Alzheimer disease classification method and system based on topology perception and group hypergraph

Abstract

The invention belongs to the related technical field of brain image processing, and provides an Alzheimer's disease classification method and system based on topology perception and a group hypergraph in order to solve the problem of inaccurate classification of the Alzheimer's disease in the prior art. Constructing a dynamic function connection network sequence through a sliding window strategy; a local topology perception encoder and a global topology perception encoder are respectively used for extracting local topology features and global topology features of each time window, deep interaction and fusion are carried out, and comprehensive feature representation of a tested level is generated; according to the method, each subject is used as a hypergraph node, hyperedges are constructed on the basis of comprehensive feature representation of a subject level and by combining feature similarity calculated by diffusion tensor imaging features and clinical embedded features of the subject, then a group hypergraph is constructed, a classification result is obtained by using a hypergraph neural network, and the early classification diagnosis accuracy of the Alzheimer's disease is effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field related to brain image processing, and particularly relates to an Alzheimer's disease classification method and system based on topology perception and group hypergraph. BACKGROUND

[0002] The statements in this section merely provide background information related to the application and do not necessarily constitute prior art.

[0003] Alzheimer's disease (AD) is a typical neurodegenerative disease, mainly manifested as memory loss, cognitive decline and other symptoms, especially in the elderly population with a high incidence. AD diagnosis is crucial as it can provide timely treatment intervention and quality of life improvement for patients. However, existing AD diagnosis methods still have certain limitations.

[0004] Traditional AD diagnosis methods mainly rely on physician's clinical symptom observation, cognitive testing and imaging examination (such as MRI, CT scan, etc.). Although these methods can assist in clinical diagnosis to some extent, they often cannot capture subtle changes in disease progression, and the accuracy and timeliness of diagnosis are limited. In addition, traditional imaging methods are mostly based on static image analysis, which cannot effectively utilize the time series information in pathological data. In recent years, diagnosis methods based on time series data have gradually become a direction to solve these problems. Time series data can reflect the continuity of physiological and pathological changes in patients, capturing subtle fluctuations in disease progression. This provides more opportunities for early diagnosis of Alzheimer's disease, especially with the help of biomedical signal processing and artificial intelligence technology, AD detection methods based on time series data have shown significant advantages.

[0005] However, current AD detection methods based on time series data still face the following challenges: 1. Data fusion problem: existing methods usually only extract features from a single perspective, ignoring the comprehensive use of multi-modal data, resulting in limitations in diagnostic performance. Although some methods use multi-source data, they lack effective feature fusion techniques, failing to fully utilize the complementary advantages of different data perspectives.

[0006] 2. Most methods are based on static graph structure learning, which cannot fully utilize the dynamic information in time series data, thus ignoring the features and relationships that change over time.

[0007] 3. Existing feature fusion methods often ignore the dynamic changes of time series information, lack effective fusion mechanisms based on time series information, and are difficult to fully capture the time dimension features in multi-source data. SUMMARY

[0008] In order to overcome the above-mentioned deficiencies of the prior art, the present application provides an Alzheimer's disease classification method and system based on topology perception and group hypergraph, which effectively improves the early diagnosis accuracy of Alzheimer's disease, has strong robustness and generalization ability, and is suitable for disease auxiliary detection and identification in intelligent medical treatment.

[0009] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides an Alzheimer's disease classification method based on topology perception and group hypergraph, comprising: extracting a time sequence from resting-state functional magnetic resonance imaging data, and constructing a dynamic functional connection network sequence through a sliding window strategy; extracting local topology features and global topology features of each time window by using a local topology perception encoder and a global topology perception encoder respectively, and generating a comprehensive feature representation of a subject level by deeply interacting and fusing the local topology perception representation and the global topology perception representation of each time window; taking each subject as a hypergraph node, constructing a hyperedge and a probabilistic hypergraph correlation matrix based on the comprehensive feature representation of the subject level and the feature similarity calculated by combining the diffusion tensor imaging features and the clinical information features of the subject, and then constructing a group hypergraph, processing the group hypergraph by using a hypergraph neural network, and obtaining an Alzheimer's disease classification result of the subject.

[0010] In a second aspect, the present application provides an Alzheimer's disease classification system based on topology perception and group hypergraph, comprising: a processing module configured to extract a time sequence from resting-state functional magnetic resonance imaging data, and construct a dynamic functional connection network sequence through a sliding window strategy; an extraction module configured to extract local topology features and global topology features of each time window by using a local topology perception encoder and a global topology perception encoder respectively, and generate a comprehensive feature representation of a subject level by deeply interacting and fusing the local topology perception representation and the global topology perception representation of each time window; a construction and classification module configured to take each subject as a hypergraph node, construct a hyperedge and a probabilistic hypergraph correlation matrix based on the comprehensive feature representation of the subject level and the feature similarity calculated by combining the diffusion tensor imaging features and the clinical information features of the subject, and then construct a group hypergraph, process the group hypergraph by using a hypergraph neural network, and obtain an Alzheimer's disease classification result of the subject.

[0011] In a third aspect, the present application provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein when the computer instructions are run by the processor, the method of the first aspect is completed.

[0012] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, which, when executed by a processor, complete the method of the first aspect.

[0013] The above one or more technical solutions have the following beneficial effects: In the present application, the dual-branch topological coding architecture of local topological perception encoder and global topological perception encoder can simultaneously capture the local detail features and global structure patterns of brain function network, realize comprehensive characterization of brain network topological characteristics; based on group hypergraph learning, the high-order similarity relationship between subjects is mined, the complementary information at the group level is fully utilized, and the classification accuracy of Alzheimer's disease is significantly improved, thereby providing strong technical support for early diagnosis and identification of Alzheimer's disease.

[0014] In the present application, the local topological perception encoder extracts local topological features from the dynamic graph sequence to generate local topological perception representation; the global topological perception encoder integrates local structure modeling and global long-range dependence capture through the cooperative mechanism of graph isomorphism network (GIN) and Transformer encoder to generate global topological perception features; and the effective integration of multi-scale information is realized through the cross-attention mechanism, thereby significantly improving the expression ability of the graph learning model.

[0015] In the present application, the multi-modal fusion strategy can be compatible with image data of different medical centers, and the organic integration of clinical information not only enriches the feature representation dimension, but also provides intuitive clinical explanation basis for the diagnosis result, which is helpful for constructing an interpretable auxiliary diagnosis system.

[0016] In the present application, the dynamic time window analysis and time attention pooling mechanism can effectively capture the dynamic characteristics of the evolution of brain function connection with time, and significantly improve the perception sensitivity of the model to pathological changes.

[0017] The advantages of the additional aspects of the present application will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0018] The drawings accompanying the specification of the present application form a part thereof, serve to provide further understanding of the present application, and together with the description of the exemplary embodiments of the present application and their description serve to explain the present application, and do not constitute improper limitations on the present application.

[0019] Figure 1 The whole algorithm flowchart in the embodiment one of the present application is shown in the following figure: Figure 2 The topological perception feature extraction flowchart in the embodiment one of the present application is shown in the following figure: Figure 3This is a flowchart detailing the implementation of the local topology-aware encoder and the global topology-aware encoder in Embodiment 1 of the present invention. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0023] Example 1 First, the relevant terms involved in this embodiment will be explained: Hyperedges can connect multiple nodes. Unlike traditional adjacency matrices, which can only reflect the relationship between pairs of nodes, they can capture higher-order relationships between multiple nodes. Note that this is based on node classification, where each node represents a certain category. The goal is to find the correlation between nodes and connect points of the same category.

[0024] Graph structure: ;in Represents the set of vertices. This represents a set of hyperedges. Unlike the adjacency matrix in a traditional graph neural network where each edge represents only the relationship between two nodes, here each hyperedge connects two or more vertices. The diagonal matrix represents the weight of each hyperedge. Each node represents an individual, and a hyperedge can connect multiple individuals, with patients having similar symptoms being connected to the same hyperedge.

[0025] Independence Matrix: Hypergraph uses the incidence matrix It indicates that the size is ,in The corresponding element is represented as: .

[0026] Degree: Vertex degree Defined as: ,in, It is a super-edge The weight represents the strength or importance of the hyperedge.

[0027] Hypergraph Laplacian Operator: Given a hypergraph ,in, It is a vertex set. is a superedge set, is a supergraph Laplacian, which is an n*n semi-positive definite matrix, where n is the number of vertices.

[0028] The eigen-decomposition of can be obtained as:

[0029] where, is a matrix containing orthogonal eigenvectors; is a diagonal matrix containing corresponding non-negative eigenvalues, and the superscript T denotes transpose.

[0030] The Fourier transform corresponding to a signal x on a supergraph is defined as: Here the eigenvectors are considered as Fourier basis. Spectral convolution: a signal is spectrally convolved with a filter as:

[0031] where, denotes element-wise Hadamard product, is a function of Fourier coefficients.

[0032] Since the computational complexity of forward and inverse Fourier transform is , the Chebyshev polynomials are used to parameterize to reduce the computational complexity.

[0033] The filter can be approximated by order Chebyshev polynomials as:

[0034] where, is a Chebyshev polynomial, and the recursive calculation formula is:

[0035] and .

[0036] The scaled Laplacian is used in the Chebyshev expansion:

[0037] where, is the largest eigenvalue of the Laplacian, is the identity matrix.

[0038] ​Finally, the convolution operation can be simplified as:

[0039] wherein, is an association matrix, and is a degree matrix, is a possible weight matrix.

[0040] The embodiment discloses an Alzheimer's disease classification method based on topology perception and group hypergraph, comprising: extracting a time sequence from resting-state functional magnetic resonance imaging data, and constructing a dynamic functional connection network sequence through a sliding window strategy; extracting local topology features and global topology features of each time window by using a local topology perception encoder and a global topology perception encoder respectively, and performing deep interaction and fusion on the local topology features and the global topology features of each time window to generate a comprehensive feature representation of a subject; taking each subject as a hypergraph node, constructing a hyperedge based on the comprehensive feature representation of the subject, and combining feature similarity calculated based on diffusion tensor imaging features and clinical information features of the subject, and further constructing a group hypergraph, processing the group hypergraph by using a hypergraph neural network to obtain an Alzheimer's disease classification result of the subject.

[0041] The embodiment can extract multi-dimensional topology features from resting-state functional magnetic resonance imaging data and generate more comprehensive brain network representation by combining a local topology perception encoder (LTAE) and a global topology perception encoder (GTAE), and a dynamic functional connection network sequence generated by a dynamic time window.

[0042] In a specific implementation, the local topology-aware encoder captures the fine-grained local topology patterns of brain networks through dynamic correlation graph statistics and structural embedding convolution, while the global topology-aware encoder combines graph isomorphism networks and Transformer architecture, simultaneously capturing global long-range dependencies and local neighborhood structures, thereby providing more rich multi-scale topology information for brain region node embedding. For DTI modality data, the method extracts multi-level white matter network features including statistical features, graph theory features, and texture features, which are complementary to fMRI functional topology features. At the same time, the method effectively models the high-order similarity relationship between subjects through group hypergraph learning. The introduction of hypergraph structure can better integrate multi-modal topology features and capture the complex pathological patterns of brain function and structure networks. In addition, through the time attention pooling mechanism, the model can adaptively weight the importance of different time windows, thereby improving the perception of key changes in dynamic functional connectivity and further improving the accuracy and robustness of classification.

[0043] In summary, the method proposed in this embodiment effectively utilizes the complementary advantages of multi-modal brain image data, and overcomes the limitations of traditional single-modality methods in feature representation integrity and pathological information coverage through joint topology coding of fMRI functional networks and DTI structural networks and group hypergraph structure learning. The method provides strong technical support for early diagnosis and identification of Alzheimer's disease.

[0044] The following will be combined Figures 1-3 The method for classifying Alzheimer's disease based on multi-view spatio-temporal topology fusion proposed in this embodiment will be described in detail, including the following parts: I. Multi-modal feature extraction.

[0045] Step 1, time series topology-aware dynamic graph representation and cross-attention fusion module.

[0046] As Figure 1 shown, the steps implemented in this embodiment include: Step 11: division of dynamic time windows.

[0047] First, for 4D fMRI (3D voxels containing time variation), the brain at each measurement time is visualized as a set of discrete brain functional regions. The BOLD signal within each region, i.e. blood oxygen level dependent signal, is represented by the average of all voxel intensities within the region, forming a time series vector. In this way, a set of BOLD signals is obtained for the entire brain, with each sequence corresponding to a specific ROI (region of interest). The BOLD signal is defined as where , is the number of time points.

[0048] In technical terms, a sliding window method is adopted to capture the temporal dynamics in the BOLD signal. This method involves sliding a fixed length window over the time series with a window length of and a step size of , resulting in a series of possibly overlapping time segments, and the BOLD signal can be divided into segments, each with a window length of and a step size of . denotes the BOLD signal of the i-th ROI in the t-th time segment.

[0049] Step 12: Constructing the dynamic graph structure.

[0050] The acquired time series data is represented as a graph model , where the node set corresponds to the pre-defined regions of interest (ROIs) in the brain, and the edge set represents the functional connections between these regions. Each node is associated with a feature vector , where the feature vector generally represents the characteristics of the time region in the time series data, here referring to the correlation-based feature vector, i.e., the vector obtained by taking the upper triangle of the FCN matrix obtained by formula (1) and flattening it, denotes the total number of ROIs.

[0051] For any two nodes and , if there is information flow from to , then , and the connection strength between the two nodes is described by the weight , and ; if there is no connection, then , where the weight is represented by different correlation matrix construction methods. If there is no information flow, only a connection relationship exists, then . Note that in the brain functional graph, there is no edge connection between a node and itself, i.e. , so .

[0052] The node feature matrix and the adjacency matrix of the graph defined by this are and , respectively, where denotes the dimension of each node feature, denotes the total number of ROIs. For the edge weight feature, i.e., the adjacency matrix The evaluation is performed by using the FCN matrix given by formula (1) under the corresponding time window and then thresholding it.

[0053] Considering computational cost and ease of calculation, the Pearson correlation coefficient PC between nodes within each time segment is calculated using formula (1). We construct the FCN matrix as an evaluation of the edge weight matrix.

[0054] Correlation coefficient Reflects on time segments Above, node With nodes The linear correlation strength between the BOLD signals: (1) in, This represents the covariance of the BOLD signal between two ROIs. It represents the standard deviation.

[0055] Therefore, for each individual, their time period The brain's functional network can be represented as While functional networks contain important information about all nodes, they also include redundant and noisy information. Therefore, by selecting edges with statistical significance and setting their weights to 1, while setting the rest to 0, a sparse adjacency matrix with strong connectivity can be formed. ,in, This indicates the selected threshold, which is 30% in this embodiment.

[0056] Considering the efficiency and simplicity of the code, for the node feature matrix The evaluation also uses FCN, but unlike the edge weight matrix evaluation strategy, the FCN matrix is ​​no longer binarized in the node evaluation. Instead, the original FCN feature matrix is ​​used to retain comprehensive information.

[0057] In summary, the graph constructed within the t-th time window can be represented as follows: As input for topology-aware feature extraction, where, The node feature matrix, It is a sparse adjacency matrix.

[0058] For each research subject, its dynamic graph sequence can be represented as a set. Including from the 1st to the 1st This representation, showing all graphs for each time window, facilitates the analysis of changes in the brain's functional networks across different time windows. Step 13: Local-Global Topology-Aware Feature Extraction Process.

[0059] To generate node embeddings that can capture both local neighborhood dynamics and global topological dependencies simultaneously , denotes the dimension of each node feature, and the embodiment proposes an innovative topology-aware feature extraction framework. The framework processes each time window graph in a dynamic graph sequence set through a local topology-aware encoder (LTAE) and a global topology-aware encoder (GTAE), and implements effective integration of multi-scale information with the aid of cross-attention mechanisms, thereby significantly improving the expression ability of graph learning models. Compared with the limitation of traditional graph neural networks (GNN) that only rely on static adjacency matrices, the method of the embodiment enhances the topology-aware ability of graph representation learning by explicitly modeling local similarity and global correlation, and injects more rich structural information into node embedding.

[0060] The local topology-aware encoder LTAE aims to extract local dynamic features from the node original time series information (non-FCN matrix). Its core is a multi-branch structure including a detail CNN encoder and a topology information supplement part; wherein the processing process of the detail CNN encoder is: firstly, a one-dimensional convolutional neural network (1D-CNN) is applied to each node time series for local time series modeling, and the convolution process is:

[0061] wherein, is the convolution kernel weight, is the convolution kernel size, is the bias term, denotes a nonlinear activation function (such as ReLU), and * denotes a one-dimensional convolution operation; i denotes the node index.

[0062] Subsequently, max-pooling operation is applied to :

[0063] wherein, is the pooling window size, and finally , denotes the local dynamic representation of the i-th node, is the dimension after pooling. The CNN feature matrix is obtained by concatenating all nodes:

[0064] ​​

[0065] The selection of CNNs fully utilizes the local topological change characteristics encoded in the original time series, complements the original fine-grained information, and enhances the model's perception of dynamic neighborhood structure.

[0066] Meanwhile, for the dynamic graph structure of each time window, the local topological perception encoder LTAE introduces a graph topology feature extractor to analyze the static topological information contained in the dynamic graph structure under each time window. The input of the graph topology feature extractor is the dynamic adjacency matrix between nodes , which comes from the node connection weight graph obtained by calculating the Pearson correlation coefficient or mutual information of the original fMRI time series in the first sliding window. For each graph , the graph topology feature extractor uses existing graph index calculation methods to extract the following types of topological features for each node: Clustering Coefficient: measures the closeness between the neighborhoods of nodes; Small-Worldness: measures whether the graph has high clustering and short path properties at the same time; PageRank value: measures the importance of the node and its information propagation potential; Eigenvector Centrality: measures the connection degree of the node to high importance nodes.

[0067] The above indexes are based on the classical definition of graph theory and can be obtained by NetworkX calculation tools, which belong to the existing technology category. Finally, the extracted structural features constitute the graph structure feature matrix , where represents the topological feature dimension, which is a supplement of the topological information as local information.

[0068] In addition, as shown in Figure 2 , the local topological perception encoder LTAE also introduces a fully connected encoder. The fully connected encoder uses a shallow multilayer perceptron (MLP) to perform nonlinear transformation on the original node features . The MLP consists of two fully connected networks combined with ReLU activation function and batch normalization operation, and outputs the initial representation , which enhances the model's expression ability for the original signal. Finally, the local topological perception encoder LTAE fuses the three types of features through concatenation and linear transformation to generate the local topological perception representation , where is the number of subjects, is the number of sliding time windows, to the final local topological-aware feature dimension.

[0069] As Figure 2 shown, the global topological-aware encoder (GTAE) adopts a graph Transformer encoder, which is a collaborative mechanism of graph isomorphism network (GIN) and Transformer encoder. This part integrates local structure modeling and global long-range dependency capturing to generate global topological-aware representation . Compared with the limitation of traditional graph convolutional network (GCN) relying on Laplacian smoothing, the global topological-aware encoder uses GIN to accurately depict the differences in local neighborhoods and model the topological correlation between nodes in the whole graph through the multi-head self-attention mechanism of Transformer, thereby significantly enhancing the node embedding's ability to perceive the graph structure. This architecture takes the node features in the dynamic graph sequence as input and combines a multi-scale feature fusion strategy to provide robust global representation for complex graph learning tasks.

[0070] The global topological-aware encoder GTAE first uses a multi-layer GIN to perform iterative neighborhood aggregation on the input graph features to capture local topological patterns.

[0071] The update rule of the ( l ) layer GIN is as follows:

[0072] where denotes the node hidden state at the ( l ) layer, is the set of adjacent nodes of node , and is a learnable topologically sensitive coefficient used to dynamically balance the contributions of the node's own features and neighborhood aggregated features.

[0073] It is composed of two linear layers and a ReLU activation function, and its parameterization capability ensures effective encoding of local topological structures. Compared with traditional GNNs, GIN has stronger graph isomorphism discrimination ability and can accurately distinguish the structural differences of local neighborhoods by introducing a learnable topologically sensitive coefficient and multi-layer nonlinear transformation.

[0074] Subsequently, the GIN finally outputs the node embedding feature matrix ; is the representation of the ( ) node in the GIN at the ( ) layer, i.e., the output of the final layer; |V| is the number of nodes in the graph; denotes the hidden dimension, i.e., the dimension of each node representation.

[0075] The output representation of GIN is input into the Transformer encoder to capture topological dependencies across the entire graph. To enhance global information integration, GTAE adds a learnable [CLS] tag before the node feature sequence, forming an enhanced input. .

[0076] The Transformer encoder models the global associations between nodes through a multi-head self-attention mechanism. The calculation process of the (h)th attention head is as follows:

[0077] in, For the ( l Layer input features, These are the projection matrices for the query, key, and value, respectively. ,in: Hidden dimensions for the input and output of the Transformer; The number of attention heads in a multi-head attention mechanism; For each dimension of key / query / value in each attention head, For attention weights.

[0078] The self-attention mechanism dynamically assigns weights, enabling each node to adaptively focus on other nodes in the entire graph, regardless of whether they are directly connected, thus effectively capturing long-range topological dependencies. The outputs of all attention heads are concatenated, linearly transformed, residually connected, and layer normalized to generate global topology-aware features.

[0079] To achieve the fusion of multi-scale topological information, GTAE adopts a hierarchical feature integration strategy, which integrates the local features of GIN nodes. Global interaction features of nodes after Transformer encoding and the [CLS] mark feature of broadcasting. The vectors are concatenated to form a comprehensive feature vector, where [CLS] is a learnable parameter vector with dimensions of . During training, the learnable parameter vector [CLS] is updated along with other node features. [CLS] serves as the aggregation center for the entire graph-level semantics and is added to the node feature sequence before the sequence, aggregating the information of the entire graph through the Transformer self-attention mechanism.

[0080] The final global representation of the nodes is generated through mapping in the fully connected layer. :

[0081] in, denotes the feature concatenation, is the output of the fully connected layer, where is the output dimension of the final global representation, which is not necessarily equal to . is the embedding dimension required by the downstream task (e.g., for classification, clustering, etc.), which is set by the last fully connected layer.

[0082] Through this local and global collaborative modeling manner, GTAE not only preserves the graph geometric properties encoded in the dynamic graph sequence, but also captures the complex interaction patterns of the whole graph topology through the Transformer, providing powerful node representations for downstream tasks.

[0083] Step 14: Cross-attention fusion of dynamic graph features.

[0084] To effectively integrate the local and global topological information extracted in the dynamic graph sequence, this embodiment designs a bidirectional cross-attention fusion mechanism (Cross-Attention Fusion Module, CAFM) to interact model the multi-scale features output by the local topology-aware encoder (LTAE) and the global topology-aware encoder (GTAE) in the time dimension. Specifically, for each subject, after the original fMRI signal is divided by the dynamic window and the graph structure is constructed, the LTAE outputs the local topology-aware representation: ; where is the batch size, is the number of time windows, is the local representation dimension.

[0085] The GTAE outputs the global topology-aware representation , is the global representation dimension; the global representation is broadcasted along the time dimension to obtain :

[0086] where is a unit vector, and generally, is equal to in the previous , so as to realize the replication of each global time slice, so that its dimension is aligned with the local sequence.

[0087] Subsequently, the local representation and the global representation are respectively mapped to the shared attention space , and the query (Q) vectors, i.e., and , and the key (K) vectors, i.e., and , value (V) vector, i.e. and :

[0088]

[0089] where, , is a linear projection matrix, is the unified attention spatial dimension. Cross-attention mechanism takes one side as query and the other side as key-value pair, realizing information interaction of two-way representation. The time-causal mask matrix , represents the number of sliding time windows, where is used to ensure the forward dependency in sequence processing: Local to global path Local→Global: taking as the query, the global topology information is fused:

[0090] Global to local path Global→Local: taking as the query, the local topology information is fused:

[0091] The above attention mechanisms can be stacked in multi-head form (Multi-head Attention) for enhancement, further improving the model's ability to model heterogeneous topology information. The dual-channel attention output is concatenated in the channel dimension, and is fused through linear transformation and layer normalization to obtain the feature :

[0092] where, is a projection matrix, is a layer normalization operation, represents the feature concatenation operation.

[0093] To further stabilize the training, residual connection is introduced: , is the linear transformation matrix in the residual connection, which adjusts the dimension matching of the residual connection. The final feature representation is:

[0094] Average pooling is performed on the time dimension to generate sample-level graph embedding representation :

[0095] wherein, denotes the number of sliding time windows; is the batch size, is the local representation dimension.

[0096] wherein, denotes the number of sliding time windows; is the batch size, is the local representation dimension.

[0097] Step 2: Node feature construction based on DTI / clinical information features.

[0098] Step 21: DTI preprocessing and structural connectivity construction.

[0099] In this step, the diffusion tensor imaging (DTI) raw data is standardized and preprocessed to obtain stable and comparable white matter microstructure measures and construct individual-level structural connectivity.

[0100] Specifically, first, eddy current and field inhomogeneity correction, head motion correction and skull stripping are performed; individual T1 spatial registration is completed using the b0 image, and then nonlinear deformation is performed to the standard MNI space, so as to strictly correspond to the brain region template adopted by fMRI, preferably AAL-116, Desikan, Schaefer, etc. can also be used. Subsequently, tensor fitting is performed based on the diffusion signal to obtain the anisotropy fraction FA, mean diffusivity (MD) of each voxel, and optionally reduced radial diffusivity RD, axial diffusivity AD.

[0101] To represent the structural connectivity, under the support of fiber tracking (deterministic or probabilistic) or spherical deconvolution model, the weighted adjacency matrix of the individual-level white matter structure network is constructed . The preferred weight definition is:

[0102] wherein, denotes the set of all fiber bundles connecting ROI and ROI , and is the average value of FA (Fractional Anisotropy) on all voxels in the set; denotes the number of fiber bundles connecting ROI i and ROI j . To further improve the stability and comparability of the connection graph, Symmetrization and sparsification (e.g., proportion thresholding or minimum spanning tree + thresholding strategy) are performed, and density matching is used to control the difference in network structure between individuals. For FA / MD images from multiple centers, ComBat batch effect elimination algorithm or distribution normalization method can also be used to alleviate the bias brought by data sources. In addition, quality control (QC) is completed by removing abnormal slices, directions, and pseudo-connected paths. Step 22: Multi-view DTI feature extraction (statistical-texture-graph integration).

[0103] To characterize the integrity, spatial complexity, and topological organization of the white matter network, multi-level features are extracted from three dimensions, and FA and MD indices are calculated separately (MD and FA are calculated in the same way). (1) Statistical features (intensity distribution within ROI) to quantify local white matter integrity and heterogeneity, including: Calculate the mean value in each ROI wherein, is the total number of voxels in the ROI, represents the FA value of the th voxel; standard deviation to measure the dispersion of FA values; skewness to measure the asymmetry of the distribution; kurtosis to measure the sharpness of the distribution; high FA volume wherein, is the pre-set FA threshold, is an indicator function, which is 1 when is 0 otherwise, represents the volume of a single voxel; entropy wherein, is the probability that the FA value falls into the th histogram interval; and three-way gradient amplitude wherein represents the gradient value of the FA image in the x, y, and z directions, respectively.

[0104] (2) Texture / morphology features (FA / MD grid level) to characterize the self-similarity and spatial complexity of white matter structure, including: based on the gray level co-occurrence matrix (GLCM) to calculate energy, contrast, homogeneity, correlation, etc. Based on local binary pattern (LBP) to extract fine-grained texture histogram wherein, the intensity value of the central pixel, the intensity value of the neighborhood pixel, is a sign function, when , otherwise , the number of neighborhood points, the radius, preferably the radius , neighborhood ; estimate the fractal dimension using box counting method , wherein, the number of boxes required to cover the image with side length .

[0105] (3) Graph theory features (based on structural connectivity ): the average degree of the whole network , wherein, V is the total number of network nodes, denotes the degree of node ; the average shortest path , wherein, is the shortest path length between node and ; the clustering coefficient , wherein, denotes the number of edges actually existing between the neighbors of node ; the global efficiency defined as the average of the inverse of the shortest path length between all pairs of nodes ; the local efficiency , which is similar to the global efficiency, but calculated for the local subgraph of each node; the betweenness centrality measuring the frequency of a node appearing in all shortest paths and the eigenvector centrality measuring the influence of a node based on the eigenvector of the adjacency matrix; and the small-world property , wherein , , and respectively represent the clustering coefficient and the average shortest path of a random network, and the modularity and the homophily are added as necessary to enhance the representation ability of the global-local organization. The above features are aggregated at the individual level and dimensionally unified to form a multi-view feature set . Step 23: DTI feature integration, time alignment and fusion strategy.

[0106] The multi-view feature set is standardized (z-score), projected to a shared hidden space by a multilayer perceptron, and the subject-level structural prior representation is obtained, wherein is the batch size, is the structural feature dimension: ​

[0107] To align with fMRI-based temporal embedding, temporal expansion is preferred to achieve a consistent mapping of the time window, where... It is a unit vector. Indicates the outer product operation:

[0108] This allows for seamless integration with the bidirectional cross-attention module in step 14, in order to Key-value pairs (K / V) serve as structural priors and interact with local / global temporal queries.

[0109] In another embodiment, It can also be used as an edge / hyperedge weight prior in a group graph / hypergraph to participate in structure learning, so as to realize the constraint and reinforcement of functional dynamic discrimination by the structure prior. If clinical covariates exist, they can be mapped to the same latent space through a projection matrix and then compared with... Alternatively, the output from step 14 can be spliced ​​and merged, and then linearly transformed to obtain... The input group hypergraph / volume integrator completes the downstream tasks. Step 24: Feature preprocessing and feature splicing of clinical information.

[0110] To introduce stable background priors at the patient level and improve the model's robustness in discriminating boundary samples, this step uniformly encodes and embeds clinical information in a low dimension, and then splices it with the image features obtained in the previous step.

[0111] Let the original clinical variable matrix be... ,in, This indicates the number of raw clinical fields before processing, which includes information such as cognitive scale scores (e.g., MMSE, MoCA, CDR), disease duration, years of education, ApoE genotype, age, and gender. First, for... Missing value imputation and outlier suppression are performed: continuous variables are standardized using z-scores, categorical variables are encoded using one-hot encoding or learnable embeddings, and finally all processed variables are concatenated to obtain the result. ,in, This represents the total feature dimension after processing. Subsequently, a lightweight multilayer perceptron is used to... Mapping to a low-dimensional latent space yields the subject-level clinical feature embeddings:

[0112] Finally, fused node features were constructed at the subject level for subsequent group graph / hypergraph learning. These features consist of three parts directly concatenated along the feature dimension:

[0113] in: These are features derived from the time pooling of the dynamic functional connection graph in step 14. It is a structural white matter feature derived from the DTI feature integration in step 23. This is the clinical feature embedding obtained in this step; the three are aligned in a unified latent space and used as node features input to the subsequent population hypergraph / graph structure learning and classification module.

[0114] Step 3: Hypergraph construction process.

[0115] The three subject-level representations obtained from the upstream in this stage serve as inputs: functional dynamic representations obtained by time pooling after dynamic window sequences are encoded using "local-global" encoding and bidirectional cross-attention fusion. DTI structure prior representation With clinical embedding features .

[0116] Subsequently, to construct hyperedge relationships, node similarity modeling is performed within the "semantic fusion space." First, the similarity between nodes is... Mapped to a unified hypergraph embedding space : It is the dimension of the hypergraph embedding space. Specifically, This indicates the dimension used when mapping the input features to this space, where, Let be a learnable linear projection weight matrix. Next, define a Gaussian kernel function to measure the nodes. and Similarity within this space:

[0117] in, Indicates the first The semantic embedding vector of each node, The kernel bandwidth is typically adaptively estimated using the median Euclidean distance between all node pairs. The similarity matrix in each node Selecting the first feature space The most similar nodes form a superedge. :

[0118] Combine the sets of hyperedges of all nodes to obtain the final set of hyperedges. The total number of corresponding hyperedges is denoted as .

[0119] Next, we construct the probabilistic hypergraph incidence matrix. , which is the first Line number Columns represent nodes For hyperedge Subordination strength:

[0120] To improve the stability of subsequent hypergraph convolutions, the correlation matrix of probabilistic hypergraphs is... Symmetric normalization is performed. Specifically, the node degree matrix is ​​first defined. With hypermarginal matrix ,in, , It is a unit column vector.

[0121] Normalized hypergraph incidence matrix for:

[0122] Simultaneously construct the Laplacian matrix of the hypergraph :

[0123] in identity matrix ， This refers to the number of nodes (i.e., the number of subjects, as opposed to batch size). ) ， Used to apply structural smoothness constraints in supervised learning.

[0124] Ultimately, two core inputs for the hypergraph neural network were obtained: node features. Or its linear transformation; structural information: normalized hypergraph incidence matrix and the Laplacian matrix of the hypergraph Among them, the hypergraph Laplacian matrix This is an optional parameter used for regular expressions.

[0125] These inputs will be fed into subsequent hypergraph convolutional networks (HyperGCN / HyperGAT) for structure-aware information propagation and classification learning. Additionally, a hypergraph Laplacian smoothing regularization term can be added to the loss function. This encourages consistency in node representation within highly similar hyperedges, thereby improving robustness to data noise and graph sparsity. The computational complexity of the entire composition process mainly comes from two... Similarity calculation and Top- Neighbor selection operation and hypergraph association matrix The normalization process, the overall calculation and storage cost is linearly related to the number of non-zero elements, which is suitable for medium-sized data sets with a number of subjects B in the tens to hundreds.

[0126] Step four: hypergraph neural network training.

[0127] After completing the multi-modal feature fusion to obtain the node feature matrix and constructing the group hypergraph structure, the end-to-end model training and optimization are performed in this phase.

[0128] First, based on the hypergraph adjacency matrix , the node degree matrix and the hyperedge degree matrix ， and introduce the learnable hyperedge weight matrix , finally construct the normalized hypergraph adjacency matrix : where is the learnable hyperedge weight matrix.

[0129] The feature matrix and the hypergraph adjacency matrix are input into the hypergraph neural network with attention mechanism, which realizes feature propagation through multi-layer attention hypergraph convolution: first, linear transformation is performed on the node features to obtain , the attention score between nodes is calculated , and the attention weight is obtained through softmax normalization , and the hypergraph adjacency matrix is dynamically adjusted using the weight to obtain where denotes element-wise multiplication (Hadamard product), is the attention weight matrix.

[0130] Finally, the feature update with residual connection is realized through the feature propagation formula , where is the learnable scaling factor, is the hypergraph adjacency matrix. The hypergraph neural network model uses He initialization method parameters, uses Adam optimizer for training, and dynamically adjusts the learning rate through a multi-step learning rate scheduler.

[0131] To solve the class imbalance problem, a weighted cross-entropy loss function based on the reciprocal square root of class frequency is used , where the class weight . Where, denotes the total number of training samples, denotes the number of classes, is the weighted coefficient of the th class, the frequency of the sample in the training set, the sample in the first the true one-hot label value of the class, the prediction probability output by the hypergraph neural network model.

[0132] During the training process, the model performance is evaluated on the validation set every round, the optimal parameters are saved for final testing, and regularization techniques such as weight decay and dropout are used to prevent overfitting, ensuring that the model obtains stable and reliable Alzheimer's disease classification performance on the test set.

[0133] The Alzheimer's disease classification method based on topology-aware local-global joint encoding and group hypergraph learning proposed in this embodiment has shown significant practical value and technical advantages in the field of intelligent diagnosis of neurodegenerative diseases. This method innovatively integrates multi-modal brain image data and clinical information to build an end-to-end intelligent diagnosis framework, providing reliable technical support for early and accurate identification of AD. In practical applications, the outstanding advantages of this method mainly lie in the following aspects: First, by constructing a dual-branch architecture of local topology-aware encoder and global topology-aware encoder, fine-grained local features and global long-range dependencies are extracted from brain functional networks, respectively, realizing multi-scale representation of brain network topology characteristics and providing rich discriminative features for disease recognition. Second, this method innovatively integrates fMRI functional connectivity data and DTI structural connectivity information collected at the same time for the same subject, systematically representing the integrity, spatial complexity, and topological organization of white matter networks through a multi-level feature extraction framework, fully leveraging the complementary advantages of multi-modal data.

[0134] It is particularly worth emphasizing that the dynamic time window analysis and time attention pooling mechanism used in this method can effectively capture the dynamic characteristics of brain functional connectivity evolution over time, significantly improving the model's sensitivity to pathological changes. The introduction of the group hypergraph learning architecture further breaks through the limitations of traditional pairwise sample relationships, realizes consensus decision-making at the group level by constructing high-order similarity relationships between subjects, and greatly enhances the model's generalization ability and robustness. In clinical practice, this method has shown important application prospects: its multi-modal fusion strategy can be compatible with image data from different medical centers, effectively alleviating the batch effect of multi-center data through standardized preprocessing and feature alignment mechanisms; at the same time, the organic integration of clinical information not only enriches the feature representation dimension, but also provides intuitive clinical interpretation basis for the diagnosis result, which is helpful for building an interpretable auxiliary diagnosis system.

[0135] With the continuous development of medical artificial intelligence technology, the method shows a broad application space. The core technical framework can be extended to key clinical scenarios such as early screening of mild cognitive impairment and disease progression prediction, and can also be applied to intelligent diagnosis of other neuropsychiatric diseases. The successful practice of this method not only provides an effective technical tool for the precise diagnosis and treatment of Alzheimer's disease, but also provides an important technical reference for the innovation and development of the entire intelligent medical field, promoting the important transformation of medical artificial intelligence from single modal analysis to multi-modal fusion, from static diagnosis to dynamic monitoring, and from individual analysis to group intelligence.

[0136] Embodiment two The purpose of this embodiment is to provide an Alzheimer's disease classification system based on topology perception and group hypergraph, comprising: The processing module is configured to extract a time sequence from resting-state functional magnetic resonance imaging data, and construct a dynamic functional connection network sequence through a sliding window strategy. The extraction module is configured to extract local topology features and global topology features of each time window by using a local topology perception encoder and a global topology perception encoder respectively, and to generate a comprehensive feature representation of the subject level by deeply interacting and fusing the local topology perception representation and the global topology perception representation of each time window. The construction and classification module is configured to take each subject as a hypergraph node, construct a hyperedge based on the comprehensive feature representation of the subject level and the feature similarity calculated by combining the diffusion tensor imaging features and the clinical information features of the subject, and then construct a group hypergraph, and obtain the Alzheimer's disease classification result of the subject by processing the group hypergraph using a hypergraph neural network.

[0137] In more embodiments, there is also provided: An electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the method described in embodiment one is completed. For brevity, this will not be repeated here.

[0138] It should be understood that in this embodiment, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSPs, application-specific integrated circuits ASICs, ready-to-program gate arrays FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0139] The memory can include read-only memory and random access memory, and provide instructions and data to the processor, a part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0140] A computer readable storage medium for storing computer instructions, which when executed by a processor, complete the method described in embodiment one.

[0141] The method in embodiment one can be directly embodied as a hardware processor to complete, or be completed by a combination of hardware and software modules in the processor. The software modules can be located in a storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, register, etc. The storage medium is located in the memory, and the processor reads the information in the memory to complete the steps of the above method in combination with the hardware. To avoid repetition, it will not be described in detail here.

[0142] Although the specific embodiments of the present application are described above with reference to the drawings, it is not a limitation on the scope of protection of the present application, and those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. Alzheimer's disease classification method based on topology awareness and group hypergraph, characterized in that, The application relates to a method for Alzheimer's disease classification based on resting-state functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) data. The method comprises the following steps: extracting time series from resting-state functional magnetic resonance imaging data and constructing a dynamic graph through a sliding window strategy; extracting local topological features and global topological features of each time window by using a local topological perception encoder and a global topological perception encoder respectively, and performing deep interaction and fusion on the local topological perception representation and the global topological perception representation of each time window to generate a comprehensive feature representation of a subject; 2. The topology-aware and community-hypergraph-based Alzheimer's disease classification method of claim 1, wherein, taking each subject as a hypergraph node, constructing a hyperedge and a probabilistic hypergraph correlation matrix based on the comprehensive feature representation of the subject, combining the feature similarity calculated based on the structural prior representation of the subject and the clinical embedding features of the diffusion tensor imaging, and then constructing a group hypergraph, and processing the group hypergraph by using a hypergraph neural network to obtain the Alzheimer's disease classification result of the subject. The local topological features of each time window are extracted by using a local topological perception encoder, and the specific process is as follows: extracting fine-grained dynamic pattern features of a local neighborhood of each node original time information by using a one-dimensional convolutional neural network, extracting node local time sequence features, performing maximum pooling processing on the node local time sequence features to obtain a local dynamic representation of the node, and splicing the local dynamic representation of each node to obtain a CNN feature matrix; calculating graph topological features by using a graph topological feature extractor on the dynamic adjacency matrix between nodes of each time window to obtain a graph structure feature matrix; processing each node original time sequence information by using a multilayer perceptron to generate an initial feature representation; 3. The topology-aware and community-hypergraph-based Alzheimer's disease classification method of claim 1, wherein, splicing and linear transformation fusion are performed on the CNN feature matrix, the graph structure feature matrix and the initial feature representation to obtain the local topological features. The global topological features of each time window are extracted by using a global topological perception encoder, and the specific process is as follows: iterative neighborhood aggregation is performed on the dynamic graph by using a graph isomorphism network to capture local topological patterns and obtain node local features; the node embedding feature matrix is input into a Transformer encoder to capture topological dependencies in the whole graph range and obtain node global interaction features; 4. The topology-aware and community-hypergraph-based Alzheimer's disease classification method of claim 1, wherein, the node local features, the node global interaction features and the broadcast label features are spliced, and global topological features are generated by mapping through a fully connected layer. The local topological perception representation and the global topological perception representation of each time window are deeply interacted and fused to generate a comprehensive feature representation of a subject, and the specific process is as follows: the local topological perception representation and the global topological perception representation are projected into a shared attention space, and corresponding query vectors, key vectors and value vectors are obtained through linear transformation; the local topological perception representation is taken as a query vector, and the global topological perception representation is fused; the global topological perception representation is taken as a query vector, and the local topological perception representation is fused; the outputs of the double-path attention mechanism are spliced in the channel dimension, and linear transformation and layer normalization are performed on the outputs to obtain a final feature representation; the final feature representation is subjected to average pooling operation in the time dimension to obtain a comprehensive feature representation of a subject.

5. The topology-aware and community-hypergraph-based Alzheimer's disease classification method of claim 1, wherein, The statistical features, texture features, morphological features and graph theory features extracted from the diffusion tensor imaging data are taken as the diffusion tensor imaging structure prior representation; wherein the statistical features include the mean, standard deviation, skewness, peak value, entropy, three-way gradient amplitude and volume of the anisotropy fraction in the region of interest; and the graph theory features include the whole-network average degree, average shortest path, clustering coefficient, global efficiency, local efficiency, betweenness centrality and eigenvector centrality of the individual-level white matter structure network, and small-world attribute; The extraction of the clinical embedding features is specifically: linear-nonlinear mapping is performed on the preprocessed original clinical variable matrix by using a perception machine to obtain the clinical embedding features.

6. The topology-aware and community-hypergraph-based Alzheimer's disease classification method of claim 1, wherein, The feature similarity is calculated based on the subject-level comprehensive feature representation, in combination with the diffusion tensor imaging structure prior representation and the clinical embedding features of the subjects, to construct a hyperedge and a probabilistic hypergraph association matrix, which is specifically: The subject-level comprehensive feature representation, the diffusion tensor imaging structure prior representation and the clinical embedding features are spliced and projected into a hypergraph embedding space; the similarity of different nodes in the hypergraph embedding space is calculated based on a Gaussian kernel function; Based on the computed similarity, the top-k most similar nodes in the hypergraph embedding space of each node are selected to form a hyperedge; Based on the computed similarity, the top-k most similar nodes in the hypergraph embedding space of each node are selected to form a hyperedge; A probabilistic hypergraph association matrix is constructed and normalized, and a hypergraph Laplacian matrix is constructed based on the normalized probabilistic hypergraph association matrix.

7. The topology-aware and community-hypergraph-based Alzheimer's disease classification method of claim 1, wherein, A weighted cross-entropy loss function based on the reciprocal square root of the class frequency is used as the loss function for training the hypergraph neural network.

8. Alzheimer's disease classification system based on topology awareness and population hypergraph, characterized in that, It comprises: The processing module is configured to extract a time sequence from resting-state functional magnetic resonance imaging data and construct a dynamic functional connection network sequence through a sliding window strategy; The extraction module is configured to extract local topological features and global topological features of each time window by using a local topological perception encoder and a global topological perception encoder respectively, and to generate a subject-level comprehensive feature representation by deeply interacting and fusing the local topological perception representation and the global topological perception representation of each time window; The construction and classification module is configured to take each subject as a hypergraph node, construct a hyperedge and a probabilistic hypergraph association matrix based on the feature similarity calculated from the subject-level comprehensive feature representation in combination with the diffusion tensor imaging features and the clinical information features of the subjects, and then construct a group hypergraph, process the group hypergraph by using a hypergraph neural network, and obtain the Alzheimer's disease classification result of the subjects.

9. An electronic device, comprising: It comprises a memory and a processor, and computer instructions stored in the memory and running on the processor, which, when executed by the processor, complete the method of any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, It is used for storing computer instructions, which, when executed by the processor, complete the method of any one of claims 1-6.

Images