A multi-view hypergraph fusion alzheimer's disease classification method

Abstract

The application belongs to the technical field of medical image processing and artificial intelligence, and particularly relates to a multi-view hypergraph fusion Alzheimer's disease classification method, structural magnetic resonance image data is acquired, brain region morphological features are extracted, and a sample feature matrix is constructed by taking a subject as a node; LASSO, sparse group LASSO and overlapping group LASSO are respectively used to construct subject correlation hypergraphs of three views; three hypergraphs are input into a parameter-independent hypergraph neural network coding branch, cross-view supervised contrast learning is performed at each network layer, and node features of different views are aligned; node-level attention weights are constructed according to node degrees of each subject node under each view, and a classification result is output after adaptive weighted fusion of node hidden representations of each view is input into a classifier; a classification loss and a contrast loss are jointly optimized to train the model, and the accuracy and stability of Alzheimer's disease classification are effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, and in particular to a multi-view hypergraph fusion method for classifying Alzheimer's disease. Background Technology

[0002] Alzheimer's disease (AD) is an irreversible neurodegenerative disease, primarily characterized by memory loss and cognitive decline. With the increasing aging of the population, early screening and auxiliary diagnosis of AD are of significant clinical importance. Structural magnetic resonance imaging (MRI) can reflect morphological changes in brain tissue, such as cortical thickness, gray matter volume, and subcortical structure volume, and has been widely used in computer-aided diagnostic research for AD.

[0003] In recent years, hypergraph neural networks have been applied to brain disease diagnosis. A single hyperedge in a hypergraph can connect multiple subject nodes simultaneously, effectively capturing high-order topological relationships between subjects and overcoming the limitation of traditional graph neural networks, which can only describe pairwise relationships. Existing technologies include AD classification methods based on hypergraph neural networks, such as AD classification methods based on multimodal hypergraph convolutional neural networks, AD classification methods based on line-hypergraph neural networks, and AD classification methods based on multi-hypergraph fusion. Among these, the hypergraph neural network model based on multi-hypergraph fusion proposed by Cao Pengjie et al. constructs three single hypergraphs using LASSO, sparse group LASSO, and covered group LASSO methods respectively based on sMRI data and fuses them. A hypergraph neural network model is then built on this foundation for the classification of AD and mild cognitive impairment.

[0004] However, existing hypergraph-based AD image classification methods still have the following shortcomings: First, the multi-hypergraph fusion stage usually adopts a strategy of fusing multiple hypergraphs into one hypergraph, or uses fixed fusion methods such as simple stitching and average pooling, failing to adaptively fuse according to the differences in topological structure of different subjects under different hypergraph views, resulting in insufficient utilization of effective features or redundant information stacking; Second, existing methods only perform feature fusion at the final output layer, ignoring the differences between the local structural information and global topological semantics contained in different network layers of the hypergraph neural network, resulting in insufficient cross-view feature consistency. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-view hypergraph fusion method for Alzheimer's disease classification, aiming to solve the problems in existing Alzheimer's disease magnetic resonance imaging hypergraph classification methods, such as coarse multi-hypergraph fusion methods, lack of hierarchical alignment of cross-view features, and lack of adaptive fusion mechanisms, which lead to insufficient mining of high-order topological information and limited classification accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a multi-view hypergraph fusion method for Alzheimer's disease classification, comprising: S1, acquiring and preprocessing structural magnetic resonance imaging data of subjects, extracting morphological features of brain regions of each subject, constructing a sample feature matrix with each subject as a node, and obtaining the corresponding category label for each subject, including Alzheimer's disease patient category label and normal control category label; S2, based on the sample feature matrix, constructing subject association hypergraphs from three perspectives using LASSO sparse representation strategy, sparse group LASSO sparse representation strategy, and overlapping group LASSO sparse representation strategy respectively, wherein each perspective subject association hypergraph contains a set of nodes and a set of hyperedges; S3, inputting the subject association hypergraphs from the three perspectives into three structurally identical and parameter-independent hypergraph neural network encoding branches, wherein each hypergraph neural network encoding branch aggregates information of multiple subject nodes within each hyperedge through hypergraph convolution, and extracts information from each perspective. The hidden node representations under different perspectives are obtained. In each network layer of the hypergraph neural network encoding branch, cross-view supervised contrastive learning is performed on the hidden node representations under different perspectives to obtain the node hidden representations of each perspective after cross-view alignment. S4. Node-level attention weights are constructed based on the node degree of each subject's node in each perspective. The node degree is the number of hyperedges associated with the subject's node. The node-level attention weights are used to adaptively weight and fuse the hidden node representations of each perspective after cross-view alignment to obtain fused node features. The fused node features are input into a classifier consisting of a fully connected layer and a normalized output layer to output the classification result of each subject belonging to the Alzheimer's disease patient category or the normal control category. The classification loss is calculated based on the classification result and the category label. The classification loss is weighted and summed with the contrastive loss generated by cross-view supervised contrastive learning to obtain the total loss. The network parameters are optimized through backpropagation to obtain the trained Alzheimer's disease classification model.

[0007] The subject association hypergraph is represented by an association matrix, which is defined as follows:

[0008] in, Represents the subject node, Indicates a superedge; if the subject node Belongs to superedge Then the elements of the correlation matrix The value is 1 if it is not 1, otherwise it is 0; Subject node node degree and super edge hyperboundary They are defined as follows:

[0009]

[0010] in, Represents the set of nodes consisting of all subject nodes. This represents the set of all hyperedges.

[0011] Subject node The node degree under the LASSO sparse representation strategy is denoted as... In the context of the LASSO sparse representation strategy for sparse groups, the node degree is denoted as... The node degree under the LASSO sparse representation strategy for overlapping groups is denoted as ; Node degree , and After splicing, the subject node is obtained. Structural feature vectors :

[0012] in, This indicates a vector concatenation operation.

[0013] In step S2, three perspectives of subject association hypergraphs are constructed using LASSO sparse representation strategies: sparse group LASSO sparse representation strategy and overlapping group LASSO sparse representation strategy. Specifically, when constructing an individual-level subject association hypergraph using the LASSO sparse representation strategy, the feature vector of each subject node is represented as a sparse linear combination of the feature vectors of other subject nodes. The objective function is optimized as follows:

[0014] in, Indicates the first Feature vectors of each subject node Indicates except the first The matrix consisting of the feature vectors of all other subject nodes besides the subject node. Represents a sparse coefficient vector. Represents the regularization parameter; according to The subject nodes corresponding to non-zero elements are used to generate hyperedges, constructing an individual-level subject association hypergraph. When constructing a group-level subject association hypergraph using the sparse group LASSO sparse representation strategy, the subject nodes are divided into multiple non-overlapping groups, and the optimization objective function is:

[0015] in, Represents a set of non-overlapping groups. Indicates the first The sparse coefficient subvectors corresponding to each group and Let represent the individual-level regularization parameter and the group-level regularization parameter, respectively. Hyperedges are generated based on the subject nodes corresponding to the non-zero sparse coefficients to construct a group-level subject association hypergraph. When constructing the overlapping group-level subject association hypergraph using the overlapping group LASSO sparse representation strategy, subject nodes are divided into multiple groups with overlapping relationships, allowing the same subject node to belong to multiple groups. The optimization objective function is:

[0016] in, Represent a set of groups with overlapping relationships; generate hyperedges based on the subject nodes corresponding to non-zero sparse coefficients, and construct a hypergraph of subject associations at the overlapping group level.

[0017] The propagation process of hypergraph convolution in step S3 is as follows:

[0018] in, Indicates the first The node representation matrix of the layer, Indicates the first The node representation matrix updated by the hypergraph convolution. Represents the correlation matrix. Represents the hyperedge weight matrix. Represents the node degree matrix, Represents the hypermarginality matrix. Indicates the first Layer learnable parameter matrix, This represents a non-linear activation function.

[0019] Step S3, cross-perspective supervised contrastive learning, specifically includes: The hidden representations of nodes in the same network layer under different viewpoints are input to a projection function composed of a multilayer perceptron, and mapped to the contrastive learning space to obtain projection features. For any two different viewpoints, the projection features of the subject node under one viewpoint are used as anchors, the projection features of the subject node corresponding to the anchor point with the same class label under the other viewpoint are used as positive samples, and all projection features with different class labels under the other viewpoint are used as negative samples. The cross-view supervised contrastive loss is then calculated.

[0020] in, Let V = {L, S, O} represent the network layer index, where L represents the view constructed based on the LASSO sparse representation strategy, S represents the view constructed based on the sparse group LASSO sparse representation strategy, and O represents the view constructed based on the overlapping group LASSO sparse representation strategy. and This represents two different perspectives. In cross-perspective supervised comparison, k ≠ k ; Indicates the total number of subject nodes. This indicates the subject node index corresponding to the anchor point. This represents the subject node index corresponding to the negative sample. This represents the subject node index corresponding to the positive sample. Indicates the node with the subject A set of positive sample nodes with the same category label. This represents the number of nodes in the set of positive sample nodes. Representing the subject node From the perspective Next The projection characteristics of the layer Represents the cosine similarity function. This represents the temperature parameter.

[0021] The total cross-view contrastive learning loss for each network layer is obtained by weighted summation:

[0022] in, This represents the total number of layers in the coding branches of a hypergraph neural network. Indicates the first The weighting coefficients of the layer contrastive loss. This represents the perspective constructed based on the LASSO sparse representation strategy. This represents the perspective of constructing a sparse representation based on the sparse group LASSO sparse representation strategy. This represents the perspective constructed based on the LASSO sparse representation strategy of overlapping groups.

[0023] The method for constructing node-level attention weights in step S4 includes: converting the structural feature vectors of each subject node... The attention scoring network, composed of multilayer perceptrons, is input and mapped using a normalized exponential function to obtain the attention weights of each subject node under each viewpoint:

[0024] in, Representing the subject node Attention weight vectors from three perspectives, and satisfying , Representing the subject node From the perspective Attention weights.

[0025] Step S4, adaptive weighted fusion, includes: for each subject node, multiplying the hidden node representation of the subject node in each viewpoint with the attention weight in the corresponding viewpoint and then summing the results to obtain the fused node features of the subject node.

[0026] in, Representing the subject node The characteristics of the fusion node Representing the subject node From the perspective The nodes below are hidden.

[0027] The classification loss is the weighted cross-entropy loss:

[0028] in, Indicates the first The true category label of each subject node Indicates the first The predicted probability that each subject node belongs to the Alzheimer's disease patient category. This represents the category weight coefficients; the total loss is:

[0029] in, This represents the weighting coefficient used to balance the classification loss and the total loss of hierarchical cross-view comparative learning.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention employs three different approaches—LASSO sparse representation, sparse group LASSO sparse representation, and overlapping group LASSO sparse representation—to construct subject association hypergraphs from three perspectives. These strategies characterize high-order associations among subjects at the individual, non-overlapping group, and overlapping group levels, respectively. These three mapping strategies reveal the intrinsic structure of the subject group at different granularities. Compared to a single mapping strategy, they effectively reduce information bias caused by specific mapping methods, providing more comprehensive and multi-dimensional topological information for subsequent classification.

[0031] 2. This invention performs cross-view supervised contrastive learning in each layer of the hypergraph neural network's encoding branch. Projected features of the same subject node with the same class label under different viewpoints are used as positive samples, and projected features with different class labels are used as negative samples. This forces the model to align the feature distribution of samples of the same class across different hypergraph viewpoints. Compared to existing techniques that only perform feature fusion at the output layer, this invention's hierarchical contrastive mechanism can align local structural information in shallow networks and global semantic information in deep networks, eliminating redundancy and discrepancies between multiple viewpoints layer by layer, significantly improving the consistency of cross-view features.

[0032] 3. This invention constructs structural feature vectors based on the node degree of different subject nodes under various perspectives, and then generates personalized cross-perspective attention weights for each subject node through an attention scoring network. Node degree reflects the topological connectivity strength of the subject under various hypergraph perspectives. Different subjects have naturally different topological structures, therefore the discriminative contribution of the same perspective to different subjects is also different. This invention uses dynamic weighted fusion based on node degree, which, compared with the uniform weighting or simple splicing methods in the prior art, can more accurately retain the effective feature information of each subject, suppress redundant noise, and enhance the discriminative power of the fused features.

[0033] 4. This invention jointly optimizes the classification loss and the total loss of hierarchical cross-view comparative learning, enabling the model to maintain cross-view feature consistency while improving classification discriminative power. The two losses work together to guide network parameter updates. The entire scheme forms a complete technical loop from sMRI data input to classification result output, which can be directly applied to early image-assisted screening of AD, effectively improving the accuracy and stability of AD classification. Attached Figure Description

[0034] Figure 1 This is a flowchart of a multi-view hypergraph fusion method for Alzheimer's disease classification provided in an embodiment of this application; Figure 2 This is a comparison diagram of the effects of the classification method provided in this application and traditional classification methods, as provided in an embodiment of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0036] See Figure 1 This invention provides a multi-view hypergraph fusion method for classifying Alzheimer's disease, comprising: S1. Obtain the structural magnetic resonance imaging data of the subjects and perform preprocessing. Extract the morphological features of the brain regions of each subject, construct a sample feature matrix with each subject as a node, and obtain the corresponding category label for each subject. The category label includes the category label for Alzheimer's disease patients and the category label for normal controls.

[0037] Structural magnetic resonance imaging (sMRI) data can be obtained from publicly available medical imaging databases, such as the Alzheimer's Disease Neuroimaging Initiative dataset. Participants include both Alzheimer's patients and healthy controls.

[0038] As one possible implementation, the preprocessing includes operations such as craniotomy, intensity normalization, spatial transformation, subcortical structure segmentation, and cortical surface reconstruction. For example, FreeSurfer software can be used to automate the processing of structural magnetic resonance imaging (MRI) images. FreeSurfer is an open-source brain imaging analysis software capable of surface reconstruction and subcortical structure segmentation of MRI images. After processing, the brain extraction results, subcortical segmentation results, and cortical boundary reconstruction results undergo quality control, retaining only data from subjects that pass quality control. Subsequently, morphological features of multiple brain regions of interest are extracted based on brain region atlases. These morphological features include one or more of the following: cortical thickness features, subcortical structure volume features, and gray matter volume features. These features all reflect the morphological changes in brain tissue caused by Alzheimer's disease and are important bases for subsequent classification tasks.

[0039] Suppose there are N subjects in total, and F-dimensional brain region features are extracted from each subject, then the sample feature matrix is ​​represented as:

[0040] in, Indicates the subject index; Indicates the first The feature vector of each subject; F represents the feature dimension; Let N represent the space of real matrices with N rows and F columns.

[0041] The category label set is represented as:

[0042] in, 0 represents the normal control category, and 1 represents the Alzheimer's disease patient category.

[0043] In one specific embodiment, eligible subjects were selected from the ADNI dataset, including several Alzheimer's disease patients and several normal controls. All subjects' structural magnetic resonance imaging (MRI) data underwent the same preprocessing procedure. Preprocessing was performed fully automatically using the `recon-all` command in FreeSurfer software, which integrates craniotomy, intensity normalization, Talairach spatial transformation, subcortical structure segmentation, and cortical surface reconstruction. After processing, the Desikan-Killiany brain region atlas was used to extract cortical thickness and gray matter volume for 68 cortical regions, as well as volumetric features of subcortical structures (including the hippocampus, amygdala, and caudate nucleus). The extracted features constituted a sample feature matrix, and the category label for each subject was recorded.

[0044] S2. Based on the sample feature matrix, three perspectives of subject association hypergraphs are constructed using LASSO sparse representation strategy, sparse group LASSO sparse representation strategy and overlapping group LASSO sparse representation strategy respectively. Each perspective of the subject association hypergraph contains a set of nodes and a set of hyperedges.

[0045] A subject association hypergraph is a special graph structure in which each subject is represented as a node, and each hyperedge can connect multiple subject nodes simultaneously, thereby characterizing the higher-order associations between subjects.

[0046] The subject association hypergraph is represented as follows:

[0047] in, Represents the set of nodes consisting of all subject nodes. This represents the set of all hyperedges.

[0048] Specifically, Indicates by A set of nodes consisting of subject nodes Indicates by A set of hyperedges consisting of 100 hyperedges. Each hyperedge... It is a set of nodes A non-empty subset, namely and .

[0049] The subject association hypergraph uses an association matrix. The incidence matrix is ​​defined as follows:

[0050] in, Represents the subject node, Indicates a superedge; if the subject node Belongs to superedge Then the elements of the correlation matrix The value is 1 if it is not 1, otherwise the value is 0.

[0051] Subject node node degree and super edge hyperboundary They are defined as follows:

[0052]

[0053] Among them, node degree Indicates a node containing a subject. The number of hyperedges reflects the number of associations that the subject node participates in from this perspective; hyperedge degree Indicates the superedge The number of subject nodes included reflects the number of nodes connected by the hyperedge. Node degree is an important basis for constructing subsequent node-level attention weights.

[0054] This invention employs three sparse representation strategies to construct subject association hypergraphs from three perspectives: an individual-level subject association hypergraph based on the LASSO sparse representation strategy, a group-level subject association hypergraph based on the sparse group LASSO sparse representation strategy, and an overlapping group-level subject association hypergraph based on the overlapping group LASSO sparse representation strategy. These three mapping strategies reveal the intrinsic structure of the subject population at different granularities, effectively reducing information bias caused by a single mapping method.

[0055] LASSO sparse representation strategy is a method that... A method for feature selection using norm regularization. When constructing an individual-level subject association hypergraph using the LASSO sparse representation strategy, the feature vector of each subject node is represented as a sparse linear combination of the feature vectors of other subject nodes. For the ... For each subject node, the objective function for optimization is:

[0056] in, Indicates the first Feature vectors of each subject node; Indicates except the first The matrix formed by the feature vectors of all other subject nodes besides the nth subject node, i.e., the nth subject node. The column is set to a vector of 0; Represents a sparse coefficient vector; This represents the regularization parameter that controls sparsity. express Norm, which is the square root of the sum of the squares of the elements of a vector; express Norm, which is the sum of the absolute values ​​of the elements of a vector; Indicates To optimize the variables, find the minimum value of the objective function.

[0057] After solving the above optimization objective function, if If some elements in the dataset are non-zero, then the subject node corresponding to the non-zero element will be linked to the first... All subject nodes are grouped into the same superedge. For example, if The Middle If the nth element is not zero, then the nth element is not zero. The subject and the first If two subjects have high similarity in the feature space, they should belong to the same hyperedge. By traversing all subject nodes in this way, an individual-level subject association hypergraph is constructed.

[0058] Sparse group LASSO is a method that introduces both individual-level and group-level sparsity constraints on top of LASSO. When constructing a group-level subject association hypergraph using the sparse group LASSO sparse representation strategy, subject nodes are first divided into multiple non-overlapping groups using clustering methods. The K-means clustering algorithm can be used, which iteratively optimizes the data into a predetermined number of clusters, where samples within each cluster have high similarity. The optimization objective function for simultaneously introducing individual-level and group-level sparsity constraints into the sparse linear representation model is:

[0059] in, This represents the individual-level sparsity regularization parameter, which controls the sparsity at the individual level. This represents the group-level sparsity regularization parameter, which controls the sparsity at the group level. Represents a set of non-overlapping groups; Indicates the first The sparse coefficient subvectors corresponding to each group, i.e. The middle belongs to the first The coefficient components of each group; Indicates the first The sparse coefficient subvectors corresponding to each group Norm.

[0060] After solving the above optimization objective function, hyperedges are generated based on the subject nodes corresponding to non-zero sparse coefficients to construct a group-level subject association hypergraph. This perspective can simultaneously capture individual-level sparse relationships and group-level structural relationships: group-level sparse constraints. It tends to select or discard features in whole groups to preserve group structure information; individual-level sparsity constraints. This ensures that sparsity is maintained within the group.

[0061] Overlapping group LASSO is an extension of sparse group LASSO, allowing overlapping relationships between different groups, meaning the same subject node can belong to multiple groups simultaneously. Considering that a subject may belong to multiple potential structural groups, this invention further employs overlapping group LASSO to construct an overlapping group-level subject association hypergraph. Based on the clustering results, overlapping subject groups are constructed, enabling the same subject node to belong to multiple groups. The sparse linear representation model is then solved under the constraints of the overlapping group structure, with the optimization objective function being:

[0062] in, The symbols represent sets of groups with overlapping relationships, and the meanings of the remaining symbols are the same as in sparse group LASSO. After solving the above optimization objective function, hyperedges are generated based on the subject nodes corresponding to non-zero sparse coefficients to construct an overlapping group-level subject association hypergraph. This perspective can model situations where subjects share activation across multiple groups. For example, a subject may simultaneously possess features of multiple cognitive subtypes. This perspective can capture such multiple attribution relationships and enhance the expressive power of the hypergraph structure.

[0063] It should be noted that the regularization parameters of the above three sparse representation strategies... , and Cross-validation can be used to select the regularization parameter value that balances model sparsity and fitting accuracy. In a specific embodiment, five-fold cross-validation is used to select the regularization parameter value that optimizes the performance of the validation set.

[0064] S3. Input the subject association hypergraphs from the three perspectives into three hypergraph neural network encoding branches with the same structure and independent parameters. Each hypergraph neural network encoding branch aggregates the information of multiple subject nodes within each hyperedge through hypergraph convolution to extract the node hidden representations from each perspective. In each network layer of the hypergraph neural network encoding branch, perform cross-view supervised contrastive learning on the node hidden representations from different perspectives to obtain cross-view aligned node hidden representations from each perspective.

[0065] Each hypergraph neural network encoding branch corresponds to a hypergraph viewpoint, with identical structures and independent parameters across branches. Identical structure means each branch has the same number of network layers, the same hidden layer dimensions, and other architectural designs; parameter independence means that the learnable parameters (such as weight matrices and bias terms) of each branch are not shared and are updated independently. This parameter independence design allows each encoding branch to learn its own unique feature representation for its respective viewpoint's hypergraph structure, avoiding interference between features from different viewpoints. Within each encoding branch, node features are updated layer by layer through multiple hypergraph convolutions.

[0066] The core idea of ​​hypergraph convolution is the "node-hyperedge-node" information propagation method: first, the node features are aggregated onto the hyperedge to which they belong, and then the aggregated features on the hyperedge are propagated back to each node, thereby realizing the aggregation of information from multiple nodes within the hyperedge. Unlike ordinary graph convolution, which can only aggregate information from adjacent nodes, hypergraph convolution can simultaneously aggregate information from all nodes connected by a hyperedge, thus making it better at depicting higher-order topological relationships.

[0067] In each layer of the hypergraph neural network's encoding branch, cross-view supervised contrastive learning is performed on the node hidden representations from different perspectives. Specifically, the node hidden representations of the same subject node in the same network layer under different perspectives are mapped to the contrastive learning space. Projected features of the same subject node with the same class label under different perspectives are used as positive sample pairs, and projected features with different class labels are used as negative sample pairs. The contrastive loss function is used to shorten the distance between positive sample pairs and widen the distance between negative sample pairs. This mechanism can simultaneously perform cross-view feature alignment in each network layer, eliminating redundancy and discrepancies between multiple perspectives layer by layer.

[0068] For example, in each coding branch, the first The propagation process of hypergraph convolution is represented as:

[0069] in, Indicates the first The node representation matrix of the layer, with each row corresponding to the feature vector of a subject node; Indicates the first The first layer obtained after hypergraph convolution update Layer nodes represent matrices; H represents the correlation matrix, with dimension 1. , For the number of nodes, is the number of hyperedges; represents the hyperedge weight matrix. It is a diagonal matrix, with the diagonal elements being the weights of each hyperedge. It can be set to the identity matrix by default. The node degree matrix is ​​a diagonal matrix, with diagonal elements representing the node degree of each subject node. ; The hyperedge degree matrix is ​​a diagonal matrix, with diagonal elements representing the hyperedge degree of each hyperedge. ; Indicates the first The layer learnable parameter matrix consists of the weight parameters that the model needs to learn during training. To represent a nonlinear activation function, a modified linear unit can be used, i.e. ; Represents the node degree matrix The negative one-half power, that is, for The diagonal matrix formed by taking the negative 1 / 2 power of the diagonal elements.

[0070] In the above propagation formula, This is the Laplacian normalization term for the hypergraph, which performs normalization aggregation of node features based on the hypergraph topology. For linear transformation, the node features are changed from the first... The latent space of layer 1 is mapped to the first layer 2. The implicit space of the layer. The overall calculation process can be understood as: first through... The node features are aggregated to the superedge, and then... The hyperedge features are backpropagated to the nodes, and after normalization and nonlinear transformation, the updated node representation is obtained.

[0071] In one specific embodiment, the number of layers in the hypergraph neural network coding branch Set to 2 layers, with hidden layer dimensions set to 64. The first hypergraph convolution layer will take the input features (dimensions...) The first layer maps the latent features to a 64-dimensional latent space, and the second layer of hypergraph convolution maps the 64-dimensional latent space features to a 128-dimensional feature space. Each hypergraph convolution is followed by a ReLU activation function and a dropout regularization layer with a dropout rate of 0.5 to prevent overfitting.

[0072] The hidden representations of nodes in the same network layer from different perspectives are represented by projection functions composed of multilayer perceptrons, and then mapped to the contrastive learning space to obtain projection features:

[0073] in, Indicates the subject node index; Indicates the hypergraph neural network layer index; Indicates the hypergraph perspective. This represents the perspective constructed based on the LASSO sparse representation strategy. This represents the perspective of constructing a sparse representation based on the sparse group LASSO sparse representation strategy. This represents the perspective constructed based on the LASSO sparse representation strategy of overlapping groups; Representing the subject node From the perspective Next The node representation vector of the layer, i.e. The first in the matrix The vector corresponding to the row; Indicates perspective Next The node representation matrix of the layer; express The projected feature vector obtained after mapping by the projection function; Represents the projection function composed of a multilayer sensing mechanism. Projection function It consists of two fully connected layers. The first layer maps the node representations to 128 dimensions, and the second layer maps the 128-dimensional features to a 64-dimensional contrastive learning space. The ReLU activation function is used between the two layers.

[0074] For any two different hypergraph perspectives and ( and ), its first The cross-layer perspective supervision contrast loss is expressed as:

[0075] in, Let V represent the total number of subject nodes; let the view set be V={L,S,O}, where: L represents the view constructed based on the LASSO sparse representation strategy, S represents the view constructed based on the sparse group LASSO sparse representation strategy, and O represents the view constructed based on the overlapping group LASSO sparse representation strategy. and This represents two different perspectives. In cross-perspective supervised comparison, k ≠ k This represents the index of the subject node corresponding to the anchor point, i.e., the subject node. From the perspective Projection features As an anchor point; N represents the number of subjects; This represents the subject node index corresponding to the negative sample; This represents the subject node index corresponding to the positive sample; Indicates the node with the subject A set of positive sample nodes that have the same category label and do not contain themselves; This represents the number of nodes in the set of positive sample nodes; Representation and Hypergraph Perspective A different hypergraph perspective; Represents the cosine similarity function. This is used to measure the similarity between two vectors in a direction; the larger the value, the more similar the two vectors are. Represented by natural constant An exponential function with base 0; Represented by natural constant A logarithmic function with base 0; This represents a temperature parameter used to control the sharpness of contrast loss, and its value typically ranges from 0.1 to 1.0.

[0076] The physical meaning of the above loss function is: for anchor points The molecular calculation is compared with the positive sample. The similarity index is calculated by dividing the denominator by the similarity index with all negative samples. ( The sum of similarity indices of ( ). By minimizing this loss, the model is encouraged to bring anchor points closer to positive samples and further away from negative samples.

[0077] The total cross-view contrastive learning loss for each network layer is obtained by weighted summation:

[0078] in, This represents the total number of layers in the coding branches of the hypergraph neural network; Indicates the first The weighting coefficients of the layer contrastive loss are used to control the contribution of each layer contrastive loss to the total loss; Indicates the first Layer perspective and perspective Cross-perspective supervisory contrast loss; Indicates the first Layer perspective and perspective Cross-perspective supervisory contrast loss; Indicates the first Layer perspective and perspective Cross-perspective monitoring and comparison loss.

[0079] The hierarchical contrastive loss design enables the model to align local structural information in shallow networks and global semantic information in deep networks, eliminating redundancy and discrepancies between multiple perspectives layer by layer. In a specific embodiment, if Then shallow weights and deep weights It can be set to different values, for example , This is to emphasize the importance of deep feature alignment.

[0080] Through cross-view supervised contrastive learning, the projected features of the same subject node under different viewpoints are brought closer together, while the projected features of different classes are pushed apart. This mechanism effectively enhances the intra-class compactness of samples of the same class under different hypergraph viewpoints, while expanding the inter-class separability between samples of different classes, thereby improving the classifier's discriminative ability. Compared with existing technologies that only perform feature fusion at the output layer, the hierarchical contrastive mechanism of this invention can perform cross-view alignment at each network layer, avoiding the problem of shallow feature inconsistencies being amplified and accumulated in deeper layers.

[0081] S4. Construct node-level attention weights based on the node degree of each subject node in each viewpoint. The node degree is the number of hyperedges associated with that subject node. Use the node-level attention weights to adaptively weight and fuse the hidden representations of nodes in each viewpoint after cross-viewpoint alignment to obtain fused node features. Input the fused node features into a classifier consisting of a fully connected layer and a normalized output layer to output the classification results of each subject belonging to the Alzheimer's disease patient category or the normal control category. Calculate the classification loss based on the classification results and category labels. Sum the classification loss with the contrastive loss generated by cross-viewpoint supervised contrastive learning to obtain the total loss. Optimize the network parameters through backpropagation to obtain the trained Alzheimer's disease classification model.

[0082] Node degree reflects the number of associations a subject's node participates in from a given perspective. A higher node degree indicates that the subject is associated with more other subjects and occupies a more central position in the topology from that perspective; a lower node degree indicates sparser associations. Since the three mapping strategies (individual-level, group-level, and overlapping group-level) characterize subject associations at different granularities, the node degree of the same subject may differ significantly across the three perspectives. For example, a subject may only be similar to a few individuals (low node degree) from an individual-level perspective, but may belong to a larger group (high node degree) from a group-level perspective. Therefore, the node degree distribution can reflect the subject's topological importance across different perspectives and is an ideal basis for constructing personalized attention weights.

[0083] This application concatenates the node degrees of each subject node in three perspectives into a structural feature vector, and then inputs the structural feature vector into an attention scoring network composed of a multilayer perceptron. After mapping by a normalized exponential function (Softmax function), the attention weights in each perspective are obtained. Different subjects receive different weight assignments due to differences in their topological structures.

[0084] We use attention weights to perform a weighted summation of the hidden representations of nodes from different perspectives to obtain fused node features. These fused features retain the effective information from each perspective while suppressing interference from redundant perspectives through personalized weights.

[0085] The fused node features are input into the classifier. The classifier consists of a fully connected layer and a normalized output layer. The fully connected layer maps the fused node features to a binary classification score space, and the normalized output layer (such as a Softmax layer or a Sigmoid layer) converts the scores into predicted probabilities of belonging to each category. The category with the higher predicted probability is taken as the final classification result.

[0086] During the model training phase, the classification loss is calculated based on the classification results and class labels. This application uses weighted cross-entropy loss as the classification loss to alleviate the class imbalance problem between Alzheimer's disease patient samples and normal control samples. The classification loss is weighted and summed with the total loss of hierarchical cross-view comparative learning to obtain the total loss. All learnable parameters of the model are jointly optimized through backpropagation until the model converges, resulting in a trained Alzheimer's disease classification model.

[0087] For example, subject node The node degree under the LASSO sparse representation strategy is denoted as... In the context of the LASSO sparse representation strategy for sparse groups, the node degree is denoted as... The node degree under the LASSO sparse representation strategy for overlapping groups is denoted as The specific calculation method for the degree of each node is as follows:

[0088] in, Indicates the view index. ; Indicates the hypergraph perspective The correlation matrix below; Indicates the hypergraph perspective The number of superedges; Indicates perspective Next subject node With super-edge The values ​​of the elements in the correlation matrix.

[0089] node degree , and After splicing, the subject node is obtained. Structural feature vectors :

[0090] in, This represents the concatenation operation, which concatenates three scalar values ​​into a single three-dimensional vector. .

[0091] The structural feature vectors of each subject node The input is an attention scoring network composed of multilayer perceptrons. After mapping with a normalized exponential function (Softmax function), the attention weights of each subject node under each viewpoint are obtained:

[0092] in, Representing the subject node Attention weight vectors from three perspectives ; This represents an attention scoring network composed of multilayer perceptrons, with a three-dimensional structural feature vector as input. The output is a three-dimensional unnormalized attention score; This represents the normalized exponential function, for the input vector. , The unnormalized scores are converted into a normalized probability distribution.

[0093] Attention weight vector satisfy:

[0094] in, Representing the subject node From the perspective Attention weights.

[0095] For each subject node, the node hidden representation of the subject node under each viewpoint is multiplied by the attention weight under the corresponding viewpoint and then summed to obtain the fused node features of the subject node:

[0096] in, Representing the subject node Features of fusion nodes; Representing the subject node From the perspective The hidden representation of the nodes below is the feature vector after cross-view comparative learning and alignment. Representing the subject node From the perspective Attention weights.

[0097] This fusion approach features subject-specific personalization: the weights of the same subject across different perspectives are determined by their own topological structure (node ​​degree distribution); even within the same perspective, different subjects may have different weights. For example, for a subject with a high node degree at the individual level and a low node degree at the group level, the attention scoring network may assign them a higher weight. Viewpoint weight and lower Viewpoint weights are assigned, and for another subject with a different topological distribution, the weight assignment is adjusted accordingly. This personalized fusion mechanism can more accurately retain the effective feature information of each subject and suppress redundant noise.

[0098] Merge node features Input classifier. The classifier consists of a fully connected layer and a normalized output layer. The fully connected layer maps the fused node features to a two-dimensional score space, and the normalized output layer uses a softmax function to convert the two-dimensional scores into predicted probabilities of belonging to the Alzheimer's patient category and the normal control category. For the ... Each subject node has a classifier outputting a predicted probability. This represents the predicted probability that the subject belongs to the Alzheimer's disease patient category. When When, the subject is classified as an Alzheimer's patient; when At that time, the subject was classified as a normal control.

[0099] For example, during the model training phase, weighted cross-entropy loss is used as the classification loss to alleviate the class imbalance problem between Alzheimer's patient samples and normal control samples:

[0100] in, Indicates the total number of samples; Indicates the first The true category label of each subject node ; Indicates the first The predicted probability that each subject node belongs to the Alzheimer's disease patient category. ; Represents the category weight coefficient. , Indicate category The number of samples. When the number of positive and negative samples is unbalanced, minority class samples receive greater weight, forcing the model to pay more attention to the classification accuracy of the minority class.

[0101] The total loss function is obtained by weighted summation of the classification loss and the total loss of hierarchical cross-view comparison learning:

[0102] in, Represents the total loss function; This represents the classification loss (weighted cross-entropy loss). This represents the total learning loss across hierarchical cross-perspective comparisons; This represents the weighting coefficient used to balance the classification loss and the total loss of hierarchical cross-view comparative learning, and is used to control the relative importance of the two losses.

[0103] For example, the value of is in the range of 0.1 to 1.0, and the optimal value can be selected through cross-validation. If If the size is too small, the effect of comparative learning is weakened; if... If the value is too large, the dominance of the classification task is weakened. It is usually set to... Use this as an initial value, and then adjust it based on the performance of the validation set.

[0104] The model's learnable parameters are jointly optimized using the backpropagation algorithm. Specifically, the total loss is calculated. For the gradient of each learnable parameter, the parameter values ​​are updated using a gradient descent optimizer such as the Adam optimizer. In a specific embodiment, the Adam optimizer is used, with an initial learning rate set to... The weight decay is set to The training epochs were set to 200, and the batch size was set to 32. Training was stopped early to prevent overfitting when the validation set loss no longer decreased after 20 consecutive training epochs.

[0105] After training, a trained Alzheimer's disease classification model is obtained. This model can preprocess and extract features from the structural magnetic resonance imaging data of the test subjects, and then perform hypergraph construction, hypergraph neural network encoding, cross-view contrastive learning, and node-level attention fusion according to the same process, ultimately outputting the classification result of whether the test subject belongs to the Alzheimer's disease patient category or the normal control category.

[0106] For example, refer to Figure 2 , Figure 2 A comparison of the effectiveness of the classification method provided in this application with traditional classification methods. Figure 2 The horizontal and vertical axes in the figure represent specificity and sensitivity, respectively. Figure 2 The two curves in the diagram represent the traditional method and the method provided in this application, respectively. According to... Figure 2 It can be seen that the method provided in this application outperforms traditional methods in terms of classification accuracy AUC, and achieves a balanced and excellent performance in both sensitivity and specificity.

[0107] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0108] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-view hypergraph fusion method for Alzheimer's disease classification, characterized in that, include: S1. Acquire and preprocess the structural magnetic resonance imaging data of the subjects, extract the morphological features of brain regions of each subject, construct a sample feature matrix with each subject as a node, and obtain the corresponding category label for each subject. The category label includes the category label for Alzheimer's disease patients and the category label for normal controls. S2. Based on the sample feature matrix, construct subject association hypergraphs from three perspectives using LASSO sparse representation strategy, sparse group LASSO sparse representation strategy, and overlapping group LASSO sparse representation strategy, respectively. The subject association hypergraph from each perspective contains a set of nodes and a set of hyperedges. S3. Input the subject association hypergraphs from the three perspectives into three structurally identical and parameter-independent hypergraph neural network encoding branches. Each hypergraph neural network encoding branch aggregates the information of multiple subject nodes within each hyperedge through hypergraph convolution to extract the node hidden representations from each perspective. In each network layer of the hypergraph neural network coding branch, cross-view supervised contrastive learning is performed on the node hidden representations under different viewpoints to obtain cross-view aligned node hidden representations for each viewpoint. S4. Construct node-level attention weights based on the node degree of each subject node under each viewpoint, where the node degree is the number of hyperedges associated with the subject node. The node-level attention weights are used to adaptively weight and fuse the hidden representations of nodes in each viewpoint after cross-viewpoint alignment to obtain fused node features. The fusion node features are input into a classifier consisting of a fully connected layer and a normalized output layer, which outputs a classification result indicating whether each subject belongs to the Alzheimer's disease patient category or the normal control category. The classification loss is calculated based on the classification result and the category label. The classification loss is weighted and summed with the contrastive loss generated by the cross-view supervised contrastive learning to obtain the total loss. The network parameters are optimized through backpropagation to obtain the trained Alzheimer's disease classification model.

2. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 1, characterized in that, The subject association hypergraph is represented by an association matrix, which is defined as follows: in, Represents the subject node, Indicates a superedge; if the subject node Belongs to superedge Then the elements of the correlation matrix The value is 1 if it is not 1, otherwise it is 0; Subject node node degree and super edge hyperboundary They are defined as follows: in, Represents the set of nodes consisting of all subject nodes. This represents the set of all hyperedges.

3. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 2, characterized in that, Subject node The node degree under the LASSO sparse representation strategy is denoted as... In the context of the LASSO sparse representation strategy for sparse groups, the node degree is denoted as... The node degree under the LASSO sparse representation strategy for overlapping groups is denoted as ; the node degree , and After splicing, the subject node is obtained. Structural feature vectors : in, This indicates a vector concatenation operation.

4. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 1, characterized in that, In step S2, three perspectives of subject association hypergraphs are constructed using LASSO sparse representation strategies: sparse group LASSO sparse representation strategy and overlapping group LASSO sparse representation strategy. Specifically, when constructing an individual-level subject association hypergraph using the LASSO sparse representation strategy, the feature vector of each subject node is represented as a sparse linear combination of the feature vectors of other subject nodes. The objective function is optimized as follows: in, Indicates the first Feature vectors of each subject node Indicates except the first The matrix consisting of the feature vectors of all other subject nodes besides the subject node. Represents a sparse coefficient vector. Represents the regularization parameter; according to The subject nodes corresponding to non-zero elements are used to generate hyperedges, constructing an individual-level subject association hypergraph. When constructing a group-level subject association hypergraph using the sparse group LASSO sparse representation strategy, the subject nodes are divided into multiple non-overlapping groups, and the optimization objective function is: in, Represents a set of non-overlapping groups. Indicates the first The sparse coefficient subvectors corresponding to each group and Let represent the individual-level regularization parameter and the group-level regularization parameter, respectively. Hyperedges are generated based on the subject nodes corresponding to the non-zero sparse coefficients to construct a group-level subject association hypergraph. When constructing the overlapping group-level subject association hypergraph using the overlapping group LASSO sparse representation strategy, subject nodes are divided into multiple groups with overlapping relationships, allowing the same subject node to belong to multiple groups. The optimization objective function is: in, Represent a set of groups with overlapping relationships; generate hyperedges based on the subject nodes corresponding to non-zero sparse coefficients, and construct a hypergraph of subject associations at the overlapping group level.

5. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 1, characterized in that, The propagation process of the hypergraph convolution described in step S3 is as follows: in, Indicates the first The node representation matrix of the layer, Indicates the first The node representation matrix updated by the hypergraph convolution. Represents the correlation matrix. Represents the hyperedge weight matrix. Represents the node degree matrix, Represents the hypermarginality matrix. Indicates the first Layer learnable parameter matrix, This represents a non-linear activation function.

6. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 1, characterized in that, The cross-view supervised contrastive learning described in step S3 specifically includes: The hidden representations of nodes in the same network layer under different viewpoints are input to a projection function composed of a multilayer perceptron, and mapped to the contrastive learning space to obtain projection features. For any two different viewpoints, the projection features of the subject node under one viewpoint are used as anchors, the projection features of the subject node corresponding to the anchor point with the same class label under the other viewpoint are used as positive samples, and all projection features with different class labels under the other viewpoint are used as negative samples. The cross-view supervised contrastive loss is then calculated. in, Let V = {L, S, O} represent the network layer index, where: L represents the view constructed based on the LASSO sparse representation strategy, S represents the view constructed based on the sparse group LASSO sparse representation strategy, and O represents the view constructed based on the overlapping group LASSO sparse representation strategy. and This represents two different perspectives. In cross-perspective supervised comparison, k ≠ k Indicates the total number of subject nodes. This indicates the index of the subject node corresponding to the anchor point. This represents the subject node index corresponding to the negative sample. This represents the subject node index corresponding to the positive sample. Indicates the node with the subject A set of positive sample nodes with the same category label. This represents the number of nodes in the set of positive sample nodes. Representing the subject node From the perspective Next The projection characteristics of the layer Represents the cosine similarity function. This represents the temperature parameter.

7. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 6, characterized in that, The weighted summation of the cross-view supervised contrastive losses of each network layer yields the total hierarchical cross-view contrastive learning loss: in, This represents the total number of layers in the coding branches of a hypergraph neural network. Indicates the first The weighting coefficients of the layer contrastive loss. This represents the perspective constructed based on the LASSO sparse representation strategy. This represents the perspective of constructing a sparse representation based on the sparse group LASSO sparse representation strategy. This represents the perspective constructed based on the LASSO sparse representation strategy of overlapping groups.

8. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 3, characterized in that, The method for constructing node-level attention weights in step S4 includes: converting the structural feature vectors of each subject node... The attention scoring network, composed of multilayer perceptrons, is input and mapped using a normalized exponential function to obtain the attention weights of each subject node under each viewpoint: in, Representing the subject node Attention weight vectors from three perspectives, and satisfying , Representing the subject node From the perspective Attention weights.

9. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 8, characterized in that, Step S4, adaptive weighted fusion, includes: for each subject node, multiplying the hidden node representation of the subject node in each viewpoint with the attention weight in the corresponding viewpoint and then summing the results to obtain the fused node features of the subject node. in, Representing the subject node The characteristics of the fusion node Representing the subject node From the perspective The nodes below are hidden.

10. The Alzheimer's disease classification method based on multi-view hypergraph fusion according to claim 1, characterized in that, The classification loss is a weighted cross-entropy loss: in, Indicates the first The true category label of each subject node Indicates the first The predicted probability that each subject node belongs to the Alzheimer's disease patient category. Represents the category weight coefficients; the total loss is: in, This represents the weighting coefficient used to balance the classification loss and the total loss of hierarchical cross-view comparative learning.

Images