomical measurements overlook the biological spatial information of brain anatomical structures. This results in a challenging spatial information gap in machine learning-based brain age prediction methods, even for leading algorithms that exhibit significant measurement errors. Because these methods lack context information from neighboring voxels and are insensitive to nonlinear relationships 16 , researchers are increasingly turning their attention to Convolutional Neural Networks (CNNs) in the field of deep learning. 17 Ballester et al. 18 proposed a multi-channel network model based on ResNet18. They used segmented gray and white matter slices as input data for two channels, constructing a convolutional neural network. The final brain age prediction was obtained by averaging the results of three independently trained neural networks. Despite improvements in the MAE evaluation metric in this study, the use of 2D slices as input data prevents the network from integrating contextual information from different brain regions. Additionally, the selection of slices is subject to human factors, leading to significant changes in prediction accuracy when different slice indices are chosen, resulting in a lack of stable predictive performance. Pardakhti et al. 19 , using a 3D ResNet, achieved the best MAE of approximately 5 years. However, this method only provides whole-brain age predictions, whereas current research on brain age prediction is moving towards regional visualization, requiring voxel-level predictions for assessing regional changes in brain aging and lesion locations related to age-related diseases. 20,21 Therefore, Popescu et al. 22 , based on the U-Net network for image segmentation followed by voxel-level brain age prediction, achieved the best MAE of around 7 years in the frontal cortex and periventricular regions. However, this method solely employs U-Net for training without effectively integrating regional information between different feature layers. Their utilization of multi-scale features is not thorough, and identical features in different brain regions may represent different information, leading to the omission of many features and the loss of substantial spatial information. To address these issues, this paper proposes a 3D network architecture, Tri-UNet, based on 3D U-Net and 3D ResNet. Brain region data is utilized as input features for the multi-input network to make it more suitable for brain age prediction tasks. The effectiveness of the proposed method is validated on the publicly available Cam-CAN dataset, demonstrating good accuracy. The main contributions of this paper are as follows: Proposed a brain age prediction method based on the 3D U-Net network structure and residual learning, addressing the issue of insufficient feature utilization in predicting brain age using the U-Net network. This method enhances the utilization of features by repeatedly reusing features through residual connections from ResNet on the foundation of U-Net. It strengthens the correlation between features at different scales, achieving maximal feature utilization and demonstrating good performance. Proposed a brain age prediction method based on a multi-channel input network, addressing the issue of information redundancy in previous studies using multi-channel input networks to predict brain age. In this paper, the original 3D brain imaging data were segmented into anatomical regions defined by experts, and the segmented brain regions were used as input data for the network. This allows the network to learn specific parameters for different brain regions, capturing region-specific atrophy patterns. Consequently, it integrates complementary information from different brain regions, improving the accuracy of brain age prediction while reducing training time. Materials and Methods Dataset The neuroimaging data utilized in this investigation emanate from the publicly accessible Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data repository. This paper utilized T1-weighted MRI data from 651 healthy participants (male/female = 322/329, mean age = 54.7 ± 18.6, age range 18–89 years) from the Cam-CAN repository, as shown in Table 1 . Table 1 Description of Cam-CAN Dataset Samples Age Average Age Agender (Men/Women) Cam-CAN 651 18.5–88.9 54.7 ± 18.6 322/329 To examine the age distribution of participants in the Cam-CAN dataset, the HC (Cam-CAN) data was categorized into groups of 5 years each, resulting in the age distribution graph shown in Fig. 1 . All MRI data were acquired from a 3T Siemens TIM Trio scanner equipped with a 32-channel head coil. The scanner utilized a 3D magnetization-prepared rapid gradient-echo (MPRAGE) sequence to obtain high-resolution 3D T1-weighted images, with the following parameters: TR = 2250 ms, TE = 2.99 ms, inversion time TI = 900 ms, flip angle FA = 9°; field of view FOV = 256×240×192 mm^3, slices = 192; voxel size = 1 \({\text{m}\text{m}}^{3}\) isotropic; GRAPPA acceleration factor = 2; scan time TA = 4 minutes 32 seconds. The data for this project adhered to exclusion criteria similar to other datasets, excluding individuals with neurological disorders (e.g., Parkinson's disease, Alzheimer's disease), psychiatric disorders (e.g., depression), hematologic disorders (e.g., anemia, leukemia), or a history of traumatic brain injury. Additionally, participants with conditions such as migraines, diabetes, or tinnitus were excluded, and only healthy individuals meeting these criteria were included. Ethical approval was obtained locally for data collection in the Cam-CAN project, and subsequently, an anonymized version of the data was made publicly available. U-Net U-Net is a convolutional neural network model initially introduced by Ronneberger et al. 23 in 2015 for image segmentation. Its structure is symmetrical, resembling the uppercase letter "U," and consists of encoding and decoding stages with skip connections in between. The encoding stage performs downsampling, while the decoding stage performs upsampling. The number of feature channels doubles in the encoding stage and halves in the decoding stage. U-Net has found widespread application in the field of image segmentation 24 . However, since medical imaging data is often in three-dimensional format, U-Net needs to slice the 3D data before training, leading to the loss of spatial contextual features in medical images. Therefore, Çiçek et al. 25 proposed 3D U-Net, a network structure similar to 2D U-Net but designed for three-dimensional data. In this study, the architecture of 3D U-Net is adopted to construct the framework for brain age prediction. Tri-UNet Currently, in the work utilizing the U-Net network for predicting brain age, there is a lack of sufficient utilization of features at different scales. Therefore, this paper proposes a network model named Tri-UNet, based on 3D ResNet and 3D U-Net, for the task of brain age prediction. The U-Net's multi-level encoding-decoding structure can combine deep and shallow features, enabling better learning of contextual semantic information from input features. This not only enhances prediction accuracy but also allows the model to learn fine-grained features. The residual structure in ResNet transforms end-to-end feature mappings into residual mappings, facilitating better learning of feature information across network layers and improving prediction accuracy. The network model structure is illustrated in Fig. 2 . This paper, drawing inspiration from the design principles of U-Net, introduces a multi-layer, multi-branch feature learning network called Tri-UNet. Subsequently, by incorporating ideas from Inception and skip connections, the paper proposes a Three-Branch Residual Block (Trible Res Block). The Trible Res Block allows each feature layer to capture information from both higher and lower layers, addressing the issue of degradation in deep networks. Finally, the tail of Tri-UNet is appended with a ResNet 34 network, utilized for predicting brain age based on the features learned by the encoding-decoding structure of Tri-UNet. Trible Res Block As indicated by the two dashed boxes in Fig. 2 , the model structure of the Trible Res Block includes three paths. The first path is the normal input path, which performs no operations, preserving the content of the input features. The second path is the upsampling path, which enlarges the size of the input features. It then undergoes two convolutional block operations with the aim of learning large-sized features after information restoration, thereby enhancing information fitting capabilities. The third path is the downsampling path, which reduces the size of the input feature map through max-pooling. It then executes two basic block operations to learn deep features. After the three paths undergo convolutional operations, the resulting feature maps are concatenated together. It is important to note that the upsampling path needs to undergo downsampling to reduce the size of the feature map before concatenation, while the downsampling path requires upsampling to increase the size of the feature map. This ensures alignment of the feature map sizes from the three paths before concatenation. Finally, the concatenated feature map set is used as the output. Following the Trible Res Block operation, the original input features carry both deep and shallow information simultaneously. The Trible Res Block merges the original input features with the upsampling and downsampling features, enhancing the correlation between upper and lower-level information. Multi-channel input network model Model structure In the multi-channel input network model, the network can learn regional features from several brain regions and use them to regressively predict brain age. Because different brain regions can provide additional complementary information 26 , the neural network in the multi-channel input network model can capture more regional information, thereby improving the accuracy of brain age prediction. Firstly, based on biological prior knowledge, this paper identifies brain regions most relevant to brain age. Then, multiple-channel inputs are fed into the network, consisting of MRI segmentation maps of brain regions that have a significant impact on the accuracy of brain age prediction. Finally, the convolutional neural network produces the ultimate predicted age. The network model is illustrated in Fig. 3 . Here, "Network" can be any convolutional neural network, and in this paper, a 3D ResNet34 network model is employed. Brain region segmentation Many studies indicate that during the process of brain aging, the extent of atrophy varies in different brain regions 27,28 . Moreover, we using measurements of anatomical regions of different sizes as input features, obtaining smaller measurements as input features leads to more accurate predictions of brain age. Therefore, instead of using the entire brain, this paper utilizes several brain regions most relevant to predicting brain age as input data for the prediction framework to obtain regional estimates of brain age. The network can focus more on specific brain regions without being influenced by other parts of the brain. Personalized brain regions can exhibit more sensitive changes with the development of diseases or in response to treatment effects, providing a more detailed representation of changes in brain structure over time 29 . This allows deep learning network models to achieve interpretable results and serves as a reference for more localized brain age predictions. Finally, by combining features from different brain regions, the multi-channel input network model produces more robust predictions of brain age. Experimental Design To validate the effectiveness of the two proposed methods in this paper, comparative experiments were conducted. First, to verify the effectiveness of the improvements on the U-Net network model, whole-brain cropped images were used as input data for the Tri-UNet model to obtain predictions of brain age. These predictions were then compared with the method proposed by Popescu et al., which is also based on the U-Net network for predicting brain age. Secondly, to assess the impact of using different brain regions as input features on age prediction, separate training was conducted using the entire brain and two specific brain regions (the hippocampus and amygdala) with a 3D ResNet 34 model. Furthermore, the hippocampus and amygdala were combined and input into the multi-channel input network model for comparative experiments. The selected brain regions were chosen based on medical prior knowledge, and numerous studies suggest a high correlation between these two brain regions and aging. The experiments were run on a server with a single Tesla V100-SXM2-32GB GPU, using CUDA version 10.3 and Python as the programming language. The model was implemented using the PyTorch 1.3.0 framework. The Adam optimizer was employed with a learning rate of 0.001, the number of epochs set to 60, and a batch size of 4. Result and Discussion Data Preprocessing Typically, in tasks involving predicting brain age based on neuroimaging data, the original images undergo preprocessing steps before being input into the network. 30 This involves slicing the original 3D images into 2D images for network input. However, this approach loses contextual information in the feature space, leading to lower prediction accuracy compared to directly inputting 3D images. Therefore, this study exclusively employs the fully automated processing pipeline "recon-all" from the medical image processing software Freesurfer for the preprocessing of original images. This includes steps such as skull stripping, image correction, image registration, image segmentation, spatial normalization, and spatial smoothing, as illustrated in Fig. 4 , which contrasts the preprocessed image with the original image. ( 1 ) Skull Stripping: In the original medical imaging data, there are non-brain tissues such as the skull, blood vessels, muscles, and cerebellum. To avoid impacting subsequent processing steps, the accuracy of brain tissue segmentation, and the final experimental results, it is common to strip non-brain structures from the image during the preprocessing operation. ( 2 ) Image Correction: The image correction step primarily involves anterior-posterior commissure (AC-PC) correction. It uses standard 256×256×256 mode for resampling and employs the N3 algorithm to correct non-uniform tissue intensity. ( 3 ) Image Registration: For quantitative analysis of several different images, strict alignment of these images is necessary, known as image registration. Medical image registration involves seeking a spatial transformation or a series of spatial transformations for a medical image so that corresponding points on this image match spatially with points on another image. This consistency refers to the same anatomical point on the human body having the same spatial position in two matching images. ( 4 ) Image Segmentation: In MRI data processing, there are instances where only the states of specific regions are of interest. This requires extracting the tissues of the target region based on the brain's anatomical structure. Once the brain regions are identified, the required brain areas are segmented as input images for the network, followed by individual and joint analysis. ( 5 ) Spatial Normalization: Spatial normalization involves registering images to the standard brain template space known as the Montreal Neurological Institute (MNI). This process unifies the coordinate space of all images. The MNI template space is a standardized brain image obtained by averaging a large amount of brain MRI data from healthy subjects and is commonly used as a template for brain image standardization. ( 6 ) Spatial Smoothing: Spatial smoothing is employed to suppress image noise, enhance the signal-to-noise ratio, and reduce inconsistencies in anatomical or functional structures between images. Typically, Gaussian kernel functions with standard deviation are used for smoothing. Comparison with other models First, cropped images with dimensions of 128×128×128 were used as input data for Tri-UNet and two baseline models (U-Net and ResNet 34). Tri-UNet achieved a minimum Mean Absolute Error (MAE) of 7.46 years. Compared to the best MAE of the brain age prediction network proposed by Popescu et al. [48] (9.5 years), Tri-UNet showed an improvement of 2.04 years. Subsequently, cropped images with dimensions of 32×32×32 were used as input data for Tri-UNet and the two baseline models, conducting ablation experiments. Finally, a comparative experiment between single-channel input networks and multi-channel input networks was performed using ResNet34. The experimental results for input data with dimensions of 128×128×128 are shown in Table 2 . Tri-UNetResNet denotes the model where ResNet34 is concatenated directly after the Tri-UNet network for brain age prediction. The results demonstrate the effectiveness of the proposed Tri-UNet method in the task of predicting brain age. Table 2 Comparison of Tri-UNet with Other Models Models Dataset MinMAE MaxMAE MeanMAE U-Net Whole Brain 15.05 43.21 17.23 ResNet34 Whole Brain 9.17 36.6 12.42 Tri-UNet Whole Brain 7.46 16.9 10.05 When the input data size is 128×128×128, as the results predicted by U-Net are far less favorable compared to ResNet34 and the proposed model Tri-UNet concatenated with ResNet34, the comparison is made only between ResNet34 and the proposed model in other evaluation metrics. The experimental results are shown in Table 3 . Table 3 Comparison between Tri-UNet and ResNet34 Models Dataset MAE RMSE MSE ResNet34 Whole Brain 9.17 11.47 131.68 Tri-UNet Whole Brain 7.46 9.32 86.96 Result of single-channel input network To validate the performance of the multi-channel network and the approach using brain region segmentation maps obtained with medical prior knowledge as input data, a comparative experiment was conducted on the 3D ResNet 34 model using single-channel input. The input data consisted of untrimmed images (size: 256×256×256), and the experimental results are shown in Table 4 . Table 4 Results of the single-channel input network (256×256×256) Dataset MinMAE MaxMAE MeanMAE Whole Brain 4.98 115.42 22.4 AM 8.26 225.95 24.91 HA 6.95 154.43 18.78 HBT 9.07 160.46 18.02 When the input data size is 32×32×32, similarly, the performance of the single-channel network was validated on the 3D ResNet 34 model, obtaining different evaluation metrics. The experimental results are shown in Table 5 . Table 5 Results of the single-channel input network (32×32×32) Dataset MAE RMSE MSE Whole Brain 15.18 17.56 308.54 HA 16.7 19.18 367.99 To validate the performance of the multi-channel input network and the method of using anatomically informed brain region segmentation maps as input data, a comparative experiment was conducted on the 3D ResNet 34 model. The entire dataset used untrimmed images (size: 256×256×256), and the experimental results are presented in Table 6 . Table 6 Results of the multi-channel input network Dataset MinMAE MaxMAE MeanMAE HA L + R 8.6 285.96 30.54 HA + AM 8.6 165.65 22.12 HA + Whole Brain 14.47 224.45 31.66 Result of different input The following figure illustrates the changes in Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Squared Error (MSE) loss metrics over training epochs when using the entire brain as input data (as shown in Fig. 5 ): The depicted graph illustrates the evolution of MAE, RMSE, and MSE metrics with the progression of training epochs, utilizing the Amygdala (AM) as input data (refer to Fig. 6 ). The following figure depicts the variations in MAE, RMSE, and MSE metrics with the increase in training epochs, employing the Hippocampus (HA) as input data (as shown in Fig. 7 ). Through the comparison of results for the whole brain, hippocampus (HA), and amygdala presented in Table 4 , it is observed that the minimum MAE for the whole brain is 4.98 years, reaching the lowest value among all experimental outcomes. In contrast, the minimum MAE for HA is 6.95 years, and for AM, it is 8.26 years. This difference may be attributed to the information richness provided by the whole brain as input data, potentially yielding more accurate results in certain rounds of training. Deep learning, which involves the neural network learning detailed features, benefits from a larger Region of Interest (ROI) to enhance global feature representation. The larger ROI or whole brain as input features leads to a larger receptive field, resulting in better predictive performance for brain age. Furthermore, from Table 4 , it is observed that the minimum MAE for HA is lower than that for AM, suggesting that HA is more correlated with age compared to AM. However, in terms of average MAE, the whole brain has an average MAE of 22.40, HA has an average MAE of 18.78, and AM has an average MAE of 24.91. This indicates that larger regions may offer more information for more accurate predictions, but sometimes the smaller region chosen brings less noise. Therefore, a balance must be struck between obtaining more information and achieving finer segmentation features of brain regions. Comparing the curves in Figs. 5 , 6 , and 7 , it can be observed that both HA and AM provide less information than the whole brain, resulting in faster convergence of the models, confirming the above conclusions. Additionally, this study used a high-resolution method to segment the hippocampus (HBT) and used it as input data, as shown in Fig. 8 displaying the loss change curve. From the figure, it is evident that the minimum MAE results for predicting brain age using high-resolution hippocampal images as input data are not as good as those obtained with low-resolution hippocampus (HA) as input data. The minimum MAE for HBT is 9.07 years, which is 2.12 years higher than the result for HA. This may be attributed to the richer information content in high-resolution images, which also introduces more redundant information. This finding aligns with the results of the comparison experiment between the whole brain and HA, providing mutual confirmation for the observed reasons. Hence, to confirm this hypothesis, the whole brain and hippocampus (HA) were used together as input for the multi-channel network to predict the brain age. The resulting loss change curve is shown in Fig. 9 : Observing the case where the whole brain and hippocampus are used together as input, the minimum MAE is 14.47 years. The result is neither as good as training the model with the whole brain alone (MAE of 4.98 years) nor as good as training the model with the hippocampus alone (MAE of 6.95 years). This validates the hypothesis in this study that in the task of predicting brain age, finer voxel predictions do not necessarily yield better results, and including more features does not guarantee better outcomes. In the presence of redundant features, the results may not be as good as those obtained with individual features. Therefore, separating the left and right sides of the hippocampus (HA L + R) as input for the multi-channel network, the loss change curves for the three evaluation metrics (MAE, RMSE, MSE) with increasing training epochs are shown in Fig. 10 . Combining the results in Table 6 , it can be observed that separating the left and right sides of the hippocampus (HA L + R) as input for the dual-channel network does not perform as well as inputting the hippocampus alone. This may be due to both providing the same amount of information, but separating the sides increases the regions with pixel intensity values of 0, resulting in more noise and, therefore, less effective results compared to inputting the hippocampus alone. Trimming off the regions with intensity values of 0 might improve the performance. Finally, using the hippocampus and amygdala (HA + AM) as input for the multi-channel network, the loss change curves for the three evaluation metrics (MAE, RMSE, MSE) with increasing training epochs are shown in Fig. 11 . Combining the information from Table 6 , it is observed that the minimum MAE when using the hippocampus and amygdala together as input for the multi-channel network is 8.60 years. This is slightly higher than the minimum MAE with the entire brain as input (4.98 years). However, the average MAE improves to 22.12 years compared to 22.40 years for the entire brain, demonstrating that a multi-channel input network can enhance stability in performance by combining different data. Nevertheless, as the input images are of the same size, the double-channel input images contain many more regions with pixel grayscale values of 0, as shown in Fig. 12 . In the future, it would be beneficial to crop out the informative regions as input, which may lead to better results. Conclusion This chapter introduces a brain age prediction model, Tri-UNet, based on 3D U-Net and 3D ResNet. The model incorporates residual connections from ResNet, reusing features multiple times on the basis of U-Net, enhancing the correlation between features at different scales, and achieving optimal feature utilization, resulting in good overall performance. Additionally, a multi-channel input network model based on 3D ResNet is proposed, predicting brain age by inputting 3D brain region data determined by anatomical prior knowledge. This approach addresses issues in existing work related to information redundancy or segmentation regions lacking anatomical principles. Experimental results demonstrate an improved performance of the proposed brain age prediction framework compared to existing methods, positioning it as a valuable auxiliary tool in clinical medical research. Declarations Competing interests The authors declare no competing interests. Author Contribution Yu Pang wrote the main manuscript and designed the experiment. Yihuai Cai revised the manuscript Data availability The datasets generated and/or analyzed during the current study are available in the Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data repository ( https://www.cam-can.com/index.php?content=dataset ) References Eliot, L., Ahmed, A., Khan, H. & Patel, J. Dump the “dimorphism”: Comprehensive synthesis of human brain studies reveals few male-female differences beyond size. Neuroscience & Biobehavioral Reviews 125 , 667-697, doi:https://doi.org/10.1016/j.neubiorev.2021.02.026 (2021). Knickmeyer, R. C. et al. A structural MRI study of human brain development from birth to 2 years. J Neurosci 28 , 12176-12182, doi:10.1523/JNEUROSCI.3479-08.2008 (2008). Good, C. D. et al. A voxel-based morphometric study of ageing in 465 normal adult human brains. Neuroimage 14 , 21-36, doi:10.1006/nimg.2001.0786 (2001). Alam, S. B., Nakano, R., Kamiura, N. & Kobashi, S. in 2014 Joint 7th International Conference on Soft Computing and Intelligent Systems (SCIS) and 15th International Symposium on Advanced Intelligent Systems (ISIS). 683-687. Franke, K. & Gaser, C. Longitudinal Changes in Individual BrainAGE in Healthy Aging, Mild Cognitive Impairment, and Alzheimer’s Disease. GeroPsych 25 , 235-245, doi:10.1024/1662-9647/a000074 (2012). Franke, K., Ziegler, G., Kloppel, S., Gaser, C. & Alzheimer's Disease Neuroimaging, I. Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters. Neuroimage 50 , 883-892, doi:10.1016/j.neuroimage.2010.01.005 (2010). Li, Y. et al. Dependency criterion based brain pathological age estimation of Alzheimer's disease patients with MR scans. Biomed Eng Online 16 , 50, doi:10.1186/s12938-017-0342-y (2017). Koutsouleris, N. et al. Accelerated brain aging in schizophrenia and beyond: a neuroanatomical marker of psychiatric disorders. Schizophr Bull 40 , 1140-1153, doi:10.1093/schbul/sbt142 (2014). Pardoe, H. R. et al. Structural brain changes in medically refractory focal epilepsy resemble premature brain aging. Epilepsy Res 133 , 28-32, doi:10.1016/j.eplepsyres.2017.03.007 (2017). Cole, J. H., Leech, R., Sharp, D. J. & Alzheimer's Disease Neuroimaging, I. Prediction of brain age suggests accelerated atrophy after traumatic brain injury. Ann Neurol 77 , 571-581, doi:10.1002/ana.24367 (2015). Wasay, M., Grisold, W., Carroll, W. & Shakir, R. World Brain Day 2016: celebrating brain health in an ageing population. Lancet Neurol 15 , 1008, doi:10.1016/S1474-4422(16)30171-5 (2016). Fleisher, A. S. et al. Chronic divalproex sodium use and brain atrophy in Alzheimer disease. Neurology 77 , 1263-1271, doi:10.1212/WNL.0b013e318230a16c (2011). Kim, J. & Shin, N. Cancer coping, healthcare professionals' support and posttraumatic growth in brain-tumor patients. Psychol Health Med 27 , 780-787, doi:10.1080/13548506.2021.1876890 (2022). Jiang, H. et al. Predicting Brain Age of Healthy Adults Based on Structural MRI Parcellation Using Convolutional Neural Networks. Front Neurol 10 , 1346, doi:10.3389/fneur.2019.01346 (2019). Bintsi, K.-M., Baltatzis, V., Kolbeinsson, A., Hammers, A. & Rueckert, D. in Machine Learning in Clinical Neuroimaging and Radiogenomics in Neuro-oncology Lecture Notes in Computer Science Ch. Chapter 10, 98-107 (2020). Feng, X., Cai, Y. & Xin, R. Optimizing diabetes classification with a machine learning-based framework. BMC Bioinformatics 24 , doi:10.1186/s12859-023-05467-x (2023). Joo, Y. et al. Brain age prediction using combined deep convolutional neural network and multi-layer perceptron algorithms. Sci Rep 13 , 22388, doi:10.1038/s41598-023-49514-2 (2023). Ballester, P. L. et al. Predicting Brain Age at Slice Level: Convolutional Neural Networks and Consequences for Interpretability. Front Psychiatry 12 , 598518, doi:10.3389/fpsyt.2021.598518 (2021). Pardakhti, N. & Sajedi, H. Brain age estimation based on 3D MRI images using 3D convolutional neural network. Multimedia Tools and Applications 79 , 25051-25065, doi:10.1007/s11042-020-09121-z (2020). Ruigrok, A. N. et al. A meta-analysis of sex differences in human brain structure. Neurosci Biobehav Rev 39 , 34-50, doi:10.1016/j.neubiorev.2013.12.004 (2014). Fernández, A. et al. Sex differences in the progression to Alzheimer’s disease: a combination of functional and structural markers. GeroScience , doi:10.1007/s11357-023-01020-z (2023). Popescu, S. G., Glocker, B., Sharp, D. J. & Cole, J. H. Local Brain-Age: A U-Net Model. Front Aging Neurosci 13 , 761954, doi:10.3389/fnagi.2021.761954 (2021). Ronneberger, O., Fischer, P. & Brox, T. in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 Lecture Notes in Computer Science Ch. Chapter 28, 234-241 (2015). Wan, C. et al. Optimized-Unet: Novel Algorithm for Parapapillary Atrophy Segmentation. Front Neurosci 15 , 758887, doi:10.3389/fnins.2021.758887 (2021). Çiçek, Ö., Abdulkadir, A., Lienkamp, S. S., Brox, T. & Ronneberger, O. in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. (eds Sebastien Ourselin et al. ) 424-432 (Springer International Publishing). Zhang, X., Lee, V. C. S., Rong, J., Liu, F. & Kong, H. Multi-channel convolutional neural network architectures for thyroid cancer detection. PLOS ONE 17 , e0262128, doi:10.1371/journal.pone.0262128 (2022). Anders M. Fjell & Kristine B. Walhovd. Structural Brain Changes in Aging: Courses, Causes and Cognitive Consequences. Reviews in the Neurosciences 21 , 187-222, doi:doi:10.1515/REVNEURO.2010.21.3.187 (2010). Fjell, A. M. et al. High consistency of regional cortical thinning in aging across multiple samples. Cereb Cortex 19 , 2001-2012, doi:10.1093/cercor/bhn232 (2009). Cui, W. et al. Personalized fMRI Delineates Functional Regions Preserved within Brain Tumors. Annals of Neurology 91 , 353-366, doi:https://doi.org/10.1002/ana.26303 (2022). Barry, R. L., Strother, S. C. & Gore, J. C. Complex and magnitude-only preprocessing of 2D and 3D BOLD fMRI data at 7 T. Magn Reson Med 67 , 867-871, doi:10.1002/mrm.23072 (2012). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editor invited by journal 31 Dec, 2023 Submission checks completed at journal 31 Dec, 2023 First submitted to journal 29 Dec, 2023 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-3820912","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":264603462,"identity":"47196ede-9d6d-4584-848e-66fc82b01f08","order_by":0,"name":"Yu Pang","email":"","orcid":"","institution":"Jilin Institute of Chemical Technology College of Science","correspondingAuthor":false,"prefix":"","firstName":"Yu","middleName":"","lastName":"Pang","suffix":""},{"id":264603463,"identity":"64f917fd-c4ee-4373-a6fd-7f347981ca27","order_by":1,"name":"Yihuai Cai","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABA0lEQVRIie3PMUsDMRTA8Rci53L11ufkV2gRDkHED+IUDuxi9htqiRRel4JrB+l9BV0OxwuBcwm4Ki7mG3Ss4GDagnS5O8eC+Q+BF96PEIBQaC9jCoQCiA7vVLXM/QXn6o8kNtrN7ZqwLrJVAHidnfbod2wumRpC92xOjuAmxd7i9iqZerLKy0aCVhAKawYENsXj8kXODVNsZj8aSR/WhAwjNktxUNZSecIZtZDEbcgl8ThF8VDLopPg9hVBUZT1tRrJxy6Cb25yJmiYUcy1U3UlnzzRbX9J7ofu/YvOL4rCKfM9GsvFq9Gfq7yZ+A5wZzCbs2rb9/HlzjDuWA6FQqH/2A/9X2Ir4yut6QAAAABJRU5ErkJggg==","orcid":"","institution":"Jilin Institute of Chemical Technology College of Science","correspondingAuthor":true,"prefix":"","firstName":"Yihuai","middleName":"","lastName":"Cai","suffix":""}],"badges":[],"createdAt":"2023-12-29 11:00:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3820912/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3820912/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":49129212,"identity":"e14a11b3-1a0a-496e-bcff-efebf15d0fe4","added_by":"auto","created_at":"2024-01-03 15:13:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":26785,"visible":true,"origin":"","legend":"\u003cp\u003eAge distribution of Cam-CAN participants\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/9a1266bf7bf477b64da46f52.png"},{"id":49129929,"identity":"0aa1f29c-de04-4ff2-aad2-22c7e16c4feb","added_by":"auto","created_at":"2024-01-03 15:21:34","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":142478,"visible":true,"origin":"","legend":"\u003cp\u003eThe architecture of Tri-UNet\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/370988b3f7c301cab6b5f3df.png"},{"id":49129211,"identity":"a93fc85c-c699-41c2-b4b9-01faa80bbfb3","added_by":"auto","created_at":"2024-01-03 15:13:34","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":15414,"visible":true,"origin":"","legend":"\u003cp\u003eMulti-channel input network architecture diagram\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/8806e07c631fbf734e88edf2.png"},{"id":49129215,"identity":"58a1a3eb-d116-4588-8696-598508cc9b71","added_by":"auto","created_at":"2024-01-03 15:13:34","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":148884,"visible":true,"origin":"","legend":"\u003cp\u003eComparison images between post-image preprocessing and the original image\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/ca7dfd3fa178bd2fd376f411.png"},{"id":49129933,"identity":"93497c11-4636-4d48-b82f-f6c316d41006","added_by":"auto","created_at":"2024-01-03 15:21:35","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":104888,"visible":true,"origin":"","legend":"\u003cp\u003eLoss curve of using the whole brain as input\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/290e6b5b25947f48aec89796.png"},{"id":49129214,"identity":"82df0109-9638-4e6f-8484-e2c2ea4f467b","added_by":"auto","created_at":"2024-01-03 15:13:34","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":80807,"visible":true,"origin":"","legend":"\u003cp\u003eLoss curve of using the amygdala as input\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/00cb34c4b723d00ddad86e02.png"},{"id":49129930,"identity":"be2ece33-2383-480f-b541-2229b7829b1f","added_by":"auto","created_at":"2024-01-03 15:21:34","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":74879,"visible":true,"origin":"","legend":"\u003cp\u003eLoss curve of using the hippocampus as input\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/4afa1c3ffc98cc1a3d21d1b6.png"},{"id":49129219,"identity":"6592c8e5-457c-4c55-8a0c-f0b2258d6e44","added_by":"auto","created_at":"2024-01-03 15:13:34","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":81592,"visible":true,"origin":"","legend":"\u003cp\u003eLoss curve of using high-resolution hippocampus as input\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/4632deb0e370bb0f4ce3a3f1.png"},{"id":49129217,"identity":"04ccd96d-673b-4a5b-b8a6-729fc11ea390","added_by":"auto","created_at":"2024-01-03 15:13:34","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":97209,"visible":true,"origin":"","legend":"\u003cp\u003eLoss curve of using both the whole brain and hippocampus as input\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/2c99c90e13bb425f3acfd38f.png"},{"id":49130181,"identity":"a4681454-f328-461e-a9bb-178f19a376b6","added_by":"auto","created_at":"2024-01-03 15:29:34","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":81728,"visible":true,"origin":"","legend":"\u003cp\u003eLoss curve of using the left and right parts of the hippocampus separately as input\u003c/p\u003e","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/7607f84f96def32234e3e639.png"},{"id":49129932,"identity":"3bacf66e-3b41-464a-98e5-7491ed7a0690","added_by":"auto","created_at":"2024-01-03 15:21:34","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":84667,"visible":true,"origin":"","legend":"\u003cp\u003eLoss curve of using the hippocampus and amygdala together as input\u003c/p\u003e","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/ce4d26b84d5decea3400e7fb.png"},{"id":49129221,"identity":"69eba97b-4e2e-44d9-a150-49d09c3bcd5f","added_by":"auto","created_at":"2024-01-03 15:13:35","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":62044,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of the whole brain with the hippocampus and amygdala anatomical structures\u003c/p\u003e","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/a84a0b762c9c138a303791da.png"},{"id":49130960,"identity":"5c582792-d96f-424c-9d60-68fd58e3aedf","added_by":"auto","created_at":"2024-01-03 15:37:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1574622,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3820912/v1/8945582a-a4ca-46b2-bdfe-8a90a48d074c.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Tri-UNet: A Brain Age Prediction Method Based on Different Scale Features of Magnetic Resonance Imaging","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe structure of the human brain undergoes gradual changes throughout the entire lifespan\u003csup\u003e1\u003c/sup\u003e. Knickmeyer et al.\u003csup\u003e2\u003c/sup\u003e, in a longitudinal study of brain structure MRI in healthy individuals, found that the brain volume of newborns is typically only half that of adults, increasing to 90% of adult size by adulthood. Research by Good et al.\u003csup\u003e3\u003c/sup\u003e indicates a gradual decrease in gray matter volume throughout adulthood, with white matter volume following an inverted \"U-shaped\" curve, reaching its peak in middle age.\u003c/p\u003e \u003cp\u003eBinte et al.\u003csup\u003e4\u003c/sup\u003e conducted a morphological analysis of brain structure using structural MRI data from 619 participants aged 19.98 to 83.62 years in the publicly available IXI dataset. The study revealed that as humans reach middle age, the brain undergoes aging with increasing age, accompanied by structural atrophy. The most significant manifestation of this aging process is a decline in cognitive abilities. Furthermore, brain aging increases the risk of brain diseases, especially age-related neurological disorders such as mild cognitive impairment, Alzheimer's disease, schizophrenia, epilepsy, traumatic brain injury.\u003csup\u003e5\u0026ndash;10\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThe various diseases that accompany brain aging impose significant burdens on individuals, families, and society as a whole. While aging is inevitable, interventions in early life can be employed to prevent or reduce brain damage, potentially preventing common neurological diseases\u003csup\u003e11\u0026ndash;13\u003c/sup\u003e. Therefore, accurately assessing the degree of individual brain aging to effectively identify the brain's health status has become a crucial research topic.\u003c/p\u003e \u003cp\u003eWith the rapid development of deep learning technology and its expanding applications, an increasing number of researchers are employing deep learning methods to predict brain age. Jiang et al.\u003csup\u003e14\u003c/sup\u003e obtained brain MRI data from a total of 1454 healthy subjects aged 18\u0026ndash;90 from five publicly available datasets. The data were split into a training set with 1303 samples and a test set with 151 samples. The researchers then segmented the brain into seven different functional networks using the cortical parcellation template (CorticalParcellation_Yeo2011). Subsequently, three methods\u0026mdash;3D CNN, Gaussian Process Regression (GPR), and Relevance Vector Regression (RVR)\u0026mdash;were employed to train models. All three methods performed best on the Frontoparietal Network (FPN), Dorsal Attention Network (DAN), and Default Mode Network (DMN) brain networks. The 3D CNN method demonstrated average Mean Absolute Errors (MAE) of 5.55 years, 5.77 years, and 6.07 years for the three networks, outperforming the machine learning methods GPR and RVR. This work suggests that Convolutional Neural Networks (CNNs) hold significant potential for predicting brain age.\u003c/p\u003e \u003cp\u003eBintsi et al.\u003csup\u003e15\u003c/sup\u003e proposed a patch-based brain age prediction framework using an ensemble approach of 3D Convolutional Neural Network (CNN) and linear regression. The model was trained on the UK Biobank dataset, yielding a final average MAE of 2.46 years. By utilizing patches, the researchers identified brain regions with the greatest impact on brain age prediction, providing interpretable results and advancing anatomical research in the context of deep learning for brain age prediction. This study confirmed the hippocampus and regions such as the ventricles as most relevant to age prediction.\u003c/p\u003e \u003cp\u003eTraditional brain age prediction methods based on anatomical measurements overlook the biological spatial information of brain anatomical structures. This results in a challenging spatial information gap in machine learning-based brain age prediction methods, even for leading algorithms that exhibit significant measurement errors. Because these methods lack context information from neighboring voxels and are insensitive to nonlinear relationships\u003csup\u003e16\u003c/sup\u003e, researchers are increasingly turning their attention to Convolutional Neural Networks (CNNs) in the field of deep learning.\u003csup\u003e17\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eBallester et al.\u003csup\u003e18\u003c/sup\u003e proposed a multi-channel network model based on ResNet18. They used segmented gray and white matter slices as input data for two channels, constructing a convolutional neural network. The final brain age prediction was obtained by averaging the results of three independently trained neural networks. Despite improvements in the MAE evaluation metric in this study, the use of 2D slices as input data prevents the network from integrating contextual information from different brain regions. Additionally, the selection of slices is subject to human factors, leading to significant changes in prediction accuracy when different slice indices are chosen, resulting in a lack of stable predictive performance. Pardakhti et al.\u003csup\u003e19\u003c/sup\u003e, using a 3D ResNet, achieved the best MAE of approximately 5 years. However, this method only provides whole-brain age predictions, whereas current research on brain age prediction is moving towards regional visualization, requiring voxel-level predictions for assessing regional changes in brain aging and lesion locations related to age-related diseases.\u003csup\u003e20,21\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eTherefore, Popescu et al.\u003csup\u003e22\u003c/sup\u003e, based on the U-Net network for image segmentation followed by voxel-level brain age prediction, achieved the best MAE of around 7 years in the frontal cortex and periventricular regions. However, this method solely employs U-Net for training without effectively integrating regional information between different feature layers. Their utilization of multi-scale features is not thorough, and identical features in different brain regions may represent different information, leading to the omission of many features and the loss of substantial spatial information.\u003c/p\u003e \u003cp\u003eTo address these issues, this paper proposes a 3D network architecture, Tri-UNet, based on 3D U-Net and 3D ResNet. Brain region data is utilized as input features for the multi-input network to make it more suitable for brain age prediction tasks. The effectiveness of the proposed method is validated on the publicly available Cam-CAN dataset, demonstrating good accuracy.\u003c/p\u003e \u003cp\u003eThe main contributions of this paper are as follows:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eProposed a brain age prediction method based on the 3D U-Net network structure and residual learning, addressing the issue of insufficient feature utilization in predicting brain age using the U-Net network. This method enhances the utilization of features by repeatedly reusing features through residual connections from ResNet on the foundation of U-Net. It strengthens the correlation between features at different scales, achieving maximal feature utilization and demonstrating good performance.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eProposed a brain age prediction method based on a multi-channel input network, addressing the issue of information redundancy in previous studies using multi-channel input networks to predict brain age. In this paper, the original 3D brain imaging data were segmented into anatomical regions defined by experts, and the segmented brain regions were used as input data for the network. This allows the network to learn specific parameters for different brain regions, capturing region-specific atrophy patterns. Consequently, it integrates complementary information from different brain regions, improving the accuracy of brain age prediction while reducing training time.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDataset\u003c/h2\u003e \u003cp\u003eThe neuroimaging data utilized in this investigation emanate from the publicly accessible Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data repository. This paper utilized T1-weighted MRI data from 651 healthy participants (male/female = 322/329, mean age = 54.7 ± 18.6, age range 18–89 years) from the Cam-CAN repository, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" 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=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDescription of Cam-CAN\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSamples\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAge\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAverage Age\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAgender (Men/Women)\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCam-CAN\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e651\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18.5–88.9\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e54.7 ± 18.6\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e322/329\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e \u003cp\u003e\u003c/p\u003e \u003cp\u003eTo examine the age distribution of participants in the Cam-CAN dataset, the HC (Cam-CAN) data was categorized into groups of 5 years each, resulting in the age distribution graph shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. All MRI data were acquired from a 3T Siemens TIM Trio scanner equipped with a 32-channel head coil. The scanner utilized a 3D magnetization-prepared rapid gradient-echo (MPRAGE) sequence to obtain high-resolution 3D T1-weighted images, with the following parameters: TR = 2250 ms, TE = 2.99 ms, inversion time TI = 900 ms, flip angle FA = 9°; field of view FOV = 256×240×192 mm^3, slices = 192; voxel size = 1\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{m}\\text{m}}^{3}\\)\u003c/span\u003e\u003c/span\u003e isotropic; GRAPPA acceleration factor = 2; scan time TA = 4 minutes 32 seconds.\u003c/p\u003e \u003cp\u003eThe data for this project adhered to exclusion criteria similar to other datasets, excluding individuals with neurological disorders (e.g., Parkinson's disease, Alzheimer's disease), psychiatric disorders (e.g., depression), hematologic disorders (e.g., anemia, leukemia), or a history of traumatic brain injury. Additionally, participants with conditions such as migraines, diabetes, or tinnitus were excluded, and only healthy individuals meeting these criteria were included. Ethical approval was obtained locally for data collection in the Cam-CAN project, and subsequently, an anonymized version of the data was made publicly available.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eU-Net\u003c/h2\u003e \u003cp\u003eU-Net is a convolutional neural network model initially introduced by Ronneberger et al.\u003csup\u003e23\u003c/sup\u003e in 2015 for image segmentation. Its structure is symmetrical, resembling the uppercase letter \"U,\" and consists of encoding and decoding stages with skip connections in between. The encoding stage performs downsampling, while the decoding stage performs upsampling. The number of feature channels doubles in the encoding stage and halves in the decoding stage. U-Net has found widespread application in the field of image segmentation\u003csup\u003e24\u003c/sup\u003e. However, since medical imaging data is often in three-dimensional format, U-Net needs to slice the 3D data before training, leading to the loss of spatial contextual features in medical images.\u003c/p\u003e \u003cp\u003eTherefore, Çiçek et al.\u003csup\u003e25\u003c/sup\u003e proposed 3D U-Net, a network structure similar to 2D U-Net but designed for three-dimensional data. In this study, the architecture of 3D U-Net is adopted to construct the framework for brain age prediction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eTri-UNet\u003c/h2\u003e \u003cp\u003eCurrently, in the work utilizing the U-Net network for predicting brain age, there is a lack of sufficient utilization of features at different scales. Therefore, this paper proposes a network model named Tri-UNet, based on 3D ResNet and 3D U-Net, for the task of brain age prediction. The U-Net's multi-level encoding-decoding structure can combine deep and shallow features, enabling better learning of contextual semantic information from input features. This not only enhances prediction accuracy but also allows the model to learn fine-grained features. The residual structure in ResNet transforms end-to-end feature mappings into residual mappings, facilitating better learning of feature information across network layers and improving prediction accuracy. The network model structure is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThis paper, drawing inspiration from the design principles of U-Net, introduces a multi-layer, multi-branch feature learning network called Tri-UNet. Subsequently, by incorporating ideas from Inception and skip connections, the paper proposes a Three-Branch Residual Block (Trible Res Block). The Trible Res Block allows each feature layer to capture information from both higher and lower layers, addressing the issue of degradation in deep networks. Finally, the tail of Tri-UNet is appended with a ResNet 34 network, utilized for predicting brain age based on the features learned by the encoding-decoding structure of Tri-UNet.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eTrible Res Block\u003c/h2\u003e \u003cp\u003eAs indicated by the two dashed boxes in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the model structure of the Trible Res Block includes three paths. The first path is the normal input path, which performs no operations, preserving the content of the input features. The second path is the upsampling path, which enlarges the size of the input features. It then undergoes two convolutional block operations with the aim of learning large-sized features after information restoration, thereby enhancing information fitting capabilities. The third path is the downsampling path, which reduces the size of the input feature map through max-pooling. It then executes two basic block operations to learn deep features.\u003c/p\u003e \u003cp\u003eAfter the three paths undergo convolutional operations, the resulting feature maps are concatenated together. It is important to note that the upsampling path needs to undergo downsampling to reduce the size of the feature map before concatenation, while the downsampling path requires upsampling to increase the size of the feature map. This ensures alignment of the feature map sizes from the three paths before concatenation. Finally, the concatenated feature map set is used as the output.\u003c/p\u003e \u003cp\u003eFollowing the Trible Res Block operation, the original input features carry both deep and shallow information simultaneously. The Trible Res Block merges the original input features with the upsampling and downsampling features, enhancing the correlation between upper and lower-level information.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eMulti-channel input network model\u003c/h2\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003eModel structure\u003c/h2\u003e \u003cp\u003eIn the multi-channel input network model, the network can learn regional features from several brain regions and use them to regressively predict brain age. Because different brain regions can provide additional complementary information\u003csup\u003e26\u003c/sup\u003e, the neural network in the multi-channel input network model can capture more regional information, thereby improving the accuracy of brain age prediction. Firstly, based on biological prior knowledge, this paper identifies brain regions most relevant to brain age. Then, multiple-channel inputs are fed into the network, consisting of MRI segmentation maps of brain regions that have a significant impact on the accuracy of brain age prediction. Finally, the convolutional neural network produces the ultimate predicted age. The network model is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Here, \"Network\" can be any convolutional neural network, and in this paper, a 3D ResNet34 network model is employed.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eBrain region segmentation\u003c/h2\u003e \u003cp\u003eMany studies indicate that during the process of brain aging, the extent of atrophy varies in different brain regions\u003csup\u003e27,28\u003c/sup\u003e. Moreover, we using measurements of anatomical regions of different sizes as input features, obtaining smaller measurements as input features leads to more accurate predictions of brain age. Therefore, instead of using the entire brain, this paper utilizes several brain regions most relevant to predicting brain age as input data for the prediction framework to obtain regional estimates of brain age. The network can focus more on specific brain regions without being influenced by other parts of the brain. Personalized brain regions can exhibit more sensitive changes with the development of diseases or in response to treatment effects, providing a more detailed representation of changes in brain structure over time\u003csup\u003e29\u003c/sup\u003e. This allows deep learning network models to achieve interpretable results and serves as a reference for more localized brain age predictions. Finally, by combining features from different brain regions, the multi-channel input network model produces more robust predictions of brain age.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eExperimental Design\u003c/h2\u003e \u003cp\u003eTo validate the effectiveness of the two proposed methods in this paper, comparative experiments were conducted. First, to verify the effectiveness of the improvements on the U-Net network model, whole-brain cropped images were used as input data for the Tri-UNet model to obtain predictions of brain age. These predictions were then compared with the method proposed by Popescu et al., which is also based on the U-Net network for predicting brain age. Secondly, to assess the impact of using different brain regions as input features on age prediction, separate training was conducted using the entire brain and two specific brain regions (the hippocampus and amygdala) with a 3D ResNet 34 model. Furthermore, the hippocampus and amygdala were combined and input into the multi-channel input network model for comparative experiments. The selected brain regions were chosen based on medical prior knowledge, and numerous studies suggest a high correlation between these two brain regions and aging.\u003c/p\u003e \u003cp\u003eThe experiments were run on a server with a single Tesla V100-SXM2-32GB GPU, using CUDA version 10.3 and Python as the programming language. The model was implemented using the PyTorch 1.3.0 framework. The Adam optimizer was employed with a learning rate of 0.001, the number of epochs set to 60, and a batch size of 4.\u003c/p\u003e \u003c/div\u003e"},{"header":"Result and Discussion","content":"\u003ch2\u003eData Preprocessing\u003c/h2\u003e\u003cp\u003eTypically, in tasks involving predicting brain age based on neuroimaging data, the original images undergo preprocessing steps before being input into the network.\u003csup\u003e30\u003c/sup\u003e This involves slicing the original 3D images into 2D images for network input. However, this approach loses contextual information in the feature space, leading to lower prediction accuracy compared to directly inputting 3D images. Therefore, this study exclusively employs the fully automated processing pipeline \"recon-all\" from the medical image processing software Freesurfer for the preprocessing of original images. This includes steps such as skull stripping, image correction, image registration, image segmentation, spatial normalization, and spatial smoothing, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, which contrasts the preprocessed image with the original image.\u003c/p\u003e\u003cp\u003e(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Skull Stripping:\u003c/p\u003e\u003cp\u003eIn the original medical imaging data, there are non-brain tissues such as the skull, blood vessels, muscles, and cerebellum. To avoid impacting subsequent processing steps, the accuracy of brain tissue segmentation, and the final experimental results, it is common to strip non-brain structures from the image during the preprocessing operation.\u003c/p\u003e\u003cp\u003e(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) Image Correction:\u003c/p\u003e\u003cp\u003eThe image correction step primarily involves anterior-posterior commissure (AC-PC) correction. It uses standard 256×256×256 mode for resampling and employs the N3 algorithm to correct non-uniform tissue intensity.\u003c/p\u003e\u003cp\u003e(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Image Registration:\u003c/p\u003e\u003cp\u003eFor quantitative analysis of several different images, strict alignment of these images is necessary, known as image registration. Medical image registration involves seeking a spatial transformation or a series of spatial transformations for a medical image so that corresponding points on this image match spatially with points on another image. This consistency refers to the same anatomical point on the human body having the same spatial position in two matching images.\u003c/p\u003e\u003cp\u003e(\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) Image Segmentation:\u003c/p\u003e\u003cp\u003eIn MRI data processing, there are instances where only the states of specific regions are of interest. This requires extracting the tissues of the target region based on the brain's anatomical structure. Once the brain regions are identified, the required brain areas are segmented as input images for the network, followed by individual and joint analysis.\u003c/p\u003e\u003cp\u003e(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) Spatial Normalization:\u003c/p\u003e\u003cp\u003eSpatial normalization involves registering images to the standard brain template space known as the Montreal Neurological Institute (MNI). This process unifies the coordinate space of all images. The MNI template space is a standardized brain image obtained by averaging a large amount of brain MRI data from healthy subjects and is commonly used as a template for brain image standardization.\u003c/p\u003e\u003cp\u003e(\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e) Spatial Smoothing:\u003c/p\u003e\u003cp\u003eSpatial smoothing is employed to suppress image noise, enhance the signal-to-noise ratio, and reduce inconsistencies in anatomical or functional structures between images. Typically, Gaussian kernel functions with standard deviation are used for smoothing.\u003c/p\u003e\u003ch2\u003eComparison with other models\u003c/h2\u003e\u003cp\u003eFirst, cropped images with dimensions of 128×128×128 were used as input data for Tri-UNet and two baseline models (U-Net and ResNet 34). Tri-UNet achieved a minimum Mean Absolute Error (MAE) of 7.46 years. Compared to the best MAE of the brain age prediction network proposed by Popescu et al. [48] (9.5 years), Tri-UNet showed an improvement of 2.04 years. Subsequently, cropped images with dimensions of 32×32×32 were used as input data for Tri-UNet and the two baseline models, conducting ablation experiments. Finally, a comparative experiment between single-channel input networks and multi-channel input networks was performed using ResNet34.\u003c/p\u003e\u003cp\u003eThe experimental results for input data with dimensions of 128×128×128 are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Tri-UNetResNet denotes the model where ResNet34 is concatenated directly after the Tri-UNet network for brain age prediction. The results demonstrate the effectiveness of the proposed Tri-UNet method in the task of predicting brain age.\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" 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\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\u003eComparison of Tri-UNet with Other Models\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModels\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMinMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMaxMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMeanMAE\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e15.05\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e43.21\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e17.23\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet34\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9.17\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e36.6\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12.42\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTri-UNet\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e7.46\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e16.9\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e10.05\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003eWhen the input data size is 128×128×128, as the results predicted by U-Net are far less favorable compared to ResNet34 and the proposed model Tri-UNet concatenated with ResNet34, the comparison is made only between ResNet34 and the proposed model in other evaluation metrics. The experimental results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" 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\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\u003eComparison between Tri-UNet and ResNet34\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModels\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMSE\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet34\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9.17\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e11.47\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e131.68\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTri-UNet\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e7.46\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e9.32\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e86.96\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003ch2\u003eResult of single-channel input network\u003c/h2\u003e\u003cp\u003eTo validate the performance of the multi-channel network and the approach using brain region segmentation maps obtained with medical prior knowledge as input data, a comparative experiment was conducted on the 3D ResNet 34 model using single-channel input. The input data consisted of untrimmed images (size: 256×256×256), and the experimental results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cdiv class=\"gridtable\"\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\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\u003eResults of the single-channel input network (256×256×256)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMinMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMaxMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMeanMAE\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWhole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e4.98\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e115.42\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e22.4\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAM\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8.26\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e225.95\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e24.91\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHA\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6.95\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e154.43\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e18.78\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHBT\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9.07\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e160.46\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e18.02\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003eWhen the input data size is 32×32×32, similarly, the performance of the single-channel network was validated on the 3D ResNet 34 model, obtaining different evaluation metrics. The experimental results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\u003cdiv class=\"gridtable\"\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\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of the single-channel input network (32×32×32)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMSE\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWhole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e15.18\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e17.56\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e308.54\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHA\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16.7\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19.18\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e367.99\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003eTo validate the performance of the multi-channel input network and the method of using anatomically informed brain region segmentation maps as input data, a comparative experiment was conducted on the 3D ResNet 34 model. The entire dataset used untrimmed images (size: 256×256×256), and the experimental results are presented in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e\u003cdiv class=\"gridtable\"\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\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of the multi-channel input network\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMinMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMaxMAE\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMeanMAE\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHA L + R\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8.6\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e285.96\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e30.54\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHA + AM\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e8.6\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e165.65\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e22.12\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHA + Whole Brain\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14.47\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e224.45\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e31.66\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003ch2\u003eResult of different input\u003c/h2\u003e\u003cp\u003eThe following figure illustrates the changes in Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Squared Error (MSE) loss metrics over training epochs when using the entire brain as input data (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e):\u003c/p\u003e\u003cp\u003eThe depicted graph illustrates the evolution of MAE, RMSE, and MSE metrics with the progression of training epochs, utilizing the Amygdala (AM) as input data (refer to Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe following figure depicts the variations in MAE, RMSE, and MSE metrics with the increase in training epochs, employing the Hippocampus (HA) as input data (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThrough the comparison of results for the whole brain, hippocampus (HA), and amygdala presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, it is observed that the minimum MAE for the whole brain is 4.98 years, reaching the lowest value among all experimental outcomes. In contrast, the minimum MAE for HA is 6.95 years, and for AM, it is 8.26 years. This difference may be attributed to the information richness provided by the whole brain as input data, potentially yielding more accurate results in certain rounds of training. Deep learning, which involves the neural network learning detailed features, benefits from a larger Region of Interest (ROI) to enhance global feature representation. The larger ROI or whole brain as input features leads to a larger receptive field, resulting in better predictive performance for brain age. Furthermore, from Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, it is observed that the minimum MAE for HA is lower than that for AM, suggesting that HA is more correlated with age compared to AM. However, in terms of average MAE, the whole brain has an average MAE of 22.40, HA has an average MAE of 18.78, and AM has an average MAE of 24.91. This indicates that larger regions may offer more information for more accurate predictions, but sometimes the smaller region chosen brings less noise. Therefore, a balance must be struck between obtaining more information and achieving finer segmentation features of brain regions. Comparing the curves in Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, and \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e, it can be observed that both HA and AM provide less information than the whole brain, resulting in faster convergence of the models, confirming the above conclusions.\u003c/p\u003e\u003cp\u003eAdditionally, this study used a high-resolution method to segment the hippocampus (HBT) and used it as input data, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e displaying the loss change curve. From the figure, it is evident that the minimum MAE results for predicting brain age using high-resolution hippocampal images as input data are not as good as those obtained with low-resolution hippocampus (HA) as input data. The minimum MAE for HBT is 9.07 years, which is 2.12 years higher than the result for HA. This may be attributed to the richer information content in high-resolution images, which also introduces more redundant information. This finding aligns with the results of the comparison experiment between the whole brain and HA, providing mutual confirmation for the observed reasons.\u003c/p\u003e\u003cp\u003eHence, to confirm this hypothesis, the whole brain and hippocampus (HA) were used together as input for the multi-channel network to predict the brain age. The resulting loss change curve is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e:\u003c/p\u003e\u003cp\u003eObserving the case where the whole brain and hippocampus are used together as input, the minimum MAE is 14.47 years. The result is neither as good as training the model with the whole brain alone (MAE of 4.98 years) nor as good as training the model with the hippocampus alone (MAE of 6.95 years). This validates the hypothesis in this study that in the task of predicting brain age, finer voxel predictions do not necessarily yield better results, and including more features does not guarantee better outcomes. In the presence of redundant features, the results may not be as good as those obtained with individual features.\u003c/p\u003e\u003cp\u003eTherefore, separating the left and right sides of the hippocampus (HA L + R) as input for the multi-channel network, the loss change curves for the three evaluation metrics (MAE, RMSE, MSE) with increasing training epochs are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e. Combining the results in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, it can be observed that separating the left and right sides of the hippocampus (HA L + R) as input for the dual-channel network does not perform as well as inputting the hippocampus alone. This may be due to both providing the same amount of information, but separating the sides increases the regions with pixel intensity values of 0, resulting in more noise and, therefore, less effective results compared to inputting the hippocampus alone. Trimming off the regions with intensity values of 0 might improve the performance.\u003c/p\u003e\u003cp\u003eFinally, using the hippocampus and amygdala (HA + AM) as input for the multi-channel network, the loss change curves for the three evaluation metrics (MAE, RMSE, MSE) with increasing training epochs are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eCombining the information from Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, it is observed that the minimum MAE when using the hippocampus and amygdala together as input for the multi-channel network is 8.60 years. This is slightly higher than the minimum MAE with the entire brain as input (4.98 years). However, the average MAE improves to 22.12 years compared to 22.40 years for the entire brain, demonstrating that a multi-channel input network can enhance stability in performance by combining different data. Nevertheless, as the input images are of the same size, the double-channel input images contain many more regions with pixel grayscale values of 0, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e. In the future, it would be beneficial to crop out the informative regions as input, which may lead to better results.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis chapter introduces a brain age prediction model, Tri-UNet, based on 3D U-Net and 3D ResNet. The model incorporates residual connections from ResNet, reusing features multiple times on the basis of U-Net, enhancing the correlation between features at different scales, and achieving optimal feature utilization, resulting in good overall performance. Additionally, a multi-channel input network model based on 3D ResNet is proposed, predicting brain age by inputting 3D brain region data determined by anatomical prior knowledge. This approach addresses issues in existing work related to information redundancy or segmentation regions lacking anatomical principles. Experimental results demonstrate an improved performance of the proposed brain age prediction framework compared to existing methods, positioning it as a valuable auxiliary tool in clinical medical research.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eYu Pang wrote the main manuscript and designed the experiment. Yihuai Cai revised the manuscript\u003c/p\u003e\n \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eData availability\u003c/h2\u003e \u003cp\u003eThe datasets generated and/or analyzed during the current study are available in the Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data repository (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.cam-can.com/index.php?content=dataset\u003c/span\u003e\u003cspan address=\"https://www.cam-can.com/index.php?content=dataset\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/div\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eEliot, L., Ahmed, A., Khan, H. \u0026amp; Patel, J. Dump the \u0026ldquo;dimorphism\u0026rdquo;: Comprehensive synthesis of human brain studies reveals few male-female differences beyond size. \u003cem\u003eNeuroscience \u0026amp; Biobehavioral Reviews\u003c/em\u003e \u003cstrong\u003e125\u003c/strong\u003e, 667-697, doi:https://doi.org/10.1016/j.neubiorev.2021.02.026 (2021).\u003c/li\u003e\n\u003cli\u003eKnickmeyer, R. C.\u003cem\u003e et al.\u003c/em\u003e A structural MRI study of human brain development from birth to 2 years. \u003cem\u003eJ Neurosci\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 12176-12182, doi:10.1523/JNEUROSCI.3479-08.2008 (2008).\u003c/li\u003e\n\u003cli\u003eGood, C. D.\u003cem\u003e et al.\u003c/em\u003e A voxel-based morphometric study of ageing in 465 normal adult human brains. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 21-36, doi:10.1006/nimg.2001.0786 (2001).\u003c/li\u003e\n\u003cli\u003eAlam, S. B., Nakano, R., Kamiura, N. \u0026amp; Kobashi, S. in \u003cem\u003e2014 Joint 7th International Conference on Soft Computing and Intelligent Systems (SCIS) and 15th International Symposium on Advanced Intelligent Systems (ISIS).\u003c/em\u003e 683-687.\u003c/li\u003e\n\u003cli\u003eFranke, K. \u0026amp; Gaser, C. Longitudinal Changes in Individual BrainAGE in Healthy Aging, Mild Cognitive Impairment, and Alzheimer\u0026rsquo;s Disease. \u003cem\u003eGeroPsych\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 235-245, doi:10.1024/1662-9647/a000074 (2012).\u003c/li\u003e\n\u003cli\u003eFranke, K., Ziegler, G., Kloppel, S., Gaser, C. \u0026amp; Alzheimer\u0026apos;s Disease Neuroimaging, I. Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e50\u003c/strong\u003e, 883-892, doi:10.1016/j.neuroimage.2010.01.005 (2010).\u003c/li\u003e\n\u003cli\u003eLi, Y.\u003cem\u003e et al.\u003c/em\u003e Dependency criterion based brain pathological age estimation of Alzheimer\u0026apos;s disease patients with MR scans. \u003cem\u003eBiomed Eng Online\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 50, doi:10.1186/s12938-017-0342-y (2017).\u003c/li\u003e\n\u003cli\u003eKoutsouleris, N.\u003cem\u003e et al.\u003c/em\u003e Accelerated brain aging in schizophrenia and beyond: a neuroanatomical marker of psychiatric disorders. \u003cem\u003eSchizophr Bull\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, 1140-1153, doi:10.1093/schbul/sbt142 (2014).\u003c/li\u003e\n\u003cli\u003ePardoe, H. R.\u003cem\u003e et al.\u003c/em\u003e Structural brain changes in medically refractory focal epilepsy resemble premature brain aging. \u003cem\u003eEpilepsy Res\u003c/em\u003e \u003cstrong\u003e133\u003c/strong\u003e, 28-32, doi:10.1016/j.eplepsyres.2017.03.007 (2017).\u003c/li\u003e\n\u003cli\u003eCole, J. H., Leech, R., Sharp, D. J. \u0026amp; Alzheimer\u0026apos;s Disease Neuroimaging, I. Prediction of brain age suggests accelerated atrophy after traumatic brain injury. \u003cem\u003eAnn Neurol\u003c/em\u003e \u003cstrong\u003e77\u003c/strong\u003e, 571-581, doi:10.1002/ana.24367 (2015).\u003c/li\u003e\n\u003cli\u003eWasay, M., Grisold, W., Carroll, W. \u0026amp; Shakir, R. World Brain Day 2016: celebrating brain health in an ageing population. \u003cem\u003eLancet Neurol\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 1008, doi:10.1016/S1474-4422(16)30171-5 (2016).\u003c/li\u003e\n\u003cli\u003eFleisher, A. S.\u003cem\u003e et al.\u003c/em\u003e Chronic divalproex sodium use and brain atrophy in Alzheimer disease. \u003cem\u003eNeurology\u003c/em\u003e \u003cstrong\u003e77\u003c/strong\u003e, 1263-1271, doi:10.1212/WNL.0b013e318230a16c (2011).\u003c/li\u003e\n\u003cli\u003eKim, J. \u0026amp; Shin, N. Cancer coping, healthcare professionals\u0026apos; support and posttraumatic growth in brain-tumor patients. \u003cem\u003ePsychol Health Med\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, 780-787, doi:10.1080/13548506.2021.1876890 (2022).\u003c/li\u003e\n\u003cli\u003eJiang, H.\u003cem\u003e et al.\u003c/em\u003e Predicting Brain Age of Healthy Adults Based on Structural MRI Parcellation Using Convolutional Neural Networks. \u003cem\u003eFront Neurol\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 1346, doi:10.3389/fneur.2019.01346 (2019).\u003c/li\u003e\n\u003cli\u003eBintsi, K.-M., Baltatzis, V., Kolbeinsson, A., Hammers, A. \u0026amp; Rueckert, D. in \u003cem\u003eMachine Learning in Clinical Neuroimaging and Radiogenomics in Neuro-oncology\u003c/em\u003e \u003cem\u003eLecture Notes in Computer Science\u003c/em\u003e Ch. Chapter 10, 98-107 (2020).\u003c/li\u003e\n\u003cli\u003eFeng, X., Cai, Y. \u0026amp; Xin, R. Optimizing diabetes classification with a machine learning-based framework. \u003cem\u003eBMC Bioinformatics\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, doi:10.1186/s12859-023-05467-x (2023).\u003c/li\u003e\n\u003cli\u003eJoo, Y.\u003cem\u003e et al.\u003c/em\u003e Brain age prediction using combined deep convolutional neural network and multi-layer perceptron algorithms. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 22388, doi:10.1038/s41598-023-49514-2 (2023).\u003c/li\u003e\n\u003cli\u003eBallester, P. L.\u003cem\u003e et al.\u003c/em\u003e Predicting Brain Age at Slice Level: Convolutional Neural Networks and Consequences for Interpretability. \u003cem\u003eFront Psychiatry\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 598518, doi:10.3389/fpsyt.2021.598518 (2021).\u003c/li\u003e\n\u003cli\u003ePardakhti, N. \u0026amp; Sajedi, H. Brain age estimation based on 3D MRI images using 3D convolutional neural network. \u003cem\u003eMultimedia Tools and Applications\u003c/em\u003e \u003cstrong\u003e79\u003c/strong\u003e, 25051-25065, doi:10.1007/s11042-020-09121-z (2020).\u003c/li\u003e\n\u003cli\u003eRuigrok, A. N.\u003cem\u003e et al.\u003c/em\u003e A meta-analysis of sex differences in human brain structure. \u003cem\u003eNeurosci Biobehav Rev\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, 34-50, doi:10.1016/j.neubiorev.2013.12.004 (2014).\u003c/li\u003e\n\u003cli\u003eFern\u0026aacute;ndez, A.\u003cem\u003e et al.\u003c/em\u003e Sex differences in the progression to Alzheimer\u0026rsquo;s disease: a combination of functional and structural markers. \u003cem\u003eGeroScience\u003c/em\u003e, doi:10.1007/s11357-023-01020-z (2023).\u003c/li\u003e\n\u003cli\u003ePopescu, S. G., Glocker, B., Sharp, D. J. \u0026amp; Cole, J. H. Local Brain-Age: A U-Net Model. \u003cem\u003eFront Aging Neurosci\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 761954, doi:10.3389/fnagi.2021.761954 (2021).\u003c/li\u003e\n\u003cli\u003eRonneberger, O., Fischer, P. \u0026amp; Brox, T. in \u003cem\u003eMedical Image Computing and Computer-Assisted Intervention \u0026ndash; MICCAI 2015\u003c/em\u003e \u003cem\u003eLecture Notes in Computer Science\u003c/em\u003e Ch. Chapter 28, 234-241 (2015).\u003c/li\u003e\n\u003cli\u003eWan, C.\u003cem\u003e et al.\u003c/em\u003e Optimized-Unet: Novel Algorithm for Parapapillary Atrophy Segmentation. \u003cem\u003eFront Neurosci\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 758887, doi:10.3389/fnins.2021.758887 (2021).\u003c/li\u003e\n\u003cli\u003e\u0026Ccedil;i\u0026ccedil;ek, \u0026Ouml;., Abdulkadir, A., Lienkamp, S. S., Brox, T. \u0026amp; Ronneberger, O. in \u003cem\u003eMedical Image Computing and Computer-Assisted Intervention \u0026ndash; MICCAI 2016.\u003c/em\u003e (eds Sebastien Ourselin\u003cem\u003e et al.\u003c/em\u003e) 424-432 (Springer International Publishing).\u003c/li\u003e\n\u003cli\u003eZhang, X., Lee, V. C. S., Rong, J., Liu, F. \u0026amp; Kong, H. Multi-channel convolutional neural network architectures for thyroid cancer detection. \u003cem\u003ePLOS ONE\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, e0262128, doi:10.1371/journal.pone.0262128 (2022).\u003c/li\u003e\n\u003cli\u003eAnders M. Fjell \u0026amp; Kristine B. Walhovd. Structural Brain Changes in Aging: Courses, Causes and Cognitive Consequences. \u003cem\u003eReviews in the Neurosciences\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 187-222, doi:doi:10.1515/REVNEURO.2010.21.3.187 (2010).\u003c/li\u003e\n\u003cli\u003eFjell, A. M.\u003cem\u003e et al.\u003c/em\u003e High consistency of regional cortical thinning in aging across multiple samples. \u003cem\u003eCereb Cortex\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 2001-2012, doi:10.1093/cercor/bhn232 (2009).\u003c/li\u003e\n\u003cli\u003eCui, W.\u003cem\u003e et al.\u003c/em\u003e Personalized fMRI Delineates Functional Regions Preserved within Brain Tumors. \u003cem\u003eAnnals of Neurology\u003c/em\u003e \u003cstrong\u003e91\u003c/strong\u003e, 353-366, doi:https://doi.org/10.1002/ana.26303 (2022).\u003c/li\u003e\n\u003cli\u003eBarry, R. L., Strother, S. C. \u0026amp; Gore, J. C. Complex and magnitude-only preprocessing of 2D and 3D BOLD fMRI data at 7 T. \u003cem\u003eMagn Reson Med\u003c/em\u003e \u003cstrong\u003e67\u003c/strong\u003e, 867-871, doi:10.1002/mrm.23072 (2012).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-3820912/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3820912/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn the process of human aging, significant age-related changes occur in brain tissue. To assist individuals in assessing the degree of brain aging, screening for disease risks, and further diagnosing age-related diseases, it is crucial to develop an accurate method for predicting brain age. This paper proposes a multi-scale feature fusion method called Tri-UNet based on the U-Net network structure, as well as a brain region information fusion method based on multi-channel input networks. These methods address the issue of insufficient image feature learning in brain neuroimaging data. They can effectively utilize features at different scales of MRI and fully leverage feature information from different regions of the brain. In the end, experiments were conducted on the Cam-CAN dataset, resulting in a minimum Mean Absolute Error (MAE) of 7.46. The results demonstrate that this method provides a new approach to feature learning at different scales in brain age prediction tasks, contributing to the advancement of the field and holding significance for practical applications in the context of elderly education.\u003c/p\u003e","manuscriptTitle":"Tri-UNet: A Brain Age Prediction Method Based on Different Scale Features of Magnetic Resonance Imaging","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-03 15:13:30","doi":"10.21203/rs.3.rs-3820912/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvited","content":"","date":"2023-12-31T12:06:35+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2023-12-31T12:04:44+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2023-12-29T10:57:06+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8d2dd25c-8d2e-460f-9b4c-700ffd8525c6","owner":[],"postedDate":"January 3rd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":27883337,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":27883338,"name":"Health sciences/Diseases"},{"id":27883339,"name":"Health sciences/Health care"},{"id":27883340,"name":"Physical sciences/Engineering"}],"tags":[],"updatedAt":"2024-06-04T06:13:37+00:00","versionOfRecord":[],"versionCreatedAt":"2024-01-03 15:13:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3820912","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3820912","identity":"rs-3820912","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}