Alzheimer's disease classification method and system based on topology awareness and group hypergraph

Abstract

The application belongs to the technical field of brain image processing, and aims to solve the problem of inaccurate classification of Alzheimer's disease. A method and system for classifying Alzheimer's disease based on topology perception and group hypergraph are proposed. The method extracts time series from resting-state functional magnetic resonance imaging data and constructs dynamic functional connectivity network sequences through a sliding window strategy. Local topology features and global topology features of each time window are extracted using local and global topology perception encoders, respectively, and then subjected to deep interaction and fusion to generate comprehensive feature representations at the subject level. Each subject is treated as a hypergraph node, and hyperedges are constructed based on the comprehensive feature representations at the subject level, combined with the feature similarity calculated from the diffusion tensor imaging features and clinical embedding features of the subjects. A group hypergraph is then constructed, and a hypergraph neural network is used to obtain the classification results, effectively improving the accuracy of early classification and diagnosis of Alzheimer's disease.

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, and 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:

[0006] 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, and fail to fully utilize the complementary advantages of different data perspectives.

[0007] 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.

[0008] 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

[0009] 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.

[0010] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0011] In a first aspect, the present application provides an Alzheimer's disease classification method based on topology perception and group hypergraph, comprising:

[0012] 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;

[0013] 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;

[0014] 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.

[0015] In a second aspect, the present application provides an Alzheimer's disease classification system based on topology perception and group hypergraph, comprising:

[0016] 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;

[0017] 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;

[0018] 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.

[0019] 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, when the computer instructions are run by the processor, the method of the first aspect is completed.

[0020] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method of the first aspect is completed.

[0021] The above one or more technical solutions have the following beneficial effects:

[0022] 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 the comprehensive characterization of brain network topological characteristics; based on group hypergraph learning to mine the high-order similarity relationship between subjects, fully utilize the complementary information at the group level, significantly improve the classification accuracy of Alzheimer's disease, and provide strong technical support for early diagnosis and identification of Alzheimer's disease.

[0023] 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; the cross-attention mechanism is used to effectively integrate multi-scale information, thereby significantly improving the expression ability of the graph learning model.

[0024] 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 to build an interpretable auxiliary diagnosis system.

[0025] 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.

[0026] 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 known by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0027] 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.

[0028] Figure 1 is the overall algorithm flowchart in embodiment one of the present application;

[0029] Figure 2 is the feature extraction flowchart based on topology awareness in embodiment one of the present application;

[0030] Figure 3 is the detailed flowchart of local topology-aware encoder and global topology-aware encoder in embodiment one of the present application. DETAILED DESCRIPTION

[0031] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0032] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application.

[0033] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0034] Embodiment one

[0035] First, the relevant terms involved in this embodiment are explained:

[0036] Hyperedge: can connect multiple nodes, unlike the traditional adjacency matrix which can only reflect the relationship between two nodes, can capture the high-order association between multiple nodes, note that here for node classification, each node represents a class, find the correlation between nodes, and connect the points in a class.

[0037] Graph structure: ; wherein represents a set of vertices, represents a set of hyperedges, unlike the adjacency matrix corresponding to the original graph neural network, each edge only represents 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 the hyperedge can connect multiple individuals, where patients with similar symptoms can be connected on the same hyperedge.

[0038] Association matrix: the hypergraph is represented by an association matrix , size , wherein The corresponding element in .

[0039] Degree: the degree of vertex ​ Defined as: ,in, It is a super-edge The weight represents the strength or importance of the hyperedge.

[0040] Hypergraph Laplacian Operator: Given a hypergraph ,in, It is a vertex set. It is a super-edge set. It is the Laplacian operator for hypergraphs, which is an n*n positive semidefinite matrix, where n is the number of vertices.

[0041] right Eigenvalue decomposition yields:

[0042]

[0043] in, It is a matrix containing orthogonal eigenvectors; It is a diagonal matrix containing the corresponding non-negative eigenvalues, and the superscript T indicates transpose.

[0044] The Fourier transform of a signal x on a hypergraph is defined as: Here the feature vector It is considered a Fourier basis.

[0045] Spectral convolution: signal With filters The spectral convolution is represented as:

[0046]

[0047] in, This represents the element-wise Hadamard product. It is a function of the Fourier coefficients.

[0048] Since the computational complexity of the forward and inverse Fourier transforms is O(n), Chebyshev polynomials are used for parameterization. This reduces computational complexity.

[0049] filter It can be used Approximation using Chebyshev polynomials:

[0050]

[0051] in, It is a Chebyshev polynomial, and the recursive calculation formula is:

[0052]

[0053] and and .

[0054] Using scaled Laplacian in Chebyshev expansion :

[0055]

[0056] where, is the largest eigenvalue of the Laplacian, is the identity matrix.

[0057] Finally, the convolution operation can be simplified as:

[0058]

[0059] where, is the affinity matrix, and are the degree matrices, is the possible weight matrix.

[0060] The embodiment discloses an Alzheimer's disease classification method based on topology awareness and group hypergraph, and comprises the following steps:

[0061] 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;

[0062] Extracting local topology features and global topology features of each time window by using a local topology awareness encoder and a global topology awareness encoder respectively, and generating a comprehensive feature representation of a subject level by deeply interacting and fusing the local topology features and the global topology features of each time window;

[0063] Taking each subject as a hypergraph node, constructing a hyperedge based on a comprehensive feature representation of a subject level and a feature similarity calculated by combining diffusion tensor imaging features and clinical information features of the subject, and further constructing a group hypergraph, and processing the group hypergraph by using a hypergraph neural network to obtain an Alzheimer's disease classification result of the subject.

[0064] The embodiment can extract multi-dimensional topology features from resting-state functional magnetic resonance imaging data and generate more comprehensive brain network representations by combining a local topology awareness encoder (LTAE) and a global topology awareness encoder (GTAE) and a dynamic functional connection network sequence generated by a dynamic time window.

[0065] In 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 network 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 of the embodiment 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 of the embodiment 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 ability of key changes of dynamic functional connectivity and further improving the accuracy and robustness of classification.

[0066] In summary, the method proposed in the 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 network and DTI structural network and group hypergraph structure learning. The method provides strong technical support for early diagnosis and identification of Alzheimer's disease.

[0067] The following will be combined Figures 1-3 The method of classifying Alzheimer's disease based on multi-view spatio-temporal topology fusion proposed in the embodiment will be described in detail, including the following parts:

[0068] I. Multi-modal feature extraction.

[0069] Step 1, time series topology-aware dynamic graph representation and cross-attention fusion module.

[0070] As Figure 1 shown, the steps implemented in the embodiment include:

[0071] Step 11: division of dynamic time window.

[0072] First, for 4D fMRI (containing 3D voxels changing over time), the brain at each measurement time is visualized as a set of discrete brain functional regions. The BOLD signal in each region, i.e. the blood oxygen level dependent signal, is represented by the average of all voxel intensities in the region, forming a time series vector. In this way, a set of BOLD signals is obtained for the entire brain, each sequence corresponding to a specific ROI, i.e. region of interest, and the BOLD signal is defined where , is the number of time points.

[0073] In technical aspect, a sliding window method is adopted to capture the temporal dynamics in BOLD signals. This method involves sliding a fixed length window over the time series with a window length of and a step size of , which results in a series of possibly overlapping time segments, and the whole BOLD signal can be divided into segments, each with a window length of and a step size of . denotes the BOLD signal of the ith ROI in the tth time segment.

[0074] Step 12: Constructing the dynamic graph structure.

[0075] 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, which here refers to the correlation-based feature vector, i.e., the vector obtained by flattening the upper triangle of the FCN matrix obtained by formula (1), denotes the total number of ROIs.

[0076] For any two nodes and , if there is information flow from to , then , the connection strength between the two nodes is described by the weight , and there is ; if there is no connection, then , where the weight represents the different expressions of constructing the correlation matrix. 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 .

[0077] 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 of the FCN matrix given by formula (1) under the corresponding time window is thresholded.

[0078] Considering the computational cost and computational simplicity, the Pearson correlation coefficient PC between nodes in each time slice is calculated by formula (1) to construct the FCN matrix as the evaluation of the edge weight matrix.

[0079] The correlation coefficient reflects the linear correlation strength between the BOLD signals of node and node in the time slice :

[0080] (1)

[0081] wherein, covariance of the BOLD signals between two ROIs, and standard deviation.

[0082] Therefore, for each individual, the brain functional network in the time period can be represented as Although all the important information between nodes is contained in the functional network, it also includes redundant information and noise information, so the sparse adjacency matrix with strong connection relationship can be formed by screening out the edges with statistical significance and setting the weight of the edge to 1 and the rest to 0. wherein, represents the selected threshold, which is 30% in the present embodiment.

[0083] Considering the efficiency and simplicity of the code, the evaluation of the node feature matrix also uses FCN, and the difference from the edge weight matrix evaluation strategy is that in the evaluation of the node, the FCN matrix is no longer binarized, but the original FCN feature matrix is used to retain comprehensive information.

[0084] In summary, the graph constructed in the tth time window can be represented as as the input of the topology-aware feature extraction, wherein, is the node feature matrix, is the sparse adjacency matrix.

[0085] For each research object, its dynamic graph sequence can be represented as a set including all graphs from the 1st to the time window. Such representation facilitates the analysis of the functional network changes of the brain in different time windows,

[0086] Step 13: Local-global topology-aware feature extraction process.

[0087] To generate node embeddings that can capture both local neighborhood dynamics and global topology 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 the set of dynamic graph sequences 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 limitations of traditional graph neural networks (GNNs) 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 relevance, and injects more rich structural information into node embeddings.

[0088] The local topology-aware encoder LTAE aims to extract local dynamic features from the node raw time series information (non-FCN matrix). Its core is a multi-branch structure including a detail CNN encoder and a topology information supplement part; among them, the processing process of the detail CNN encoder is as follows: first, apply a one-dimensional convolutional neural network (1D-CNN) to each node's time series for local time series modeling, and the convolution process is as follows:

[0089]

[0090] 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.

[0091] Subsequently, max-pooling operation is applied to :

[0092]

[0093] wherein, is the pooling window size, and finally , denotes the local dynamic representation of the th node, is the dimension after pooling.

[0094] The CNN feature matrix is ​​obtained by concatenating all nodes:

[0095]

[0096] The choice of CNN fully utilizes the local topological change characteristics encoded in the original time series, supplements the original fine-grained information, and enhances the model's ability to perceive dynamic neighborhood structures.

[0097] Simultaneously, for the dynamic graph structure within each time window, the Local Topology Aware Encoder (LTAE) introduces a graph topology feature extractor to analyze the static topological information contained within the dynamic graph structure under each time window. Its input is the dynamic adjacency matrix between nodes. This matrix is ​​derived from the original fMRI time-series signal at the 1st... The node connection weight graph is obtained by calculating the Pearson correlation coefficient or mutual information in a sliding window. For each graph... The graph topology feature extractor uses existing graph theory metric calculation methods to extract the following types of topology features for each node:

[0098] Clustering coefficient: measures the degree of closure between the neighborhoods of nodes;

[0099] Small-Worldness: Measures whether a graph simultaneously possesses high clustering and short path properties;

[0100] PageRank value: measures the importance of a node and its potential for information dissemination;

[0101] Eigenvector centrality: measures the degree of connection between a node and nodes with high importance.

[0102] The above metrics are all based on classical graph theory definitions and can be obtained using the NetworkX computing tool, falling within the scope of existing technology. Ultimately, the extracted structural features constitute a graph structure feature matrix. ,in It represents the topological feature dimension, serving as a supplement to the topological information of local information.

[0103] In addition, such as Figure 2 As shown, the Local Topology-Aware Encoder (LTAE) also introduces a fully connected encoder, which employs a shallow multilayer perceptron (MLP) to process the original node features. The MLP undergoes a nonlinear transformation, consisting of two fully connected layers, and incorporates the ReLU activation function and batch normalization. The initial output representation is as follows: This enhances the model's ability to represent the original signal. Finally, the Local Topology-Aware Encoder (LTAE) fuses the three types of features through concatenation and linear transformation to generate a local topology-aware representation. ,in, For the number of subjects, This represents the number of sliding time windows. This represents the final local topology-aware feature dimension.

[0104] like Figure 2 As shown, the Global Topology-Aware Encoder (GTAE) employs a Graph Transformer encoder, which is a collaborative mechanism between the Graph Isomorphic Network (GIN) and the Transformer encoder. This part integrates local structure modeling and global long-range dependency capture to generate a global topology-aware representation. Compared to the limitations of traditional Graph Convolutional Networks (GCNs) that rely on Laplacian smoothing, the Global Topology-Aware Encoder (GIN) accurately characterizes local neighborhood differences and models the topological relationships between nodes in the entire graph through the Transformer's multi-head self-attention mechanism, thereby significantly enhancing the node embedding's ability to perceive graph structure. This architecture uses node features from dynamic graph sequences as input and combines a multi-scale feature fusion strategy to provide robust global representations for complex graph learning tasks.

[0105] The Global Topology Aware Encoder (GTAE) first utilizes multi-layer GIN to analyze the graph features of the input. Perform iterative neighborhood aggregation to capture local topological patterns.

[0106] No.( l The update rules for layer GIN are as follows:

[0107]

[0108] in, Represents a node In the ( l The hidden state of the layer. For nodes The set of adjacent nodes, It is a learnable topology sensitivity coefficient used to dynamically balance the contributions of node's own characteristics and neighborhood aggregation characteristics.

[0109] Composed of two linear layers and a ReLU activation function, its parameterization capability ensures efficient encoding of local topological structures. Compared to traditional GNNs, GINs introduce learnable topology-sensitive coefficients. With multi-layer nonlinear transformation, it has a stronger graph isomorphism discrimination capability and can accurately distinguish the structural differences in local neighborhoods.

[0110] Subsequently, the GIN final output node embedding feature matrix ; is the representation of the th node at the th layer of GIN, i.e., the final layer output; |V| is the number of nodes in the graph; represents the hidden dimension, i.e., the dimension of the representation of each node.

[0111] The output representation of GIN is input into the Transformer encoder to capture the topological dependencies in the full graph range. To enhance the ability of global information integration, GTAE adds a learnable [CLS] token in front of the node feature sequence to form an enhanced input .

[0112] The Transformer encoder models the global association between nodes through a multi-head self-attention mechanism. The calculation process of the

[0113]

[0114] where is the input feature of the l th layer, are the projection matrices of the query, key and value, respectively, where: is the hidden dimension of the Transformer input and output; is the number of attention heads in the multi-head attention mechanism; is the dimension of the key / query / value in each attention head, is the attention weight.

[0115] The self-attention mechanism dynamically allocates weights to enable each node to adaptively focus on other nodes in the full graph, regardless of whether they are directly connected, thereby effectively capturing long-range topological dependencies. After the outputs of all attention heads are concatenated, linearly transformed, residual connected and layer normalized, global topologically aware features are generated.

[0116] To realize the fusion of multi-scale topological information, GTAE adopts a hierarchical feature integration strategy to concatenate the node local features of GIN, the node global interaction features encoded by the Transformer, and the broadcasted [CLS] token features to form a comprehensive feature vector, where [CLS] is a learnable parameter vector with a dimension ofThe parameter vector [CLS] can be updated together with other node features during training. [CLS] is added to the sequence of node features at the beginning of the sequence as the aggregation center of the whole graph-level semantics, and the information of the whole graph is aggregated through the Transformer self-attention mechanism.

[0117] The final global representation of the node is generated by mapping through the fully connected layer :

[0118]

[0119] wherein, represents feature splicing, is a fully connected layer, and the output is , the output dimension of the final global representation, which is not necessarily equal to . represents the embedding dimension required by the downstream task (for example, for classification, clustering, etc.), which is set by the last fully connected layer.

[0120] Through this local and global collaborative modeling method, GTAE not only retains the graph geometric characteristics 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.

[0121] Step 14: Cross-attention fusion of dynamic graph features.

[0122] 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 interactively model the multi-scale features output by the local topological perception encoder (LTAE) and the global topological perception encoder (GTAE) in the time dimension.

[0123] Specifically, for each subject, after the original fMRI signal is divided by a dynamic window and the graph structure is constructed, the LTAE outputs a local topological perception representation: ; wherein, is the batch size, is the number of time windows, is the local representation dimension.

[0124] The GTAE outputs a global topological perception representation , is the global representation dimension; the global representation is broadcasted and expanded along the time dimension to obtain :

[0125]

[0126] where, is a unit vector, generally, is equal to in the previous , so as to achieve replication of each global time slice, so that its dimension is aligned with the local sequence.

[0127] Subsequently, the local representation and the global representation are mapped to the shared attention space , respectively, to obtain the query (Q) vector, i.e. and , the key (K) vector, i.e. and , and the value (V) vector, i.e. and :

[0128]

[0129]

[0130] where, , is a linear projection matrix, is the uniform attention space dimension.

[0131] The cross-attention mechanism takes one side as the query and the other side as the key-value pair to realize the information interaction of two-way representation. A timing causal mask matrix is introduced, represents the number of sliding time windows, where is used to ensure the forward dependency in sequence processing:

[0132] Local to global path Local→Global: taking as the query, the global topology information is fused:

[0133]

[0134] Global to local path Global→Local: taking as the query, the local topology information is fused:

[0135]

[0136] The above attention mechanisms can be stacked in multi-head form (Multi-head Attention) for N layers to enhance the modeling ability of the model for heterogeneous topology information.

[0137] The dual-channel attention output is spliced in the channel dimension, and is fused through linear transformation and layer normalization to obtain the feature :

[0138]

[0139] where, is the projection matrix, is the layer normalization operation, denotes the feature concatenation operation.

[0140] To further stabilize the training, a residual connection is introduced: , is the linear transformation matrix in the residual connection, which serves to adjust the dimensionality of the residual connection. The final feature representation is:

[0141]

[0142] The temporal dimension is averaged-pooled to generate the sample-level graph embedding representation :

[0143]

[0144] where, denotes the number of sliding time windows; is the batch size, is the local representation dimension.

[0145] where, denotes the number of sliding time windows; is the batch size, is the local representation dimension.

[0146] Step 2: Node feature construction based on DTI / clinical information features.

[0147] Step 21: DTI preprocessing and structural connection construction.

[0148] In this step, the original diffusion tensor imaging (DTI) data is standardized and preprocessed to obtain stable and comparable white matter microstructure measurements and construct individual-level structural connections.

[0149] 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 to strictly correspond to the brain region template used by fMRI in space, 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.

[0150] To characterize structural connectivity, a weighted adjacency matrix of individual-level white matter structural network is constructed under the support of fiber tracking (deterministic or probabilistic) or spherical deconvolution model . The preferred weighted definition is:

[0151]

[0152] where, denotes the set of all fiber tracts connecting ROI and ROI , and is the mean of FA (Fractional Anisotropy) over all voxels in the set; denotes the number of fiber tracts connecting ROI i and ROI j . To further improve the stability and comparability of the connectivity graph, the is symmetrized, sparsified (e.g. proportion thresholding or minimum spanning tree + thresholding joint strategy), 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.

[0153] Step 22: Multi-view DTI feature extraction (integration of statistics, texture and graph theory).

[0154] To systematically characterize the integrity, spatial complexity and topological organization of 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).

[0155] (1) Statistical features (intensity distribution within ROI), to quantify local white matter integrity and heterogeneity, including:

[0156] Calculate the mean within each ROI, where, is the total number of voxels within the ROI, denotes the FA value of the th voxel;

[0157] Standard deviation , to measure the dispersion of FA values;

[0158] Skewness , to measure the asymmetry of the distribution;

[0159] Kurtosis , to measure the sharpness of the distribution;

[0160] High FA volume wherein, is a pre-set FA threshold, is an indicator function, when is 1 and 0 otherwise, denotes the volume of a single voxel;

[0161] Entropy wherein, is the probability that the FA value falls in the histogram bin;

[0162] and the three-directional gradient magnitude wherein denote the gradient values of the FA image in x, y, z directions, respectively.

[0163] (2) Texture / morphology features (FA / MD grid level) to characterize the self-similarity and spatial complexity of white matter structures, including: based on the gray level co-occurrence matrix (GLCM) to calculate energy, contrast, homogeneity, correlation and other indicators; based on local binary pattern (LBP) to extract fine-grained texture histogram wherein, is the intensity value of the center pixel, is the intensity value of the neighborhood pixel, is the sign function, when , otherwise , is the number of neighborhood points, is the radius, preferably the radius , neighborhood ; the box counting method is used to estimate the fractal dimension wherein, denotes the number of boxes with side length required to cover the image.

[0164] (3) Graph theory features (based on structural connectivity ):

[0165] Calculate the average degree of the whole network wherein, V is the total number of network nodes, denotes the degree of node ; average shortest path wherein, is the shortest path length between node and ; clustering coefficient wherein, denotes node ​the number of edges actually existing between the neighbors of a node; global efficiency defined as the average of the inverse shortest path length between all pairs of nodes ; local efficiency , which is similar to global efficiency but computed for the local subgraph of each node; betweenness centrality, which measures the frequency of occurrence of a node in all shortest paths, and eigenvector centrality, which measures the influence of a node based on the adjacency matrix eigenvectors; and small-world property , where , , and represent the clustering coefficient and the average shortest path of a random network, respectively, with modularity and homophily added if necessary to enhance the representational capacity of the global-local organization. The above features are aggregated at the individual level and dimensionally unified to form a multi-view feature set .

[0166] Step 23: DTI feature integration, time alignment, and fusion strategy.

[0167] The multi-view feature set is standardized (z-score) and projected to a shared hidden space by a multilayer perceptron to obtain a subject-level structural prior representation, where is the batch size, is the dimension of the structural feature:

[0168]

[0169] is aligned with the fMRI-based temporal embedding, preferably using time dimension expansion to achieve consistent mapping of the time window, where is a unit vector, denotes the outer product operation:

[0170]

[0171] Thus, it can be seamlessly connected with the bidirectional cross-attention module of step 14 to serve as the K / V of the structural prior, interacting with local / global temporal queries.

[0172] In another embodiment, can also participate in structural learning as the edge / hyperedge weight prior of the group graph / hypergraph to realize the constraint and reinforcement of the structural prior on the discrimination of functional dynamics. If there is a clinical covariate, it can be mapped to the same hidden space after projection matrix and spliced and fused with or the output of step 14, and linearly transformed to , which is input into the group hypergraph / graph convolution classifier to complete the downstream task.

[0173] Step 24: Feature preprocessing and feature splicing of clinical information.

[0174] To introduce stable contextual priors and improve the robustness of the model in discriminating boundary samples, this step encodes and embeds the clinical information into a low-dimensional space, and concatenates it with the image features obtained in the previous step.

[0175] Let the original clinical variable matrix be wherein, represents the number of original clinical fields before processing, which includes cognitive scale scores (such as MMSE, MoCA, CDR, etc.), disease duration, education years, ApoE genotype, age, and gender information. First, fill in the missing values and suppress outliers: z-score standardization for continuous variables, one-hot encoding or learnable embedding for categorical variables, and finally concatenate all processed variables to obtain wherein, represents the total feature dimension after processing. Then, a lightweight multilayer perceptron is used to map to a low-dimensional hidden space to obtain the subject-level clinical feature embedding:

[0176]

[0177] Finally, the fusion node features for subsequent group graph / hypergraph learning are constructed at the subject level. The feature is directly concatenated by three parts in the feature dimension:

[0178]

[0179] wherein: is the feature from step 14 after time pooling of dynamic functional connectivity graph. is the structural white matter feature from step 23 DTI feature integration. is the clinical feature embedding obtained in this step; the three are aligned in a unified hidden space and used as node features to input the subsequent group hypergraph / graph structure learning and classification module.

[0180] Step 3: Hypergraph construction process.

[0181] In this phase, the three subject-level representations obtained upstream are input: the functional dynamic representation obtained by "local-global" encoding and bidirectional cross-attention fusion of dynamic window sequences, time pooling, DTI structural prior representation , and clinical embedding features .

[0182] Then, to construct the hyperedge relationship, the similarity between nodes is modeled in the "semantic fusion space". First, map to a unified hypergraph embedding space : is the dimension of the hypergraph embedding space. Specifically, denotes the dimension size used when mapping the input features into this space, where is a learnable linear projection weight matrix. Then, a Gaussian kernel function is defined to measure the similarity between nodes and in this space:

[0183]

[0184] where denotes the semantic embedding vector of the th node, is the kernel bandwidth, which is usually adaptively estimated by the median of the Euclidean distance between all node pairs. According to the similarity matrix in , the top most similar nodes in the feature space of each node are selected to form the hyperedge :

[0185]

[0186] The hyperedge set of all nodes is merged to obtain the final hyperedge set , and the corresponding total number of hyperedges is denoted as .

[0187] Next, a probabilistic hypergraph affinity matrix is constructed, which is the th row and the th column of the matrix denotes the membership strength of the hyperedge :

[0188]

[0189] To improve the stability of subsequent hypergraph convolution, the probabilistic hypergraph affinity matrix is symmetrically normalized. Specifically, first define the node degree matrix and the hyperedge degree matrix , where , are unit column vectors.

[0190] The normalized hypergraph affinity matrix is:

[0191]

[0192] At the same time, the hypergraph Laplacian matrix is constructed:

[0193]

[0194] 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.

[0195] 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.

[0196] 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.

[0197] 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 has an overall computational and storage cost that is linearly related to the number of non-zero elements, making it suitable for medium-sized datasets with a subject count B ranging from tens to hundreds.

[0198] Step 4: Hypergraph Neural Network Training.

[0199] After completing multimodal feature fusion to obtain the node feature matrix After constructing the group hypergraph structure, this stage involves end-to-end model training and optimization.

[0200] First, based on the hypergraph association matrix Calculate the node degree matrix With hypermarginal matrix ， And a learnable hyperedge weight matrix is ​​introduced. Finally, a normalized hypergraph adjacency matrix is ​​constructed. : ,in, is a learnable hyperedge weight matrix.

[0201] The feature matrix is input into the attention mechanism-equipped hypergraph neural network, which realizes feature propagation through multi-layer attention hypergraph convolution: each layer first performs linear transformation on the node features to obtain , calculates the attention score between nodes , and obtains the attention weight by softmax normalization. The weight is used to dynamically adjust the hypergraph adjacency matrix to obtain , where represents element-wise multiplication (Hadamard product), is the attention weight matrix.

[0202] Finally, the feature update with residual connection is realized through the feature propagation formula , where is a learnable scaling factor, is the hypergraph adjacency matrix. The hypergraph neural network model adopts the He initialization method parameter, uses the Adam optimizer for training, and dynamically adjusts the learning rate through a multi-step learning rate scheduler.

[0203] 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, represents the total number of training samples, represents the number of classes, is the weighted coefficient of the th class, is the sample frequency of the class in the training set, is the sample real one-hot label value in the th class, is the predicted probability output by the hypergraph neural network model.

[0204] During the training process, the model performance is evaluated on the validation set every round, and the optimal parameters are saved for final testing. 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.

[0205] ​The Alzheimer's disease classification method based on topology-aware local-global joint coding 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. The method innovatively integrates multi-modal brain image data and clinical information to construct an end-to-end intelligent diagnosis framework, providing reliable technical support for early and accurate identification of AD. In practical application, 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, the 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.

[0206] It is particularly emphasized 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 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, enabling consensus decision-making at the group level by constructing high-order similarity relationships between subjects, greatly enhancing 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.

[0207] With the continuous development of medical artificial intelligence technology, this method has shown a broad application space. Its 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.

[0208] Embodiment Two

[0209] The purpose of this embodiment is to provide an Alzheimer's disease classification system based on topology awareness and group hypergraph, which includes:

[0210] a processing module configured to extract a time sequence from resting-state functional magnetic resonance imaging data, and construct a dynamic functional connectivity network sequence through a sliding window strategy;

[0211] a extracting module configured to extract local topology features and global topology features of each time window by using a local topology-aware encoder and a global topology-aware encoder respectively, and generate a comprehensive feature representation of a subject level by deep interaction and fusion of local topology-aware representation and global topology-aware representation of each time window;

[0212] a constructing and classifying module configured to take each subject as a hypergraph node, construct a hyperedge based on a feature similarity calculated from the comprehensive feature representation of the subject level in combination with diffusion tensor imaging features and clinical information features of the subject, and further construct a group hypergraph, and obtain an Alzheimer's disease classification result of the subject by processing the group hypergraph using a hypergraph neural network.

[0213] In more embodiments, there are also provided:

[0214] An electronic device includes a memory and a processor, and computer instructions stored on the memory and run 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.

[0215] It should be understood that in the embodiments, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, ready-to-program gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0216] The memory can include read-only memory and random access memory, and provide instructions and data to the processor, and a part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0217] A computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method described in embodiment one is completed.

[0218] The method in the embodiment one can be directly embodied by a hardware processor or 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 a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or the like. The storage medium is located in a memory, and a processor reads information in the memory and combines hardware to complete the steps of the above method. To avoid repetition, no further detailed description is given here.

[0219] Although the specific embodiments of the present application are described above with reference to the drawings, the description is not a limitation to the scope of protection of the present application, and those skilled in the art should understand that various modifications or changes made 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. An Alzheimer's disease classification method based on topological perception and population hypergraph, characterized in that, include: Time series data were extracted from resting-state functional magnetic resonance imaging data, and dynamic graphs were constructed using a sliding window strategy. Local and global topological features for each time window are extracted using local and global topological encoders, respectively. The local and global topological representations for each time window are then deeply interacted and fused to generate a comprehensive feature representation at the subject level. Specifically: The local and global topology-aware representations are projected onto a shared attention space, and the corresponding query vector, key vector, and value vector are obtained through linear transformation. The query vector is represented by a local topology-aware representation and then fused with a global topology-aware representation. The query vector is represented by a global topology-aware representation, which is then fused with a local topology-aware representation. The outputs of the dual-path attention mechanism are concatenated in the channel dimension, fused with linear transformation and layer normalization, and residual connections are introduced to obtain the final feature representation. The final feature representation is subjected to average pooling in the time dimension to obtain the comprehensive feature representation at the subject level; Using each subject as a hypergraph node, a hyperedge and a probabilistic hypergraph association matrix are constructed based on the comprehensive feature representation at the subject level, the feature similarity calculated by combining the subject's diffusion tensor imaging structure prior representation and clinical embedding features, and then a population hypergraph is constructed. The population hypergraph is then processed using a hypergraph neural network to obtain the Alzheimer's disease classification results for the subjects.

2. The Alzheimer's disease classification method based on topological perception and population hypergraph as described in claim 1, characterized in that, The local topological features of each time window are extracted using a local topology-aware encoder, specifically as follows: For the original temporal information of each node, fine-grained dynamic pattern features of the local neighborhood are extracted through a one-dimensional convolutional neural network. Local temporal features of the node are extracted, and max pooling is performed on the local temporal features of the node to obtain the local dynamic representation of the node. The local dynamic representations of each node are concatenated to obtain the CNN feature matrix. For the dynamic adjacency matrix between nodes in each time window, the graph topology features are calculated using a graph topology feature extractor to obtain the graph structure feature matrix; The raw temporal information of each node is processed by a multilayer perceptron to generate an initial feature representation; The CNN feature matrix, graph structure feature matrix, and initial feature representation are concatenated and fused using linear transformation to obtain local topological features.

3. The Alzheimer's disease classification method based on topological perception and population hypergraph as described in claim 1, characterized in that, The global topological features for each time window are extracted using a global topology-aware encoder, specifically as follows: Graph isomorphic networks are used to iteratively aggregate neighborhoods in dynamic graphs to capture local topological patterns and obtain local node features. The node embedding feature matrix is ​​input into the Transformer encoder to capture the topological dependencies across the entire graph and obtain the global interaction features of the nodes. The local features of nodes, the global interaction features of nodes, and the labeled features of broadcasts are concatenated and then mapped through a fully connected layer to generate global topological features.

4. The Alzheimer's disease classification method based on topological perception and population hypergraph as described in claim 1, characterized in that, The statistical features, texture features, morphological features, and graph theory features extracted from the diffusion tensor imaging data are used as the prior representation of the diffusion tensor imaging structure. The statistical features include the mean, standard deviation, skewness, peak value, entropy, triaxial gradient magnitude, and volume of the anisotropy score within the region of interest. The graph theory features include the average degree, average shortest path, clustering coefficient, global efficiency, local efficiency, betweenness centrality, eigenvector centrality, and small-world properties of the individual-level white matter structure network. The extraction of the clinical embedding features specifically involves: preprocessing the original clinical variable matrix and then using a perceptron to perform a linear-nonlinear mapping to obtain the clinical embedding features.

5. The Alzheimer's disease classification method based on topological sensing and population hypergraph as described in claim 1, characterized in that, Based on the comprehensive feature representation at the subject level, combined with the subject's diffusion tensor imaging structural prior representation and clinical embedding features, a hyperedge and probabilistic hypergraph correlation matrix is ​​constructed, specifically as follows: The comprehensive feature representation of the subject level, the prior representation of the diffusion tensor imaging structure, and the clinical embedding features are concatenated and projected into the hypergraph embedding space; the similarity of different nodes in the hypergraph embedding space is calculated based on the Gaussian kernel function. Based on the calculated similarity, the top-ranked nodes are selected in the hypergraph embedding space of each node. The most similar nodes form a superedge; Construct a probabilistic hypergraph incidence matrix and normalize it. Then, construct the hypergraph Laplacian matrix based on the normalized probabilistic hypergraph incidence matrix.

6. The Alzheimer's disease classification method based on topological perception and population hypergraph as described in claim 1, characterized in that, The weighted cross-entropy loss function based on the inverse square root of the class frequency is used as the loss function for training the hypergraph neural network.

7. An Alzheimer's disease classification system based on topological sensing and population hypergraph, characterized in that, include: The processing module is configured to: extract time series data from resting-state functional magnetic resonance imaging data and construct dynamic functional connectivity network sequences using a sliding window strategy; The extraction module is configured to: extract local and global topological features for each time window using both a local and global topological encoder; and perform deep interaction and fusion of the local and global topological representations for each time window to generate a comprehensive feature representation at the subject level. Specifically: The local and global topology-aware representations are projected onto a shared attention space, and the corresponding query vector, key vector, and value vector are obtained through linear transformation. The query vector is represented by a local topology-aware representation and then fused with a global topology-aware representation. The query vector is represented by a global topology-aware representation, which is then fused with a local topology-aware representation. The outputs of the dual-path attention mechanism are concatenated in the channel dimension, fused with linear transformation and layer normalization, and residual connections are introduced to obtain the final feature representation. The final feature representation is subjected to average pooling in the time dimension to obtain the comprehensive feature representation at the subject level; The construction and classification module is configured as follows: taking each subject as a hypergraph node, constructing hyperedges and probabilistic hypergraph association matrices based on the comprehensive feature representation at the subject level, the feature similarity calculated by combining the subject's diffusion tensor imaging features and clinical information features, and then constructing a population hypergraph. The population hypergraph is then processed using a hypergraph neural network to obtain the Alzheimer's disease classification results for the subjects.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-5.

Images