Alzheimer's disease classification and key brain area determination method based on random attention and counterfactual contrastive learning
By optimizing the graph convolutional network model through random attention and counterfactual contrastive learning, the black box nature of decision-making and unreliable interpretation in existing Alzheimer's disease classification models are solved, achieving credible interpretation and high-accuracy classification of key brain regions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DALIAN UNIV OF TECH
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-10
AI Technical Summary
Existing Alzheimer's disease classification models based on graph neural networks suffer from problems such as black box decision-making, unreliable ex post facto interpretations, and a lack of validation loops for interpretation validity. They are difficult to intuitively present the specific brain regions and functional connections of abnormal classifications, and the interpretation results are difficult for clinicians to accept in medical scenarios.
We employ a method based on stochastic attention and counterfactual contrastive learning. By acquiring functional magnetic resonance imaging data, we construct a brain network graph and use graph neural networks to jointly model the graph structure and node features. By combining stochastic brain region node selection and counterfactual node masking, we calculate cross-entropy, fact-counterfactual contrastive loss, and node-level KL loss to optimize the graph convolutional network model and determine key brain region nodes.
It significantly improves the classification accuracy and stability of the model. The key brain region nodes output are highly consistent with clinical medical priors, and can intuitively present the specific brain regions and functional connections that lead to abnormal classifications. This is consistent with the real pathological mechanism and enhances the credibility of the interpretation results.
Smart Images
Figure CN122067028B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image analysis technology, and in particular to a method for classifying Alzheimer's disease and identifying key brain regions based on random attention and counterfactual contrast learning. Background Technology
[0002] In the early diagnosis of Alzheimer's disease (AD), brain functional network analysis based on functional magnetic resonance imaging (fMRI) data combined with graph neural networks (GNNs) has become the mainstream approach. However, existing GNN-based classification models have significant drawbacks:
[0003] Decision-making black box: Traditional deep learning models have a highly nonlinear structure and often only focus on macro classification indicators (such as accuracy), making it difficult to intuitively present the specific brain regions and functional connections that lead to abnormal classifications.
[0004] Post-hoc interpretations are unreliable: existing interpretability methods often infer judgment criteria by backtracking gradients or attention weights after the model training is fixed. This kind of "post-hoc interpretation" lacks explicit constraints during the training process, and the "key brain regions" identified may just be data shortcuts (spurious correlations) and do not reflect the actual pathological mechanisms.
[0005] The validity of the explanation lacks a closed-loop verification mechanism: Existing technologies cannot verify within the same framework whether the model is truly highly dependent on the selected key brain regions, making it difficult for clinicians to accept the explanation results in medical scenarios where safety requirements are extremely high. Summary of the Invention
[0006] Therefore, it is necessary to propose a method for classifying Alzheimer's disease and identifying key brain regions based on random attention and counterfactual contrast learning to address the above problems.
[0007] A method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrast learning, the method comprising:
[0008] The functional magnetic resonance imaging (fMRI) data of the subjects were acquired, and the fMRI data were preprocessed to obtain target fMRI data. The target fMRI data was then divided into standard brain regions, and the time series of each region of interest was extracted. The region of interest is a regional unit divided according to the standard brain region division template, and each region of interest corresponds to a brain region node.
[0009] The functional connectivity strength between brain region nodes is calculated based on the extracted time series data, and the functional connectivity matrix, adjacency matrix, edge set between brain region nodes, and node feature matrix are calculated based on the functional connectivity strength. The brain region node set is determined based on the functional magnetic resonance imaging data. The adjacency matrix, edge set between brain region nodes, node feature matrix, and brain region node set constitute a brain network graph.
[0010] The brain network graph is input into a graph neural network encoder to perform joint modeling of graph structure information and node feature information to obtain a basic node embedding representation. A node selection module composed of a multilayer perceptron is introduced to map the basic node embedding representation to a set of importance probabilities of brain region nodes in the first interval. The set of importance probabilities is composed of each importance probability vector, and the importance probability vector is the probability of a brain region node being selected as a key node.
[0011] Based on the importance probability vector in the node importance probability set, Bernoulli sampling is performed independently on each brain region node to generate a corresponding binary mask; multiple binary masks constitute a random brain region node selection mask;
[0012] A discriminant subgraph adjacency matrix is obtained by structural gating using a random brain region node selection mask docking matrix; the discriminant subgraph edge set is determined from the discriminant subgraph adjacency matrix, and the discriminant subgraph adjacency matrix, the discriminant subgraph edge set, the brain region node set, and the node feature matrix are then used. Construct a discriminant subgraph;
[0013] A graph-level representation vector is obtained by sequentially extracting features from the discriminant subgraph and aggregating graph-level readout functions through a graph neural network encoder. The graph-level representation vector is then input into a classification predictor and mapped to the category probability space to obtain the classification result. The cross-entropy classification loss is calculated based on the classification prediction result and the true category label.
[0014] The probability values are directly selected from the top-K values using Top-K calculation. The set of brain region nodes is used as the set of explanatory nodes, and a counterfactual node mask complementary to the set of explanatory nodes is constructed based on it. The counterfactual node feature matrix, counterfactual adjacency moments, and counterfactual subgraph edge set are determined based on the counterfactual node mask. The set of brain region nodes, the set of counterfactual subgraph edges, the counterfactual adjacency moments, and the counterfactual node feature matrix constitute a counterfactual subgraph. The fact-counterfactual contrast loss is calculated based on the discriminative subgraph and the counterfactual subgraph.
[0015] The node-level KL loss is determined based on the importance probability vector and the total number of nodes in the brain region;
[0016] The total training objective loss is determined based on the cross-entropy classification loss, the fact-counterfact comparison loss, and the node-level KL loss. The graph convolutional network model is then trained using the total training objective loss to obtain an optimized graph convolutional network model.
[0017] The connection relationships and node features corresponding to the explanatory node set are extracted from the brain network graph, and the induced edge set of the explanatory node set in the brain network graph is determined. The connection relationships, the node features, the explanatory node set and the induced edge set constitute a sample-level explanatory subgraph.
[0018] Based on the sample-level explanatory subgraph, the frequency of explanatory node sets for samples of the same category is trained and statistically analyzed by optimizing the graph convolutional network model. This determines the frequency of occurrence of each brain region node in that category of samples. The higher the frequency, the stronger the brain region node. The more consistently a sample in this category is identified as a key brain region node by the optimized graph convolutional network model, the better.
[0019] In one embodiment, the calculation of functional connectivity strength between brain region nodes based on the extracted time series, and the calculation of the functional connectivity matrix, adjacency matrix, edge set between brain region nodes, and node feature matrix based on the functional connectivity strength, are achieved through the following expression:
[0020]
[0021]
[0022]
[0023] in, It is a time series; This represents the fMRI signal value of the i-th region of interest at time point T; For functional connection matrix; The region of interest (ROI) represents the functional connectivity strength between region of interest (ROI) i and region of interest j, and N represents the total number of ROIs. and Let i and j represent the fMRI signals of regions of interest (ROI) and ROI at time point t, respectively. and These represent the mean of the corresponding time series, and T represents the total number of time points; This represents the functional connectivity strength between region of interest i and region of interest j.
[0024] In one embodiment, the calculation of the adjacency matrix based on the functional connectivity strength is achieved through the following expression:
[0025]
[0026]
[0027] in, It is an adjacency matrix; This represents the functional connectivity between brain region node i and brain region node j. The functional connectivity strength between region of interest i and region of interest j; The threshold used to control the sparsity of the adjacency matrix.
[0028] In one embodiment, the calculation of the edge set and node feature matrix between brain region nodes based on the functional connectivity strength is described. This can be achieved using the following expression:
[0029]
[0030]
[0031]
[0032] in, and Representing the sets of brain region nodes respectively The i-th and j-th brain region nodes in the brain; Let be the node features of the i-th brain region node; The node feature matrix; The functional connectivity strength between the i-th region of interest and the N-th region of interest.
[0033] In one embodiment,
[0034] The embedding of basic nodes represents Mapped to the set of importance probabilities of brain region nodes in the first interval This can be achieved using the following expression:
[0035]
[0036]
[0037] in, Based on the embedding representation of the basic node, The hidden dimension embedded in the node; For graph neural network encoders; These are the learnable parameters of the graph neural network encoder; For the Sigmoid function; It is a multilayer perceptron; Select learnable parameters for the module to be used in the node.
[0038] In one embodiment,
[0039] The use of random brain region node selection mask docking matrix Structural gating is performed to obtain the adjacency matrix of the discriminant subgraph; the edge set of the discriminant subgraph is determined by the adjacency matrix of the discriminant subgraph, and the set of edges of the discriminant subgraph, the set of brain region nodes, and the node feature matrix constitute the discriminant subgraph, which is implemented by the following expression:
[0040]
[0041]
[0042]
[0043] in, To determine the adjacency matrix of a subgraph; Select a mask for random brain region nodes; It is an adjacency matrix; This is element-wise multiplication; and Represents the set of brain region nodes The first in The brain region node and the first Each brain region node; Represents the adjacency matrix of the discriminant subgraph The Middle Line number Column elements; To distinguish subgraphs; A set of brain region nodes; The node feature matrix; To distinguish subgraphs;
[0044] The graph-level representation vector is obtained by sequentially extracting features from the discriminant subgraph and aggregating graph-level readout functions through a graph neural network encoder. This graph-level representation vector is then input into a classification predictor, and the classification result is obtained by mapping it to the category probability space using the following expression:
[0045]
[0046]
[0047] in, Vectors are represented at the graph level; For graph-level readout functions; For graph neural network encoders, These are learnable parameters that are fully shared with the original model. For classification predictors; Its learnable parameters; The graph convolutional network module predicts the category based on the output of the discriminant subgraph, which is Alzheimer's disease (AD), normal cognitive impairment (NC), or mild cognitive impairment (MCI).
[0048] The formula for calculating the cross-entropy classification loss is as follows:
[0049]
[0050] in, Cross-entropy is used for classification loss; Indicates the number of categories; Indicates the true label in the first place Indicator value on class; Indicates the prediction result in the first... Probability output on the class.
[0051] In one embodiment,
[0052] The counterfactual node feature matrix, counterfactual adjacency moments, and counterfactual subgraph edge set are implemented by the following expression:
[0053]
[0054]
[0055]
[0056]
[0057]
[0058]
[0059] in, To explain the set of nodes; V is the importance probability vector; V is the set of brain region nodes; The preset explanation of the budget parameters; For counterfactual node masks; The counterfactual node mask for the Nth brain region node; and Represents the set of brain region nodes The first in The brain region node and the first Each brain region node; The feature matrix of the counterfactual nodes; It is a counterfactual adjacency matrix; For counterfactual node mask; The node feature matrix; It is an adjacency matrix; This represents element-wise multiplication. This represents the structure-gated matrix composed of counterfactual node masks;
[0060] The calculation of the fact-counterfactual contrast loss based on the discriminant subgraph and the counterfactual subgraph is achieved by the following expression:
[0061]
[0062]
[0063]
[0064] in, Vectors are represented at the graph level; For graph-level readout functions; This represents a graph neural network encoder; This represents the learnable parameters that are fully shared with the original model; For counterfact subgraphs; A graph-level representation of a counterfactual subgraph; For fact-counterfact comparison loss; Represents the Sigmoid function; This represents the similarity measurement function.
[0065] In one embodiment, the determination of the node-level KL loss based on the importance probability vector and the total number of brain region nodes is achieved by the following expression:
[0066]
[0067] in, For node-level KL loss; This represents the total number of nodes in the brain region. For the first The probability of a brain region node being selected as a key node; The preset node retention ratio is a fixed hyperparameter.
[0068] In one embodiment, the determination of the total training objective loss based on the cross-entropy classification loss, the fact-counterfactual comparison loss, and the node-level KL loss is achieved by the following expression:
[0069]
[0070] in, Loss due to overall training objective; Cross-entropy is used for classification loss; For fact-counterfact comparison loss; For node-level KL loss; The weighting coefficients represent the loss in the fact-counterfact comparison.
[0071] In one embodiment, the frequency of occurrence of each brain region node in this category of samples is expressed by the following expression:
[0072]
[0073] in, For the number of training sessions; Indicates the number of samples in this category; For indicator functions; The frequency of occurrence of each brain region node in this category of samples. The larger the value, the more likely it is to be a brain region node. The more consistently the optimized graph convolutional network model identifies a node as a key brain region in this category of samples; For the first The set of explanatory nodes obtained from the training statistics .
[0074] In the AD vs. NC task on the real ADNI dataset, this invention optimizes the graph convolutional network model, which significantly outperforms existing GNN baseline models in terms of accuracy, overall performance, and stability. The key brain region nodes output (such as the hippocampus and posterior cingulate cortex) are highly consistent with clinical medical priors. It can intuitively present the specific brain regions and functional connections that lead to abnormal classification. While conforming to the real pathological mechanism, it verifies within the same framework whether the model is truly highly dependent on the selected key brain region structures, so that the interpretation results can be accepted by clinicians in medical scenarios with extremely high safety requirements. Attached Figure Description
[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0076] in:
[0077] Figure 1 This is a flowchart of an Alzheimer's disease classification and key brain region identification method based on random attention and counterfactual contrast learning in one embodiment;
[0078] Figure 2 This is a data flow graph for optimizing a graph convolutional network model in one embodiment;
[0079] Figure 3 This is a heatmap showing the importance of brain region nodes in one embodiment;
[0080] Figure 4 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation
[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0082] like Figure 1 and Figure 2 As shown, in one embodiment, a method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning specifically includes the following steps:
[0083] S10: Acquire the subject's functional magnetic resonance imaging (fMRI) data, and preprocess the fMRI data to obtain target fMRI data; divide the target fMRI data into standard brain regions, and extract the time series of each region of interest. The region of interest is a regional unit divided according to the standard brain region division template, and each region of interest corresponds to a brain region node.
[0084] S20: Calculate the functional connectivity strength between brain region nodes based on the extracted time series, and calculate the functional connectivity matrix, adjacency matrix, edge set between brain region nodes, and node feature matrix based on the functional connectivity strength; determine the brain region node set based on the functional magnetic resonance imaging data; the adjacency matrix, edge set between brain region nodes, node feature matrix, and brain region node set constitute a brain network graph;
[0085] S30: Input the brain network graph into the graph neural network encoder to perform joint modeling of graph structure information and node feature information to obtain the basic node embedding representation; introduce a node selection module composed of a multilayer perceptron to map the basic node embedding representation to the importance probability set of brain region nodes in the first interval, the importance probability set is composed of each importance probability vector, the importance probability vector is the probability of a brain region node being selected as a key node;
[0086] S40: Based on the importance probability vector in the node importance probability set, Bernoulli sampling is performed independently on each brain region node to generate a corresponding binary mask; multiple binary masks constitute a random brain region node selection mask;
[0087] S50: A discriminant subgraph adjacency matrix is obtained by structural gating using a random brain region node selection mask docking matrix; the discriminant subgraph edge set is determined from the discriminant subgraph adjacency matrix. The discriminant subgraph adjacency matrix, the discriminant subgraph edge set, the brain region node set, and the node feature matrix are then used to determine the discriminant subgraph edge set. Construct a discriminant subgraph;
[0088] S60: The graph neural network encoder sequentially extracts features from the discriminant subgraph and aggregates graph-level readout functions to obtain a graph-level representation vector. The graph-level representation vector is then input into a classification predictor and mapped to the category probability space to obtain a classification result. Based on the classification prediction result and the true category label, the cross-entropy classification loss is calculated.
[0089] S70: Directly select the top probability values using Top-K calculation. The set of brain region nodes is used as the set of explanatory nodes, and a counterfactual node mask complementary to the set of explanatory nodes is constructed based on it. The counterfactual node feature matrix, counterfactual adjacency moments, and counterfactual subgraph edge set are determined based on the counterfactual node mask. The set of brain region nodes, the set of counterfactual subgraph edges, the counterfactual adjacency moments, and the counterfactual node feature matrix constitute a counterfactual subgraph. The fact-counterfactual contrast loss is calculated based on the discriminative subgraph and the counterfactual subgraph.
[0090] S80: Determine the node-level KL loss based on the importance probability vector and the total number of nodes in the brain region;
[0091] S90: Determine the total training objective loss based on the cross-entropy classification loss, fact-counterfact comparison loss, and node-level KL loss, and train the graph convolutional network model using the total training objective loss to obtain an optimized graph convolutional network model;
[0092] S100: Extract the connection relationships and node features corresponding to the explanatory node set from the brain network graph, and determine the induced edge set of the explanatory node set in the brain network graph. The connection relationships, the node features, the explanatory node set, and the induced edge set constitute a sample-level explanatory subgraph.
[0093] S110: Based on the sample-level explanatory subgraph, frequency training statistics are performed on the set of explanatory nodes for samples of the same category by optimizing the graph convolutional network model, thereby determining the frequency of occurrence of each brain region node in samples of that category. The higher the frequency, the stronger the brain region node. The more consistently the optimized graph convolutional network model identifies a node as a key brain region (such as the hippocampus, posterior cingulate cortex, etc.) in this category of samples, the more likely it is to be identified.
[0094] Specifically, functional magnetic resonance imaging (fMRI) data of the subjects were acquired to form a raw image sample set for subsequent brain network construction. The raw image sample set includes fMRI data of 480 subjects, including 160 Alzheimer's disease (AD) patients, 160 patients with mild cognitive impairment (MCI) and 160 cognitively normal controls.
[0095] The functional magnetic resonance imaging (fMRI) data is preprocessed to obtain target fMRI data, thereby improving image data quality, enhancing spatial consistency among different subjects, and providing standardized input for subsequent brain region segmentation, time-series extraction, and brain network construction. The process is as follows:
[0096] (1) Initial time point elimination: The first 10 time points of the functional magnetic resonance imaging (fMRI) scan are eliminated to eliminate field drift and noise interference caused by the magnetic field of the magnetic resonance instrument not being stable in the initial scanning stage, so that the data sequence entering the subsequent analysis reaches a stable signal baseline.
[0097] (2) Time-layer correction and head motion correction: Since functional magnetic resonance imaging data are collected sequentially by slice, there are sampling differences in time between different slices. Therefore, time-layer correction is performed on all slices with a specified slice as a reference to achieve alignment in the time dimension. At the same time, the head motion generated during the scanning process of the subject is corrected to reduce the impact of motion artifacts on subsequent temporal correlation analysis.
[0098] (3) Spatial standardization and resampling: The corrected functional magnetic resonance imaging data were nonlinearly registered to the standard space of the Montreal Neuroscience Institute (MNI) and resampled using a uniform voxel resolution of 3 mm × 3 mm × 3 mm to improve the comparability of brain images of different subjects in a uniform space, which facilitates subsequent brain network construction and population statistical analysis.
[0099] (4) Spatial smoothing: The standardized functional magnetic resonance imaging data were smoothed using a three-dimensional Gaussian kernel. A smoothing kernel of 4mm full width at half maximum (FWHM) was used to effectively reduce random noise and improve the stability of statistical results.
[0100] After the above standardized preprocessing, higher quality, more standardized structure and cross-subject comparability of target functional magnetic resonance imaging data are obtained, which serve as input for subsequent brain segmentation and brain network sample construction.
[0101] For each subject, the time series based on the i-th region of interest (ROI) is as follows:
[0102]
[0103] in, It is a time series; This represents the fMRI signal value of the i-th region of interest at time point T.
[0104] The functional connectivity strength between any two regions of interest (ROIs) is calculated using the Pearson correlation coefficient, and the functional connectivity matrix R is constructed as follows:
[0105]
[0106]
[0107] in, For functional connection matrix; The region of interest (ROI) represents the functional connectivity strength between region of interest (ROI) i and region of interest j, and N represents the total number of ROIs. and Let i and j represent the fMRI signals of regions of interest (ROI) and ROI at time point t, respectively. and These represent the mean of the corresponding time series, and T represents the total number of time points; This represents the functional connectivity strength between region of interest i and region of interest j.
[0108] Fisher-Z transform and adaptive thresholding are applied to the functional connectivity matrix R to generate adjacency matrices corresponding to the functional connectivity relationships between brain region nodes. :
[0109]
[0110]
[0111] in, It is an adjacency matrix; This represents the functional connectivity between brain region node i and brain region node j. The functional connectivity strength between region of interest i and region of interest j; The threshold used to control the sparsity of the adjacency matrix.
[0112] Based on the adjacency matrix Further determine the set of edges between brain region nodes. :
[0113]
[0114] in, and Representing the sets of brain region nodes respectively The i-th and j-th brain region nodes are represented. The set of edges E between brain region nodes is used to characterize the functional connections preserved between brain region nodes.
[0115] Simultaneously, the i-th row of the functional connectivity matrix R is used as the feature vector of the i-th brain region node to obtain the node feature matrix. :
[0116]
[0117]
[0118] in, Let be the node features of the i-th brain region node; The node feature matrix; The functional connectivity strength between the i-th region of interest and the N-th region of interest.
[0119] Through the brain region node set corresponding to each subject Edge set Adjacency matrix and node feature matrix This allows for the construction of connection-based brain network maps corresponding to individual subjects. :
[0120]
[0121] in, A set of brain region nodes, Let the set of edges between brain region nodes be defined by the adjacency matrix. This is the adjacency matrix corresponding to the functional connectivity relationships between brain regions. This is the node feature matrix. The brain network diagram... It serves as input for subsequent graph neural network encoding, node importance estimation, discriminative subgraph construction, interpretive subgraph generation, and counterfactual subgraph verification.
[0122] The subjects' functional magnetic resonance imaging (fMRI) data were used to construct a connected brain network map G=(V,E,A,X). Here, the set of brain region nodes V represents the region of interest (ROI), the set of edges between multiple brain region nodes is the edge set E, and the adjacency matrix is used. Represents the strength of functional connectivity between brain region nodes (constructed based on Pearson correlation coefficient). The node feature matrix (the definition of parameters in the brain network graph can be deleted); the connection-based brain network graph Input a graph neural network encoder to perform joint modeling of graph structure information and node feature information, and obtain the basic node embedding representation. :
[0123]
[0124] in, Based on the embedding representation of the basic node, The hidden dimension embedded in the node; For graph neural network encoders; These are the learnable parameters of the graph neural network encoder.
[0125] A node selection module composed of a multilayer perceptron (MLP) is introduced to embed basic nodes into the representation. Mapped to the first interval Importance probability vector of brain region nodes :
[0126]
[0127] in, , , For the Sigmoid function; It is a multilayer perceptron; Learnable parameters for selecting modules for nodes. Importance probability vectors of brain region nodes. elements in The physical meaning is: the first The probability of a brain region node being selected as a key node; MLP stands for Multilayer Perceptron.
[0128] Bernoulli random sampling: To avoid structural overfitting to a few fixed brain regions during training, graph convolutional network models use importance probability vectors of brain region nodes. elements in Bernoulli sampling is performed independently on each brain region node to generate the first... Brain region nodes binary mask The binary masks corresponding to all brain region nodes constitute the random brain region node selection mask. ;
[0129]
[0130]
[0131] Among them, when When, it indicates the first Each brain region node was selected into the discriminant subgraph during this forward propagation; when When the node's feature matrix is not selected into the discriminant subgraph during this forward propagation, it indicates that the node's feature matrix will not be selected.
[0132] The significance of introducing randomness lies in the fact that during model training, if a deterministic node selection strategy is adopted (such as directly and rigidly selecting the local nodes with the highest probability), the model is prone to getting trapped in local optima, focusing only on and becoming fixed on a few specific combinations of brain regions, thus leading to severe structural overfitting. Conversely, injecting random perturbations into the brain network topology space through Bernoulli sampling forces the model to continuously explore and evaluate different combinations of brain region substructures based on probability during each forward propagation. This proactive exploration mechanism effectively expands the search space for potentially pathogenic brain networks, preventing the model from over-relying on a few fixed neural connections, thereby ensuring that the final extracted discriminative subgraph has a stronger global perspective and generalization robustness when facing unknown data.
[0133] Discriminant subgraph ( Construction: Using a random brain region node selection mask docking matrix Structural gating was performed, preserving the probability values with higher values. Edges associated with corresponding brain region nodes (i.e., edges formed by high-importance brain region nodes) suppressed edges associated with low-importance brain region nodes:
[0134] in, To determine the adjacency matrix of a subgraph; Select a mask for random brain region nodes; This is the adjacency matrix corresponding to the functional connectivity relationships between nodes in brain regions; This is element-wise multiplication.
[0135] From the adjacency matrix Determine the set of edges of the discriminant subgraph ,in, and Represents the set of brain region nodes The first in The brain region node and the first Each brain region node, Represents the adjacency matrix of the discriminant subgraph The Middle Line 1 Column elements; thus obtaining the discriminant subgraph used for classification training. :
[0136]
[0137] in, To distinguish subgraphs; A set of brain region nodes; To determine the adjacency matrix of a subgraph; The node feature matrix; To determine the set of edges in a subgraph; to determine the subgraph Edges induced by highly important nodes are preserved, while redundant edges are suppressed. That is, only edges where both endpoints are selected are retained, ensuring that the discriminant subgraph consists entirely of highly important nodes and their interconnections.
[0138] Classification prediction: First, the discriminant subgraph is processed by a graph neural network encoder. Feature extraction and graph-level readout function aggregation are performed sequentially to obtain the graph-level representation vector. The process is represented as:
[0139]
[0140] in, Vectors are represented at the graph level; For graph-level readout functions; For graph neural network encoders, Learnable parameters are fully shared with the original model; all brain region nodes are embedded and aggregated into a single graph-level representation vector. This graph-level representation of vectors Condensed discriminant subgraph The global structural information. Then, the graph-level representation vector of the discriminant. The input is a classification predictor, which maps the data to the class probability space to obtain the classification result. :
[0141]
[0142] in, It is a classification predictor (usually a multilayer perceptron, MLP). Given its learnable parameters, the graph-level representation vector is... Mapping to the category probability space, outputting the classification result. . This represents the class prediction result output by the model based on the discriminant subgraph. (Classification result) This refers to Alzheimer's disease (AD), normal cognitive impairment (NC), or mild cognitive impairment (MCI).
[0143] Based on the classification prediction results With real category labels Calculate the cross-entropy classification loss :
[0144]
[0145] in, Cross-entropy is used for classification loss; Indicates the number of categories. Indicates the true label in the first place Indicator value on class, Indicates the prediction result in the first... Probability output on the class.
[0146] The classification loss As the main supervisory signal during the training phase of the model (graph convolutional network model), it is used to drive the joint update of the graph neural network encoder parameters, classification predictor parameters, and node importance probability distribution, so that the model can still maintain its effective ability to distinguish disease categories under structural constraints.
[0147] Explanation of node set determination: Given a subject's connection-based brain network and the node importance probability vectors corresponding to the nodes in its brain regions This invention directly selects the top probability values using Top-K calculation. The set of brain region nodes V is used as the explanatory node set. :
[0148]
[0149] in, To explain the set of nodes; V is the importance probability vector; V is the set of brain region nodes; The preset interpretation budget parameters are used to control the size of the interpretation subgraph. In a preferred embodiment of the invention, This involves selecting the 10 most important brain regions as the basis for interpretation. This parameter can be adjusted according to actual diagnostic needs to achieve a balance between interpretive simplicity and information completeness.
[0150] Contrary fact node mask construction: based on the set of explanatory nodes Construct a counterfactual node mask that complements it. :
[0151]
[0152]
[0153] in, For counterfactual node masks; The counterfactual node mask for the Nth brain region node; The i-th brain region node in the set of brain region nodes; the counterfactual node mask The code is used to simulate the state of brain network structure after "removal of key brain region nodes".
[0154] Counterfactual subgraph construction: using counterfactual node masks Gating the node features and connection structure of the original brain network input graph generates a counterfactual node feature matrix corresponding to the counterfactual subgraph. and counterfactual adjacency matrix :
[0155]
[0156]
[0157] in, The feature matrix of the counterfactual nodes; It is a counterfactual adjacency matrix; For counterfactual node masks; The node feature matrix; It is an adjacency matrix; This represents element-wise multiplication. This represents the structure gating matrix composed of counterfact node masks.
[0158] Furthermore, according to the counterfactual adjacency matrix Determine the set of edges of the counterfactual subgraph ;
[0159]
[0160] Thus, a counterfactual subgraph is constructed. The counterfactual subgraph This indicates the removal of the set of explanatory nodes. The brain network structure obtained after the corresponding key brain region node is used to characterize the counterfactual structural state "if the key brain region node does not participate in the current classification decision".
[0161] Fact-counterfact graph representation learning and contrast constraints: discriminative subgraphs With counterfactual subgraphs By inputting the parameter-shared graph neural network encoder and the graph-level readout function respectively, the corresponding graph-level representations are obtained:
[0162]
[0163]
[0164] in, Vectors are represented at the graph level; For graph-level readout functions; This represents a graph neural network encoder; This represents the learnable parameters that are fully shared with the original model; For counterfact subgraphs; A graph-level representation of a counterfactual subgraph;
[0165] To enhance the model's reliance on key brain region nodes and their connectivity structures, a graph-level representation of the discriminant subgraph is developed. Graph-level representation of counterfactual subgraphs Imposing contrast constraints to separate the two in the representation space defines the fact-counterfact contrast loss. for:
[0166]
[0167] in, For fact-counterfact comparison loss; This represents the Sigmoid function. This represents a similarity metric function used to measure the degree of similarity between two graph-level representations.
[0168] Through the aforementioned fact-counterfactual comparison constraint This enables the model to learn to preserve discriminant subgraphs of key brain region nodes. Should be consistent with the current classification results Maintain consistency, while removing key brain region nodes from the counterfactual subgraph. It should be significantly distinguishable from factual structures in the representation space, thereby verifying and enhancing the discriminative contribution of the identified key brain region nodes.
[0169] Node-level KL regularization terms To avoid node selection distribution To address the issues of excessive density or degeneration into a "select all" structure (marking all brain region nodes as "important"), this invention introduces a node-level KL regularization term based on the information bottleneck concept to constrain the distribution of brain region node selection. Node-level KL loss. Defined as:
[0170]
[0171] in, For node-level KL loss; This represents the total number of nodes in the brain region. For the first The probability of a brain region node being selected as a key node; The preset node retention ratio is a fixed hyperparameter.
[0172] Taking into account the above constraints, the total training objective loss of the model in this invention is defined as follows:
[0173]
[0174] in: The cross-entropy classification loss, calculated based on the classification prediction results of the discriminant subgraph and the true labels, is used to ensure the disease classification performance of the model. To represent the fact-counterfactual contrast loss, it is used to differentiate between the discriminative subgraph and the counterfactual subgraph in the representation space, thereby strengthening the model's dependence on key brain region nodes and their connection structures; The node-level KL loss is used to control the sparsity of the node selection probability distribution. The weighting coefficients representing the fact-counterfact contrast loss are used to adjust the contribution of the contrast loss to the total loss function; The weight coefficients of the node-level KL divergence constraint terms are used to adjust the contribution of the sparse constraints to the total loss function.
[0175] Classification loss guarantees prediction performance, contrastive loss verifies the validity of the explanation (forcing the model to rely on key brain region nodes), and KL regularization controls the sparsity of the explanation (ensuring that only truly important brain region nodes are selected). The overall training objective loss is based on this. The learnable parameters of the graph neural network encoder, node selection module, and classification predictor are jointly optimized to ensure that the graph convolutional network model meets the requirements of disease classification accuracy, key brain region structure discrimination, and node selection sparsity during training.
[0176] Based on the optimized graph convolutional network model after training, forward computation is performed on the input brain network graph G to obtain the corresponding node importance probability vector (this is the final node importance probability, which is also the most accurate).
[0177]
[0178] in, Represents the probability vector of node importance The Middle The component is used to characterize the first component. The probability of a brain region node being identified as a critical node by the model.
[0179] To obtain stable and reproducible sample-level interpretation results, the node importance probability vector P is sorted by numerical value, and the top [values] are selected. From the brain region nodes, we obtain the set of explanatory nodes. :
[0180]
[0181] in, , This indicates the preset number of explanation nodes.
[0182] Furthermore, from the aforementioned brain network diagram Extract the set of explanation nodes Corresponding connection relationship and node features Construct sample-level explanatory subgraphs :
[0183]
[0184] in, This represents an interpretation of the subgraph; Represents the set of explanatory nodes In brain network diagrams The set of induced edges in; This represents the adjacency matrix corresponding to the subgraph. This represents the node feature matrix corresponding to the subgraph.
[0185] The Explanation Subgraph This is used to characterize the key brain region nodes and their connecting substructures upon which the current subject's classification decision relies. Among them, the explanatory node set... The subgraph reflects the key brain regions that the model considers to contribute most to the current classification result. This further preserves the connectivity between the key brain region nodes, thus forming a sample-level interpretable result output.
[0186] The graph convolutional network model outputs node importance probability vectors after training. And generate an explanation node set through a deterministic Top-K readout strategy. and Explanation of Subgraphs This enables a stable interpretation of the brain network classification results for individual subjects.
[0187] In the explanation subgraph After generation, statistical analysis and consistency verification are performed on the interpretation results at the category level, and the impact of the fact-counterfact contrast learning mechanism on the model's classification performance and interpretability is evaluated through ablation experiments that remove fact-counterfact contrast learning.
[0188] Explanatory subgraphs based on generated sample levels The set of explanatory nodes for samples of the same category Frequency statistics were performed to calculate the frequency of occurrence of each brain region node in this category of samples. :
[0189]
[0190] in, For the number of training sessions; This indicates the number of samples in that category; For indicator functions; The frequency of occurrence of each brain region node in this category of samples. The larger the value, the more likely it is to be a brain region node. The more consistently the optimized graph convolutional network model identifies a node as a key brain region in this category of samples; For the first The set of explanatory nodes obtained from the training statistics .
[0191] Furthermore, based on the aforementioned probability of occurrence The nodes in each brain region were sorted to obtain the final Top-K set of key brain region nodes, and then compared for consistency with the recognized key brain region nodes in Alzheimer's disease research. This was combined with... Figure 3 The importance score of the region of interest in the brain region node importance heatmap can be used to visualize the key brain region nodes identified by the model. The key brain region nodes obtained by the final statistics have a high degree of overlap with the recognized key brain region nodes such as the hippocampus, posterior cingulate cortex, and precuneus, indicating that the interpretation results output by the method have good medical consistency and biological rationality.
[0192] To verify the effect of fact-counterfact contrastive learning on improving the model's classification performance and the ability to distinguish key brain region nodes, while keeping the rest of the model structure and training settings unchanged, the fact-counterfact contrastive learning module was removed, that is, the contrastive learning constraint term of the optimized graph convolutional network model was removed to obtain the SACE-GN (w / o-CL) model, and its performance was compared and analyzed with the full model SACE-GN (Full) (i.e., the optimized graph convolutional network model).
[0193] According to the ablation experiment results in Table 1, all performance metrics of the graph convolutional network model decreased after removing fact-counterfactual contrastive learning. Specifically, the classification accuracy of SACE-GN (Full) was 82.81%, while the classification accuracy of SACE-GN (w / o-CL) after removing fact-counterfactual contrastive learning decreased to 79.69%, a decrease of 3.12 percentage points; precision decreased from 80.71% to 76.66%, a decrease of 4.05 percentage points; recall decreased from 87.50% to 85.62%, a decrease of 1.88 percentage points; specificity decreased from 78.12% to 73.75%, a decrease of 4.37 percentage points; and F1-score decreased from 83.72% to 80.83%, a decrease of 2.89 percentage points.
[0194] Table 1
[0195]
[0196] The results show that fact-counterfactual contrastive learning can effectively improve the overall classification performance of optimized graph convolutional network models, especially in terms of accuracy, specificity, and F1-score. This indicates that after introducing fact-counterfactual contrastive learning, the model can not only improve its ability to identify Alzheimer's disease samples, but also more effectively distinguish between key brain region connectivity structures and non-key structures, thereby enhancing the model's dependence on discriminative brain region substructures and reducing the interference of irrelevant brain region connectivity on classification results.
[0197] The beneficial effects of this application are as follows:
[0198] Significantly improved classification accuracy: In the AD vs. NC task on the real ADNI dataset, the accuracy of the optimized graph convolutional network model reached 82.81%, and the F1 score reached 83.72%. The overall performance and stability are significantly better than the existing GNN baseline model.
[0199] Excellent structural interpretability and causal verification: the key brain region nodes output (such as the hippocampus, posterior cingulate cortex, etc.) are highly consistent with clinical medical priors.
[0200] Extremely strong explanatory validity: In a rigorous counterfactual deletion test, after forcibly removing a very small number of the top-10 core brain region nodes selected by the model, the absolute accuracy of the model plummeted (by 6.56%), which provides irrefutable evidence that the substructures extracted by this invention do indeed monopolize most of the key information that determines the course of the disease.
[0201] Figure 4 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 4 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program that, when executed by the processor, enables the processor to implement an age recognition method. The internal memory may also store a computer program that, when executed by the processor, enables the processor to implement the age recognition method. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0202] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0203] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0204] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning, characterized in that, The method includes: The functional magnetic resonance imaging (fMRI) data of the subjects were acquired, and the fMRI data were preprocessed to obtain target fMRI data. The target fMRI data was then divided into standard brain regions, and the time series of each region of interest was extracted. The region of interest is a regional unit divided according to the standard brain region division template, and each region of interest corresponds to a brain region node. The functional connectivity strength between brain region nodes is calculated based on the extracted time series data, and the functional connectivity matrix, adjacency matrix, edge set between brain region nodes, and node feature matrix are calculated based on the functional connectivity strength. The brain region node set is determined based on the functional magnetic resonance imaging data. The adjacency matrix, edge set between brain region nodes, node feature matrix, and brain region node set constitute a brain network graph. The brain network graph is input into a graph neural network encoder to perform joint modeling of graph structure information and node feature information to obtain a basic node embedding representation. A node selection module composed of a multilayer perceptron is introduced to map the basic node embedding representation to a set of importance probabilities of brain region nodes in the first interval. The set of importance probabilities is composed of each importance probability vector, and the importance probability vector is the probability of a brain region node being selected as a key node. Based on the importance probability vector in the node importance probability set, Bernoulli sampling is performed independently on each brain region node to generate a corresponding binary mask; multiple binary masks constitute a random brain region node selection mask; A discriminant subgraph adjacency matrix is obtained by structural gating using a random brain region node selection mask docking matrix; the discriminant subgraph edge set is determined by the discriminant subgraph adjacency matrix, and the discriminant subgraph adjacency matrix, the discriminant subgraph edge set, the brain region node set, and the node feature matrix constitute the discriminant subgraph; A graph-level representation vector is obtained by sequentially extracting features from the discriminant subgraph and aggregating graph-level readout functions through a graph neural network encoder. The graph-level representation vector is then input into a classification predictor and mapped to the class probability space to obtain the classification result. The cross-entropy classification loss is calculated based on the classification result and the true class label. The probability values are directly selected from the top-K values using Top-K calculation. The set of brain region nodes is used as the set of explanatory nodes, and a counterfactual node mask complementary to the set of explanatory nodes is constructed based on it. The counterfactual node feature matrix, counterfactual adjacency moments, and counterfactual subgraph edge set are determined based on the counterfactual node mask. The set of brain region nodes, the set of counterfactual subgraph edges, the counterfactual adjacency moments, and the counterfactual node feature matrix constitute a counterfactual subgraph. The fact-counterfactual contrast loss is calculated based on the discriminative subgraph and the counterfactual subgraph. The node-level KL loss is determined based on the importance probability vector and the total number of nodes in the brain region; The total training objective loss is determined based on the cross-entropy classification loss, the fact-counterfact comparison loss, and the node-level KL loss. The graph convolutional network model is then trained using the total training objective loss to obtain an optimized graph convolutional network model. The connection relationships and node features corresponding to the explanatory node set are extracted from the brain network graph, and the induced edge set of the explanatory node set in the brain network graph is determined. The connection relationships, the node features, the explanatory node set and the induced edge set constitute a sample-level explanatory subgraph. Based on the sample-level explanatory subgraph, the frequency of explanatory node sets for samples of the same category is trained and statistically analyzed by optimizing the graph convolutional network model. This determines the frequency of occurrence of each brain region node in that category of samples. The higher the frequency, the stronger the brain region node. The more consistently a sample in this category is identified as a key brain region node by the optimized graph convolutional network model, the better.
2. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 1, characterized in that, The calculation of functional connectivity strength between brain region nodes based on the extracted time series, and the calculation of the functional connectivity matrix, adjacency matrix, edge set between brain region nodes, and node feature matrix based on the functional connectivity strength, are achieved through the following expressions: in, It is a time series; This represents the fMRI signal value of the i-th region of interest at time point T; For functional connection matrix; The region of interest (ROI) represents the functional connectivity strength between region of interest (ROI) i and region of interest j, and N represents the total number of ROIs. and Let fMRI signals of regions of interest i and j be represented at time point t, respectively. and These represent the mean of the corresponding time series, and T represents the total number of time points; This represents the functional connectivity strength between region of interest i and region of interest j.
3. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 2, characterized in that, The adjacency matrix calculated based on the functional connection strength is achieved through the following expression: in, It is an adjacency matrix; This represents the functional connectivity between brain region node i and brain region node j. The functional connectivity strength between region of interest i and region of interest j; The threshold used to control the sparsity of the adjacency matrix.
4. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 3, characterized in that, The set of edges and node feature matrices between brain region nodes are calculated based on the functional connectivity strength. This can be achieved using the following expression: in, It is the set of edges between nodes in brain regions; and Representing the sets of brain region nodes respectively The i-th and j-th brain region nodes in the brain; Let be the node features of the i-th brain region node; The node feature matrix; The functional connectivity strength between the i-th region of interest and the N-th region of interest.
5. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 1, characterized in that, The embedding of basic nodes represents Mapped to the set of importance probabilities of brain region nodes in the first interval This can be achieved using the following expression: in, Based on the embedding representation of the basic node, V represents the hidden layer dimension of node embedding, and V is the set of brain region nodes. For graph neural network encoders; These are the learnable parameters of the graph neural network encoder; For the Sigmoid function; It is a multilayer perceptron; Select learnable parameters for the module to be used in the node.
6. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 5, characterized in that, The use of random brain region node selection mask docking matrix Structural gating is performed to obtain the adjacency matrix of the discriminant subgraph; The discriminant subgraph edge set is determined by the adjacency matrix of the discriminant subgraph. The discriminant subgraph, consisting of the adjacency matrix, the edge set, the brain region node set, and the node feature matrix, is implemented by the following expression: in, To determine the adjacency matrix of a subgraph; Select a mask for random brain region nodes; It is an adjacency matrix; This is element-wise multiplication; and Represents the set of brain region nodes The first in The brain region node and the first Each brain region node; Represents the adjacency matrix of the discriminant subgraph The Middle Line 1 Column elements; To distinguish subgraphs; A set of brain region nodes; The node feature matrix; To distinguish subgraphs; The graph-level representation vector is obtained by sequentially extracting features from the discriminant subgraph and aggregating graph-level readout functions through a graph neural network encoder. This graph-level representation vector is then input into a classification predictor, and the classification result is obtained by mapping it to the category probability space using the following expression: in, Vectors are represented at the graph level; For graph-level readout functions; For graph neural network encoders, These are learnable parameters that are fully shared with the original model. For classification predictors; Its learnable parameters; The graph convolutional network module outputs a classification result based on a discriminant subgraph, where the classification result is Alzheimer's disease (AD), normal cognitive impairment (NC), or mild cognitive impairment (MCI). The formula for calculating the cross-entropy classification loss is as follows: in, Cross-entropy is used for classification loss; Indicates the number of categories; Indicates the true label in the first place Indicator value on class; Indicates the prediction result in the first... Probability output on the class.
7. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 1, characterized in that, The counterfactual node feature matrix, counterfactual adjacency moments, and counterfactual subgraph edge set are implemented by the following expression: in, To explain the set of nodes; V is the importance probability vector; V is the set of brain region nodes; The preset explanation of the budget parameters; For counterfactual node masks; The counterfactual node mask for the Nth brain region node; and Represents the set of brain region nodes The first in The brain region node and the first Each brain region node; The feature matrix of the counterfactual nodes; It is a counterfactual adjacency matrix; For counterfactual node masks; The node feature matrix; It is an adjacency matrix; This represents element-wise multiplication. This represents the structure-gated matrix composed of counterfactual node masks; For equal operations; This represents the operation of selecting the top K brain region nodes according to their importance probability from high to low; Represents the counterfactual node mask for the i-th brain region node; Represents the set of edges in a counterfactual subgraph; The calculation of the fact-counterfactual contrast loss based on the discriminant subgraph and the counterfactual subgraph is achieved by the following expression: in, Vectors are represented at the graph level; For graph-level readout functions; This represents a graph neural network encoder; This represents the learnable parameters that are fully shared with the original model; For counterfact subgraphs; A graph-level representation of a counterfactual subgraph; For fact-counterfact comparison loss; Represents the Sigmoid function; This represents the similarity measurement function; This represents the discriminant subgraph.
8. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 1, characterized in that, The determination of the node-level KL loss based on the importance probability vector and the total number of brain region nodes is achieved through the following expression: in, For node-level KL loss; This represents the total number of nodes in the brain region. For the first The probability of a brain region node being selected as a key node; The preset node retention ratio is a fixed hyperparameter.
9. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 1, characterized in that, The determination of the total training objective loss based on the cross-entropy classification loss, the fact-counterfactual comparison loss, and the node-level KL loss is achieved through the following expression: in, Loss due to overall training objective; Cross-entropy is used for classification loss; For fact-counterfact comparison loss; For node-level KL loss; The weighting coefficients representing the loss in the fact-counterfact comparison; These are the weighting coefficients for node-level KL loss.
10. The method for classifying Alzheimer's disease and identifying key brain regions based on stochastic attention and counterfactual contrastive learning according to claim 1, characterized in that, The frequency of occurrence of each brain region node in this category of samples is expressed by the following formula: in, For the number of training iterations; Indicates the number of samples in this category; For indicator functions; The frequency of occurrence of each brain region node in this category of samples. The larger the value, the more likely it is to be a brain region node. The more consistently the optimized graph convolutional network model identifies a node as a key brain region in this category of samples; For the first The set of explanatory nodes obtained from the training statistics .