Depression diagnosis method based on K-GAT network

Abstract

The invention discloses a depression diagnosis method based on a K-GAT network, and the method comprises the following steps: 1, obtaining the brain region thickness of each brain region based on a brain magnetic resonance image; constructing a brain topological graph by taking the brain regions as nodes, taking correlation connection between the brain regions as edges between the nodes and taking the extracted thicknesses of the brain regions as features of the nodes; the brain topological graph comprises a brain topological graph of a depression patient and a brain topological graph of a healthy person; step 2, constructing a K-GAT network for classifying the brain magnetic resonance images; screening whether there are many brain region connections, and if there are many brain region connections, determining that the brain magnetic resonance image of the patient with depression is the brain magnetic resonance image; otherwise, a healthy human brain magnetic resonance image is obtained; and step 3, classifying the brain magnetic resonance images by adopting the trained K-GAT network, and distinguishing a depression patient from a non-depression patient. The method comprises the following steps of: obtaining a brain topological graph, inputting the brain topological graph into a K-GAT network, and training the K-GAT network to obtain a trained K-GAT network.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a depression diagnosis method based on a K-GAT network. Background Art

[0002] In recent years, with the in-depth study of depression, people have gradually realized that it is not just a simple problem of low mood, but a brain dysfunction involving structural and functional abnormalities in multiple brain regions. Neuroimaging studies have shown that patients with depression have obvious organic lesions in specific brain areas, such as those related to emotion processing and cognitive executive functions. However, these changes are usually only detectable after the disease has persisted for a long time. Therefore, traditional detection methods based on structural magnetic resonance imaging (MRI) have significant limitations in the early identification of depression.

[0003] Functional magnetic resonance imaging (fMRI) provides a new way to explore depression. By detecting the functional connectivity differences of the brain in a resting state, researchers can reflect the functional connectivity characteristics of the brain's neural network. With the rapid development of deep learning technology, researchers have tried to apply graph neural networks (GCN) to the analysis of functional brain networks (FBNs) to extract potential graph representations, thereby more effectively distinguishing patients with depression from healthy individuals.

[0004] Although the brain network analysis method based on graph neural network GCN has shown initial potential in depression diagnosis, it still has some shortcomings:

[0005] 1. Noise problem: When constructing functional brain networks, the commonly used Pearson Correlation (PC) method will produce a denser connection network. However, these connection networks often contain a lot of noise and false connections, which mainly come from technical limitations, operator performance and environmental factors. Unfiltered noise not only reduces the classification accuracy of the model, but also may lead to overfitting of the model, thereby increasing computational complexity.

[0006] 2. Insufficient model interpretability: When the existing graph neural network GCN model is used for brain image classification, the model's multi-layer nonlinear transformation and complex graph structure representation make the model's decision-making process difficult to explain. This "black box" feature aggregation process makes it difficult for clinicians to clearly understand which brain area connections play a key role in depression diagnosis, thus limiting the model's application in actual clinical practice.

[0007] 3. Lack of training data: The effectiveness of deep learning models usually relies on a large amount of high-quality training data. However, the current depression-related fMRI datasets used for training are relatively limited. In particular, when different medical institutions collect fMRI data, the heterogeneity of the data increases due to differences in the selection of brain regions. This problem limits the generalization ability of the model, making it difficult to show consistent performance on different datasets. Summary of the invention

[0008] In order to solve the defects of the prior art, the present invention proposes a depression diagnosis method based on graph attention network. It includes the following steps:

[0009] Step 1, obtaining the brain thickness of each brain region based on brain magnetic resonance images;

[0010] Taking brain regions as nodes, correlation connections between brain regions as edges between nodes, and the extracted thickness of each brain region as a feature of the node, a brain topology map is constructed; the brain topology map includes brain topology maps of patients with depression and brain topology maps of healthy people; the healthy people refer to people who do not suffer from depression;

[0011] Step 2: construct a K-GAT network for classifying brain magnetic resonance images; input the brain topology map into the K-GAT network, train the K-GAT network, and obtain a trained K-GAT network.

[0012] Step 3: Use the trained K-GAT network to classify brain magnetic resonance images.

[0013] Furthermore, in step 1, the brain thickness of each brain region is obtained based on the brain magnetic resonance image, which specifically includes the following steps:

[0014] Step 1.1, obtaining a whole brain gray matter image based on a brain magnetic resonance image;

[0015] The gray matter image includes gray matter volume GMV and gray matter concentration GMC;

[0016] Step 1.2, based on the gray matter image, obtaining the gray matter thickness GMT and gray matter surface area SA corresponding to each brain region of the whole brain;

[0017] Step 1.3, constructing a first brain map based on the gray matter volume GMV, gray matter concentration GMC, gray matter thickness GMT and gray matter surface area SA;

[0018] Step 1.4, calculating the low frequency amplitude ALFF, the fractional low frequency amplitude fALFF and the local consistency ReHo for the first brain map to obtain the brain thickness of each brain region.

[0019] Furthermore, in the brain topology map, brain region thickness is used as a node feature Where n is the number of nodes, i.e. the number of brain regions, and d in is the feature dimension; the connection relationship between nodes uses the adjacency matrix Indicates that A ij Represents the connection relationship between node i and node j, the adjacency matrix A ij Each element in is assigned a value according to the threshold, 1 if it is greater than the threshold and 0 if it is less than the threshold.

[0020] Furthermore, the K-GAT network includes a first linear layer, a K-GAT layer, and a second linear layer.

[0021] The node feature X in the brain topology map is input into the first linear layer in the K-GAT network to obtain the hidden feature d hidden Represents the feature dimension of the linear layer;

[0022] Hidden feature H (0) It is expressed as:

[0023] H (0) =W in X

[0024] in, is the weight matrix of the linear layer, X represents the thickness of the brain area;

[0025] Hidden feature H (0) With the adjacency matrix A ij Input to the K-GAT layer, for node i and its neighbor node j, update node i to obtain the updated node feature H (L) .

[0026] The characteristic H of each node (L) After the second linear layer, the feature H (L) The dimension is from d hidden Compress to the final output dimension d out ;

[0027] Z=W out H (L)

[0028] in, is the final output result, d out is the output category;

[0029] The output result Z applies log_softmax to convert the output of each node into the logarithmic probability P of the category;

[0030] P = log\_softmax(Z)

[0031] in, is the log probability of the node in each category.

[0032] Furthermore, the hidden feature H (0) and the adjacency matrix A ij Input to the K-GAT layer, for node i and its neighbor node j, update node i to obtain the updated node feature H (L) , specifically including the following steps;

[0033] Step 2.1 Linearly transform the hidden features

[0034] Hidden features for each node Perform a linear transformation:

[0035]

[0036] in, is the weight matrix, l represents the number of iterations;

[0037] Step 2.2 When A ij = 1, calculate the attention coefficient e between node i and its neighbor node j ij :

[0038]

[0039] in is a learnable attention vector;

[0040] ‖ represents the concatenation operation of features;

[0041] LeakyReLU(·) represents the activation function;

[0042] Step 2.3, for the attention coefficient e ij Normalize and get the normalized attention weight α ij ;

[0043]

[0044] in Represents the neighbor set of node i

[0045] Step 2.4: Perform weighted summation on the neighbor node features of node i to obtain the new hidden feature representation of node i

[0046]

[0047] Among them, σ is a nonlinear activation function, W′ is a shared weight matrix;

[0048] Step 2.5: Use the multi-head attention mechanism to the new hidden features of node i The concatenation is expressed as:

[0049]

[0050] Among them, K represents the number of attention heads, ‖ represents the concatenation operation, and W (k) are the attention weight and linear transformation matrix of the kth head respectively.

[0051] Step 2.6: Set the feature of node i Input to the KAN layer, after further feature extraction, let l = l + 1, return to step 2.1, and finally get the feature H of each node (L) .

[0052] Furthermore, step 2 also includes pruning the trained K-GAT model to prune the network into a smaller sub-network.

[0053] Beneficial effects: Compared with the traditional graph neural network GCN model, the K-GAT network model of the present invention can more effectively deal with the noise problem in the brain network by introducing the attention mechanism. The Kolmogorov-Arnold-graph attention network K-GAT model assigns weights to different node connections, and by learning the importance of connections between different brain regions, it is closer to the real brain map structure, reduces the impact of noise, and thus improves the classification accuracy and accuracy of the model.

[0054] In addition, the K-GAT model also has significant advantages in improving the interpretability of the model. Due to the introduction of the Kolmogorov-Arnold network layer (KAN layer), the model can more intuitively show which brain regions play a key role in the diagnosis of depression by analyzing the frequency of retained features after pruning during the classification stage, thereby providing clinicians with more powerful decision support.

[0055] At the same time, the present invention effectively alleviates the problem of limited existing data sets by performing multimodal analysis on patients' functional magnetic resonance imaging, and improves the generalization ability and diagnostic accuracy of the model. It not only improves the early diagnosis ability of depression, but also provides new tools and methods for the formulation of personalized treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 This is a schematic diagram of the network architecture of K-GAT.

[0058] Figure 2 This is a schematic diagram of pruning after adding the KAN layer.

[0059] Figure 3 It is a diagram of the key brain regions that are preserved and their connections.

[0060] Figure 4 It is a comparison chart of the brain topology of patients and healthy people. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the present invention clearer, the specific implementation of the present invention is described below in combination with the implementation mode and the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and are not intended to limit the scope of the invention claimed for protection. All other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] Step 1: Cooperate with a hospital to obtain fMRI images of more than 1,000 real patients with depression, and perform format conversion, denoising, and registration to obtain the patients' whole brain gray matter volume GMV and gray matter concentration GMC. ① Convert the original functional magnetic resonance imaging fMRI DICOM format file into NIFTI format that can be used for data analysis;

[0063] ② Use SPM12 to divide, standardize and modulate the structural phase, or only divide and standardize to obtain gray matter volume GMV and gray matter concentration GMC brain map, that is, gray matter image, and resample to 1mm x 1mm x 1mm size;

[0064] ③ After visual inspection of the gray matter images, all subjects' gray matter images were taken and correlation analysis was performed with the average gray matter image of all subjects. Images with a correlation coefficient lower than 0.8 were considered abnormal and were removed from the following analysis.

[0065] ④ Create a mask with voxel value > 0.2 to filter and remove other possible noise signals.

[0066] ⑤ Registration. fMRI image registration is the process of aligning or superimposing one or more groups of fMRI images in order to spatially unify images of different time points or different subjects. The importance of this process lies in the fact that fMRI images may be inconsistent in position and posture during acquisition due to factors such as slight movements of the subjects, different equipment, and differences in scanning angles, which may affect subsequent data analysis and interpretation. Through image registration, these spatial differences can be eliminated, so that different fMRI images can be accurately overlapped in the same spatial reference system, so as to conduct time series analysis, comparison of brain region functions, and inter-group statistical analysis.

[0067] Step 2: The brain regions were divided according to the Desikan-Killiany (DK) template. Based on existing studies, 68 brain regions with a high correlation with brain diseases such as depression and autism were selected to obtain the gray matter thickness GMT and surface area SA of the 68 brain regions.

[0068] ① Refer to the Freesurfer script used and published by the ENIGMA working group to preprocess the gray matter images;

[0069] ②This script will generate a CSV file. The first line is the title, describing the names of the extracted brain regions. Each line after the first line gives the gray matter thickness GMT and surface area SA of each brain region found in the FreeSurfer directory;

[0070] ③ Perform quality control on the gray matter thickness GMT and surface area SA results of different brain regions obtained above in R-studio, using a method similar to the aforementioned association analysis. At the same time, potential outliers will be marked to remind you to pay attention in further quality control;

[0071] ④ Perform inner surface quality control in Matlab and outer surface quality control in Freesurfer, referring to the relevant steps of ENIGMA.

[0072] Step 3: Troubleshoot interference factors.

[0073] The above-mentioned brain structure indicators are residualized, and the structural indicators refer to gray matter volume GMV, gray matter concentration GMC, gray matter thickness and gray matter surface area SA; first, the effects of age and gender on brain structure are removed by regression analysis, and then the first brain map is reconstructed based on gray matter volume GMV, gray matter concentration GMC, gray matter thickness GMT and gray matter surface area SA to prepare for the next step of analysis. This part of the analysis will be performed in Matlab through self-written scripts and programs. After constructing the first brain map, the course of disease, equivalent dose of psychiatric drugs, and potential gender-age interactions are analyzed across the whole brain to exclude the above factors from having a serious impact on brain structure.

[0074] Step 4: Preprocessing and calculation of low frequency amplitude (ALFF), fractional low frequency amplitude (fALFF) and local consistency (ReHo)

[0075] ① The functional magnetic resonance imaging (fMRI) data of the first 10 time points were removed to avoid the influence of MRI machine field uniformity and subjects' discomfort on the results, and the last 230 time points were included in the subsequent analysis;

[0076] ② Correct the acquisition time difference between different scanning layers at the same time point (time correction) and the subject's head movement effect at different time points during the entire data acquisition process (spatial correction). After spatial correction, check the head movement parameters of each subject and exclude subjects whose head movement in any direction is >3mm or any angle is >3° during the scanning process;

[0077] ③ After head motion correction, the brain images of each subject were normalized to the standardized brain MNI space developed by the Montreal Neurological Institute (MNI) according to the EPI template and using an optimized 12-parameter affine transformation; the voxel size was further resampled to 3

[0078] ×3×3mm3;

[0079] ④ Smooth according to the 6 mm smoothing core;

[0080] ⑤ Remove data linear drift and perform 0.01-0.1Hz bandpass filtering on the data to reduce the impact of high-frequency physiological noise;

[0081] ⑥ The white matter signal, cerebrospinal fluid signal, whole brain average signal and 24 head movement parameters were regressed and removed from the original time series voxel by voxel as covariates.

[0082] ⑦ Calculate the low-frequency amplitude ALFF, fractional low-frequency amplitude fALFF and local consistency ReHo in the brain imaging data processing and analysis toolbox DPABI to obtain the specific thickness of each brain region.

[0083] Step 5: Organize the above specific brain map data, build brain area connections, and create a data set.

[0084] The thickness of brain regions is sorted into csv format according to the sample individuals and different brain regions, and the correlation analysis of the data is performed according to the correlation coefficients between each brain region, and presented in the form of a matrix. Then, the threshold of correlation is set through experiments to determine the connection relationship between brain regions and construct an adjacency matrix.

[0085] Brain regions are used as nodes, correlation connections between brain regions are used as edges between nodes, the extracted thickness of each brain region is used as the feature of the node, and the edge connections between nodes are constructed according to the Pearson correlation coefficient between each brain region, thereby realizing the construction of a brain topology map.

[0086] In the present invention, a single case sample is taken as an example of graph input. The input brain topology graph consists of node features and structural information. The node features are a 68×1 matrix, where 68 is the number of nodes (68 brain regions) and 1 is the feature dimension of the node (brain region thickness).

[0087] Step 6: Build the K-GAT network and use the brain map as the graph input for graph classification

[0088] In view of the low accuracy and poor interpretability of existing models and their inability to focus on key brain areas, we propose the following Figure 1 A graph attention network architecture with a KAN layer added is shown.

[0089] First, brain region thickness is used as input node feature Where n is the number of nodes, i.e. the number of brain regions, and d in As the input feature dimension, in this specific implementation, there is only one feature dimension, the brain region thickness feature. The correlation between brain regions is used as the input edge, with the adjacency matrix , which represents the connection relationship between nodes (brain regions), where A ij Indicates the connection relationship between node i and node j. Each element in the adjacency matrix is ​​assigned a value according to the threshold (1 is assigned if it is greater than the threshold, and 0 is assigned if it is less than the threshold). The element value can be regarded as whether it constitutes a connection. If it is set to 1, it is connected, and if it is set to 0, it is not connected. Through the linear layer (lin_in), the input feature X is mapped to the hidden feature space That is: H (0) =W in X, where is the weight matrix of the linear layer, H (0) is the initial hidden layer feature, d hidden represents the feature dimension of a linear layer, wherein the hidden layer is a linear layer;

[0090] Then we enter the middle layer, the graph attention layer. The graph attention mechanism is the core of GAT, which performs weighted aggregation on the neighbors of each node. During aggregation, the weights of different neighbors are controlled by the learned attention coefficient. Specifically, given node i and its neighbor node j, the update process of GAT is divided into the following steps:

[0091] (a) Linear transformation:

[0092] Hidden features for each node Perform a linear transformation:

[0093]

[0094] in, is the weight matrix, d ′ is the dimension after linear transformation, d represents the feature dimension, and l represents the number of iterations.

[0095] (b) Attention coefficient calculation:

[0096] Next, the attention coefficient between node i and its neighbor j is calculated. ij Indicates the importance of node j to node i. The calculation formula is as follows:

[0097]

[0098] in, is a learnable attention vector, ‖ represents the concatenation operation of features, e ij is the unnormalized attention coefficient, and LeakyReLU(·) represents the activation function.

[0099] (c) Softmax normalization:

[0100] In order to get the normalized attention weight α ij , apply the Softmax function to all neighbor nodes j of each node i:

[0101]

[0102] in Represents the neighbor set of node i.

[0103] (d) Weighted aggregation:

[0104] Finally, the weighted sum of the neighbor node features of node i is performed to obtain the new hidden feature representation of node i

[0105]

[0106] Where σ is a nonlinear activation function (such as ReLU) and W′ is a shared weight matrix.

[0107] (e) Multi-head attention mechanism:

[0108] In order to increase the expressive power of the model, GAT uses a multi-head attention mechanism. Through multiple independent attention mechanisms, each head generates a hidden feature representation, and then these representations are spliced ​​together:

[0109]

[0110] Among them, K represents the number of attention heads, ‖ represents the concatenation operation, and W (k) are the attention weight and linear transformation matrix of the kth head respectively;

[0111] After each GAT feature extraction, the output results Passed to KANLayer for further processing, the learnable univariate function G in KAN j,i (x) is parameterized by B-spline. Fourier functions are selected as spline functions, i.e., univariate functions; each univariate function operates on a predefined interval [a, b], which is evenly divided into g-1 subintervals (where g represents the number of grid points). In each subinterval, the function consists of q-order (i.e., q-1 degree) polynomial fragments that are smoothly connected at the grid points. Each univariate function G j,i (x) can be expressed as a linear combination of g+q-1 B-spline basis functions:

[0112]

[0113] Among them, w l is a learnable weight parameter, B l,q (x) is a standard B-spline basis function. By learning these weight parameters, the model can adaptively adjust the shape of each univariate function to capture the complex nonlinear relationship between the input features.

[0114] This B-spline-based representation method not only ensures the smoothness of the function, but also provides sufficient flexibility to approximate various nonlinear mappings. KANLayer takes the features of 68 brain regions extracted by graph neural network processing as input. After multiple layers of GAT and KANLayer, the final hidden feature H (L) After the second linear layer, the dimension is reduced from d hidden Compress to the final output dimension d out ; Z = W out H (L) ,in, is the final output result, d out is the number of output categories (for binary classification tasks, d out =2). Finally, log_softmax is applied to the output results to convert the output of each node into the logarithmic probability of the category: P = log\_softmax(Z), where is the log probability of the node in each category.

[0115] like Figure 1As shown in Figure 1, the input includes node features and edge connection information. The edge is updated through the graph attention mechanism, where the thickness of the edge represents the weight of the edge. The larger the weight, the thicker the edge. Taking the directed edge CA as an example, the edge aggregation attention mechanism is jointly determined by all the edges and node features connected to C. The update of node features is implemented through the K-GAT network, which is affected by all nodes directly connected to the node. Figure 1 Taking the update of node F as an example, we first perform neighborhood aggregation. F is directly connected to nodes E, D, and G, and the aggregation process is determined by these connections. Specifically, the contributions of edges FE, FD, and FG are and Where α is the weight in the normalized adjacency matrix, h (l) is the node feature after the lth update. These contributions are summed up to form the aggregation result Then, the aggregation results As input into the KAN layer, it passes through a series of learnable univariate functions G j,i (x) processing. Assume that the aggregation result is d hidden dimensional vector, the output of the KAN layer is m-dimensional. The KAN layer applies m functions to each dimension i. The jth component of the KAN layer output is Finally, the new features of node F are determined by the output of the KAN layer: Among them, y m Represents the value of the mth dimension of the KAN layer output;

[0116] Where: i ranges from 1 to d hidden , represents the dimension of the input feature; is the i-th component of the aggregated input feature vector; G j,i is a learnable univariate function of the j-th output dimension on the i-th input dimension; therefore, y m Specifically refers to the last component of the KAN layer output vector, which is obtained by summing all input dimensions after the corresponding univariate function transformation. j Together they form a new feature representation of the node [y1,y2,…,y m ].

[0117] During training, the univariate function G in the KAN layer j,i The (x) parameters are optimized through back propagation, so that the model can learn the most suitable nonlinear transformation. This architecture first aggregates the graph structure and then applies nonlinear transformation, effectively combining graph structure information and node features, and enhancing the expressiveness of the model through learnable nonlinear functions.

[0118] Step 7: Set the threshold and calculate and verify the degree of association based on the pruning results (feature frequency).

[0119] After obtaining the trained K-GAT model, in order to visualize the importance of each brain region, the following steps are required: Figure 2 The pruning operation shown in the figure is used to highlight the key brain areas, and the threshold and neurons to be retained are set to prune the trained K-GAT model, and the nodes (brain areas) with smaller input or output connections are pruned and deleted to retain the important brain areas. Specifically: KAN is sparse at the node level. For example, for each node, the incoming fraction I of the u-th neuron in the KAN layer in the w-th layer is defined as w,u and outgoing score O l,i They are:

[0120] I w,u =max r (|φ w-1,r,u |1)

[0121] O w,u =max j (|φ w+1,j,u |1)

[0122] φ w-1,r,u represents the activation function connecting the rth and uth neurons in the w-1th layer;

[0123] φ w+1,j,u represents the activation function connecting the jth and uth neurons in the lth layer.

[0124] If the score I is passed in w,u and outgoing score O w,u If both are smaller than the set threshold, the neuron is considered to be unimportant and is pruned.

[0125] According to the characteristic frequencies of the brain regions that are finally retained, the above steps are combined to finally obtain the following Figure 3 The 17 key brain regions shown in the figure also showed significant differences in thickness and connectivity between patients (MDD) and healthy controls (HC), such as Figure 4 As shown, the size of the node (ball) represents the thickness of the brain region, and the connection of the edge represents the connection between brain regions. It can be clearly observed Figure 4 (a) The brain regions of patients with MDD have significantly more connections than Figure 4 Middle (b) HC of a healthy person, and there are also certain differences in the thickness of brain regions such as the left precuneus.

[0126] The K-GAT model proposed in this invention has significant improvements in multiple indicators of graph classification (accuracy, recall, etc.) compared with existing models and traditional machine learning, as shown in Table 1.

[0127] Table 1 Evaluation indicators of different models

[0128]

[0129] like Figure 2 As shown, the model proposed in the present invention uses the graph attention network to extract features, and then inputs the extracted features into the KAN layer for classification. By setting the threshold and pruning operations, the node feature frequency is obtained. The node feature frequency refers to the number of times the node appears, and the frequency (number of times) of the remaining brain regions after pruning. The more it appears, the higher the frequency, and the more important the brain region, thereby reflecting the degree of association between the brain region and depression.

[0130] According to the characteristic frequencies reflected by the final pruning results, the degree of correlation between each brain region and depression is shown in Table 2.

[0131] Table 2 Correlation between brain regions and disease

[0132]

[0133]

[0134]

[0135] By analyzing the disease associations and actual structural and functional characteristics of brain regions, the following conclusions were drawn:

[0136] ①Regional perspective.

[0137] The degree of association with the disease is from high to low: parietal lobe, temporal lobe, frontal lobe, limbic system, cingulate gyrus, parahippocampal gyrus, lingual gyrus, temporal pole, triangular part, inferior frontal gyrus, and paracentral lobule.

[0138] Among the above regions, only some sub-regions are related to the classification of depression, such as the relevant regions in the parietal lobe are the left precuneus (L_precuneus), the left postcentral gyrus (L_postcentral), the left superior parietal lobule (L_superiorparietal), and the right precuneus (R_precuneus); the relevant regions in the temporal lobe are the left fusiform gyrus (L_fusiform), the left entorhinal cortex (L_entorhinal), and the right parahippocampal gyrus (R_parahippocampal); the relevant regions in the frontal lobe are the right frontal pole (R_frontalpole), the anterior part of the right middle frontal gyrus (R_rostralmiddlefrontal), and the left orbital frontal cortex (L_parsorbitalis).

[0139] These associated subregions are mainly responsible for self-referential processing, introspection and episodic memory retrieval; facial recognition, emotion processing; higher cognitive functions, emotion regulation and reward processing; emotion processing, pain perception and cognitive control, etc., which are functions related to depression.

[0140] Abnormalities in the above areas may lead to symptoms such as attention deficit disorder, difficulty forming and recalling positive memories, changes in body sensation, and psychomotor retardation.

[0141] The areas of the parietal lobe with the greatest association with the disease are the left precuneus (L_precuneus), the left postcentral gyrus (L_postcentral), the left superior parietal lobule (L_superiorparietal), and the right precuneus (R_precuneus). The left precuneus is a core node of the default mode network (DMN), which is closely related to self-referential processing, introspection, and episodic memory retrieval. Overactivity of the DMN may lead to excessive self-focus, rumination, negative self-evaluation, and difficulty in detaching from negative thoughts. The postcentral gyrus and superior parietal lobule are involved in spatial cognition and sensory integration, and their abnormalities may lead to attention allocation disorders and changes in body sensation, which may be related to attention problems and physical symptoms of depression.

[0142] The areas of the temporal lobe that were second most strongly associated with the disease were the left fusiform gyrus (L_fusiform), the left entorhinal cortex (L_entorhinal), and the right parahippocampal gyrus (R_parahippocampal). The fusiform gyrus is involved in facial recognition and emotional processing, and its abnormalities may lead to social emotion processing disorders and difficulties in recognizing facial expressions. The entorhinal cortex and the parahippocampal gyrus are closely related to memory formation and retrieval, and abnormalities in these areas may lead to problems with episodic and autobiographical memory, increased negative memories of past experiences, and difficulty forming and recalling positive memories.

[0143] The third frontal lobe-related regions associated with the disease are the right frontal pole (R_frontalpole), the right middle frontal gyrus (R_rostralmiddlefrontal), and the left orbital frontal cortex (L_parsorbitalis). These regions are involved in higher cognitive functions, emotion regulation, and reward processing. Abnormalities in the frontal pole and the middle frontal gyrus may lead to executive dysfunction, reduced emotion regulation ability, and reduced cognitive flexibility. Abnormalities in the orbital frontal cortex may lead to a weakened response to rewards, causing anhedonia and difficulty in emotion regulation.

[0144] The main related areas of the cingulate gyrus include the left caudalanteriorcingulate gyrus (L_caudalanteriorcingulate) and the right rostralanteriorcingulate gyrus (R_rostralanteriorcingulate). The cingulate gyrus is an important part of the limbic system and is involved in emotion processing, pain perception, and cognitive control. Its abnormalities may lead to decreased emotion regulation ability, excessive response to negative stimuli, imbalance between cognitive control and emotion processing, and changes in pain perception.

[0145] The parahippocampal gyrus includes the left parahippocampal gyrus (L_parahippocampal) and the right parahippocampal gyrus (R_parahippocampal). The parahippocampal gyrus is closely related to memory formation and spatial navigation. Its abnormalities may lead to problems with episodic and situational memory, spatial cognitive impairment, and abnormalities in emotion-related memory processing.

[0146] The lingual gyrus mainly involves the left lingual gyrus (L_lingual). The lingual gyrus is primarily involved in visual processing and visual memory. Its abnormalities may lead to changes in visual information processing, problems with vision-related memory, and may affect the processing of emotion-related visual stimuli.

[0147] The temporal pole primarily involves the left temporal pole (L_temporalpole). The temporal pole is involved in social and emotional processing. Its abnormalities may lead to impairments in social cognition, changes in emotional processing, and problems with autobiographical memory retrieval.

[0148] The parstriangularis primarily involves the left parstriangularis. As part of Broca's area, it is involved in language processing and executive function. Its abnormalities may lead to subtle changes in language expression, decreased executive control, and may affect internal language and thought processes.

[0149] The paracentral lobule mainly involves the left paracentral lobule (L_paracentral). The paracentral lobule is associated with motor and sensory functions. Its abnormalities may cause subtle changes in motor control, altered body sensations, and may be associated with psychomotor retardation symptoms of depression.

[0150] ②Left and right brain angles.

[0151] From a quantitative analysis, there are 12 relevant areas in the left brain and 5 in the right brain. The left brain is significantly more than the right brain, which to a certain extent indicates that the left brain plays a more important role in the neuropathology of depression.

[0152] Table 3 Brain regions related to left and right brain diseases

[0153] Left Brain Key Brain Region L_Role_Brain_Region R_Role_Brain_Region Left precuneus Right frontal pole left fusiform gyrus Right anterior cingulate gyrus Left entorhinal cortex Right middle frontal gyrus Left postcentral gyrus Right parahippocampal gyrus Left superior parietal lobe Right precuneus Left caudal anterior cingulate gyrus Left parahippocampal gyrus left lingual gyrus Left temporal pole Left triangular gyrus Left paracentral lobule left orbital gyrus

[0154] Functionally, the left and right brain areas have different functions.

[0155] The relevant brain areas in the left brain space are mainly responsible for language processing (left triangular inferior frontal gyrus), logical thinking and analysis (left superior parietal lobule, left precuneus), memory formation (left entorhinal cortex, left parahippocampal gyrus), and emotion regulation (left orbital frontal cortex, left caudal anterior cingulate gyrus).

[0156] Therefore, the depressive symptoms that may be caused by changes in the left brain are mainly negative self-talk and thinking patterns, affecting logical reasoning and problem-solving ability, affecting memory formation and retrieval, and may aggravate negative memory preferences. Abnormalities in the left emotion regulation network may lead to difficulties in emotion regulation.

[0157] The relevant brain areas in the right brain space are mainly responsible for processing attention (right precuneus), emotion processing and regulation (right frontal pole, right anterior cingulate gyrus), intuition and creative thinking (anterior part of right middle frontal gyrus), and episodic memory (right parahippocampal gyrus).

[0158] Therefore, the depressive symptoms that may be caused by changes in the right brain are mainly problems with spatial cognition and attention allocation, affecting emotional processing and regulation, which may be the main source of core depressive symptoms, affecting intuition and creative thinking, which may lead to decreased problem-solving ability and rigid thinking, and affecting situational memory, which may lead to an increase in negative memories of past experiences.

[0159] Overall, more extensive left-brain abnormalities may lead to more cognitive symptoms, such as distorted thinking, attention, and memory problems. Right-brain abnormalities, while less common, may more directly affect core emotional symptoms. Imbalances between the left and right brains may lead to an imbalance in emotional and cognitive processing, exacerbating depressive symptoms.

[0160] ③Network perspective

[0161] The associated brain regions are important components of the network closely related to emotion regulation, cognitive control, and episodic memory retrieval.

[0162] The frontal lobe is an important component of the cognitive control network and the prefrontal-limbic network. Dysfunction of these networks may lead to difficulties in emotion regulation and cognitive control in patients with depression.

[0163] The precuneus is a core node of the default mode network. This network is active in the resting state and is associated with self-referential processing, theory of mind, and episodic memory retrieval. Abnormalities in the default mode network may lead to excessive self-focus and rumination in patients with depression.

[0164] Parts of the temporal lobe are important components of the emotion processing network and the memory network. Abnormalities in them may contribute to the difficulties in emotional experience and memory retrieval that people with depression experience.

[0165] The cingulate gyrus of the limbic system is an important node in the emotion processing network and cognitive control network. Its abnormality may lead to an imbalance between the two networks and affect the integration of emotion regulation and cognitive control.

[0166] In addition, the default mode network (DMN) involves the precuneus, cingulate gyrus and other areas. DMN abnormalities may lead to abnormal self-referential processing and increased rumination in patients with depression.

[0167] The emotion processing network includes temporal lobe structures and the limbic system. Abnormalities in this network may lead to difficulties in experiencing and regulating emotions. The cognitive control network primarily involves frontal lobe regions. Dysfunction in this network may contribute to the difficulties in attention, decision-making, and emotion regulation seen in patients with depression.

[0168] In summary, the experimental results reveal the degree of correlation between different brain regions and depression, the functional abnormalities caused by different brain regions and their symptom reflections, and the impact on specific brain network structure and function.

Claims

1. A classification method for brain magnetic resonance imaging based on graph neural network, characterized in that: The steps include: Step 1, obtaining the brain thickness of each brain region based on brain magnetic resonance images; The brain regions are used as nodes, the correlation connections between brain regions are used as edges between nodes, and the thickness of each brain region is used as the feature of the node to construct a brain topology map. The brain topology map includes the brain topology map of depression patients and the brain topology map of healthy people; Step 2, constructing a K-GAT network for classifying brain magnetic resonance images, inputting the brain topology map into the K-GAT network, training the K-GAT network, and obtaining a trained K-GAT network; Step 3: Use the trained K-GAT network to classify brain magnetic resonance images.

2. According to claim 1, a classification method for brain magnetic resonance imaging based on graph neural network is characterized in that: In step 1, the thickness of each brain region is obtained based on the brain magnetic resonance image, which specifically includes the following steps: Step 1.1, obtaining a whole brain gray matter image based on a brain magnetic resonance image; The gray matter image includes gray matter volume GMV and gray matter concentration GMC; Step 1.2, obtaining the gray matter thickness GMT and gray matter surface area SA corresponding to each brain region of the whole brain based on the gray matter image; Step 1.3, constructing a first brain map based on the gray matter volume GMV, gray matter concentration GMC, gray matter thickness GMT and gray matter surface area SA; Step 1.4, calculating the low frequency amplitude ALFF, the fractional low frequency amplitude fALFF and the local consistency ReHo for the first brain map to obtain the brain thickness of each brain region.

3. According to claim 1, a classification method for brain magnetic resonance imaging based on graph neural network is characterized in that: In the brain topology map, brain region thickness is used as a node feature Where n is the number of nodes, i.e. the number of brain regions, and d in is the feature dimension; The connection relationship between nodes is based on the adjacency matrix Indicates that A ij Represents the connection relationship between node i and node j, the adjacency matrix A ij Each element in is assigned a value according to the threshold, 1 if it is greater than the threshold and 0 if it is less than the threshold.

4. According to claim 3, a classification method for brain magnetic resonance imaging based on graph neural network is characterized in that: The K-GAT network includes the first linear layer, the K-GAT layer, and the second linear layer. The node feature X in the brain topology map is input into the first linear layer in the K-GAT network to obtain the hidden feature d hidden Represents the feature dimension of the linear layer; Hidden feature H (0) It is expressed as: H (0) =W in X in, is the weight matrix of the linear layer, X represents the thickness of the brain region; Hidden feature H (0) With the adjacency matrix A ij Input to the K-GAT layer, for node i and its neighbor node j, update node i to obtain the updated node feature H (L) ; The characteristic H of each node (L) After the second linear layer, the feature H (L) The dimension is from d hidden Compress to the final output dimension d out ; Z=W out H (L) in, is the final output result, d out is the output category; The output result Z applies log_softmax to convert the output of each node into the logarithmic probability P of the category; P = log\_softmax(Z) in, is the log probability of the node in each category.

5. According to claim 4, a classification method for brain magnetic resonance imaging based on graph neural network is characterized in that: Hidden feature H (0) and the adjacency matrix A ij Input to the K-GAT layer, for node i and its neighbor node j, update node i to obtain the updated node feature H (L) , specifically including the following steps; Step 2.1 Linearly transform the hidden features Hidden features for each node Perform a linear transformation: in, is the weight matrix, l represents the number of iterations; Step 2.2 When A ij = 1, calculate the attention coefficient e between node i and its neighbor node j ij : in is a learnable attention vector; ‖ represents the concatenation operation of features; LeakyReLU(·) represents the activation function; Step 2.3, for the attention coefficient e ij Normalize and get the normalized attention weight α ij ; in Represents the neighbor set of node i Step 2.4: Perform weighted summation on the neighbor node features of node i to obtain the new hidden feature representation of node i Among them, σ is a nonlinear activation function, W′ is a shared weight matrix; Step 2.5: Use the multi-head attention mechanism to the new hidden features of node i The concatenation is expressed as: Among them, K represents the number of attention heads, ‖ represents the concatenation operation, and W (k) are the attention weight and linear transformation matrix of the kth head respectively; Step 2.6: Set the feature of node i Input to the KAN layer, after further feature extraction, let l = l + 1, return to step 2.1, and finally get the feature H of each node (L) .

6. According to claim 1, a classification method for brain magnetic resonance imaging based on graph neural network is characterized in that Step 2 also includes pruning the trained K-GAT model.

Images