Method and system for classifying multi-modal brain image data based on dual-graph autoencoder

Abstract

This invention discloses a multimodal brain imaging data classification method and system based on a dual-graph autoencoder. The method includes the following steps: Step S01. Acquire multimodal brain medical imaging data of different target individuals and extract standard uptake ratio data and similarity matrix data; Step S02. Construct individual brain region maps and population maps; Step S03. Extract individual embedding features using an individual graph autoencoder and extract population embedding features using a population graph autoencoder; Step S04. Use the standard uptake ratio, brain region similarity matrix, individual embedding features, and population embedding features as views for enhanced multi-view subspace clustering; Step S05. Select key brain regions and discriminative brain region connections to participate in the next round of clustering as enhanced views until a preset convergence condition is met. This invention can achieve high-precision, highly interpretable data classification and simultaneously output biologically interpretable discriminative brain regions and connections.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a multimodal brain image data classification method and system based on a dual-image autoencoder. Background Technology

[0002] With the rapid development of neuroimaging technology, brain state analysis based on multimodal brain imaging data has become a research hotspot in the field of intelligent medical data analysis. Brain imaging data is highly heterogeneous; for example, brain imaging data of individuals with brain function-related diseases such as Alzheimer's disease, schizophrenia, and autism spectrum disorder all exhibit high heterogeneity. This means that the imaging representations of different individuals under the same pathological state show significant differences, posing a significant challenge to accurate data classification. Accurate classification and analysis of brain imaging data is of great value for achieving personalized diagnosis and treatment, revealing potential biomarkers, and guiding clinical decision-making. Current classification and analysis techniques for brain imaging data typically rely on clinical experience or cluster analysis based on single-modal imaging features. For example, grouping based solely on brain region volume extracted by structural magnetic resonance imaging or solely on the ratio of standard uptake values ​​obtained from positron emission tomography (PET) scans. Such single-modal methods cannot comprehensively characterize the multidimensional pathological features of diseases, leading to one-sided and insufficiently robust classification results.

[0003] In recent years, multi-view clustering technology has been attempted to integrate image information from different modalities to compensate for the shortcomings of single-modal methods. However, most multi-view clustering methods still employ simple feature stitching or linear weighted fusion, failing to effectively uncover nonlinear correlations and higher-order structural relationships between different views, thus affecting classification accuracy. Furthermore, traditional clustering methods often only output group labels, unable to explain discriminative biomarkers between groups, resulting in poor interpretability of clustering results and difficulty in guiding practical biological analysis or data screening. While graph neural networks have shown great potential in brain network analysis, existing graph neural network-based brain network analysis methods mostly focus on the representation learning of individual brain networks, neglecting the group similarity structure between different individuals. They fail to fully utilize group-level topological constraints to optimize individual feature representation. Moreover, existing graph neural network-based brain network analysis methods generally lack mechanisms for co-optimizing feature selection and clustering processes, leading to a disconnect between feature selection and clustering objectives, resulting in final classification results that are neither accurate nor biologically interpretable.

[0004] In summary, there is an urgent need to provide a multimodal brain imaging data classification method that can integrate multimodal brain imaging data, make full use of individual brain network topology and group similarity relationships, achieve high-precision and highly interpretable data classification under unsupervised conditions, and output biologically interpretable discriminative brain regions and connections. Summary of the Invention

[0005] The technical problem to be solved by this invention is: In view of the technical problems existing in the prior art, this invention provides a multimodal brain imaging data classification method and system based on a dual-graph autoencoder for relevant biomarkers with high clustering accuracy and strong interpretability. It can fuse multimodal brain imaging data, make full use of individual brain network topology and group similarity relationship, achieve high-precision and highly interpretable data classification under unsupervised conditions, and simultaneously output biologically interpretable discriminative brain regions and connections.

[0006] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A multimodal brain imaging data classification method based on a dual-graph autoencoder, comprising the following steps: Step S01. Multimodal data acquisition: Acquire multimodal brain medical imaging data of different target individuals, and extract the standard uptake ratio data of each brain region of different target individuals and the brain region similarity matrix data of different target individuals; Step S02. Graph Construction: Based on the standard uptake ratio data and the brain region similarity matrix data, individual brain region maps and group maps are constructed respectively. In the individual brain region map, each brain region is used as a node, the standard uptake ratio is used as the node feature, and the similarity matrix is ​​used as the edge weight. In the group map, each target individual is used as a node, the standard uptake ratio vector is used as the node feature, and the feature similarity between target individuals is used to construct edges. Step S03. Feature extraction based on dual graph autoencoder: Individual embedding features of each target individual are extracted from each of the individual brain region maps using an individual graph autoencoder, and group embedding features of each target individual are extracted from the group map using a group graph autoencoder. The encoder of the individual graph autoencoder includes a two-layer graph convolutional network, and the encoder of the group graph autoencoder includes a two-layer graph attention network. Step S04. Multi-view subspace clustering: The standard uptake ratio data, the brain region similarity matrix data, the individual embedding features, and the group embedding features are respectively used as views for enhanced multi-view subspace clustering. The enhanced multi-view subspace clustering is performed by calculating the affinity matrix of each view, generating the attention weight of each target individual on each view based on the standard uptake ratio data of each target individual, fusing the affinity matrices of each view according to the attention weights to obtain a fused affinity matrix, and performing spectral clustering based on the fused affinity matrix to obtain the clustering grouping results. Step S05. Iteration Control: Select key brain regions based on the current clustering results, select discriminant brain region connections based on the selected key brain regions, use the selected key brain regions and discriminant brain region connections as an enhanced view and return to step S04 to participate in the next round of clustering, until the preset convergence condition is met and the final clustering results and the corresponding key brain regions and discriminant brain region connections are output.

[0007] Further, in step S03, the extraction of individual embedding features of each target individual from each individual brain region map using an individual graph autoencoder includes: inputting each brain region map into the individual graph autoencoder for encoding and decoding; extracting the embedding features of each node in the individual brain region map through an encoder composed of a two-layer graph convolutional network; obtaining the individual embedding features of each target individual through global average pooling; reconstructing the adjacency matrix of each brain region map by the decoder through the inner product of the embedding features of each node and through the sigmoid activation function σ; and reconstructing the node features of each brain region map through graph-level embedding and fully connected layers.

[0008] Further, in step S03, the step of extracting the group embedding features of each target individual from the group graph using a group graph autoencoder includes: inputting the group graph into the group graph autoencoder for encoding and decoding; extracting the embedding features of each node in the group graph through an encoder composed of two layers of graph attention network to obtain the group embedding features of each target individual; and reconstructing the adjacency matrix of the group graph by the decoder through the inner product of the group embedding features of each node and the sigmoid activation function, and reconstructing the node features of the group graph through a fully connected layer using the group embedding features of each node.

[0009] Further, in step S04, the enhanced multi-view subspace clustering is executed by configuring the enhanced multi-view subspace clustering module. This module initializes the self-representation coefficient matrix for each view, forces the diagonal to zero, and applies sparsity regularization. It stacks the self-representation coefficient matrices of each view along the third dimension into a third-order tensor and calculates the tensor kernel norm as a low-rank constraint. The affinity matrix of each view is obtained by taking the absolute value and symmetric processing from the self-representation coefficient matrices. The standard uptake ratio data of each target individual is input into the sample-level attention network to generate the attention weights of each target individual on each view. The affinity matrices of each view are then weighted row-by-row according to the attention weights and symmetric to obtain the fused affinity matrix. .

[0010] Furthermore, the sample-level attention network generates the attention weights of each target individual on each view through a two-layer fully connected network, wherein the first layer fully connected network calculates: ,in, This represents the output of the first layer of the fully connected network for the l-th target individual. , For learnable parameters, Hiding dimensions for attention This is the ratio of the standard intake value for the l-th target individual. Indicating the activation function, the second fully connected network calculates the unnormalized attention score: ,in, , For learnable parameters, This represents the output of the second fully connected layer network for the l-th target individual. Attention weights on each view are obtained using the softmax function. ,in And satisfy and , =1~4 are used to represent the sequence number of each view.

[0011] Further, in step S05, the discriminant score of each brain region is calculated based on the current clustering results. Key brain regions are then selected based on their discriminant scores. The calculation expression for the discriminant score of each brain region is as follows: , in, For the first Preset weighting coefficients for each brain region For the first Mutual information between brain regions For all brain regions The maximum value, This represents the maximum value of mutual information across all brain regions. For traversing the node indexes of each brain region, For the first The Fisher discrimination ratio for each brain region is calculated using the following expression: , in, The number of clusters into which the target individual is divided. For the first The target set of individuals in a cluster For the first The number of target individuals within a cluster. For the first The target individual in the first The ratio of standard uptake values ​​for each brain region For the first The target individual in the first Global mean of each brain region For the first No. within the cluster The mean of each brain region.

[0012] Further, in step S05, the screening of discriminative brain region connections based on the selected key brain regions includes: A hypergraph is constructed based on the selected key brain regions: using the set of selected key brain regions For a node set, for each brain region node within the node set... Select the node with the highest similarity to the current node at the group level. The nodes together form a hyperedge To obtain the superedge set Construct the association matrix of the hypergraph For any brain region node and hyperset any hyperedge in If node Contained in hyperedge In, that is Then the correlation matrix elements in Otherwise ; Calculate the Laplacian matrix of the hypergraph: Calculate the node degree matrix The node degree matrix diagonal elements Calculate the hypermarginality matrix The hyper-marginality matrix diagonal elements According to the node degree matrix and hyperedge degree matrix Calculate the Laplacian matrix of the hypergraph ; The topological importance score of the connections between each brain region is calculated based on the Laplacian matrix of the hypergraph. The strength differences of connectivity between brain regions in each cluster group were calculated based on the brain region similarity matrix data of each target individual: Discriminative brain region connections were selected based on the topological importance score and the strength differences of each brain region connection in each cluster group.

[0013] Furthermore, the expression for calculating the topological importance score of connections between brain regions is as follows: , in, Represents any two brain region nodes Topological importance score of connections; The expression for calculating the Laplacian matrix of a hypergraph is: , The expression for calculating the strength differences of connectivity between different brain regions in each cluster group is as follows: , in, For target individuals The brain region similarity matrix corresponds to two brain region nodes elements, The global mean. Indicates the total number of target individuals. The number of clusters into which the target individual is divided. For the first The target set of individuals in a cluster For the first The number of target individuals within a cluster.

[0014] Further, steps S04 and S05 are iteratively optimized through joint training. In each iteration, the features of each view and the fusion affinity matrix are first calculated based on the current model parameters. Then, the key brain regions and discriminant brain regions connectivity screening results are updated based on the spectral clustering results of the fusion affinity matrix. The screening results are used as the enhanced view input to the enhanced multi-view subspace clustering module. The feature vector of each target individual in the enhanced view is composed of the ratio of the standard uptake values ​​of the selected key brain regions and the connectivity strength of the discriminant brain regions. The overall loss of the joint training is: , in, This represents the total loss of the enhanced multi-view subspace clustering module. This represents the reconstruction loss of the individual graph autoencoder. This represents the reconstruction loss of the group graph autoencoder. , Preset weighting coefficients; The total loss of the enhanced multi-view subspace clustering module By reducing reconstruction error loss Sparse regularization loss Low-rank constraint loss and fusion consistency loss The low-rank constraint loss is obtained by weighted summation. By stacking the four coefficient matrices into a third-order tensor ,in , The self-representation coefficient matrix of the ν-th view is transformed to the frequency domain by Fast Fourier Transform. The sum of the singular values ​​of the matrix is ​​calculated for each frequency domain slice, and the average is taken to obtain the tensor nuclear norm. ,Right now , The preset weighting coefficients are used; the fusion consistency loss is: ,in Affinity matrix for each view On average, For preset weighting coefficients, To merge the affinity matrices, the affinity matrices of each view are... The calculation expression is: , .

[0015] A computer system includes a processor and a memory, the memory being used to store a computer program, and the processor being used to execute the computer program to perform the method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention employs a dual-graph autoencoder architecture that uses individual brain region maps and patient group maps for collaborative learning. It utilizes individual and group graph autoencoders respectively to extract individual brain network topology and metabolic features, as well as the similarity structure between target individuals, forming multi-view features. Combined with multi-view subspace clustering, it achieves multimodal brain imaging data classification. This approach integrates multimodal brain imaging data, fully utilizing individual brain network topology and group similarity relationships. Simultaneously, it employs a closed-loop iterative collaborative optimization mechanism where clustering and feature selection are mutually driven. The clustering results and feature selection are coupled through a closed-loop pathway into a dynamic, mutually reinforcing process, significantly improving clustering performance and providing inherent interpretability to the output results. This allows for high-precision, highly interpretable classification under unsupervised conditions, fully utilizing individual brain network topology and group similarity relationships. This overcomes the shortcomings of traditional single-modal methods, which offer limited information and cannot simultaneously characterize individual microstructure and group macrostructure.

[0017] 2. In the process of multi-view subspace clustering, this invention learns a self-representation coefficient matrix for each view and applies sparse regularization and tensor kernel norm low-rank constraints, which can effectively capture high-order nonlinear correlations between multiple views. At the same time, a sample-level attention network is used to realize adaptive weighted fusion of views, avoiding the limitations of equal weight average fusion, thereby significantly improving the sufficiency of multimodal information fusion and clustering accuracy.

[0018] 3. This invention further calculates brain region discriminant scores based on current clustering results to screen key brain regions, constructs a hypergraph and screens discriminant brain region connections. The hypergraph model can capture high-order cooperative relationships between multiple brain regions. The topological importance score calculated by the hypergraph Laplacian operator can effectively identify connections that play a key role in the hub position of complex networks. Furthermore, combining the topological importance score with the inter-group strength difference to screen discriminant brain region connections can further improve the accuracy and interpretability of classification, ensuring high-precision and highly interpretable classification.

[0019] 4. This invention further optimizes the individual graph autoencoder, the group graph autoencoder and the enhanced multi-view subspace clustering module end-to-end through joint training, so that each module can promote each other and evolve together, thereby further improving the accuracy and robustness of classification. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the implementation process of the multimodal brain imaging data classification method based on a dual-image autoencoder in this embodiment.

[0021] Figure 2 This is a detailed schematic diagram illustrating the principle and process of multimodal brain imaging data classification in this embodiment. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0023] like Figure 1 , Figure 2 As shown, the steps of the multimodal brain imaging data classification method based on dual-graph autoencoders in this embodiment include: Step S01. Multimodal data acquisition: Acquire multimodal brain medical imaging data of different target individuals, and extract the standard uptake ratio data of each brain region of different target individuals and the brain region similarity matrix data of different target individuals from the multimodal brain medical imaging data.

[0024] In this embodiment, the multimodal brain medical imaging data includes positron emission tomography (PET) images and magnetic resonance imaging (MRI) images. After preprocessing and feature extraction, the multimodal brain medical imaging data yields standardized uptake ratio data and brain region similarity matrix data. The standardized uptake ratio data is the ratio between the average radiotracer uptake value of each brain region and the uptake value of a reference region. The brain region similarity matrix data includes the correlation coefficient between brain region features of any two brain regions. The specific extraction process is as follows: (1) Extraction of Standard Uptake Ratio (SUVR) data: Brain PET images of the target individuals were acquired, and after head movement correction and spatial normalization to a standard brain template space (such as MNI space), brain regions were divided in conjunction with anatomical brain atlases (such as AAL atlases). Subsequently, using a specific brain region (such as the cerebellum or the whole brain mean) as a reference region, the ratio of the average radiotracer uptake value of each target brain region to the uptake value of the reference region was calculated to obtain the Standard Uptake Ratio (SUVR) of each brain region, thereby constructing the Standard Uptake Ratio data of each target individual to characterize the local metabolic activity of the brain region.

[0025] (2) Extraction of brain region similarity matrix data: Obtain brain MRI images (including structural magnetic resonance imaging (sMRI) or functional magnetic resonance imaging (fMRI) of the target individuals. If it is structural magnetic resonance imaging (sMRI), extract the morphological features of each divided brain region (such as cortical thickness and gray matter volume); if it is functional magnetic resonance imaging (fMRI), extract the blood oxygen level dependent (BOLD) signal of each divided brain region in the time series. Subsequently, calculate the correlation coefficient (such as Pearson correlation coefficient) between any two brain region features to measure the structural covariance or functional connectivity between brain regions, thereby constructing the brain region similarity matrix data of each target individual.

[0026] Let the total number of target individuals be The number of brain regions is For the first For each target individual, the standard intake ratio vector is denoted as... The similarity matrix of its brain regions is denoted as It is also a symmetric matrix.

[0027] Step S02. Graph Construction: Based on the standard uptake ratio data and the brain region similarity matrix data, individual brain region maps and group maps are constructed respectively. In the individual brain region map, each brain region is a node, the connection between each brain region is an edge, the standard uptake ratio is a node feature, and the brain region similarity matrix is ​​an edge weight. In the group map, each target individual is a node, the standard uptake ratio vector is a node feature, and the feature similarity between target individuals is used to construct edges.

[0028] In this embodiment, individual brain region maps and population maps are constructed based on standard uptake ratio data and brain region similarity matrix data, respectively, to capture the brain region topology, metabolic information, and similarity between individuals. In the individual brain region map, nodes are defined as individual brain regions, node features are defined as corresponding standard uptake ratio values, edges are defined as anatomical or functional connections between brain regions, and edge weights are defined as the corresponding element values ​​in the brain region similarity matrix. In the population map, nodes are defined as target individuals, node features are defined as standard uptake ratio vectors (composed of standard uptake ratio data from each brain region of the target individual), and edges are defined as feature similarities between target individuals. Feature similarity between target individuals can be determined by calculating the cosine similarity between the standard uptake ratio vectors of each target individual. A K-Nearest Neighbor Graph is constructed based on the cosine similarity, connecting each node to the K nodes with the highest cosine similarity in the population via edges, and using the calculated cosine similarity values ​​as the edge weights of the corresponding edges in the population graph.

[0029] Step S03. Feature extraction based on dual-graph autoencoder: Individual embedding features of each target individual are extracted from each brain region map using individual graph autoencoder, and group embedding features of each target individual are extracted from the group map using group graph autoencoder. The encoder of the individual graph autoencoder includes a two-layer graph convolutional network, and the encoder of the group graph autoencoder includes a two-layer graph attention network.

[0030] In this embodiment, a dual-graph autoencoder architecture is adopted, which uses individual brain region maps and patient group maps for collaborative learning. The individual graph autoencoder encodes the topological and metabolic information of the brain regions of each target individual using a graph convolutional network, extracting embedding features that reflect the microscopic connection patterns within the individual. The group graph autoencoder encodes the similarity relationships between target individuals using a graph attention network, extracting embedding features that reflect the macroscopic structural position of the group. The two branches are jointly optimized under a joint training framework. By using isomorphic graph neural networks to jointly model patient data at both intra-individual and inter-individual scales, the shortcomings of traditional methods that only focus on single-dimensional information can be overcome. Individual embedding features capture the unique brain function or metabolic network configuration of each individual, while group embedding features characterize the relative position and neighborhood relationships of individuals in the overall sample distribution. By inputting both as independent views into the downstream clustering module, a multi-scale representation with high complementarity between microscopic and macroscopic, local and global perspectives can be formed.

[0031] In this embodiment, the individual graph autoencoder includes an encoder and a decoder. The encoder consists of a two-layer graph convolutional network (GCN). Extracting individual embedding features of each target individual from various brain region maps using the individual graph autoencoder includes: inputting each brain region map into the individual graph autoencoder for encoding and decoding; extracting the embedding features of each node in the individual brain region map through the encoder composed of the two-layer graph convolutional network; and obtaining the individual embedding features of each target individual through global average pooling. The decoder reconstructs the adjacency matrix of each somatic brain region map by using the inner product of the embedded features of each node and passing it through the sigmoid activation function σ, and reconstructs the node features of each somatic brain region map through graph-level embedding and fully connected layers. The individual graph autoencoder can enhance the embedding fidelity by reconstructing the adjacency matrix and node features simultaneously during decoding.

[0032] As an optional implementation, let the first... The individual brain region map node feature matrix of each target individual is as follows: The adjacency matrix is After adding self-loops and normalizing, it is denoted as Then the output of the first layer of the graph convolutional network in the encoder of the individual graph autoencoder is: (1) in, This is the first layer learnable weight matrix. To hide dimensions, This is the LeakyReLU activation function.

[0033] The output of the second layer graph convolutional network is: (2) in, This is the second-layer learnable weight matrix. As the latent dimension, the output node embedding matrix Individual embedding features are obtained through global average pooling. , This represents the total number of brain regions in the target individual. Indicates the index of brain region nodes. Represents the node embedding matrix The Row vector (i.e., the first row vector) (Embedding features of individual brain region nodes). The decoder of the individual graph autoencoder includes an adjacency matrix reconstruction branch and a node feature reconstruction branch. The adjacency matrix reconstruction branch uses the inner product of the embedding features of each node and passes them through the sigmoid activation function. Reconstruct the adjacency matrix, that is: (3) in, This is the reconstructed adjacency matrix.

[0034] The node feature reconstruction branch reconstructs node features through graph-level embedding and a fully connected layer, i.e.: (4) in, This is a learnable weight matrix.

[0035] In this embodiment, the individual brain region map adjacency matrix obtained by the above reconstruction With node features Instead of being directly output to the downstream clustering module as input features, these features are used in subsequent iterative joint training phases (step S05) to calculate the reconstruction error by comparing the similarity matrix with the original input and the standard uptake value ratio data, thus constructing the individual graph reconstruction loss term. This approach forces the encoder of the individual graph autoencoder to extract the optimal individual embedding features that can highly restore the original micro-connection patterns and metabolic activities of the individual through an unsupervised reconstruction process.

[0036] In this embodiment, extracting the group embedding features of each target individual from the group graph using a group graph autoencoder includes: inputting the group graph into the group graph autoencoder for encoding and decoding, and extracting the embedding features of each node in the group graph through an encoder composed of a two-layer graph attention network (GAT) to obtain the group embedding features of each target individual. The decoder reconstructs the adjacency matrix of the group graph by the inner product of the group embedding features of each node and the sigmoid activation function, and reconstructs the node features of the group graph by the group embedding features of each node through a fully connected layer.

[0037] As an optional implementation, the node feature matrix of the group graph is , of which Behavior The adjacency matrix is ​​obtained by constructing a K-nearest neighbor graph based on cosine similarity. And retain the fully connected cosine similarity matrix As the target of adjacency reconstruction, among which The swarm graph autoencoder selects the fully connected cosine similarity matrix as the target for adjacency matrix reconstruction, which, compared to traditional sparse K-nearest neighbor graphs, preserves richer information on the gradual changes in similarity among patients. The encoder of the swarm graph autoencoder consists of a two-layer graph attention network. The first layer employs a multi-head attention mechanism with a number of heads... The output is: (5) in, For the first layer of graph attention network The output of ELU represents the activation function.

[0038] Output dimension is The output of the second layer graph attention network is: (6) in, This represents the second-layer graph attention network. The output dimension of the second-layer graph attention network is... The output is the group embedding feature matrix. , its first Behavioral target individuals group embedding .

[0039] The decoder for the population graph includes an adjacency matrix reconstruction branch and a node feature reconstruction branch. The adjacency matrix reconstruction branch uses the inner product of the population embedding features of each node and then applies the sigmoid activation function. Reconstruct the adjacency matrix of the community graph, i.e.: (7) in, This is the reconstructed adjacency matrix of the group graph.

[0040] The node feature reconstruction branch embeds the feature matrix through groups of nodes. The features of the community graph nodes are reconstructed through a fully connected layer, namely: (8) in, For learnable weight matrix, Features of the reconstructed group graph nodes.

[0041] Step S04. Multi-view subspace clustering: Standard uptake ratio data, brain region similarity matrix data, individual embedding features, and group embedding features are used as views for enhanced multi-view subspace clustering. Enhanced multi-view subspace clustering is performed by calculating the affinity matrix of each view and generating the attention weight of each target individual on each view based on the standard uptake ratio data of each target individual. The affinity matrices of each view are fused according to the attention weights to obtain a fused affinity matrix. Spectral clustering is then performed based on the fused affinity matrix to obtain the clustering grouping results.

[0042] In this embodiment, multi-view subspace clustering is performed by configuring an enhanced multi-view subspace clustering module. The original standard uptake ratio data of each target individual is used as the first view; the brain region similarity matrix of each target individual after dimensionality reduction is used as the second view; the individual embedding features of each target individual are used as the third view; and the group embedding features of each target individual are used as the fourth view. All these are input into the enhanced multi-view subspace clustering module. The specific execution steps of the enhanced multi-view subspace clustering module include: Step S401. Initialize the self-representation coefficient matrix for each view and force the diagonal to zero and apply sparsity regularization.

[0043] Set view number , No. The feature matrix of each view is denoted as ,in , The dimension after dimensionality reduction. , In subsequent joint training optimization, the reconstruction error loss of the multi-view subspace self-representation is calculated based on the feature matrix of each view (i.e., the calculation of...). and (the error between them), to guide the learning of the self-representation coefficient matrix.

[0044] Initialize a learnable self-representation coefficient matrix for each view And force the diagonal elements to be zero, that is Specifically, an independent random initialization strategy (such as Xavier uniform distribution initialization or Gaussian distribution initialization) can be used to initialize each view independently, that is, the self-representation coefficient matrix of each view is initialized. They are all different, enabling them to fit independent data subspace structures under different views.

[0045] Step S402. Stack the self-representation coefficient matrices of each view along the third dimension into a third-order tensor, and apply the tensor kernel norm as a low-rank constraint.

[0046] In this example, the self-representation coefficient matrices learned independently from the four views are... (in Stacked along the third dimension (view dimension), it is constructed as a third-order tensor. .

[0047] For this third-order tensor The specific technical solution for applying low-rank constraints, using the tensor nuclear norm (TNN) as the low-rank constraint, is as follows: along the third-order tensor... The third dimension (i.e., the view dimension) is subjected to a one-dimensional discrete fast Fourier transform (FFT) to convert the spatial domain tensor to a frequency domain tensor. For the obtained frequency domain tensor, each frontal slice matrix is ​​extracted, and the nuclear norm (i.e., the sum of all singular values ​​of the matrix) of each frequency domain slice matrix is ​​calculated through singular value decomposition. The tensor nuclear norm is obtained by averaging the nuclear norms of all frequency domain slice matrices. The kernel norm of this tensor is used as a low-rank constraint loss term. It is incorporated into the total loss function for end-to-end optimization. By forcibly minimizing the tensor rank composed of multiple view coefficient matrices in the frequency domain, it can effectively mine and maintain the high-order nonlinear consistency correlation between multiple view data subspaces, mathematically overcoming the information fragmentation problem caused by simple linear splicing in traditional methods.

[0048] This embodiment stacks the self-representation coefficient matrices of each view into a third-order tensor and applies a low-rank constraint on the tensor's nuclear norm, thereby uncovering higher-order consistency between the subspace structures of different views. The low-rank constraint treats the self-representation matrices of multiple views as a single, unified tensor, and forces the sample similarity structure of different views to converge by minimizing its nuclear norm. This effectively captures nonlinear higher-order correlations between views, addressing the problem of insufficient information utilization in traditional methods that rely on simple concatenation or linear weighted fusion.

[0049] Step S403. Obtain the affinity matrix of each view by taking the absolute value and symmetry processing from the self-representation coefficient matrix of each view.

[0050] As an optional implementation, the affinity matrix of each view can be obtained from the self-representation coefficient matrix by taking the absolute value and symmetricizing, i.e. , .

[0051] Step S404. Input the standard uptake ratio data of each target individual into the sample-level attention network to generate the attention weights of each target individual in each view. Then, perform row-wise weighted fusion of the affinity matrices of each view according to the attention weights and symmetricize them to obtain the fused affinity matrix. .

[0052] In this embodiment, a sample-level attention network is used to generate four view fusion weights specific to each target individual, based on their original image data. This enables adaptive weighting of each view affinity matrix row by row, allowing the system to dynamically adjust the importance of each view according to the data characteristics of each patient. For example, for individuals with atypical network characteristics, the influence of the individual embedded view is automatically weakened while the contribution of the group embedded view or the original data view is enhanced. This fine-grained adaptive fusion strategy can greatly improve the accuracy and robustness of the fusion affinity matrix.

[0053] As an optional implementation, the sample-level attention network generates the attention weights of each target individual on each view through a two-layer fully connected network. The sample-level attention network takes the original standard uptake ratio data s_ι∈R^M of each target individual as input, wherein the first layer of the fully connected network is calculated according to the following formula: (9) in, Indicates the first The output of the first fully connected layer of the target individual's network. , For learnable parameters, Hiding dimensions for attention For the first Standard intake ratio data for each target individual This represents the activation function. , indicating the sequence number of each view.

[0054] The second fully connected network calculates the unnormalized attention score: (10) in , The output of the second fully connected network is a learnable parameter. Attention weights on each view are obtained using the softmax function. ,in And satisfy and .

[0055] Then, a fusion affinity matrix is ​​obtained through weighted fusion for each target individual. The fused affinity row vector is the weighted sum of the corresponding rows of the affinity matrices of each view, that is: (11) The final fusion affinity matrix is ​​then obtained by symmetry processing: (12) in This is the average of the affinity matrices for each view.

[0056] Step S05. Iteration Control: Select key brain regions based on the current clustering results, select discriminant brain region connections based on the selected key brain regions, use the selected key brain regions and discriminant brain region connections as views and return to step S04 to participate in the next round of multi-view subspace clustering, until the preset convergence condition is met and the final clustering results and corresponding key brain regions and discriminant brain region connections are output. The feature vector of the enhanced view is composed of the standard uptake ratio of the selected key brain regions and the discriminant brain region connection strength.

[0057] In this embodiment, discriminative brain region connectivity refers to brain region association pathways in multimodal brain networks that occupy core hub positions in complex network topologies (possessing high topological importance scores) and exhibit significant differences in connectivity strength between different cluster subtypes. Discriminative brain region connectivity can serve as a biologically interpretable biomarker for distinguishing different disease or physiological subtypes. By employing a closed-loop iterative collaborative optimization mechanism that drives clustering and feature selection, the clustering results and the "basis" for feature selection are coupled into a dynamic mutually reinforcing process through a closed-loop pathway. This enables more accurate grouping, which in turn guides the selection of more discriminative image features. Furthermore, re-injecting the selected high-quality features into the clustering module further optimizes the precision of data subspace partitioning. This breaks the paradigm of independent clustering and feature selection in traditional methods, allowing the system to spontaneously converge to an equilibrium state with more accurate grouping and more reliable feature interpretation during iteration. This significantly improves clustering performance and endows the output results with inherent interpretability from a mechanistic perspective. As a result, it can fully utilize the individual brain network topology and group similarity relationship to achieve high-precision and highly interpretable data classification under unsupervised conditions.

[0058] Assuming the current target individual subtype grouping results will be The target individuals are divided into The cluster, the first The target set of individuals of a cluster is denoted as The number of target individuals within this cluster is For the first The brain region that records the first... The standard uptake ratio of the target individual in this brain region is: The global mean of this brain region is , No. The mean within the cluster is: (13) As an optional implementation, a discriminant score for each brain region is calculated based on the current clustering results, and key brain regions are selected based on these scores. Specifically, the discriminant score for each brain region can be calculated using a weighted sum of Fisher's discriminant ratio and mutual information, as shown in the following expression: (14) in, For the first Preset weighting coefficients for each brain region For the first Mutual information between brain regions For all brain regions The maximum value, This represents the maximum value of mutual information across all brain regions. For traversing the node indexes of each brain region, For the first Fisher discrimination ratio for each brain region, Fisher discrimination ratio The calculation expression is: (15) in, The number of clusters into which the target individual is divided. For the first The target set of individuals in a cluster For the first The target number of individuals in the cluster For the first The target individual in the first The ratio of standard uptake values ​​for each brain region For the first The target individual in the first Global mean of each brain region For the first No. within the cluster The Fisher discriminant ratio formula, which represents the mean of each brain region, has its numerator representing the sum of inter-cluster variances and its denominator representing the sum of intra-cluster variances. By maximizing this ratio, brain regions with the highest discriminative power among different cluster groups can be evaluated and selected.

[0059] As shown in Equation (14), the discriminant score of each brain region is obtained by weighted calculation of Fisher's discriminant ratio and mutual information, which can take into account both linear separability and nonlinear correlation, and improve the screening accuracy of key brain regions.

[0060] As an optional implementation method, mutual information The calculation expression is: (16) in, To make the first The values ​​of the brain regions are discretized into probability distributions of Q intervals. This is the joint distribution with the cluster label y.

[0061] Then, key brain regions are selected based on the discriminant scores of each brain region. For example, one could select... Top brain regions as a set of key brain regions , This is a preset ratio value.

[0062] As an optional implementation, a discriminative brain region connectivity screening method based on hypergraphs is adopted. On the selected key brain region node set, a hyperedge is constructed by forming a hyperedge around each node with several nodes most similar to it at the group level. The hypergraph Laplacian matrix is ​​then calculated, and the higher-order topological importance score of each brain region connectivity is calculated. This score is combined with the strength differences of connectivity between different clusters for joint screening, achieving discriminative brain region connectivity screening and identification. Hypergraph models can capture higher-order collaborative relationships between multiple brain regions, surpassing the limitations of traditional graph models that can only describe binary interactions between pairs of nodes. Furthermore, a single hyperedge of a hypergraph can connect multiple brain regions simultaneously, making it naturally suitable for characterizing modular and collaborative working patterns in brain functional networks. Therefore, the topological importance score calculated using the hypergraph Laplacian operator can effectively identify connections that play a key role in the hub positions of complex networks. The discriminative brain region connectivity screened by combining this score with the strength differences between groups not only achieves statistical differentiation between different groups but also provides clear biological explanations, further improving the accuracy and interpretability of classification, and achieving high-precision and highly interpretable classification of multimodal brain imaging data.

[0063] Specifically, the steps for identifying discriminative brain region connectivity based on the selected key brain regions include: (1) Constructing a hypergraph based on the selected key brain regions: using the selected set of key brain regions Let be a set of nodes, denoted as For each brain region node in the node set Select the node with the highest similarity to the current node at the group level. The nodes together form a hyperedge ,Right now For nodes The hyperedges generated around the center are the hyperedge set. Construct the association matrix of the hypergraph Traverse all nodes and all hyperedges for any brain region node. and hyperset any hyperedge in If node Contained in hyperedge In, that is Then the correlation matrix elements in Otherwise ; (2) Calculate the Laplacian matrix of the hypergraph: Calculate the degree matrix of the nodes The node degree matrix diagonal elements Calculate the hypermarginality matrix The hyper-marginality matrix diagonal elements According to the node degree matrix and hyperedge degree matrix Calculate the Laplacian matrix of the hypergraph .

[0064] The Laplacian matrix of a hypergraph can be calculated using the following formula: (17) (3) Calculate the strength difference of each brain region connection in each cluster group based on the brain region similarity matrix data of each target individual.

[0065] For any two brain regions Based on the Laplacian matrix of the hypergraph, the topological importance score of connections between brain regions can be calculated using the following formula: (18) in, Represents any two brain region nodes Topological importance score of connections.

[0066] (4) Discriminative brain region connectivity was selected based on topological importance scores and the strength differences of connectivity between brain regions in each cluster group.

[0067] For example, the expression for calculating the strength difference of connectivity between different brain regions in each cluster group is: (19) in, For target individuals The brain region similarity matrix corresponds to two brain region nodes The element that represents the target individual. brain region nodes With brain region nodes Feature similarity between them. The global mean. Indicates the total number of target individuals. The number of clusters into which the target individual is divided. For the first The target set of individuals in a cluster For the first The number of target individuals within a cluster.

[0068] Furthermore, discriminative brain region connections are screened based on topological importance scores and the strength differences of connectivity between brain regions in each cluster group. For example, it is possible to screen out... and All ranked in the top The connections serve as a discriminative brain region connectivity set. , This is a preset threshold.

[0069] As an optional implementation, the convergence condition for the iteration can be configured such that the rate of change of cluster labels is less than a preset threshold in two consecutive iterations. ,Right now ,in For the first Target individual in round iteration Clustering labels, This is an indicator function.

[0070] As an optional implementation, steps S04 and S05 are iteratively optimized through joint training. In each iteration, the features and fusion affinity matrix of each view are first calculated based on the current model parameters. Then, the screening results of key brain regions and discriminant brain regions are updated based on the spectral clustering results of the fusion affinity matrix. The screening results are then used as the new fifth view and input into the enhanced multi-view subspace clustering module. The feature vector of each target individual in the fifth view is composed of the ratio of the standard uptake values ​​of the selected key brain regions and the connection strength of the discriminant brain regions. After the fifth view is input into the enhanced multi-view subspace clustering module, the calculation method of its affinity matrix, the initialization of the self-representation coefficient matrix, and the application of sparse regularization can all adopt the same method as the first to fourth views. The updated five views are uniformly incorporated into the self-representation learning and multi-view fusion framework to participate in the next round of clustering until the preset convergence condition is reached. The final clustering grouping results and the corresponding key brain regions and discriminant brain regions are then output. As an optional implementation, the convergence condition of the iteration can be configured as the rate of change of cluster labels in two consecutive iterations being less than a preset threshold. .

[0071] To enable the individual graph autoencoder, the group graph autoencoder, and the enhanced multi-view subspace clustering module to promote each other and evolve collaboratively, this embodiment constructs a joint loss function to perform end-to-end iterative optimization of the entire network model. It is important to note that the calculation of the overall joint training loss and each sub-loss function described below is performed only during the joint training phase (i.e., parameter learning) of the model, used to update the learnable parameters of the network model through the backpropagation algorithm. In the actual application phase after model training (i.e., inputting new target individual data for inference and classification), the network will directly output the classification result without further loss calculation or parameter updates.

[0072] As an optional implementation, the overall loss for joint training is: (20) in, This represents the total loss of the enhanced multi-view subspace clustering module. This represents the reconstruction loss of the individual graph autoencoder. This represents the reconstruction loss of the group graph autoencoder. , These are preset weighting coefficients.

[0073] As an optional implementation, the total loss of the enhanced multi-view subspace clustering module is... By reducing reconstruction error loss Sparse regularization loss Low-rank constraint loss and fusion consistency loss The weighted summation yields the low-rank constraint loss. By stacking the four coefficient matrices into a third-order tensor ,in , The self-representation coefficient matrix of the ν-th view is transformed to the frequency domain by Fast Fourier Transform. The sum of the singular values ​​of the matrix is ​​calculated for each frequency domain slice, and the average is taken to obtain the tensor nuclear norm. ,Right now , The preset weighting coefficients are used; the fusion consistency loss is: ,in Affinity matrix for each view On average, For preset weighting coefficients, To merge the affinity matrices, the affinity matrices of each view are... The calculation expression is: , .

[0074] As an optional implementation, the reconstruction loss of the individual graph autoencoder is calculated. The expression is: (twenty one) in , For preset weighting coefficients, The original node feature vector, It is the Frobenius norm. It is the Euclidean norm.

[0075] As an optional implementation, the reconstruction loss of the group graph autoencoder is calculated as follows: (twenty two) in , These are preset weighting coefficients.

[0076] As an optional implementation method, the formula for calculating the reconstruction error loss is: (twenty three) As an optional implementation, the sparse regularization loss is: (twenty four) in, Let L be the L1 norm of the matrix.

[0077] Low-rank constrained loss is achieved by stacking four coefficient matrices into a third-order tensor. ,in After converting to the frequency domain using a Fast Fourier Transform, the sum of the singular values ​​of the matrix for each frequency domain slice is calculated and averaged to obtain the tensor nuclear norm. ,Right now , These are preset weighting coefficients.

[0078] As an optional implementation, the fusion consistency loss is: (25) in The average of the affinity matrices for each view. These are preset weighting coefficients.

[0079] The total loss of the multi-view subspace clustering module is: (26) in represents the sparse regularization weight coefficient.

[0080] Under the constraint of the total loss of joint training, the two autoencoders (individual graph autoencoder and group graph autoencoder) can effectively guide the learning of discriminative features that are more conducive to distinguishing groups, thereby achieving deep coupling between representation learning and clustering objectives.

[0081] This embodiment further provides a computer system, including a processor and a memory, wherein the memory is used to store a computer program and the processor is used to execute the computer program to perform the method as described above.

[0082] It is understood that the method described in this embodiment can be executed by a single device, such as a computer or server, or it can be applied to a distributed scenario where multiple devices cooperate to complete the task. In a distributed scenario, one of the multiple devices may execute only one or more steps of the method described in this embodiment, and the multiple devices interact to complete the method. The processor can be implemented using a general-purpose CPU, microprocessor, application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the method described in this embodiment. The memory can be implemented using read-only memory (ROM), random access memory (RAM), static storage devices, and dynamic storage devices. The memory can store the operating system and other applications. When the method described in this embodiment is implemented through software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0083] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A multimodal brain imaging data classification method based on a dual-graph autoencoder, characterized by the following steps: include: Step S01. Multimodal data acquisition: Acquire multimodal brain medical imaging data of different target individuals, and extract the standard uptake ratio data of each brain region of different target individuals and the brain region similarity matrix data of different target individuals; Step S02. Graph Construction: Based on the standard uptake ratio data and the brain region similarity matrix data, individual brain region maps and group maps are constructed respectively. In the individual brain region map, each brain region is used as a node, the standard uptake ratio is used as the node feature, and the elements in the brain region similarity matrix are used as the edge weights. In the group map, each target individual is used as a node, the standard uptake ratio vector is used as the node feature, and the feature similarity between target individuals is used to construct edges. The standard uptake ratio vector is composed of the standard uptake ratio data of each brain region of the target individual. Step S03. Feature extraction based on dual graph autoencoder: Individual embedding features of each target individual are extracted from each of the individual brain region maps using an individual graph autoencoder, and group embedding features of each target individual are extracted from the group map using a group graph autoencoder. The encoder of the individual graph autoencoder includes a two-layer graph convolutional network, and the encoder of the group graph autoencoder includes a two-layer graph attention network. Step S04. Multi-view subspace clustering: The standard uptake ratio data, the brain region similarity matrix data, the individual embedding features, and the group embedding features are respectively used as views for enhanced multi-view subspace clustering. The enhanced multi-view subspace clustering is performed by calculating the affinity matrix of each view, generating the attention weight of each target individual on each view based on the standard uptake ratio data of each target individual, fusing the affinity matrices of each view according to the attention weights to obtain a fused affinity matrix, and performing spectral clustering based on the fused affinity matrix to obtain the clustering grouping results. Step S05. Iteration Control: Select key brain regions based on the current clustering results, select discriminant brain region connections based on the selected key brain regions, use the selected key brain regions and discriminant brain region connections as an enhanced view and return to step S04 to participate in the next round of clustering, until the preset convergence condition is met and the final clustering results and the corresponding key brain regions and discriminant brain region connections are output.

2. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to claim 1, characterized in that, In step S03, the step of extracting individual embedding features of each target individual from each individual brain region map using an individual graph autoencoder includes: inputting each brain region map into the individual graph autoencoder for encoding and decoding; extracting the embedding features of each node in the individual brain region map through an encoder composed of a two-layer graph convolutional network; obtaining the individual embedding features of each target individual through global average pooling; reconstructing the adjacency matrix of each brain region map by the decoder through the inner product of the embedding features of each node and the sigmoid activation function σ; and reconstructing the node features of each brain region map through graph-level embedding and fully connected layers.

3. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to claim 2, characterized in that, In step S03, the step of extracting the group embedding features of each target individual from the group graph using a group graph autoencoder includes: inputting the group graph into the group graph autoencoder for encoding and decoding; extracting the embedding features of each node in the group graph through an encoder composed of a two-layer graph attention network to obtain the group embedding features of each target individual; and reconstructing the adjacency matrix of the group graph by the decoder through the inner product of the group embedding features of each node and the sigmoid activation function, and reconstructing the node features of the group graph through a fully connected layer using the group embedding features of each node.

4. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to claim 1, characterized in that, In step S04, the enhanced multi-view subspace clustering module is configured to perform the enhanced multi-view subspace clustering. This module initializes the self-representation coefficient matrix for each view, forces the diagonal to zero, and applies sparsity regularization. It stacks the self-representation coefficient matrices of each view along the third dimension into a third-order tensor and calculates the tensor kernel norm as a low-rank constraint. The affinity matrix of each view is obtained by taking the absolute value and symmetric processing from the self-representation coefficient matrices. The standard uptake ratio data of each target individual is input into the sample-level attention network to generate the attention weights for each target individual on each view. The affinity matrices of each view are then weighted row-by-row according to the attention weights and symmetric to obtain the fused affinity matrix. .

5. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to claim 4, characterized in that, The sample-level attention network generates the attention weights of each target individual on each view through a two-layer fully connected network, wherein the first layer of the fully connected network calculates: ,in, Indicates the first The output of the first fully connected layer of the target individual's network. , For learnable parameters, Hiding dimensions for attention For the first Standard intake ratio data for each target individual Indicating the activation function, the second fully connected network calculates the unnormalized attention score: ,in, , For learnable parameters, Indicates the first The output of the second fully connected layer of the target individual, the output of the second fully connected layer. Attention weights on each view are obtained using the softmax function. ,in And satisfy and , The number used to indicate the sequence of each view.

6. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to any one of claims 1 to 5, characterized in that, In step S05, the discriminant score of each brain region is calculated based on the current clustering results. Key brain regions are then selected based on their discriminant scores. The formula for calculating the discriminant score of each brain region is as follows: ， in, For the first Preset weighting coefficients for each brain region For the first Mutual information between brain regions For all brain regions The maximum value, This represents the maximum value of mutual information across all brain regions. For traversing the node indexes of each brain region, For the first The Fisher discrimination ratio for each brain region is calculated using the following expression: ， in, The number of clusters into which the target individual is divided. For the first The target set of individuals in a cluster For the first The number of target individuals within a cluster. For the first The target individual in the first The ratio of standard uptake values ​​for each brain region For the first The target individual in the first Global mean of each brain region For the first No. within the cluster Mean of each brain region.

7. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to any one of claims 1 to 5, characterized in that, In step S05, the process of screening discriminative brain region connections based on the selected key brain regions includes: A hypergraph is constructed based on the selected key brain regions: using the set of selected key brain regions For a node set, for each brain region node within the node set... Select the node with the highest similarity to the current node at the group level. The nodes together form a hyperedge To obtain the superedge set Construct the association matrix of the hypergraph For any brain region node and hyper-edge set any hyperedge in If node Contained in hyperedge In, that is Then the correlation matrix elements in Otherwise ; Calculate the Laplacian matrix of the hypergraph: Calculate the node degree matrix The node degree matrix diagonal elements Calculate the hypermarginality matrix The hyper-marginality matrix diagonal elements According to the node degree matrix and hyperboundary degree matrix Calculate the Laplacian matrix of the hypergraph ; The topological importance score of the connections between each brain region is calculated based on the Laplacian matrix of the hypergraph. The strength differences of connectivity between brain regions in each cluster group were calculated based on the brain region similarity matrix data of each target individual: Discriminative brain region connections were selected based on the topological importance score and the strength differences of each brain region connection in each cluster group.

8. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to claim 7, characterized in that, The expression for calculating the topological importance score of connections between brain regions is as follows: ， in, Represents any two brain region nodes Topological importance score of connections; The expression for calculating the Laplacian matrix of a hypergraph is: ， The expression for calculating the strength differences of connectivity between different brain regions in each cluster group is as follows: ， in, For target individuals The brain region similarity matrix corresponds to two brain region nodes elements, The global mean. Indicates the total number of target individuals. The number of clusters into which the target individual is divided. For the first The target set of individuals in a cluster For the first The number of target individuals within a cluster.

9. The multimodal brain imaging data classification method based on a dual-graph autoencoder according to any one of claims 1 to 5, characterized in that, Steps S04 and S05 are iteratively optimized through joint training. In each iteration, the features of each view and the fusion affinity matrix are first calculated based on the current model parameters. Then, the key brain regions and discriminant brain regions connectivity screening results are updated based on the spectral clustering results of the fusion affinity matrix. The screening results are used as the enhanced view input to the enhanced multi-view subspace clustering module. The feature vector of each target individual in the enhanced view is composed of the ratio of the standard uptake values ​​of the selected key brain regions and the connectivity strength of the discriminant brain regions. The overall loss of the joint training is: ， in, This represents the total loss of the enhanced multi-view subspace clustering module. This represents the reconstruction loss of the individual graph autoencoder. This represents the reconstruction loss of the group graph autoencoder. , Preset weighting coefficients; The total loss of the enhanced multi-view subspace clustering module By reducing reconstruction error loss Sparse regularization loss Low-rank constraint loss and fusion consistency loss The low-rank constraint loss is obtained by weighted summation. By stacking the four coefficient matrices into a third-order tensor ,in , The self-representation coefficient matrix of the ν-th view is transformed to the frequency domain by Fast Fourier Transform. The sum of the singular values ​​of the matrix is ​​calculated for each frequency domain slice, and the average is taken to obtain the tensor nuclear norm. ,Right now , The preset weighting coefficients are used; the fusion consistency loss is: ,in Affinity matrix for each view On average, For preset weighting coefficients, To merge the affinity matrices, the affinity matrices of each view are... The calculation expression is: , .

10. A computer system comprising a processor and a memory, the memory being used to store computer programs, characterized in that, The processor is used to execute the computer program to perform the method as described in any one of claims 1 to 9.

Images