Depression auxiliary diagnosis method based on HGN-MAMBA algorithm

Abstract

The invention belongs to the field of artificial intelligence and biomedical engineering research, and particularly relates to a method for classifying functional nuclear magnetic resonance imaging data and searching for a depression-related biomarker on the basis of the functional nuclear magnetic resonance imaging data. Comprising the following steps: step 1, data preprocessing; step 2, constructing a deep HGN-MAMBA algorithm model, and training the deep HGN-MAMBA algorithm model; step 3, predicting classification of patients and normal persons and predicting possible related biomarkers; according to the method, in the fMRI data analysis process, a brain region topological structure is reserved, advanced phenotypic data are included, and the fMRI data classification accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and biomedical engineering research, and particularly relates to an algorithmic technology for classifying functional magnetic resonance imaging data and searching for biomarkers related to depression based on this classification. Background Art

[0002] Depression is one of the most devastating mental disorders in contemporary society, with a highly complex pathological mechanism. Currently, medical practitioners face the challenge of misdiagnosis and underdiagnosis in clinical assessments, making the development of highly accurate diagnostic tools for depression crucial. Currently, a variety of diagnostic tools are available to explore brain functional activity, such as electroencephalography (EEG), magnetoencephalography (MEG), functional magnetic resonance imaging (fMRI), and positron emission tomography (PET). fMRI, due to its non-invasive nature and excellent spatial resolution, is considered the most suitable tool for studying depression. Traditional machine learning methods, such as support vector machines and k-nearest neighbors, were initially used to analyze fMRI data. However, these methods are not end-to-end, rely on the importance of features extracted in the first stage, and suffer from significant limitations in extracting spatial and temporal features, resulting in relatively poor classification results. Currently, deep learning techniques, such as deep convolutional networks, Transformer models, and graph neural networks, are increasingly being applied to brain fMRI analysis. Given the inherently complex topological information of brain networks, deep graph neural networks are more suitable for fMRI data than deep convolutional networks and Transformer models.

[0003] In the existing technology, Noman et al. from Monash University in Malaysia introduced the use of graph autoencoders in "Graph Autoencoders for Embedding Learning in Brain Networks and Major Depressive Disorder Identification" to embed the topological structure and node content of large-scale networks into low-dimensional hidden layer representations, achieving a significant accuracy rate of 72.5% in classifying MDD patients from healthy individuals.

[0004] In the paper "Fusing multi-scale fMRI features using a brain-inspired multi-channel graph neural network for major depressive disorder diagnosis", Liu Shuai et al. from Tongji University proposed a multi-channel fusion graph convolutional network (MFGCN) and used two channels of MFGCN to extract features of high-order and low-order brain maps. The average classification accuracy of this method was 77.6%.

[0005] Traditional machine learning models, such as support vector machines (SVMs), logistic regression (LR), and deep learning models like convolutional neural networks (CNNs), often ignore the complex topological information and high-level properties inherent in brain networks, as well as the temporal characteristics of fMRI, when analyzing functional magnetic resonance imaging (fMRI) data. This can lead to poor performance in identifying and analyzing brain diseases. Current graph neural networks, when studying fMRI data, only consider spatial topological features, ignoring temporal characteristics. They also rarely consider natural differences in high-level phenotypic attributes such as gender, age, and education level between subjects. Summary of the Invention

[0006] Abstract: In view of the problems existing in the prior art, the present invention provides a depression auxiliary diagnosis method based on the HGN-MAMBA algorithm. In order to study the influence of depression disease on a certain brain function region of interest (ROI) of the brain, and whether there is a mutual correlation influence between different ROI brain regions of the brain, the MAMBA algorithm is adopted to map the fMRI data of each ROI brain region to a high-dimensional space. In the high-dimensional space after mapping, each ROI brain region contains the influencing factors of other brain regions and its own time attribute. At the same time, using the high-dimensional space as data, the adjacent matrix of each ROI brain region is used as the edge weight to form an individual brain network graph. Then, the demographic information of different personnel, including phenotypic information such as gender, age and years of education, is used as the edge weight of the heterogeneous graph to form a group heterogeneous graph. This heterogeneous graph is classified by the heterogeneous graph neural network algorithm to distinguish between patients with major depressive disorder (MDD) and healthy individuals. The HGN-MAMBA algorithm can make fMRI data retain the brain region topology during the analysis process, and also include high-level phenotypic data, thereby improving the accuracy of fMRI data classification.

[0007] Technical Solution

[0008] A method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm, comprising the following steps:

[0009] Step 1, data preprocessing;

[0010] First, fMRI data (including fMRI data of patients with major depressive disorder (MDD) and healthy individuals) were preprocessed. The three-dimensional brain regions were divided into regions of interest (ROIs) using brain atlas templates, and the feature values ​​of the ROI brain regions were extracted.

[0011] The adjacency matrix of each individual brain network graph is obtained through the correlation between brain regions;

[0012] The phenotypic data of the people involved in the fMRI data collection, including gender, age, and years of education, are provided to the subsequent steps for constructing the adjacency matrix of the heterogeneous network graph.

[0013] Step 2: Construct a deep HGN-MAMBA algorithm model and train it;

[0014] Design a deep HGN-MAMBA algorithm model, including the MAMBA part and the HGN part.

[0015] The MAMBA part extracts eigenvalues ​​of individual ROI brain regions in the mapped high-dimensional space. The eigenvalues ​​are combined with the individual brain network graph adjacency matrix obtained in step 1 to calculate the brain network eigenvalues ​​of each individual in the group through a graph neural network. The HGN part integrates the individual phenotypic data (adjacency matrix) and the eigenvalues ​​of a single individual brain network into heterogeneous graph structure data.

[0016] Step 3: predict the classification of patients and normal subjects and predict possible related biomarkers;

[0017] The final heterogeneous graph node data can obtain high-precision classification of fMRI data through binary classification.

[0018] During the HGN-MAMBA operation, ROI brain regions with higher node pooling weight scores may be related to depression, thereby assisting doctors in finding biomarkers that may be related to depression.

[0019] Beneficial effects

[0020] Traditional machine learning models such as support vector machines, logistic regression, and deep learning (e.g., convolutional neural networks, transformers, etc.) often ignore the complex topological information and high-level properties inherent in brain networks when analyzing magnetic resonance imaging data. This can lead to poor performance in identifying and analyzing brain diseases. The present invention proposes a new algorithm based on the fusion of heterogeneous graph neural networks and structured state sequence models (HGN-MAMBA algorithm for short). The MAMBA algorithm is used to map the fMRI data of each ROI brain region into a high-dimensional space. In the mapped high-dimensional space, each ROI brain region contains the influencing factors of other brain regions and its own temporal properties. At the same time, using the high-dimensional space as data, the adjacent matrix of each ROI brain region is used as the edge weight to form an individual brain network graph. The demographic information of different people, including phenotypic information such as gender, age, and years of education, is then used as the edge weight of the heterogeneous graph to form a group heterogeneous graph. This heterogeneous graph is then classified using the heterogeneous graph neural network algorithm to distinguish between patients with major depressive disorder (MDD) and healthy individuals. The HGN-MAMBA algorithm preserves brain topology and includes high-level phenotypic data during fMRI data analysis, improving the accuracy of fMRI data classification. Furthermore, the neural pooling algorithm used during training and inference assists physicians in identifying biomarkers potentially associated with depression. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the processing flow of the method of the present invention;

[0022] Figure 2 Module division and data flow diagram of the present invention;

[0023] Figure 3 Schematic diagram of extracting fMRI time series according to an embodiment of the present invention;

[0024] Figure 4 Adjacency matrix visualization diagram of an embodiment of the present invention;

[0025] Figure 5 Schematic diagram of the MAMBA architecture of the HGN-MAMBA model according to an embodiment of the present invention;

[0026] Figure 6 Schematic diagram of the HGN architecture in the HGN-MAMBA model according to an embodiment of the present invention:

[0027] Figure 7 The accuracy graph of the HGN-MAMBA model training according to the embodiment of the present invention;

[0028] Figure 8 Loss function loSS graph of HGN-MAMBA model training according to an embodiment of the present invention;

[0029] Figure 9 A map of possible relevant ROI area locations predicted by experiments in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The technical solution provided by this application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of this application will become more apparent with reference to the following description.

[0031] A method for auxiliary diagnosis of depression based on HGN-MAMBA algorithm, such as Figure 1 As shown, the following steps are included:

[0032] Step 1, data preprocessing;

[0033] First, fMRI data (including fMRI data of patients with major depressive disorder (MDD) and healthy individuals) were preprocessed. The three-dimensional brain regions were divided into regions of interest (ROIs) using brain atlas templates, and the feature values ​​of the ROI brain regions were extracted.

[0034] The adjacency matrix of each individual brain network graph is obtained through the correlation between brain regions;

[0035] The phenotypic data of the people involved in the fMRI data collection, including gender, age, and years of education, are provided to the subsequent steps for constructing the adjacency matrix of the heterogeneous network graph.

[0036] Step 2: Construct a deep HGN-MAMBA algorithm model and train it;

[0037] Design a deep HGN-MAMBA algorithm model, including the MAMBA part and the HGN part.

[0038] The MAMBA part extracts eigenvalues ​​from individual ROI brain regions in the mapped high-dimensional space. The eigenvalues ​​are combined with the adjacency matrix of the individual brain network graph obtained in step 1 through a graph neural network to obtain the brain network eigenvalues ​​of all individuals. The HGN part integrates the individual phenotypic data (adjacency matrix) and the eigenvalues ​​of individual brain networks into heterogeneous graph structure data.

[0039] Step 3: predict the classification of patients and normal subjects and predict possible related biomarkers;

[0040] The final heterogeneous graph node data can obtain high-precision classification of fMRI data through binary classification.

[0041] During the HGN-MAMBA operation, ROI brain regions with higher node pooling weight scores may be related to the impact of depression, thereby assisting doctors in finding biomarkers that may be related to depression.

[0042] The module division and data flow diagram of the present invention are as follows Figure 2 shown.

[0043] Specifically,

[0044] In step 1, the preprocessing aims to localize the content of the fMRI data on a structural image with higher spatial resolution. The preprocessing includes temporal correction, registration, covariate removal, spatial normalization, and 0.01-0.1 Hz bandpass filtering to minimize the heterogeneity of the preprocessing methods.

[0045] The steps of dividing the brain region of interest and extracting the ROI brain region feature values ​​are as follows:

[0046] 1) After the fMRI data are preprocessed, the three-dimensional brain image is divided into regions using a brain atlas template to obtain brain regions with different functions.

[0047] 2) Map each voxel point of the fMRI data to the corresponding brain region, and calculate the characteristic value of each brain region based on the results of voxel weighted average or clustering method. This characteristic value is the characteristic value of the ROI brain region.

[0048] Specifically, the weighted average calculation formula of the ROI brain region voxels is as follows:

[0049]

[0050] where n i is the number of voxels in the ith brain region of the ROI, x ij is the value of the jth voxel in the i brain area, x i (t) is the eigenvalue of brain region i.

[0051] The adjacency matrix generation steps of the individual brain network graph are as follows:

[0052] 1) If Figure 3 As shown in FIG, the original fMRI data is four-dimensional data composed of 3D voxels and time, and the time series of characteristic values ​​of each ROI brain region is extracted.

[0053] 2) The partial correlation coefficient is used to measure the functional connectivity between two ROI brain regions. Specifically, the partial correlation coefficient is calculated using the Pearson coefficient. The partial correlation coefficient r ij The calculation formula is as follows:

[0054]

[0055] Among them, x i 、x j Refers to the time series of the i-th and j-th ROI respectively, They represent the time series means of the i-th and j-th ROI brain regions respectively.

[0056] 3) Calculate the correlation coefficient r between all ROI brain regions ij , constructing a correlation matrix, i.e., a functional connectivity matrix, and removing pseudo connections with correlation coefficients below a set threshold, thereby obtaining an adjacency matrix of the individual brain network graph. As an example, the threshold is set to 0.35.

[0057] like Figure 4 This is a visual diagram of the adjacency matrix of the individual brain network graph in this embodiment.

[0058] In step 2, the deep HGN-MAMBA algorithm model is an algorithm model based on the fusion of heterogeneous graph neural network and structured state sequence model (MAMBA). The algorithm model includes MAMBA part and HGN part, and also includes graph convolutional neural network and Top k pooling part.

[0059] Among them, the MAMBA part is as follows Figure 5 shown.

[0060] The MAMBA part maps the characteristic value (time series) x(t) of each brain region to the high-dimensional space y(t). The formula for calculating the characteristic value of the high-dimensional space y(t) is as follows:

[0061]

[0062] y(t)=Ch(t)

[0063]

[0064] in, is the discrete state matrix, and C are mapping parameters, h(t) is the intermediate hidden state, x(t) is the time series, y is the high-dimensional feature value after mapping, and I represents the number of iterations.

[0065] Furthermore, the MAMBA component includes multiple stacked SSM layers. In the SSM layer, the input data is first processed by a normalization layer and then divided into two branches. In the first branch, the input passes through a linear layer and an activation function. In the second branch, the input passes through a linear layer, a depthwise separable convolution, and an activation function before being sent to a two-dimensional selective scanning module (SS2D). This further integrates the feature extraction of each ROI region in time and space, and finally merges the two branches. The MAMBA component has a significant computational advantage on long-range sequences.

[0066] The feature y(t) obtained by the MAMBA part is calculated by the graph convolutional neural network (GCN) through the adjacency matrix of the individual brain network graph, and the correlation between each ROI is used as the edge of the graph to find the correlation between each ROI.

[0067] Furthermore, the core of the GCN computation is the neighborhood aggregation mechanism, which essentially extends the convolution operator to the irregular data domain. The neighborhood aggregation mechanism aggregates the information of all neighboring nodes to the central node in order to update and change the central node's own information.

[0068] Assumptions For the I+1 layer feature, it is composed of the features of the previous layer And the fusion of its neighboring node features is calculated as follows:

[0069]

[0070] Among them, D is the degree matrix, A is the adjacency matrix of the graph, and W i 、W j is the weight of the network. After GCN and TopK pooling, all node features are aggregated and concatenated to obtain the feature value of an individual (MDD or healthy).

[0071] The HGN part of the HGN-MAMBA algorithm is as follows Figure 6 shown.

[0072] First, the individual features generated by the graph convolutional neural network (GCN) are used as graph nodes, and the phenotypic data in step 1 are used as node-edge relationships to form a group graph G h .

[0073] The phenotypic data includes three types of data, namely gender, age, and years of education. The corresponding edge relationships between nodes are gender (SSS) edge, age (SAS) edge, and years of education (SES) edge, so the group graph G is formed. h For a heterogeneous graph, the three types of edge association weights are used to find the correlation between MDD patients, healthy individuals, and MDD and healthy individuals.

[0074] Group graph G h Including: individual characteristic values ​​as nodes V, phenotypic data as edges E, defined as follows:

[0075] G h ={V, E} (5)

[0076] The feature aggregation part of heterogeneous graph neural network is different from the traditional homogeneous graph neural network (such graph edge types are the same). There are different types of edges between nodes, and the authority of each edge is β i It is obtained by combining the edge type adjacency matrix and node features through graph calculation to obtain the authority value. The calculation is as follows:

[0077]

[0078] where w i The weights of the graph neural network for each type of graph.

[0079] The final feature z of each node is formed by adding and merging all node features related to the edge type. The calculation formula is as follows:

[0080] Z i =∑ i β i Z i (7).

[0081] The nodes in the heterogeneous graph neural network are sorted from high to low according to the ROI node pooling scores obtained during the HGN-MAMBA algorithm operation (Top k pooling). Specifically, the pooling score calculation formula is as shown in formula (8), where p is the important ROI node and h is the feature of the pooling layer. The index value i of the first k important nodes in the pooling layer is calculated as:

[0082]

[0083] In step 3, the heterogeneous graph node data is binary classified through a classifier (fully connected function) to distinguish healthy individuals from depression individuals, and the cross entropy function is used as the loss function.

[0084] Step 2 gives the top k node features sorted by importance, reflecting the k most important features of the entire image (individual) during the pooling calculation process; the values ​​of these k nodes are mapped to the ROI areas on the brain map, and these ROI areas are possible depression-related biomarkers.

[0085] Example

[0086] This embodiment is based on publicly available resting-state functional magnetic resonance imaging (fMRI) data from the REST-meta-MDD dataset (http: / / rfmri.org / REST-meta-MDD). The model ultimately uses a classifier (fully connected function) to achieve binary classification, that is, to distinguish between healthy individuals and depressed individuals, using the cross-entropy function as the loss function.

[0087] The fMRI data selected for the experiment are from the S20 site in the public REST-meta-MDD dataset, which has 533 samples, including 282 individuals diagnosed with major depressive disorder (MDD) and 251 healthy individuals. In addition to fMRI data, there are also phenotypic data including gender, age, and years of education. The HGN-MAMBA model was used for training, with a 5-fold cross-validation method, a learning rate of 0.001, 8 training blocks per time, 100 epochs, and 5300 iterations. After 80 epochs, the model gradually converged, and the accuracy and loss value of the model gradually stabilized, and the change trends converged. Figure 7 and Figure 8 The following is a graph showing the accuracy and loss function changes during model training. The HGN-MAMBA model achieved an accuracy of 95.6%, far outperforming other methods in the field of fMRI data classification.

[0088] During the HGN-MAMBA algorithm operation, the ROI nodes with higher pooling scores are sorted, such as Figure 9 The figure shows the first 20 ROI nodes out of 264 predicted experimentally. Different colors are used to represent different ROI nodes to indicate their location in the brain. The estimated ROIs for the front regions are 91, 92, and 87, corresponding to the prefrontal lobe, limbic lobe, and posterior cingulate gyrus. This suggests that major depression affects the prefrontal lobe, limbic lobe, and posterior cingulate gyrus.

[0089] The above description is only a description of the preferred embodiments of the present application and does not limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical content should be regarded as equivalent valid embodiments and fall within the scope of protection of the technical solution of the present application.

Claims

1. A method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm, characterized in that: The following steps are involved: Step 1, data preprocessing; First, the fMRI data is preprocessed, and the three-dimensional brain area is divided into the brain region of interest (ROI) using the brain atlas template, and the feature value of the ROI brain area is extracted; The adjacency matrix of each individual brain network graph is obtained through the correlation between brain regions; Collect phenotypic data of people involved in fMRI data, including gender, age, and years of education, which are provided to the subsequent steps for constructing the adjacency matrix of the heterogeneous network graph; Step 2: Construct a deep HGN-MAMBA algorithm model and train it; Design a deep HGN-MAMBA algorithm model, including MAMBA and HGN parts; The MAMBA part extracts eigenvalues ​​from individual ROI brain regions in the mapped high-dimensional space. These eigenvalues ​​are combined with the individual brain network graph adjacency matrix obtained in step 1 through a graph neural network to obtain the brain network eigenvalues ​​of all individuals. The HGN part integrates the individual phenotypic data and the eigenvalues ​​of a single individual brain network into heterogeneous graph structure data. Step 3: predict the classification of patients and normal subjects and predict possible related biomarkers; The final heterogeneous graph node data can obtain high-precision classification of fMRI data through binary classification; During the HGN-MAMBA operation, ROI brain regions with higher node pooling weight scores may be related to depression, thereby assisting doctors in finding biomarkers that may be related to depression.

2. The method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm according to claim 1, characterized in that: In step 1, the steps of dividing the brain region of interest and extracting the ROI brain region feature values ​​are as follows: 1) After preprocessing the fMRI data, the three-dimensional brain image is divided into regions using a brain atlas template to obtain different functional brain regions; 2) Map each voxel point of the fMRI data to the corresponding brain region, and calculate the characteristic value of each brain region based on the results of voxel weighted average or clustering method. This characteristic value is the characteristic value of the ROI brain region; Specifically, the weighted average calculation formula of the ROI brain region voxels is as follows: where n i is the number of voxels in the ith brain region of the ROI, x ij is the value of the jth voxel in the i brain area, x i (t) is the eigenvalue of brain region i.

3. The method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm according to claim 1, characterized in that: In step 1, the adjacency matrix of the individual brain network graph is generated by: 1) The original fMRI data is a four-dimensional data consisting of 3D voxels and time, and the time series of the characteristic values ​​of each ROI brain region is extracted; 2) The partial correlation coefficient is used to measure the functional connectivity between two ROI brain regions. Specifically, the partial correlation coefficient is calculated using the Pearson coefficient. The partial correlation coefficient r ij The calculation formula is as follows: Among them, x i 、x j Refers to the time series of the i-th and j-th ROI respectively, They represent the time series means of the i-th and j-th ROI brain regions respectively; 3) Calculate the correlation coefficient r between all ROI brain regions ij , constructing a correlation matrix, namely the functional connectivity matrix, and eliminating pseudo connections with correlation coefficients lower than the set threshold, thus obtaining the adjacency matrix of the individual brain network diagram.

4. The method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm according to claim 1, characterized in that: In step 2, the deep HGN-MAMBA algorithm model is an algorithm model based on the fusion of heterogeneous graph neural network and structured state sequence model; it includes MAMBA part and HGN part, and also includes graph convolutional neural network and Top k pooling part.

5. The method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm according to claim 4, characterized in that: The MAMBA part maps the characteristic value x(t) of each brain region to the high-dimensional space y(t); The MAMBA part includes multiple stacked SSM layers. In the SSM layer, the input data is first processed by a normalization layer and then divided into two branches. In the first branch, the input passes through a linear layer and an activation function. In the second branch, the input passes through a linear layer, a depthwise separable convolution, and an activation function, and is sent to a two-dimensional selective scanning module to further fuse the feature extraction of each ROI area in time and space. Finally, the two branches are merged. The feature y(t) obtained by the MAMBA part is calculated by the graph convolutional neural network (GCN) through the adjacency matrix of the individual brain network graph. The correlation between each ROI is used as the edge of the graph to find the correlation between each ROI. The core of the GCN calculation is the neighborhood aggregation mechanism, which aggregates the information of all neighboring nodes to the central node in order to update and change the information of the central node itself. Assumptions For the I+1 layer feature, it is composed of the features of the previous layer And the fusion of its neighboring node features is calculated as follows: Among them, D is the degree matrix, A is the adjacency matrix of the graph, and W i 、W j is the weight of the network; After GCN and Top K pooling, all node features are aggregated and spliced ​​to obtain the feature value of an individual.

6. The method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm according to claim 4, characterized in that: The HGN part is processed as follows: First, the individual features generated by the graph convolutional neural network GCN are used as graph nodes, and the phenotypic data in step 1 are used as node-edge relationships to form a group graph G. h ; The phenotypic data includes three types of data, namely gender, age, and years of education. The corresponding edge relationships between nodes are gender edges, age edges, and years of education edges, so the group graph G is formed. h For a heterogeneous graph, the three types of edge association weights are used to find the association between MDD patients, healthy individuals, and MDD and healthy individuals; Group graph G h Including: individual characteristic values ​​as nodes V, phenotypic data as edges E, defined as follows: G h ={V,E} (5) There are different types of edges between nodes in heterogeneous graphs, and the authority of each edge is β i It is obtained by combining the edge type adjacency matrix and node features through graph calculation to obtain the authority value. The calculation is as follows: where w i The weights of the graph neural network for each type of graph; The final feature z of each node is formed by adding and merging all node features related to the edge type. The calculation formula is as follows: WITH i =∑ i β i WITH i (7)。 7. The method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm according to claim 4, characterized in that: The nodes in the heterogeneous graph neural network are sorted from high to low according to the ROI node pooling scores obtained during the HGN-MAMBA algorithm operation (Topk pooling); specifically, the pooling score calculation formula is as shown in formula (8), where p is an important ROI node and h is the feature of the pooling layer. The index value i of the top k important nodes in the pooling layer is calculated as:

8. The method for auxiliary diagnosis of depression based on the HGN-MAMBA algorithm according to claim 1, characterized in that: In step 3, the heterogeneous graph node data is binary classified through a fully connected classifier function to distinguish healthy individuals from depression individuals, and the cross entropy function is used as the loss function; The top k node features ranked by importance given in step 2 reflect the k most important features of the entire image during the pooling calculation process. If the values ​​of these k nodes are mapped to ROI regions on the brain map, then these ROI regions are possible depression-related biomarkers.

Images