3D Convolutional Neural Network Model for Detection of Major Depressive Disorder from Grey Matter Images

Abstract

Major depressive disorder is a mental health condition characterized by ongoing feelings of sadness, trouble focusing or making decisions, and a frequent sense of fatigue or hopelessness that lasts for a prolonged period. If left undiagnosed, it can have serious consequences, including suicide. This study proposes a 3D convolutional neural network model to detect major depressive disorder using 3D grey matter images from magnetic resonance imaging. The proposed 3D convolutional architecture comprises multiple hierarchical convolutional and pooling layers, designed to automatically learn spatial patterns from magnetic resonance imaging data. The model was optimized via Bayesian hyperparameter tuning, achieving an accuracy of 72.26%, an area under the receiver operating characteristic curve of 0.80, and an area under the precision–recall curve of 0.81 on a large multisite dataset comprising 1276 patients and 1104 healthy controls. Gradient-weighted class activation mapping is utilized to find brain regions associated with major depressive disorder. From this study, six regions were identified, namely, the frontal lobe, parietal lobe, temporal lobe, thalamus, insular cortex and corpus callosum which may be affected by major depressive disorder.

1. Introduction

Mental health is a vital part of overall well-being, influencing how one handles stress, relates to others, and makes choices in life. Major depressive disorder (MDD) is a serious mental health condition characterized by a consistently low mood, less interest in usual activities, and disturbances in daily functioning, such as altered sleep patterns, appetite changes, and reduced energy levels. MDD may affect anyone at any age, but it is usually prevalent during early adulthood. MDD affects women about twice as often as it does men and one in six people will experience it at some point in their lives [1]. Research shows that MDD is one of the main risk factors for suicide and about 60% of people who lose their life have a history of MDD [2]. Additionally, MDD is commonly co-morbid with physical illnesses like cancer, stroke, and acute coronary syndrome, which can negatively impact quality of life and increase mortality rates [3]. Given the association of MDD with high rates of mortality and comorbidity, it is important to detect it as early as possible. MDD diagnosis typically involves clinical evaluation, including interviews and questionnaires. The gold standard for diagnosing MDD is clinical interviews and rating scales like the Hamilton Depression Scale [4] and use of the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-V) [5]. But these rely on subjective reports and clinician judgment, which can cause variability or delayed diagnoses. Alternatively, magnetic resonance imaging (MRI) can be used to distinguish affected individuals from healthy controls (HCs) [6]. Grey matter (GM), white matter (WM), and cerebrospinal fluid (CSF) are three essential elements of the brain that can be identified and extracted from MRI scans [7,8]. GM primarily consists of neural cell bodies, dendrites, and synapses, forming the outer layer of the brain. WM is made up of groups of nerve fibers covered in a fatty layer called myelin, which makes it look white. CSF is a clear liquid that encloses and protects the brain and spinal cord, filling the spaces within and around the central nervous system to act as a protective cushion. On MRI, these structures are differentiated based on their signal intensities [9]. Several MRI studies have found significant and region-specific changes in grey matter volume in MDD patients which correlate with MDD symptoms [10,11,12]. These structural changes serve as features for detection of MDD using 3D grey matter images.

Recent research employing machine learning and deep learning techniques has shown encouraging results in the identification of MDD. Deep learning models are increasingly used for classifying diseases based on MRI, as they can automatically extract features rather than relying on manually extracted features [13,14,15]. In traditional machine learning approaches like support vector machines and random forests, manually extracted features are used as input for classification. These approaches have demonstrated effectiveness in MDD classification but might be limited by their dependence on predefined feature extraction methods. A study by Wang et al. employed 3D DenseNet for classification of MDD and HCs, obtaining an accuracy of 77.42% [16]. Lin et al. in their late-life depression study on functional magnetic resonance imaging with 77 participants, using a 3D convolutional neural network (3D CNN), achieved an accuracy of 85% [17]. Li et al. utilized fMRI scans and employed a deep neural network to extract features, which were then classified using a kernel extreme learning machine, obtained a classification accuracy of 89% [18]. Chemin and colleagues used a 3D convolutional neural network on 83 resting state fMRI images to forecast suicidal tendencies in elderly individuals suffering from depression with an accuracy of 78.5% [19]. Belov et al. classified MDD from MRI data using a support vector machine algorithm and achieved an accuracy of 62% [20]. A study conducted by Alotaibi et al. employed a multi-atlas ensemble graph neural network for classification of MDD using functional MRI, obtaining an accuracy of 75.80% [21]. A study by Wu et al. employed a multi-task learning approach to simultaneously predict patient survival outcomes and performed semi-supervised segmentation of gliomas from brain MRI scans [22]. Yan et al. used deep learning signatures with diffusion tensor imaging data for improved glioma survival prediction [23]. Qu et al. carried out a study on MDD classification using MRI data, where they applied an attention-guided unified deep convolutional neural network and achieved an accuracy of 76.54% [24]. Qianqian et al. classified MDD by employing an adaptive multimodal neuroimage integration framework for classification using MRI and functional MRI images and obtained an accuracy of 65% [25].

A review of the literature shows that the majority of studies have employed two-dimensional CNN models for the classification of MDD. Since MRI is a 3D imaging modality, this study proposes an MDD-Net 3D convolutional neural network model to detect major depressive disorder from grey matter images. The 3D CNN model is capable of capturing spatial information from an image, allowing it to learn complex features relevant for classification. The proposed MDD-Net was trained on a large-scale multisite dataset comprising a total of 2380 subjects obtained from the REST-meta-MDD project of the DIRECT consortium [26]. Further, gradient-weighted class activation mapping was employed to identify the brain regions affected by major depressive disorder.

This paper is structured as follows: Section 2 provides a detailed description of the dataset and the proposed MDD-Net model developed specifically for the classification of major depressive disorder. Section 3 presents the experimental results and discusses the main findings. The main conclusions derived from the study are presented in Section 4.

2. Materials and Methods

2.1. Dataset

The dataset used in this study is obtained from the REST-meta-MDD project of the Depression Imaging REsearch ConsorTium (DIRECT) [26,27]. The dataset consists of a total of 2380 subjects from 17 hospitals, which include 1104 healthy controls and 1276 individuals with major depressive disorder.

Table 1. Demographic information of the dataset used in the study.

	MDD	Healthy Controls
No. of Subjects	1276	1104
Male	463	462
Female	813	642
Average Age	36 years Range (14–80)	36 years Range (12–82)

presents the demographic information of the dataset used in this study.

2.2. Data Preprocessing

This study utilizes preprocessed data that are already available in the REST-meta-MDD project [27]. The project adopted a standardized preprocessing pipeline at all participating sites to minimize heterogeneity in preprocessing methods. An overview of the MRI preprocessing pipeline is shown in Figure 1. Structural T1-weighted MRI scans were acquired using standard imaging protocols and preprocessed with Processing Assistant for Resting-State fMRI, Statistical Parametric Mapping version 8 and the Voxel-Based Morphometry toolbox version 8. Each T1 image was segmented into grey matter, white matter and cerebrospinal fluid, followed by spatial normalization to the Montreal Neurological Institute (MNI) space using the Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) algorithm. The normalized images were then smoothed with an 8 mm full-width at half-maximum (FWHM) Gaussian kernel to reduce noise and account for inter-individual variability. Quality control was conducted through assessment of sample homogeneity via covariance checks and visual inspection of single slices across all subjects. A total of 2380 preprocessed 3D grey matter images were obtained. These 3D grey matter images were then used as input to the proposed MDD-Net model for detecting major depressive disorder.

2.3. MDD-Net for Detection of Major Depressive Disorder

The overall framework for the detection of major depressive disorder is shown in Figure 2. Initially, 3D grey matter images are used as input to a 3D convolutional neural network model for the classification of MDD and HCs. The model consists of six convolutional blocks designed for hierarchical feature extraction, a 3D max-pooling layer for spatial downsampling, and a fully connected layer which integrates features captured across the entire input to produce the final classification. Feature maps from the trained model are used to visually highlight the specific brain regions using the gradient-weighted class activation mapping technique.

2.3.1. MDD-Net Model

The proposed MDD-Net model is fed with a 3D grey matter MRI image of dimension 121 × 145 × 121 (depth × height × width), allowing it to learn important features directly from the image. The model comprises six sequential convolutional blocks, each designed for hierarchical feature extraction. Each block is made up of two consecutive 3D convolutional layers, both using a 3 × 3 × 3 kernel with padding set to 1, and each layer is directly followed by a ReLU (Rectified Linear Unit) activation function. After these convolutional layers, a 3D max-pooling layer with a 2 × 2 × 2 kernel and a stride of 2 is applied to perform spatial downsampling, which reduces computational complexity and increases the receptive field in deeper layers. The network starts with 8 filters in its first convolutional layer and then doubles the number of filters in each new block, so the layers use 8, 16, 32, 64, 128, and finally, 256 filters. In the last two convolutional layers, 128 filters are used in block 5 and 256 filters are used in block 6 for experimental purposes. This gradual increase allows the network to capture more intricate feature representations as the depth of the model grows. Following the fourth convolutional block, the resulting 3D feature maps are flattened into a 28,224-size one-dimensional vector. This vector is passed into the classifier module, beginning with a fully connected layer that decreases the feature dimension to 128. A one-dimensional batch normalization layer is applied, followed by a ReLU activation and a dropout layer with a 0.7 dropout rate, to help prevent overfitting. After a fully connected layer reduces the feature size from 128 to 64, a ReLU activation function is employed to incorporate non-linearity into the model, enabling the network to more effectively learn and identify key patterns within the data. The proposed 3D CNN model is as shown in Figure 3. A detailed summary of each convolutional layer’s operations, output dimensions, filter counts and parameter counts is provided in

Table 2. The parameters used in the MDD-Net model.

Layer (Type)	Output Shape	Parameters
Input Layer	(B, 1, 121, 145, 121)	0
Block 1
Conv3D (8 filters, 3 × 3 × 3)	(B, 8, 121, 145, 121)	224 (216 + 8)
Conv3D (8 filters, 3 × 3 × 3)	(B, 8, 121, 145, 121)	1736 (1728 + 8)
MaxPool3D (2 × 2 × 2)	(B, 8, 60, 72, 60)	0
Block 2
Conv3D (16 filters, 3 × 3 × 3)	(B, 16, 60, 72, 60)	3472 (3456 + 16)
Conv3D (16 filters, 3 × 3 × 3)	(B, 16, 60, 72, 60)	6928 (6912 + 16)
MaxPool3D (2 × 2 × 2)	(B, 16, 30, 36, 30)	0
Block 3
Conv3D (32 filters, 3 × 3 × 3)	(B, 32, 30, 36, 30)	13,856 (13,824 + 32)
Conv3D (32 filters, 3 × 3 × 3)	(B, 32, 30, 36, 30)	27,680 (27,648 + 32)
MaxPool3D (2 × 2 × 2)	(B, 32, 15, 18, 15)	0
Block 4
Conv3D (64 filters, 3 × 3 × 3)	(B, 64, 15, 18, 15)	55,360 (55,296 + 64)
Conv3D (64 filters, 3 × 3 × 3)	(B, 64, 15, 18, 15)	110,656 (110,592 + 64)
MaxPool3D (2 × 2 × 2)	(B, 64, 7, 9, 7)	0
Block 5
Conv3D (128 filters, 3 × 3 × 3)	(B, 128, 7, 9, 7)	221,312 (221,184 + 128)
Conv3D (128 filters, 3 × 3 × 3)	(B, 128, 7, 9, 7)	442,496 (442,368 + 128)
MaxPool3D (2 × 2 × 2)	(B, 128, 3, 4, 3)	0
Block 6
Conv3D (256 filters, 3 × 3 × 3)	(B, 256, 3, 4, 3)	884,992 (884,736 + 256)
Conv3D (256 filters, 3 × 3 × 3)	(B, 256, 3, 4, 3)	1,769,728 (1,769,472 + 256)
MaxPool3D (2 × 2 × 2)	(B, 256, 1, 2, 1)	0
Fully Connected Layers
Fully connected 1 (512 → 128)	(B, 128)	65,664 (65,536 + 128)
BatchNorm1D (128)	(B, 128)	256 (128 + 128)
Dropout	(B, 128)	0
Fully connected 2 (128 → 64)	(B, 64)	8256 (8192 + 64)
Fully connected 3 (64 → 2)	(B, 2)	130 (128 + 2)

. All parameters in the 3D CNN model are trainable and are updated during backpropagation.

2.3.2. Loss Function and Convergence

The classifier terminates in a fully connected layer that yields two raw output scores (logits) $z=(z_1,z_2)$ for the two output classes. For training, the standard approach involves using the $c\in \{1,2\}$ multi-class cross-entropy loss function, which typically combines a log-softmax operation with the negative log-likelihood loss. For a single instance with true class c (where $c\in \{1,2\}$) and logits z, this loss is commonly formulated as:

$$L_{\mathrm{CE}}(z,c)=-log{\displaystyle \frac{e^{z_c}}{{\sum }_{j=1}^2{e}^{z_j}}}$$

(1)

where $z_c$ is the logit corresponding to the true (actual) class, $e^{z_c}$ gives the exponential score for the true class and ${\sum }_{j=1}^2{e}^{z_j}$ gives the sum of the exponentials for all classes.

With this direct formulation, training consistently stalled at $L\approx 0.69$. To restore learning, the logits were first transformed to class probabilities

$$p_i={\displaystyle \frac{e^{z_i}}{{\sum }_{j=1}^2{e}^{z_j}}},i=1,2$$

(2)

where $e^{z_i}$ is the exponential transformation of the logit for class, and ${\sum }_{j=1}^2{e}^{z_j}$ is the normalization term (sum over exponentials of all logits), ensuring all output probabilities sum to 1. These probabilities are then passed to the CrossEntropyLoss,

$$L_{\mathrm{custom}}(z,c)=\mathrm{CrossEntropy}(\mathrm{Softmax}\left(z\right),c).$$

(3)

Empirically, utilizing this custom loss formulation enabled the training process to successfully converge, leading to the results reported in this study. While the standard cross entropy loss is generally preferred for its numerical stability and direct application to logits, our findings indicate that in this specific experimental context, the standard approach failed, whereas the custom loss formulation provided a path to effective model training. The CrossEntropyLoss internally performs a log-softmax, supplying the explicit soft-maxed outputs, resulting in reliable convergence.

2.3.3. Hyperparameter Optimization

To systematically determine the optimal model configurations and key hyperparameters, a hyperparameter optimization (HPO) based on Bayesian optimization is employed. Bayesian optimization is a probabilistic optimization technique designed to efficiently find the minimum or maximum of expensive, black-box functions, especially when evaluations are costly or noisy. It works by building a surrogate probabilistic model of the objective function and using an acquisition function to select the most promising next point to evaluate, balancing exploration and exploitation [28]. The main goal of HPO is to reduce the loss measured on the validation dataset. During HPO, the study explored several important hyperparameters, including model depth (comparing architectures with either four or six convolutional blocks), optimizer choice (evaluating both Adam and AdamW), initial learning rate (testing values of 0.01, 0.001, and 0.0001), and loss function formulation (contrasting the standard cross-entropy loss with a custom two-step loss, as described in Section 2.3.2). To maintain consistency and reliability throughout the experiments, specific parameters were kept constant in all runs. The original sequence is preserved across the convolutional layers by applying the rectified linear unit (ReLU) activation function and Kaiming initialization is used as a weight initialization method. A batch size of 16 is maintained for both training and validation, and all models were trained for a fixed duration of 50 epochs. Furthermore, to reduce the risk of overfitting, dropout layers are included in the fully connected head, with a dropout rate of 0.7 determined from the initial experimental results. This experimental setup provided a robust and reproducible framework for evaluating the performance of various 3D CNN configurations.

Table 3. Hyperparameters used in the MDD-Net model.

Activation function	Rectified linear unit
Weight initialization	Kaiming initialization
Optimizer	AdamW
Model depth	4
Dropout rate	0.7
No. of epoch	50
Learning Rate	0.001
Loss function	Custom Categorical Cross-Entropy Loss
Batch size	16

provides information about the hyperparameters used for training the final model.

3. Results

3.1. Experimental Results

The experiments were conducted on a large multisite dataset from the REST-meta-MDD project of the DIRECT consortium, consisting of 2380 subjects, including 1104 healthy controls and 1276 individuals with major depressive disorder. All experiments were conducted using PyTorch version 1.13.1 as the deep learning framework. The model training was performed on a computer equipped with a 13th Generation Intel Core i7-13620H processor (2.40 GHz), 16.0 GB of RAM, and an NVIDIA RTX 4050 GPU 6 GB VRAM. The dataset was divided into three separate subsets, namely, blind test, validation and training sets. A proportion of 10% was set aside as a blind test set, while 90% of the data were assigned at random for training. In our experiments, 10-fold cross-validation without stratification was used. The validation set plays a crucial role in the tuning of hyperparameters, helping to reduce the likelihood of overfitting and improving the overall performance of the model. After model training and hyperparameter tuning, the final model was evaluated using a blind test set that was kept separate from both the training and validation stages. This approach ensured an unbiased assessment of the model’s generalization capability. Model performance on the blind test set was quantified using standard metrics, such as accuracy, precision and recall, thereby providing a comprehensive evaluation of the model’s effectiveness on previously unseen data. The blind test set, which was separated from the original dataset prior to model training, was utilized exclusively for the evaluation of the final model. A variety of evaluation metrics were employed to rigorously assess the performance and robustness of the proposed model. Advanced performance measurements, including the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC), were computed in addition to standard classification metrics like accuracy, sensitivity, specificity, and the F1-score. The models are designed to output non-normalized prediction scores (logits) for the two target classes, providing a direct measure of their discriminative capability. The loss function acted as the main optimization metric, indicating how effectively the model could differentiate between classes using its outputs.

A comprehensive comparison of selected model configurations, focusing on the effects of model depth and loss function choice, is presented in

Table 4. Configuration of models from hyperparameter tuning.

Config	Model Depth	Optimizer	Learning Rate ($\mathit{\eta }$)	Loss Function Formulation
Config-A	4	Adam	0.01	$L_{custom}$
Config-B	4	AdamW	0.001	$L_{custom}$
Config-C	4	AdamW	0.001	$L_{standard}$
Config-D	6	AdamW	0.001	$L_{custom}$
Config-E	6	Adam	0.0001	$L_{custom}$
Config-F	6	AdamW	0.001	$L_{standard}$

. Config-A achieves high sensitivity (0.93), meaning it is very effective at identifying positive cases, but low specificity (0.36), so it misclassifies negatives as positives, resulting in an accuracy of 66.95%. Config-B offers a much better balance, whereby it maintains a good sensitivity (0.83), improved specificity (0.59) and achieves the highest accuracy of 72.26%, along with the highest AUROC (0.80) and AUPRC (0.81) values, making it the most reliable configuration. Config-C, in contrast, has extremely high specificity (0.99) but very low sensitivity (0.13) and an accuracy of 52.38%, indicating it rarely identifies positives correctly and mostly predicts negative outcomes. Config-D provides high sensitivity (0.75) and specificity (0.67) as well as an accuracy of 71.43%. Config-E is similar to Config-D, with good specificity (0.76), lower sensitivity (0.67) and accuracy of 72.00%. Config-F, much like Config-C, has high specificity (0.98) but low sensitivity (0.13) and ends up with the lowest accuracies (53.78%). Among all the configurations, Config-B is found to be the best choice with AUROC of 0.80 (95% Confidence Interval: 0.748–0.843) and AUPRC of 0.81 (95% Confidence Interval: 0.761–0.861), as it offers the most balanced performance across key classification metrics, as shown in

Table 5. Classification performance MDD-Net model.

Config	Accuracy	Sensitivity	Specificity	F1 Score	AUROC	AUPRC
Config-A	66.95%	0.93	0.36	0.75	0.74	0.78
Config-B	72.26%	0.83	0.59	0.76	0.80	0.81
Config-C	52.38%	0.13	0.99	0.22	0.77	0.77
Config-D	71.43%	0.75	0.67	0.74	0.76	0.76
Config-E	72.00%	0.67	0.76	0.72	0.80	0.81
Config-F	53.78%	0.13	0.98	0.24	0.75	0.77

. These results underscore the model’s generalization ability and effectiveness in accurately classifying MDD subjects and healthy controls.

Additional experiments were conducted to compare the proposed 3D convolutional neural network model with the 3D-ResNet 50 model pretrained on medical images. During training, 3D-ResNet 50 exhibited significant convergence issues, with the loss function consistently plateauing around 0.693 across multiple epochs despite extensive hyperparameter tuning. Due to these convergence challenges, when adapting the pretrained model to our specific dataset characteristics, direct performance comparison was not feasible. However, the classification performance of our proposed 3D convolutional neural network model was evaluated against results from previous studies and a comparative summary is provided in

Table 6. Performance comparison.

SI No.	Authors	Sample Size	Model Used	Accuracy
1.	Yuqi et al. [29]	MDD = 441, HC = 395	DFH	57.7%
2.	Wang et al. [30]	MDD = 54, HC = 62	Ensemble model	72.4%
3.	Lin et al. [17]	MDD = 49, HC = 28	3D CNN	85%
4.	Hong et al. [31]	MDD = 34, HC = 34	3D FRN-ResNet	86.7%
5.	Proposed	MDD = 1276, HC = 1104	3D CNN	72.26%

. Yuqi et al. in their MDD classification study proposed a model dual-expert fMRI harmonization (DFH) and obtained an accuracy of 57.7% [29]. Wang et al. utilized 128 MRI images for classification of MDD using a multimodal fusion model that incorporates a CNN (high-frequency encoder based on SFCN), Transformer (low-frequency encoder) and multilayer perceptron along with ensemble learning, obtained an accuracy of 72.4% [30]. Lin et al. conducted a study involving 77 fMRI subjects to distinguish late-life depression from HCs, achieving an accuracy of 85% using a 3D CNN model [17]. Hong et al. utilized a dataset of 68 MRI scans to classify MDD using a 3D feature map reconstruction network built upon the ResNet architecture (3D FRN-ResNet), achieving an accuracy of 86.7% [31]. From Table 6, it is observed that the previous studies have used a lower number of subjects, which may limit the generalizability of the models. The proposed 3D convolutional neural network model was specifically trained from scratch (not fine-tuned from pretrained weights) on a large multisite dataset of 2380 subjects, allowing it to learn features directly relevant to distinguishing major depressive disorder patients from healthy controls. This dedicated training on MDD specific grey matter patterns, combined with the custom loss function that enabled successful convergence, allows the model to achieve its 72.26% classification accuracy. The proposed MDD-Net provides a generalizable solution for major depressive disorder classification, offering clear methodological and sample size advantages over previous studies.

3.2. Identification of Brain Regions Affected by Major Depressive Disorder

Gradient-weighted class activation mapping (Grad-CAM) is an explainable AI technique for visualizing class-specific activation maps by leveraging gradients to highlight important regions in the input image. This approach helps in explaining convolutional neural network model predictions [32,33]. Grad-CAM offers insights into how deep learning models process visual information and make decisions. It enhances trust and accountability in artificial intelligence systems by making their predictions more interpretable [34,35]. In this study, Grad-CAM is utilized on the final convolutional layer of the model to identify the brain regions associated with MDD. It overlays a heatmap on the image, using warm colors (reds and yellows) to mark the regions that contributed most to the model’s decision. Identification of these regions is accomplished by estimating the target prediction’s gradients with respect to the models final layer’s feature mappings. The feature maps are weighted according to the gradients, then combined and subjected to a ReLU activation that highlights only positive effects. The MDD brain regions identified by using Grad-CAM are the frontal lobe, temporal lobe, parietal lobe, thalamus, corpus callosum and insular cortex. Figure 4 and Figure 5 display the regions highlighted through the Grad-CAM analysis, revealing the regions that contributed most to the model’s decision. Figure 4a shows the parietal lobe and insular cortex regions and Figure 4b shows the temporal lobe. Figure 5a shows the thalamus and corpus callosum regions and Figure 5b shows the frontal lobe.

Table 7. Identified brain regions affected by major depressive disorder and their functions.

SI No.	Brain Regions	Functionalities
1.	Frontal Lobe	Planning, decision-making, problem-solving, voluntary motor control, behavior, emotional regulation, speech and hearing
2.	Parietal Lobe	Integration of somatosensory information like touch, pain, temperature, proprioception, spatial orientation, perception, aspects of language and mathematical processing
3.	Temporal Lobe	Primary auditory processing, language comprehension, encoding and retrieval of declarative memory, visual object and face recognition and emotional processing
4.	Thalamus	Regulation of consciousness, memory, emotion, alertness, sleep-wake cycles, relaying sensory and motor signals to the cerebral cortex
5.	Insular Cortex	Involved in interceptive awareness, taste processing, automatic control, pain perception and emotional regulation
6.	Corpus Callosum	Communication and coordination of motor, sensory and cognitive functions

lists the identified brain regions affected by major depressive disorder along with their corresponding functions.

The frontal lobe regions identified from this study are responsible for for a wide range of essential brain functions, particularly those related to higher-order cognition, behavior, and motor control [36]. Numerous studies show that the prefrontal cortex, including regions like the dorsolateral prefrontal cortex, orbitofrontal cortex, and subgenual prefrontal cortex, exhibit a reduction in grey matter volume among patients suffering from MDD [37,38]. These reductions are associated with impaired emotional regulation, decision-making, and motivation. The temporal lobe regions identified in this study are responsible for processing auditory information, language comprehension, memory formation and retrieval, emotion regulation and visual object recognition [39]. Several MRI studies have indicated significant reductions in grey matter volume in the temporal lobes, particularly the hippocampus and amygdala regions which are often associated with mood regulation [40,41]. The parietal lobe identified in this study is primarily responsible for processing sensory information and supporting spatial awareness, coordination, math skills, and language comprehension [42]. Chen et al. reported a reduced magnetization transfer ratio in the left superior parietal lobule of MDD patients, potentially linked to cognitive and attentional impairments [43]. According to a review by Zhang et al., MDD causes structural and functional alterations in the parietal lobe, which hampers learning and information processing [38]. Kili et al. reported reduced parietal lobe volume in MDD patients compared to healthy individuals, linking it to cognitive deficits like poor attention and working memory common in mood disorders [44].

The thalamus region identified in this study plays a key role in relaying motor and sensory signals to the cortex and is involved in regulating sleep, alertness, emotion, memory and attention [45,46]. Several studies have examined thalamic volume deficits using MRI and found that thalamic atrophy could be a structural marker in MDD [47,48]. Further, the findings highlighted altered thalamic activity and connections with other brain regions involved in mood regulation. The corpus callosum identified as a key region in this study is responsible for linking the left and right hemispheres of the brain, facilitating communication between them and coordinating motor, sensory, and cognitive activities [49]. A study by Lee et al. indicated that the corpus callosum, a significant brain structure linking the two hemispheres, exhibits changes in individuals with MDD, which may reflect brain dysfunction associated with the disorder [50]. Piras et al. reported a study that identified structural alterations in the corpus callosum in individuals with MDD, indicating notable thinning or atrophy, especially among those with chronic forms of the condition [51]. The insular cortex region identified in this study integrates sensory, emotional and autonomic information, which is crucial for self-awareness, cognitive control and autonomic regulation [52]. Multiple studies have identified reduced functional connectivity between the left insula and the amygdala, which may disrupt mood regulation [53,54]. According to a study by Takahashi et al. changes in the insular cortex’s general structure may increase the likelihood that MDD patients develop neurological defects and might also make their symptoms more severe [55].

Grad-CAM analysis was also conducted to provide a direct visual comparison of brain regions between MDD and healthy controls. The healthy control brain regions identified by using Grad-CAM are the parietal lobe, and the thalamocapsular and insular cortex regions. Figure 6 shows the brain regions highlighted using Grad-CAM for healthy controls. Figure 6a shows the parietal lobe and thalamocapsular regions. Figure 6b shows the parietal lobe and insular cortex regions. The parietal lobe in healthy controls is mainly responsible for integrating sensory information, supporting spatial awareness and movement coordination, while the thalamocapsular region functions as a relay center for sensory and motor signals, which plays a role in maintaining alertness and consciousness. The insular cortex region is responsible for bodily sensations, emotional awareness and social experiences in healthy controls.

The findings from the study indicate that the brain regions affected in MDD include the frontal lobe, temporal lobe, parietal lobe, thalamus, corpus callosum and insular cortex. These brain regions play a role in thinking, emotional regulation, memory and how different parts of the brain work together. Problems in the frontal lobe can affect decision-making and mood control, while changes in the parietal lobe can negatively affect attention and planning. The temporal lobe is linked to memory and feelings, and the thalamus helps pass information around the brain. The corpus callosum connects the two halves of the brain and the insular cortex helps a person feel emotions. Identifying brain regions which are affected in major depressive disorder enables clinicians to diagnose the condition more confidently, anticipate its progression and provide effective treatments.

4. Conclusions

The MDD-Net is proposed in this study to classify individuals with major depressive disorder from healthy controls using magnetic resonance imaging data. The model is trained on 3D grey matter MRI images for classification of major depressive disorder, achieving an accuracy of 72.26%. In addition, the proposed 3D convolutional neural network model attained an area under the receiver operating characteristic curve of 0.80 (95% Confidence Interval: 0.748–0.843), indicating strong discriminative ability and area under the precision–recall curve of 0.81 (95% Confidence Interval: 0.761–0.861), further highlighting the model’s reliability in identifying true positive cases of major depressive disorder. By employing gradient-weighted class activation mapping, the study identified six brain regions, namely, the frontal lobe, temporal lobe, parietal lobe, thalamus, corpus callosum and insular cortex which may be affected by major depressive disorder. The functions of these brain regions including self-awareness, emotional regulation and cognitive processing are mainly associated with major depressive disorder. The findings from the study demonstrate the effectiveness of the proposed 3D convolutional neural network model for major depressive disorder detection and highlight its potential in clinical decision support.