A schizophrenia classification method based on connection gradient
By using a graph convolutional neural network model based on connection gradients and fusing bimodal feature information, the problem of lacking stable neuroimaging biomarkers in existing technologies is solved, and high-accuracy classification of schizophrenia is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2025-07-30
- Publication Date
- 2026-08-04
AI Technical Summary
Existing multimodal fusion models lack stable neuroimaging biomarkers in the classification of schizophrenia, have low classification accuracy, and cannot effectively utilize the nonlinear propagation characteristics of brain connectivity patterns.
A graph convolutional neural network model based on connectivity gradients is adopted. By constructing functional connectivity gradients and morphological similarity gradients, combined with a Top-K pooling enhancement model, the dual-modal feature information is fully integrated to construct a graph convolutional network classification model for the identification of schizophrenia.
It improved the classification accuracy of schizophrenia, achieved higher diagnostic accuracy and model interpretability, and enhanced the ability to classify and identify schizophrenia.
Smart Images
Figure CN120953671B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a method for classifying schizophrenia based on connectivity gradients. Background Technology
[0002] Schizophrenia is a chronic, severe mental disorder characterized by cognitive, behavioral, and emotional dysfunction. It is characterized by symptoms such as hallucinations, delusions, disorganized thinking, and impaired cognitive function, severely impacting patients' quality of life. Currently, diagnosis of this disease primarily relies on clinical symptom assessment, lacking objective and stable biomarkers. This subjectivity leads to a high rate of clinical misdiagnosis and a diagnostic delay. Therefore, artificial intelligence methods to assist in the diagnosis of schizophrenia have become a key research focus in the field of clinical psychiatry.
[0003] The human brain, as a typical complex system, follows the topological properties of complex networks in its function and structure. Furthermore, the development of brain diseases is often accompanied by the dynamic reconstruction of these network topological properties. Based on the natural compatibility between the brain and complex networks, disease diagnosis can be modeled as a graph classification problem. Graph neural networks, by modeling the nonlinear propagation characteristics between nodes through message passing mechanisms, can extract high-order topological features for classification; therefore, graph neural network methods are commonly used in the classification research of brain diseases.
[0004] Structural magnetic resonance imaging (SMRI) and functional magnetic resonance imaging (fMRI) offer complementary insights into the physiology of brain health and disease. SMRI provides high-resolution images, and morphological parameters can be obtained through T1-weighted imaging analysis, allowing for further analysis of differences in these parameters. In contrast, fMRI captures patterns of brain activity, identifying regions active during specific tasks or states, which is crucial for studying diseases such as schizophrenia, where connectivity patterns are frequently disrupted.
[0005] Data fusion techniques in deep learning have been used to explore the correlation between brain imaging structure and function, and to diagnose brain diseases using multimodal information. These studies highlight the improved accuracy of using multimodal techniques in neuroscience to detect brain diseases, demonstrating the effective integration of information from different imaging modalities. However, existing multi-connectivity fusion models fuse different connectivity information in a one-off manner. The fusion only focuses on simple linear relationships between different brain connections. In summary, while graph convolutional networks are widely used to extract classification features to distinguish schizophrenia, current research indicates a lack of stable neuroimaging biomarkers for classification experiments. Summary of the Invention
[0006] The purpose of this invention is to provide a schizophrenia classification method based on connection gradients. By utilizing a graph convolutional neural network classification model based on connection gradients, fully integrating bimodal feature information, and using Top-K pooling to enhance the interpretability of the model, the method achieves schizophrenia classification and recognition, thereby improving classification accuracy.
[0007] To achieve the above objectives, this invention provides a schizophrenia classification method based on connectivity gradients, comprising the following steps:
[0008] Step S1: Acquire magnetic resonance imaging data and perform preprocessing;
[0009] Step S2: Divide the preprocessed magnetic resonance imaging data into brain regions and construct a functional connectivity network and a morphological similarity network;
[0010] Step S3: Based on the functional connectivity network and the morphological similarity network, construct the functional connectivity gradient and the morphological similarity gradient using the connectivity gradient algorithm;
[0011] Step S4: Threshold the functional connection network and morphological similarity network as edge features, and use the bimodal connection gradient as node features to construct a graph convolutional network classification model.
[0012] Step S5: Train and test the graph convolutional network classification model.
[0013] Preferably, in step S1, magnetic resonance imaging data is acquired and preprocessed, as follows:
[0014] Step S11: Acquire magnetic resonance imaging data, specifically including functional magnetic resonance imaging data and structural magnetic resonance imaging data;
[0015] Step S12, functional magnetic resonance imaging data preprocessing, specifically includes:
[0016] First, discard the data from the initial ten time points;
[0017] Secondly, temporal correction, head movement correction, removal of noise covariates, spatial standardization, and spatial smoothing are performed.
[0018] Finally, a bandpass filter is used to filter the data;
[0019] Step S13: Structural magnetic resonance imaging data preprocessing, specifically including skull dissection, tissue segmentation, surface reconstruction, metric reconstruction, and estimation of spherical normalization parameters;
[0020] After preprocessing the structural magnetic resonance imaging data, five morphological parameters were obtained, including: gray matter volume, cortical thickness, surface area, intrinsic curvature, and mean curvature.
[0021] Preferably, in step S2, a functional connectivity network and a morphological similarity network are constructed, and the specific process is as follows:
[0022] Step S21: The preprocessed resting-state functional magnetic resonance imaging data is mapped to specific brain regions in the Schaefer 400 brain atlas for subdivision.
[0023] Step S22: Extract the corresponding time series according to the different brain regions, and average the time series to obtain the average time series;
[0024] Step S23: Calculate the Pearson correlation coefficient for the average time series of different brain regions to obtain the functional connectivity network;
[0025] Step S24: Take the average of the five morphological indices obtained from the preprocessing of the structural magnetic resonance imaging data to obtain the average morphological indices of different brain regions.
[0026] Step S25: Map the average morphological indices of different brain regions to specific brain regions in the Schaefer400 brain atlas, calculate the morphological Pearson correlation coefficient between different brain regions, and obtain the morphological similarity network.
[0027] Preferably, in step S3, based on the functional connectivity network and the morphological similarity network, the connectivity gradient and morphological similarity gradient are constructed using the connectivity gradient algorithm, as follows:
[0028] Step S31: The connection gradient construction algorithm takes the connection network as input, calculates the cosine similarity cos(θ) of the connection network, and obtains the affinity matrix Cosine Similarity, as shown below:
[0029]
[0030] Among them, X i ·X j Represents the dot product of two vectors; ||X i ‖ and ‖X j ‖ represents their respective Euclidean norms; ‖X i The definition of ‖ is as follows:
[0031]
[0032] Among them, X i,k This represents the value of the k-th component of the i-th vector;
[0033] Step S32: Use the diffusion map embedding algorithm to reduce the dimensionality of the affinity matrix to obtain a low-dimensional manifold representation, i.e., the gradient, as shown below:
[0034]
[0035] Among them, P α It is the diffusion matrix; α∈[0,1] are the anisotropic diffusion parameters used by the diffusion operator; W α =D -1 / α AD -1 / α The affinity matrix is constructed by normalizing it based on the diffusion parameters; A is the original affinity matrix; D is the degree matrix of the original affinity matrix; D a W a The derived degree matrix;
[0036] Step S33: Use Procrustes rotation to align the connection gradient of each subject to the group connection gradient template.
[0037] Preferably, in step S4, the functional connectivity network and the morphological similarity network are thresholded and used as edge features, and the bimodal connectivity gradients are used as node features to construct a graph convolutional network classification model. The specific process is as follows:
[0038] Step S41: Input the M*M dimension connection network as edge features into the graph convolutional network classification model; where M represents the number of brain regions; input the M*N dimension connection gradient as node features into the graph convolutional network classification model; where N represents the connection gradient components.
[0039] Step S42: The first layer of the graph convolutional network classification model represents the specific modality convolutional layer and the shared parameter convolutional layer of the graph convolutional network classification model;
[0040] Step S43: The second layer of the graph convolutional network classification model represents the feature splicing layer of the graph convolutional network classification model;
[0041] Step S44: The third layer of the graph convolutional network classification model represents the hierarchical pooling layer of the graph convolutional network classification model. This layer uses Top-K hierarchical pooling to retain classification nodes.
[0042] Step S45: Repeat steps S41-S44 three times to form a graph convolutional network classification model based on connection gradients.
[0043] Step S46: After each Top-K pooling layer, global max pooling and global average pooling are performed to read the results. Finally, the results of the three layers are summed and input into the fully connected layer, which is then input into the classifier.
[0044] Preferably, in step S42, a specific modality convolution is constructed in the specific modality convolution layer, as shown below:
[0045]
[0046] in, It is the (l+1)th convolution of a structure-specific modality. th The feature matrix of the layer; σ is the activation function of the structure-specific mode convolution; It is the degree matrix of structure-specific mode convolution; It is a structure-specific modal convolution with a self-connected adjacency matrix; It is the structure-specific modality convolution of the lth generation th The feature matrix of the layer; It is the structure-specific modality convolution of the lth generation th The trainable weight matrix of the layer;
[0047] Feature-specific modality convolution is as follows:
[0048]
[0049] in, It is the (l+1)th function-specific modality convolution. th The feature matrix of the layer; It is the degree matrix of the function-specific modality convolution; It is a function-specific modality convolution with a self-connected adjacency matrix; It is the lth function-specific modality convolution. th The feature matrix of the layer; It is the lth function-specific modality convolution. th The trainable weight matrix of the layer;
[0050] In shared-parameter convolutional layers, modality-shared convolutions are performed as follows:
[0051]
[0052] in, and These represent the (l+1)th... th The structural feature matrix of the layer and the (l+1)th layer th The functional feature matrix of the layer; and These represent the degree matrices of structural features and functional features, respectively. and These represent adjacency matrices with self-connected structural features and adjacency matrices with self-connected functional features, respectively. and They represent the lth th The structural feature matrix and functional feature matrix of the layer; and They represent the lth th The trainable weight matrix of the layer.
[0053] Preferably, in step S43, in the feature concatenation layer, specific modality convolution and modality-shared convolution are used to concatenate features, as shown below:
[0054] H′ f =Concat(H f H f,com );
[0055] H s =Concat(H) s H s,com );
[0056] Among them, H′ f and H s ′ represents the functional and structural features concatenated from features obtained after convolution of specific and shared modalities, respectively; H f and H s H represents the functional and structural characteristics of a specific mode, respectively. f,com and H s,com These represent the functional features and structural features obtained through convolution with shared parameters, respectively; Concat represents the concatenation operation.
[0057] The spliced functional and structural features are normalized as follows:
[0058]
[0059] Among them, Sim s and Sim f These represent the structural and functional features obtained after normalization, respectively; ||·||2 represents the L2 norm; T represents the transpose operation;
[0060] Restricted shared feature learning, as shown below:
[0061]
[0062] Among them, L C This represents the loss function.
[0063] Preferably, in step S44, the hierarchical pooling layer uses Top-K hierarchical pooling to retain classification nodes, and the specific process is as follows:
[0064] Based on the trainable projection vector p, the scalar projection vector y of each node is obtained, as shown below:
[0065]
[0066] i=top k (y (l) );
[0067] Where l represents the l-th layer; y (l) Indicate l th The scalar projection vector of the node; H (l) Indicate l th Features; p (l) represents the projection vector; i represents the index used to find the k largest elements in vector y;
[0068] The new feature matrix is calculated using the ranking scores of the l-th layer, as shown below:
[0069] H (l+1) =(H (l) ⊙tahn(y (l) )) i ;
[0070] Among them, H (l+1) Represents (l+1) th The feature matrix of the layer; tahn represents the hyperbolic tangent activation function; ⊙ represents element-wise multiplication;
[0071] The new adjacency matrix is as follows:
[0072]
[0073] Among them, A (l+1) Represents (l+1) th The adjacency matrix of the layer; This means that only the submatrix corresponding to the top-k index i is retained.
[0074] Preferably, in step S5, the graph convolutional network classification model is trained and tested. The graph convolutional model is trained using the connection network and connection gradient calculated in step S3. Part of the training data and the data to be tested are input into the trained graph convolutional model, and the predicted label of the data to be tested is output.
[0075] Specifically, the cross-entropy loss function is used to calculate the error between the predicted output of the graph convolutional network classification model and the true label. The parameters of each layer in the network are updated through backpropagation of the error and the stochastic gradient descent algorithm.
[0076] Therefore, this invention employs the aforementioned schizophrenia classification method based on connection gradients. By utilizing a graph convolutional neural network classification model based on connection gradients, it fully integrates bimodal feature information and uses Top-K pooling to enhance the interpretability of the model, thereby achieving schizophrenia classification and recognition, improving classification accuracy, and increasing application value. This invention effectively solves the problems of lacking stable neuroimaging biomarkers and low classification accuracy in schizophrenia identification and classification methods, and is suitable for the classification of schizophrenia.
[0077] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0078] Figure 1 This is a flowchart of a schizophrenia classification method based on connectivity gradients according to the present invention;
[0079] Figure 2 This is a research technology roadmap for a schizophrenia classification method based on connectivity gradients, as described in this invention.
[0080] Figure 3 These are the top 10% of functional brain regions that contribute the most to the classification results extracted in this invention; where (a) is the number of functionally significant brain regions in the subnet; and (b) is the number of functionally significant brain regions in the left and right hemispheres.
[0081] Figure 4 These are the top 10% of structural brain regions that contribute the most to the classification results extracted in this invention; where (a) is the number of structurally significant brain regions in the subnet; and (b) is the number of structurally significant brain regions in the left and right hemispheres. Detailed Implementation
[0082] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0083] like Figure 1 As shown, a schizophrenia classification method based on connectivity gradients includes the following steps:
[0084] Step S1: Acquire magnetic resonance imaging data and perform preprocessing;
[0085] Step S2: Divide the preprocessed magnetic resonance imaging data into brain regions and construct a functional connectivity network and a morphological similarity network;
[0086] Step S3: Based on the functional connectivity network and the morphological similarity network, construct the functional connectivity gradient and the morphological similarity gradient using the connectivity gradient algorithm;
[0087] Step S4: Threshold the functional connection network and morphological similarity network as edge features, and use the bimodal connection gradient as node features to construct a graph convolutional network classification model.
[0088] Step S5: Train and test the graph convolutional network classification model. Use the connection network and connection gradient calculated in step S3 to train the graph convolutional model. Input some training data and test data into the trained graph convolutional model and output the predicted label of the test data.
[0089] Example
[0090] like Figure 2As shown, this invention provides a schizophrenia classification method based on connectivity gradients. First, functional connectivity networks and morphological similarity networks are constructed based on the subjects' functional magnetic resonance imaging (fMRI) data and structural magnetic resonance imaging (SMRI) data, respectively. Then, a connectivity gradient calculation algorithm is used to construct functional connectivity gradients and morphological similarity gradients. Thresholding is applied to the functional connectivity networks and morphological similarity networks to obtain sparse matrices, which are used as edge feature inputs to the graph convolution model. Finally, the connectivity gradients are used as node feature inputs to the graph convolution model.
[0091] This embodiment uses data from the publicly available COBRE dataset of the Center for Excellence in Biomedical Research as an example. This dataset includes 75 normal controls and 72 patients with schizophrenia, which are randomly divided into a training set (70%) and a test set (30%).
[0092] Step S1: Acquire magnetic resonance imaging data and perform preprocessing.
[0093] Step S11: Acquire magnetic resonance imaging data, specifically including functional magnetic resonance imaging data and structural magnetic resonance imaging data.
[0094] Step S12, functional magnetic resonance imaging data preprocessing, specifically includes:
[0095] First, discard the data from the initial ten time points;
[0096] Secondly, temporal correction, head movement correction, removal of noise covariates, spatial standardization, and spatial smoothing are performed.
[0097] Finally, a bandpass filter (0.01–0.1 Hz) is used to filter the data.
[0098] Step S13: Structural magnetic resonance imaging data preprocessing, specifically including: skull dissection, tissue segmentation, surface reconstruction, metric reconstruction, and estimation of spherical normalization parameters. After structural magnetic resonance imaging data preprocessing, five morphological indices are obtained, specifically including: gray matter volume, cortical thickness, surface area, intrinsic curvature, and mean curvature.
[0099] Step S2: Based on the selected standardized brain atlas, divide the preprocessed resting-state functional magnetic resonance imaging (fMRI) data and structural magnetic resonance imaging (SMRI) data into brain regions, and then extract the average time series and average morphological indices of each brain region to construct a functional connectivity network and a morphological similarity network.
[0100] Step S21: Map the preprocessed functional magnetic resonance imaging (fMRI) data to specific brain regions in the Schaefer 400 brain atlas, and divide the resting-state fMRI data into brain regions.
[0101] Step S22: Extract the corresponding time series according to the different brain regions, and average the time series to obtain the average time series.
[0102] Step S23: Calculate the Pearson correlation coefficient for the average time series of different brain regions to obtain the functional connectivity network.
[0103] Step S24: Take the average of the five morphological indices obtained from the preprocessing of the structural magnetic resonance imaging data to obtain the average morphological indices of different brain regions.
[0104] Step S25: Map the average morphological indices of different brain regions to specific brain regions in the Schaefer400 brain atlas, calculate the morphological Pearson correlation coefficient between different brain regions, and obtain the morphological similarity network.
[0105] Step S3: Based on the functional connectivity network, construct the functional connectivity gradient using the connectivity gradient algorithm; based on the morphological similarity network, construct the morphological similarity gradient using the connectivity gradient algorithm.
[0106] Step S31: The connection gradient construction algorithm takes the connection network as input, calculates the cosine similarity of the connection network, and obtains the affinity matrix, as shown below:
[0107]
[0108] Among them, X i ·X j Represents the dot product of two vectors; ||X i ‖ and ‖X j ‖ represents their respective Euclidean norms (i.e., the magnitude of the vector); ‖X i The definition of ‖ is as follows:
[0109]
[0110] Among them, X i,k This represents the value of the k-th component of the i-th vector.
[0111] Step S32: Use the diffusion map embedding algorithm to reduce the dimensionality of the affinity matrix to obtain a low-dimensional manifold representation (i.e., gradient), as shown below:
[0112]
[0113] Among them, P α It is the diffusion matrix; α∈[0,1] are the anisotropic diffusion parameters used by the diffusion operator; W α =D -1 / α AD -1 / α The affinity matrix is constructed by normalizing it based on the diffusion parameters; A is the original affinity matrix; D is the degree matrix of the original affinity matrix; Da W a The derived degree matrix.
[0114] Step S33: Use Procrustes rotation to align the connection gradient of each subject to the group connection gradient template.
[0115] Step S4: Threshold the functional connection network and morphological similarity network to obtain edge features, and use the bimodal connection gradient as node features to construct a graph convolutional network classification model.
[0116] Step S41: Input the M*M dimension connection network as edge features into the graph convolutional network classification model; where M represents the number of brain regions; input the M*N dimension connection gradient as node features into the graph convolutional network classification model; where N represents the connection gradient components.
[0117] Introducing connectivity gradients into graph convolutional network classification models brings the global connectivity patterns of the brain into the classification model. Compared to using connectivity networks as node features, connectivity gradients have been shown to show significant differences between the two groups of subjects. At the same time, it can also fully extract changes in the connectivity patterns of the brains of schizophrenia patients, ultimately improving the diagnostic accuracy of schizophrenia patients.
[0118] Step S42: The first layer of the graph convolutional network classification model represents the specific modality convolutional layer and the shared parameter convolutional layer of the graph convolutional network classification model.
[0119] In a specific modality convolutional layer, a specific modality convolution is constructed as follows:
[0120]
[0121] in, It is the (l+1)th convolution of a structure-specific modality. th The feature matrix of the layer; σ is the activation function of the structure-specific mode convolution; It is the degree matrix of structure-specific mode convolution; It is a structure-specific modal convolution with a self-connected adjacency matrix; It is the structure-specific modality convolution of the lth generation th The feature matrix of the layer; It is the structure-specific modality convolution of the lth generation th The trainable weight matrix of the layer.
[0122] Feature-specific modality convolution is as follows:
[0123]
[0124] in, It is the (l+1)th function-specific modality convolution.th The feature matrix of the layer; It is the degree matrix of the function-specific modality convolution; It is a function-specific modality convolution with a self-connected adjacency matrix; It is the lth function-specific modality convolution. th The feature matrix of the layer; It is the lth function-specific modality convolution. th The trainable weight matrix of the layer.
[0125] In shared-parameter convolutional layers, modality-shared convolutions are performed as follows:
[0126]
[0127] in, and These represent the (l+1)th... th The structural feature matrix of the layer and the (l+1)th layer th The functional feature matrix of the layer; and These represent the degree matrices of structural features and functional features, respectively. and These represent adjacency matrices with self-connected structural features and adjacency matrices with self-connected functional features, respectively. and They represent the lth th The structural feature matrix and functional feature matrix of the layer; and They represent the lth th The trainable weight matrix of the layer.
[0128] Step S43: The second layer of the graph convolutional network classification model represents the feature splicing layer of the graph convolutional network classification model.
[0129] In the feature concatenation layer, specific modality convolutions and modality-shared convolutions are used to concatenate features, as shown below:
[0130] H′ f =Concat(H f H f,com );
[0131] H s =Concat(H) s H s,com );
[0132] Among them, H′ f and H s ′ represents the functional and structural features concatenated from features obtained after convolution of specific and shared modalities, respectively; H f and Hs H represents the functional and structural characteristics of a specific mode, respectively. f,com and H s,com These represent the functional features and structural features obtained through convolution with shared parameters, respectively; Concat represents the concatenation operation.
[0133] The spliced functional and structural features are normalized as follows:
[0134]
[0135] Among them, Sim s and Sim f represents the structural and functional features obtained after normalization, respectively; ||·||2 represents the L2 norm; T represents the transpose operation.
[0136] Restricted shared feature learning, as shown below:
[0137]
[0138] Among them, L C This represents the loss function.
[0139] The shared parameter convolution introduced in step S42 can fully extract modal complementary features, and through multiple feature fusions, it avoids the single linear modal feature fusion of existing studies.
[0140] Step S44: The third layer of the graph convolutional network classification model represents the hierarchical pooling layer of the graph convolutional network classification model. This layer uses Top-K hierarchical pooling to retain the most effective classification nodes.
[0141] Based on the trainable projection vector p, the scalar projection vector y of each node is obtained, as shown below:
[0142]
[0143] i=top k (y (l) );
[0144] Where l represents the l-th layer; y (l) Indicate l th The scalar projection vector of the node; H (l) Indicate l th Features; p (l) y represents the projection vector; i represents the index used to find the k largest elements in vector y.
[0145] The new feature matrix is calculated using the ranking scores of the l-th layer, as shown below:
[0146] H(l+1) =(H (l) ⊙tahn(y (l) )) i ;
[0147] Among them, H (l+1) Represents (l+1) th The feature matrix of the layer; tahn represents the hyperbolic tangent activation function; ⊙ is element-wise multiplication.
[0148] The new adjacency matrix is as follows:
[0149]
[0150] Among them, A (l+1) Represents (l+1) th The adjacency matrix of the layer; This indicates that only the submatrix corresponding to the top-k index i is retained; (.) i,j It is an index operation that retrieves an element at row index i and column index j (if there is no index j, it means all indexes are used).
[0151] Using Top-K pooling can identify the brain regions that contribute most to the classification results. The saliency score of each node is obtained by calculating the ratio of the frequency of occurrence of the brain region to the total number of brain regions, which serves as the specificity of the brain region. This step enhances the interpretability of the model proposed in this invention.
[0152] Step S45: Repeat steps S41-S44 three times to form a graph convolutional network classification model based on connection gradients.
[0153] Step S46: After each Top-K pooling layer, global max pooling and global average pooling are performed to read the results. Finally, the results of the three layers are summed and input into the fully connected layer, which is then input into the classifier.
[0154] Step S5: Train and test the graph convolutional network classification model. Use the connection network and connection gradient calculated in step S3 to train the graph convolutional model. Input some training data and test data into the trained graph convolutional model and output the predicted label of the test data.
[0155] During the training phase, each patient's MRI data was preprocessed to obtain a 400×400 connection network and a 400×10 connection gradient, which were then randomly divided into training and testing sets. The training set comprised 30% of the total test samples, and the testing set comprised 70%.
[0156] The convolutional network classification model uses the connection network and connection gradient of the training set as feature inputs. First, specific modality convolution is used to extract specific modal features, and then shared parameter convolution is used to extract features. The two types of features are then concatenated. Finally, Top-K pooling is used to select nodes, and global max pooling and global average pooling are read out after the pooling layers.
[0157] After three convolution and pooling operations, the features from the readout layer after the three pooling operations are summed, and the summed features from different modalities are concatenated and input into the fully connected layer, which is then further input into the classifier for classification.
[0158] The cross-entropy loss function is used to calculate the error between the predicted output of the graph convolutional network classification model and the true label. The parameters of each layer in the network are updated through backpropagation of the error and stochastic gradient descent algorithm.
[0159] During the testing phase, test data is input into a graph convolutional model that has been trained to obtain optimal parameters. Evaluation metrics for the test data are obtained, and the experiment is repeated multiple times. Finally, the average result of the test data is taken as the final result.
[0160] Using COBRE data to classify schizophrenia patients, the experimental results showed a classification accuracy of 82.73%, a sensitivity of 78.1%, and a specificity of 87.27%.
[0161] The method of this invention was compared with the most advanced methods at home and abroad, and the comparison results are shown in Table 1. From the classification results of 146 subjects, the method of this invention has a high classification accuracy and specificity.
[0162] Table 1 Comparison of classification performance of different classification methods on the COBRE dataset
[0163] Zeng et al. (2018) 71 / 74 73.6 68.0 76.6 - Ji et al. (2019) 62 / 62 77 82 71 0.81 Huang et al. (2020) 53 / 67 82.4 91.3 72.5 - Ioannis et al. (2021) 71 / 74 79.3 85.1 72.9 - Kim et al. (2023) 62 / 73 64.4 - - - Wang et al. (2024) 71 / 74 79.62 78.52 79.71 0.7961 The method in this paper 72 / 74 82.73 78.1 87.27 0.827
[0164] Table 2 shows the brain region numbers and scores obtained using Top-K for functional channels. The table also includes information on the number of ROIs contained in subnetworks and the number of ROIs contained in the left and right hemispheres. Figure 3 As shown, the networks with higher contribution levels are mainly involved in visual and emotional processing; among the top 40 significant ROIs, the number of significant ROIs in the left brain is significantly greater than that in the right brain.
[0165] Table 2 Functionally significant brain regions
[0166] 1 0.01267 LH_Vis_1 311 0.01062 RH_SalVentAttn_Med_1 2 0.0125 LH_Vis_2 72 0.01027 LH_DorsAttn_Post_4 13 0.01233 LH_Vis_13 176 0.01027 LH_Default_PFC_11 16 0.01233 LH_Vis_16 91 0.0101 LH_DorsAttn_PrCv_2 400 0.01233 RH_Default_pCunPCC_9 99 0.0101 LH_SalVentAttn_FrOperIns_3 4 0.01199 LH_Vis_4 156 0.0101 LH_Default_Temp_8 11 0.01182 LH_Vis_11 166 0.0101 LH_Default_PFC_1 8 0.01147 LH_Vis_8 200 0.0101 LH_Default_pCunPCC_11 12 0.0113 LH_Vis_12 277 0.00993 RH_DorsAttn_Post_7 14 0.0113 LH_Vis_14 301 0.00993 RH_SalVentAttn_PrC_1 18 0.01113 LH_Vis_18 67 0.00976 LH_SomMot_36 105 0.01113 LH_SalVentAttn_FrOperIns_9 108 0.00976 LH_SalVentAttn_Med_2 117 0.01113 LH_Limbic_OFC_4 143 0.00976 LH_Cont_PFCv_1 101 0.01096 LH_SalVentAttn_FrOperIns_5 151 0.00976 LH_Default_Temp_3 109 0.01096 LH_SalVentAttn_Med_3 170 0.00976 LH_Default_PFC_5 135 0.01096 LH_Cont_PFCl_1 266 0.00976 RH_SomMot_36 138 0.01096 LH_Cont_PFCl_4 278 0.00976 RH_DorsAttn_Post_8 181 0.01079 LH_Default_PFC_16 348 0.00976 RH_Cont_PFCl_8 204 0.01079 RH_Vis_4 355 0.00976 RH_Cont_PFCl_11 120 0.01062 LH_Limbic_TempPole_2 361 0.00976 RH_Cont_PFCmp_2
[0167] Table 3 shows the brain region numbers and scores obtained using Top-K for morphological channels. The table also includes information on the number of ROIs contained in subnetworks and the number of ROIs contained in the left and right hemispheres. Figure 4As shown, networks with higher contribution levels are mainly involved in cognitive processing; no significant lateralization was found among the top 40 significant ROIs.
[0168] Table 3 Brain regions with significant structural features
[0169] 216 0.01678 RH_Vis_16 210 0.01336 RH_Vis_10 234 0.0161 RH_SomMot_4 224 0.01336 RH_Vis_24 189 0.01507 LH_Default_PFC_24 187 0.01301 LH_Default_PFC_22 205 0.01507 RH_Vis_3 207 0.01284 RH_Vis_7 177 0.0149 LH_Default_PFC_12 392 0.01284 RH_Default_pCunPCC_1 194 0.0149 LH_Default_pCunPCC_5 168 0.01267 LH_Default_PFC_3 182 0.01438 LH_Default_PFC_17 169 0.01267 LH_Default_PFC_4 233 0.01438 RH_SomMot_3 179 0.01267 LH_Default_PFC_14 184 0.01421 LH_Default_PFC_19 197 0.01267 LH_Default_pCunPCC_8 229 0.01421 RH_Vis_29 203 0.01267 RH_Vis_3 192 0.01404 LH_Default_pCunPCC_3 212 0.01267 RH_Screw_12 232 0.01404 RH_SomMot_2 221 0.01267 RH_Screw_21 238 0.01387 RH_SomMot_8 230 0.01267 RH_Screw_30 176 0.0137 LH_Default_PFC_11 196 0.01233 LH_Default_pWithPCC_7
[0170] Therefore, this invention adopts the above-mentioned schizophrenia classification method based on connectivity gradients. By utilizing a graph convolutional network classification model and fusing bimodal connectivity gradient features, schizophrenia classification is achieved, thereby improving the classification accuracy.
[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A schizophrenia classification method based on connectivity gradients, characterized in that, Includes the following steps: Step S1: Acquire magnetic resonance imaging data and perform preprocessing; Step S2: Divide the preprocessed magnetic resonance imaging data into brain regions and construct a functional connectivity network and a morphological similarity network; Step S3: Based on the functional connectivity network and the morphological similarity network, construct the functional connectivity gradient and the morphological similarity gradient using the connectivity gradient algorithm. The specific process is as follows: Step S31: The connection gradient construction algorithm takes the connection network as input and calculates the cosine similarity of the connection network. The affinity matrix is obtained. As shown below: ； in, Represents the dot product of two vectors; and Each is its own Euclidean norm; The definition is as follows: ； in, This represents the value of the k-th component of the i-th vector; Step S32: Use the diffusion map embedding algorithm to reduce the dimensionality of the affinity matrix to obtain a low-dimensional manifold representation, i.e., the gradient, as shown below: ； in, It is the diffusion matrix; These are the anisotropic diffusion parameters used by the diffusion operator; The affinity matrix is constructed by normalizing the diffusion parameters. It is the original affinity matrix; It is the degree matrix of the original affinity matrix; It is by The derived degree matrix; Step S33: Use Procrustes rotation to align the connection gradient of each subject to the group connection gradient template; Step S4: Threshold the functional connection network and morphological similarity network as edge features, and use the bimodal connection gradient as node features to construct a graph convolutional network classification model. Step S5: Train and test the graph convolutional network classification model.
2. The schizophrenia classification method based on connectivity gradient according to claim 1, characterized in that, In step S1, magnetic resonance imaging data is acquired and preprocessed, as follows: Step S11: Acquire magnetic resonance imaging data, specifically including functional magnetic resonance imaging data and structural magnetic resonance imaging data; Step S12, functional magnetic resonance imaging data preprocessing, specifically includes: First, discard the data from the initial ten time points; Secondly, temporal correction, head movement correction, removal of noise covariates, spatial standardization, and spatial smoothing are performed. Finally, a bandpass filter is used to filter the data; Step S13: Structural magnetic resonance imaging data preprocessing, specifically including skull dissection, tissue segmentation, surface reconstruction, metric reconstruction, and estimation of spherical normalization parameters; After preprocessing the structural magnetic resonance imaging data, five morphological parameters were obtained, including: gray matter volume, cortical thickness, surface area, intrinsic curvature, and mean curvature.
3. The schizophrenia classification method based on connectivity gradient according to claim 1, characterized in that, In step S2, a functional connectivity network and a morphological similarity network are constructed, and the specific process is as follows: Step S21: The preprocessed resting-state functional magnetic resonance imaging data is mapped to specific brain regions in the Schaefer 400 brain atlas for subdivision. Step S22: Extract the corresponding time series according to the different brain regions, and average the time series to obtain the average time series; Step S23: Calculate the Pearson correlation coefficient for the average time series of different brain regions to obtain the functional connectivity network; Step S24: Take the average of the five morphological indices obtained from the preprocessing of the structural magnetic resonance imaging data to obtain the average morphological indices of different brain regions. Step S25: Map the average morphological indices of different brain regions to specific brain regions in the Schaefer400 brain atlas, calculate the morphological Pearson correlation coefficient between different brain regions, and obtain the morphological similarity network.
4. The schizophrenia classification method based on connectivity gradient according to claim 1, characterized in that, In step S4, the functional connectivity network and morphological similarity network are thresholded and used as edge features, and the bimodal connectivity gradients are used as node features to construct a graph convolutional network classification model. The specific process is as follows: Step S41: Input the M*M dimension connection network as edge features into the graph convolutional network classification model; where M represents the number of brain regions; input the M*N dimension connection gradient as node features into the graph convolutional network classification model; where N represents the connection gradient components. Step S42: The first layer of the graph convolutional network classification model represents the specific modality convolutional layer and the shared parameter convolutional layer of the graph convolutional network classification model; Step S43: The second layer of the graph convolutional network classification model represents the feature splicing layer of the graph convolutional network classification model; Step S44: The third layer of the graph convolutional network classification model represents the hierarchical pooling layer of the graph convolutional network classification model. This layer uses Top-K hierarchical pooling to retain classification nodes. Step S45: Repeat steps S41-S44 three times to form a graph convolutional network classification model based on connection gradients; Step S46: After each Top-K pooling layer, global max pooling and global average pooling are performed to read the results. Finally, the results of the three layers are summed and input into the fully connected layer, which is then input into the classifier.
5. The schizophrenia classification method based on connectivity gradient according to claim 4, characterized in that, In step S42, a specific modality convolution is constructed in the specific modality convolution layer, as shown below: ； in, It is the structure-specific modal convolution of the first The feature matrix of the layer; It is the activation function for structure-specific mode convolution; It is the degree matrix of structure-specific mode convolution; It is a structure-specific modal convolution with a self-connected adjacency matrix; It is the structure-specific modal convolution. The feature matrix of the layer; It is the structure-specific modal convolution. The trainable weight matrix of the layer; Feature-specific modality convolution is as follows: ； in, It is the first function-specific modality convolution. The feature matrix of the layer; It is the degree matrix of the function-specific modality convolution; It is a function-specific modality convolution with a self-connected adjacency matrix; It is the first function-specific modality convolution. The feature matrix of the layer; It is the first function-specific modality convolution. The trainable weight matrix of the layer; In shared-parameter convolutional layers, modality-shared convolutions are performed as follows: ； ； in, and They represent the first The structural feature matrix of the layer and the first The functional feature matrix of the layer; and These represent the degree matrices of structural features and functional features, respectively. and These represent adjacency matrices with self-connected structural features and adjacency matrices with self-connected functional features, respectively. and They represent the first The structural feature matrix and functional feature matrix of the layer; and They represent the first The trainable weight matrix of the layer.
6. The schizophrenia classification method based on connectivity gradient according to claim 5, characterized in that, In step S43, in the feature concatenation layer, specific modality convolutions and modality-shared convolutions are used to concatenate features, as shown below: ； ； in, and These represent the functional and structural features, respectively, concatenated from features obtained after convolution of specific and shared modalities. and These respectively represent the functional and structural characteristics of a specific mode; and These represent the functional features and structural features obtained through convolution with shared parameters, respectively. Indicates a splicing operation; The spliced functional and structural features are normalized as follows: ； ； in, and These represent the structural and functional features obtained after normalization, respectively. Represents the L2 norm; Indicates the transpose operation; Restricted shared feature learning, as shown below: ； in, This represents the loss function.
7. The schizophrenia classification method based on connectivity gradient according to claim 6, characterized in that, In step S44, the hierarchical pooling layer uses Top-K hierarchical pooling to retain the classification nodes. The specific process is as follows: Based on trainable projection vectors Obtain the scalar projection vector for each node. As shown below: ； ； in, Indicates the l-th layer; express Scalar projection vector of the node; express Features; Represents the projection vector; Indicates the vector used for lookup The indices of the k largest elements in the array; The new feature matrix is calculated using the ranking scores of the l-th layer, as shown below: ； in, express The feature matrix of the layer; Represents the hyperbolic tangent activation function; It is element-wise multiplication; The new adjacency matrix is as follows: ； in, express The adjacency matrix of the layer; This means that only the submatrix corresponding to the top-k index i is retained.
8. The schizophrenia classification method based on connectivity gradient according to claim 1, characterized in that, In step S5, the graph convolutional network classification model is trained and tested. The graph convolutional model is trained using the connection network and connection gradient calculated in step S3. Part of the training data and the data to be tested are input into the trained graph convolutional model, and the predicted label of the data to be tested is output. Specifically, the cross-entropy loss function is used to calculate the error between the predicted output of the graph convolutional network classification model and the true label. The parameters of each layer in the network are updated through backpropagation of the error and the stochastic gradient descent algorithm.