Brain disease imbalance data classification method based on feature difference enhancement

Abstract

The invention discloses a brain disease unbalanced data classification method based on feature difference enhancement. The recognition capability of minority class samples is cooperatively improved through contribution degree driven difference enhancement and a dynamic cost sensitive mechanism. The method comprises the following steps: preprocessing rs-fMRI data, and constructing a functional connection matrix; pre-training an auto-encoder according to categories to perform feature reconstruction, and generating a contribution degree matrix by using layer-by-layer correlation propagation; calculating an inter-class statistical difference, and constructing a difference matrix to enhance the feature expression of discriminative connection in the graph neural network; inputting the enhanced features into a graph classification model to obtain a preliminary result; a self-adaptive cost matrix algorithm is adopted, the loss weight is dynamically adjusted by integrating the intra-class dispersion degree and the inter-class separation degree, and the model is guided to focus on minority classes in cooperation with a difference enhancement mechanism. According to the method, a systematic scheme for unbalanced resting state functional magnetic resonance data classification is constructed, and the recognition capability of the model for minority class samples and the overall classification performance are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of imbalanced brain disease data technology, and in particular to a brain disease imbalanced data classification method based on feature difference enhancement. Background Technology

[0002] With the increasing prevalence of brain dysfunctions such as depression and Alzheimer's disease globally, these conditions have become a significant public health challenge. Resting-state functional magnetic resonance imaging (rs-fMRI) has been widely used in the diagnosis of these diseases because it can reveal subtle functional changes between brain regions and reflect underlying connectivity patterns by capturing blood oxygen level-dependent (BOLD) signals. However, medical imaging datasets collected in real-world clinical settings often suffer from inherent class imbalances, especially in research on the diagnosis of brain dysfunctions. For example, nearly half of the collection sites in the depression dataset (REST-meta-MDD) exhibited varying degrees of data imbalance, with some sites having a sample ratio as high as 5:1. In medical image analysis, significant data imbalance can severely impact the accuracy of model evaluation. On the one hand, imbalanced data may lead to decreased accuracy in assisted diagnosis, thus affecting clinical treatment decisions. On the other hand, data imbalance can also interfere with the accurate identification of biomarkers, which play a crucial role in early intervention, personalized treatment plans, prognostic assessment, and overall disease management in mental illness.

[0003] To address this critical challenge, researchers have recently explored various techniques capable of fully revealing the underlying data structure. Graph Neural Networks (GNNs) have garnered significant attention in the diagnosis of brain dysfunction due to their unique ability to handle complex connection patterns and integrate global information. These methods typically construct individual graph functional connectomes (FCs), enabling comprehensive extraction of disease-related biomarkers. Furthermore, some methods construct population graphs to integrate inter-individual relationships, thereby improving overall classification performance. By combining local feature aggregation with global context modeling, GNNs help overcome the limitations of traditional class imbalance mitigation techniques. However, most GNN-based brain diagnostic models still assume balanced training data during model development, leading to degraded minority class detection performance when applied to real-world medical image datasets with significant class imbalance. An existing brain dysfunction diagnostic method that considers class imbalance, CSWCL-GCN, mitigates the class imbalance problem by employing a cost-sensitive weighted contrastive learning mechanism that leverages the weights of negative samples and their distance from class prototypes. However, while this method balances the training contributions of different classes to some extent, it fails to identify biomarkers, limiting its application in precision medicine. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention proposes a brain disease imbalance data classification method based on feature difference enhancement, and constructs a systematic scheme for classifying imbalanced resting-state functional magnetic resonance imaging data, which significantly improves the model's ability to identify minority class samples and its overall classification performance.

[0005] This invention provides the following technical solution: a brain disease imbalance data classification method based on feature difference enhancement, which improves the model's ability to identify minority class samples by synergistically combining a contribution-driven difference enhancement mechanism with a dynamically adaptive cost-sensitive mechanism; the method includes the following steps: S1. Data preparation and feature initialization: The resting-state functional magnetic resonance imaging data are preprocessed, the average time series of each brain region is extracted based on the preset brain atlas, and the functional connectivity matrix between brain regions is calculated as input features. S2. Category-Specific Feature Reconstruction and Contribution Analysis: For samples of different categories, pre-trained autoencoders are used for feature reconstruction; based on the hierarchical correlation propagation mechanism, the contribution of each brain region in the reconstruction process is analyzed, and a contribution matrix for each subject is generated. S3. Contribution-driven difference enhancement: Based on the contribution matrix obtained in step S2, the statistical differences in brain functional connectivity patterns of different categories of samples are calculated, a difference matrix is ​​constructed, and the feature learning process of the graph neural network is weighted using the difference matrix to amplify the feature expression of the connectivity of key brain regions for class discrimination. S4. Graph classification and loss calculation under differential enhancement: Construct a graph neural network classification model, input the features after differential enhancement in step S3 into the model, and obtain the preliminary classification results and loss. S5. Dynamic Cost-Sensitive Learning: An adaptive cost matrix algorithm is adopted, whose optimization objective combines the dispersion of samples within a class with the separation of samples between classes. It dynamically adjusts the class weights in the classification loss function, thereby working in conjunction with the difference enhancement mechanism in step S3 to guide the model to focus on the minority class samples that are difficult to identify. In step S3, the difference enhancement is applied to the feature learning stage of forward propagation, and in step S5, the cost-sensitive learning is applied to the weight update stage of backpropagation. The two work together to optimize the model parameters and improve the overall classification performance of the model on imbalanced data.

[0006] In step S1, preprocessing is performed using the DPARSF tool, which includes slice time correction, motion correction, global intensity normalization, and spatial smoothing. Preprocessing also includes spatial alignment of rs-fMRI data using registered T1 structural images to ensure spatial consistency of data from different batches.

[0007] In step S1, the preset brain atlas is the AAL brain atlas, and the functional connectivity matrix is ​​constructed by calculating the Pearson correlation coefficient between the average time series of each brain region.

[0008] As an improvement, the feature reconstruction of the pre-trained autoencoder in step S2 specifically includes: introducing random mask perturbation into the input functional connectivity matrix in each round of training. The random mask perturbation is to randomly mask some brain region connectivity elements in the functional connectivity matrix. The perturbation-reconstruction learning process is constructed through the encoder and decoder of the autoencoder, so that the autoencoder of each category learns the category-specific robust hidden representation of the sample in the functional connectivity matrix.

[0009] As an improvement, step S2, which analyzes the contribution of each brain region based on the hierarchical correlation propagation mechanism, specifically includes: sequentially masking the connections between each brain region of interest and other brain regions, and inputting the masked functional connectivity matrix into the pre-trained autoencoder of the corresponding category for reconstruction; During forward propagation, the activation values ​​of neurons in each layer of the autoencoder are recorded. During backward propagation, the reconstruction correlation of the output layer is traced back to the input layer layer by layer using the layer-by-layer correlation propagation mechanism, and the contribution of each brain region to the reconstruction of the global functional connectivity matrix is ​​calculated. The formula for calculating the contribution of the hierarchical correlation propagation mechanism is as follows: in, This represents the activation value of neuron j in layer l. , Let represent the positive and negative weights from neuron j to neuron k, respectively. It is the stability constant. Table 1 The correlation score of layer neuron k.

[0010] As an improvement, in step S2, the contribution matrix of subject i is... , where R is the number of brain regions, and the contribution matrix is ​​obtained by concatenating the contribution vectors of each brain region to all other brain regions. The r-th row of the matrix corresponds to the correlation contribution of the r-th brain region to all other brain regions.

[0011] As an improvement, step S3 calculates the statistical differences in brain functional connectivity patterns among different categories of samples and constructs the difference matrix. Specifically, this includes: for each brain region connection in the contribution matrix, i.e., the jk-th brain region connection, calculating the mean contribution of each category of samples to that connection. and standard deviation , where m is the category identifier; Define the degree of difference The degree of difference in connectivity between categories in this brain region is quantified using the following formula: Where c is the total number of sample categories. , These represent the mean and standard deviation of the contribution of category n to the connectivity of brain regions j and k, respectively. The differences in connectivity among all brain regions constitute a difference matrix.

[0012] As an improvement, step S3 uses the difference matrix to weight the feature learning process of the graph neural network, specifically including: constructing a mapping function, and weighting and amplifying the corresponding features based on the difference degree of connection between each brain region in the difference matrix, with the brain region connection with higher difference degree receiving greater weight; The feature learning process also includes: performing preliminary graph convolution and pooling on the weighted features, introducing Chebyshev convolution to capture deeper local and cross-hop neighborhood connection patterns of nodes, using SAGPooling for local pooling and combining the pooling features from global aggregation with node selection information to obtain the optimized feature representation.

[0013] As an improvement, the graph neural network classification model in step S4 includes a population graph modeling module, a multi-layer graph convolution module, and a multi-layer perceptron; The construction process of the group graph modeling module is as follows: stack the optimized feature representations of all subjects to form the node feature matrix of the group graph; use the subjects' gender, age, education level and site information to calculate the similarity between subjects as the edge weights of the group graph, summarize to obtain the adjacency matrix, and normalize and threshold the adjacency matrix to retain effective connections. The multilayer graph convolution module is used to fuse node features and edge weight information of the group graph and to pass messages. The multilayer perceptron is used to receive the fused features and output preliminary classification results.

[0014] As an improvement, the process of calculating the dispersion of samples within a class in step S5 is as follows: Based on the preliminary classification results in step S4, extract the logits representation set of samples from each class. , where c is the category identifier; Calculate the mean vector of category c ,in The number of samples in category c; Calculate the covariance matrix of category c: Intra-class dispersion: in, The logtis representation for a node. The trace of the covariance matrix, which reflects the degree of dispersion of samples within a category, is formed by the intra-class dispersion matrix of all categories. .

[0015] As an improvement, the calculation process for the separation degree of inter-class samples in step S5 is as follows: For any two classes i and j, if i ≠ j, then the inter-class separation degree is... = , Where i represents the Euclidean distance; if i = j, then =1; the separation degree between all categories constitutes the inter-class separation matrix. .

[0016] As an improvement, the adaptive cost matrix algorithm in step S5 specifically includes: Combined with numerical smoothing constant Construct the target matrix ; Initialize the cost matrix using the sample frequencies of each category in the training data. The cost matrix is ​​dynamically updated using gradient descent. ,make With the target matrix Maintain consistency; the classification loss function is a double loss function consisting of cross-entropy loss and cost matrix loss, where the cost matrix loss is calculated using the following formula: in, Describing the Frobenius norm, For coefficients, For the target matrix The element in the i-th row and j-th column, Cost matrix The element in the i-th row and j-th column.

[0017] Compared with the prior art, the advantages of the present invention are as follows: By deeply synergizing two core mechanisms—difference enhancement and cost sensitivity—this study systematically addresses the weakness in minority class identification in imbalanced resting-state functional magnetic resonance imaging (fMRI) data classification. A feature-enhanced classification scheme for imbalanced brain diseases is constructed, comprehensively improving model classification performance and application value, achieving optimal performance in relevant classification tasks on publicly available datasets. The data preprocessing workflow lays a high-quality foundation for subsequent feature learning. A functional connectivity matrix constructed from pre-defined brain atlases and Pearson correlation coefficients accurately reflects brain functional association patterns, providing reliable input for the category-specific autoencoder. The combination of random mask perturbation and the autoencoder yields a category-specific robust hidden representation, strengthening the anti-interference ability of features and providing a precise carrier for quantifying the contribution of the hierarchical correlation propagation mechanism. This enables the brain region contribution matrix to capture the intra-category brain functional association characteristics with fine granularity, while simultaneously screening out key brain regions highly relevant to the classification task, providing core support for a significant improvement in classification performance. The contribution matrix provides core data support for the construction of the difference matrix. Statistical difference calculation accurately quantifies the differences in brain functional connectivity patterns between classes. This method of enhancing inter-class feature differences, as a data augmentation technique, effectively overcomes the limitations of current neuroimaging feature extraction. The mapping function's weighted amplification of key discriminative brain region connections, synergistically with Chebyshev convolution and SAGPooling feature optimization, further strengthens inter-class discriminativeness, providing highly discriminative features for graph classification models. Population graph modeling integrates individual features and inter-subject similarity information. The fusion of multi-layer graph convolution and multi-layer perceptron achieves preliminary classification, and its output classification results provide optimization basis for dynamic cost-sensitive learning. The calculation of intra-class dispersion and inter-class separation allows the adaptive cost matrix to accurately adapt to the data distribution. By dynamically adjusting class imbalance, intra-class sample compactness, and inter-class sample separation, the learning strategy is optimized. The initialization and gradient descent update mechanism keeps the cost matrix consistent with the target matrix, significantly improving the model's performance on imbalanced datasets. The collaborative optimization of dual loss functions dynamically adjusts class weights, focusing on minority class samples. This forms a closed-loop synergy with the difference enhancement in the forward propagation stage. The discriminative features amplified by difference enhancement make cost-sensitive learning more targeted, while the weight bias of cost-sensitive learning towards the minority class guides the model to further focus on key differences in the feature learning stage. The two mutually promote and reinforce each other. The various innovations in the entire technical solution are interconnected and work synergistically to improve overall classification accuracy, significantly enhance the model's ability to identify minority class samples, and endow the model with stronger robustness and generalization ability. It is suitable for various scenarios such as auxiliary diagnosis of brain diseases based on imbalanced resting-state functional magnetic resonance imaging data. Attached Figure Description

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Figure 1This is a model diagram of a brain disease imbalance data classification method based on feature difference enhancement. Detailed Implementation

[0019] A method for classifying imbalanced brain disease data based on feature difference enhancement, the specific steps of which are as follows: Step S1: Data Preparation and Feature Initialization S1.1 s-fMRI Data Preprocessing: The raw rs-fMRI image data were standardized using the DPARSF tool, including slice time correction, motion correction, global intensity normalization, and spatial smoothing. Simultaneously, the preprocessed rs-fMRI data were spatially aligned using registered T1-weighted structural images to ensure spatial consistency across batches. S1.2 Brain region time series extraction and functional connectivity matrix construction: The AAL brain map is selected as the preset brain map. The time series of all voxels in each brain region are extracted. The Pearson correlation coefficient between the average time series of any two brain regions is calculated, and the functional connectivity (FC) matrix is ​​constructed as the initial input features of the model. Step S2: Category-Specific Feature Reconstruction and Contribution Analysis S2.1 Autoencoder Structure Design and Pre-training: For each category in the dataset, a dedicated autoencoder is constructed. The autoencoder adopts an encoder-decoder symmetric structure, and the feature reconstruction learning process is constructed through the encoder and decoder.

[0020] S2.2 Random Mask Perturbation and Reconstruction Training: In each round of training, a random mask perturbation is introduced into the input FC matrix. Through the perturbation-reconstruction learning process, each class-specific autoencoder learns the class-specific robust hidden representation of the class samples on the FC matrix.

[0021] S2.3, LRP Mechanism Implementation and Contribution Matrix Generation: Based on a pre-trained category-specific autoencoder, the contribution of each brain region is quantified through the Layer-by-Layer Relevance Propagation (LRP) mechanism. The connections between each region of interest (ROI) and other brain regions are sequentially masked. The masked full-joint (FC) matrix is ​​then input into the corresponding category's autoencoder for reconstruction. During forward propagation, the activation values ​​of neurons in each layer are recorded. During backward propagation, LRP is used to trace the reconstruction relevance of the output layer back to the input layer, calculating the contribution of each brain region. The contribution calculation formula for the LRP mechanism is as follows: in, This represents the activation value of neuron j in layer l. , Let represent the positive and negative weights from neuron j to neuron k, respectively. It is the stability constant. Table 1 The correlation score of layer neuron k; Masking-reconstruction-LRP analysis was performed sequentially on all brain regions for each subject to obtain the contribution vector of each brain region to all other brain regions. The contribution vectors were then concatenated to generate the subject's contribution matrix. , where R is the number of brain regions, and the r-th row of the matrix corresponds to the correlation contribution of the r-th brain region with all other brain regions; Step S3: Contribution-Driven Difference Enhancement S3.1, Constructing the Difference Matrix: Based on the contribution matrix of all subjects, for each brain region connection, calculate the mean contribution of each category of samples to that connection. and standard deviation (where m is the category identifier), the degree of difference between categories in the connectivity of this brain region is quantified by the difference calculation formula, which is: Where c is the total number of sample categories. , These represent the mean and standard deviation of the contribution of category n to the connectivity of brain regions j and k, respectively. The difference matrix is ​​formed by the degree of difference in connectivity among all brain regions. S3.2 Feature Weighting and Optimization: Construct a mapping function to weight and amplify the corresponding features based on the difference in the connectivity of each brain region in the difference matrix; perform preliminary graph convolution and pooling on the weighted features, introduce Chebyshev convolution to capture the deeper local and cross-jump neighborhood connectivity patterns of nodes, use SAGPooling for local pooling and combine the pooling features of global convergence with node selection information to obtain the optimized feature representation; Step S4: Graph Classification and Loss Calculation under Differentiation Enhancement S4.1 Group Graph Modeling: Stack the optimized feature representations of all subjects to form the node feature matrix of the group graph; use the subjects' gender, age, education level and site information to calculate the similarity between subjects as the edge weights of the group graph, summarize to obtain the adjacency matrix, and normalize and threshold the adjacency matrix to retain effective connections. S4.2 Training of Graph Neural Network Classification Model: Construct a graph neural network classification model that includes a group graph modeling module, a multi-layer graph convolution module, and a multi-layer perceptron (MLP). The multi-layer graph convolution module fuses the node features and edge weight information of the group graph and performs message passing. The MLP receives the fused features and outputs the preliminary classification results. The classification loss is calculated. Step S5: Dynamic Cost-Sensitive Learning S5.1 Calculation of intra-class scatter and inter-class separation: Based on the preliminary classification results, extract the logits representation set of samples from each class. , where c is the category identifier; Calculate the mean vector of category c ,in The number of samples in category c; Calculate the covariance matrix of category c: Intra-class dispersion: in, The logtis representation for a node. The trace of the covariance matrix, which reflects the degree of dispersion of samples within a category, is formed by the intra-class dispersion matrix of all categories. ; The inter-class separability of category i and category j can be expressed as: in, Wherein represents Euclidean distance; S5.2 Adaptive Cost Matrix Construction and Optimization: Combining Numerical Smoothing Constants Construct the target matrix ; Initialize the cost matrix using the sample frequencies of each category in the training data. The cost matrix is ​​dynamically updated using gradient descent. ,make With the target matrix Maintain consistency; S5.3, Dual-Loss Collaborative Optimization: A dual-loss function is adopted, consisting of cross-entropy loss and cost matrix loss. The formula for calculating the cost matrix loss is as follows: in, Describing the Frobenius norm, For coefficients, For the target matrix The element in the i-th row and j-th column, Cost matrix The element in the i-th row and j-th column. A dual loss function dynamically adjusts the class weights, focusing on minority class samples. The difference enhancement in the forward propagation stage and the cost-sensitive learning in the backpropagation stage work synergistically to continuously optimize the model parameters.

[0022] The dataset used in this invention is derived from the publicly available depression (REST-meta-MDD dataset), which covers rs-fMRI data of 565 patients with depression and 110 healthy controls, as well as rich non-imaging phenotypic data. A five-fold cross-validation strategy was used to test the effectiveness of the proposed method. Specifically, the entire dataset was first evenly divided into five parts, and one part was selected as the test set in each cycle, while the remaining parts were merged and used as the training set. In each cross-validation cycle, 20% of the training set was also divided as the validation set to determine the optimal parameter configuration of the model. The brain network method proposed in this invention will be compared with the following methods: (1) EV-GCN: This model is based on variational edges to construct an adaptive population graph, integrates medical images and phenotypic data, and uses Monte Carlo edges to discard the uncertainty of the graph structure, thereby achieving disease prediction. (2) BrainGNN: This model first constructs a functional connectivity graph for each subject, maps each region of the brain to nodes in the graph, and then uses specific node pooling layers to filter and extract the most discriminative regional features. (3) FC-HGNN: This is an end-to-end heterogeneous graph neural network model. The model fully learns the features of brain regions within and between hemispheres, and constructs a population graph based on gender and multimodal data to achieve adaptive feature aggregation and classification prediction.

[0023] From the data performance, this invention achieves significant superiority in all indicators: F1-macro reaches 88.31%±4.60, which is about 6 percentage points higher than the best comparison method FC-HGNN's 82.35%±5.01; AUC is as high as 93.35%±3.34, which is 6.45 percentage points higher than FC-HGNN's 86.90%±6.64; bACC (balanced accuracy) reaches 85.88%±3.05, which is 8.1 percentage points higher than FC-HGNN's 77.78%±6.68; F1-weighted is 93.91%±2.34, which is also higher than FC-HGNN's 91.07%±2.18.

[0024] The data results are highly consistent with the technological innovations of this invention: First, the significant improvement of F1-macro, a core indicator sensitive to minority class sample recognition performance, directly confirms the synergistic effectiveness of the two core mechanisms of difference enhancement and cost sensitivity—inter-class feature difference enhancement amplifies the differences in functional connectivity patterns between the minority and majority classes, and the adaptive cost matrix dynamically adjusts class weights, accurately solving the key problem of weak minority class recognition in imbalanced data; Second, the substantial improvement in AUC and bACC demonstrates the superiority of the feature extraction of this invention. The key brain regions selected through the layer-by-layer correlation propagation (LRP) mechanism, Chebyshev convolution, and SAGPooling feature optimization provide the model with highly discriminative feature representations, effectively overcoming the limitations of existing neuroimaging feature extraction methods; Third, all four indicators maintain a high level with small fluctuations (standard deviations are generally lower than those of the comparison methods), proving the robustness and generalization ability of the technical solution of this invention. This is due to the effective integration of individual features and subject similarity by population graph modeling, and the stable convergence of model parameters brought about by the synergistic optimization of the dual loss functions.

[0025] In contrast, compared to other methods, BrainGNN, lacking a dedicated mechanism for class imbalance, achieved an F1-macro score of only 53.36% ± 4.55, significantly lower than the present invention. Although EV-GCN integrates multimodal data, it lacks accurate capture of inter-class differences and dynamic cost adjustment, resulting in all indicators falling short of the present invention. While FC-HGNN performed relatively best, it failed to achieve deep synergy between difference enhancement and cost sensitivity, and did not quantify brain region contribution through the LRP mechanism. Therefore, it still lags behind the present invention in terms of recognition accuracy and balance performance.

[0026] After model training, classification performance was evaluated on the test set, and comparative experiments were conducted to verify the synergistic effect of each innovation. The results show that the synergistic mechanism of difference enhancement and cost sensitivity can effectively improve the identification ability of minority class samples. The contribution analysis of key brain regions and feature optimization further enhance the inter-class discrimination. The adaptive cost matrix dynamically adapts to the data distribution, enabling the model to have stronger robustness and generalization ability on imbalanced datasets.

[0027] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0028] The units described in some embodiments of this disclosure can be implemented in software or in hardware. The described units can also be located in a processor, and the functions described above can be performed at least in part by one or more hardware logic components.

[0029] The above description only illustrates the preferred embodiments of the present invention and should not be construed as limiting the scope of the claims. The present invention is not limited to the above embodiments, and variations in its specific structure are permitted. All modifications made within the scope of the independent claims of this invention are also within the scope of protection of this invention.

Claims

1. A method for classifying imbalanced brain disease data based on feature difference enhancement, characterized in that, The method combines a contribution-driven differential enhancement mechanism with a dynamically adaptive cost-sensitive mechanism to improve the model's ability to identify minority class samples; the method includes the following steps: S1. Data preparation and feature initialization: The resting-state functional magnetic resonance imaging data are preprocessed, the average time series of each brain region is extracted based on the preset brain atlas, and the functional connectivity matrix between brain regions is calculated as input features. S2. Category-Specific Feature Reconstruction and Contribution Analysis: For samples of different categories, pre-trained autoencoders are used for feature reconstruction; based on the hierarchical correlation propagation mechanism, the contribution of each brain region in the reconstruction process is analyzed, and a contribution matrix for each subject is generated. S3. Contribution-driven difference enhancement: Based on the contribution matrix obtained in step S2, the statistical differences in brain functional connectivity patterns of different categories of samples are calculated, a difference matrix is ​​constructed, and the feature learning process of the graph neural network is weighted using the difference matrix to amplify the feature expression of the connectivity of key brain regions for class discrimination. S4. Graph classification and loss calculation under differential enhancement: Construct a graph neural network classification model, input the features after differential enhancement in step S3 into the model, and obtain the preliminary classification results and loss. S5. Dynamic Cost-Sensitive Learning: An adaptive cost matrix algorithm is adopted, whose optimization objective combines the dispersion of samples within a class with the separation of samples between classes. It dynamically adjusts the class weights in the classification loss function, thereby working in conjunction with the difference enhancement mechanism in step S3 to guide the model to focus on the minority class samples that are difficult to identify. The difference enhancement in step S3 acts on the feature learning stage of forward propagation, and the cost-sensitive learning in step S5 acts on the weight update stage of backpropagation. The two work together to optimize the model parameters to improve the overall classification performance of the model on imbalanced data.

2. The brain disease imbalance data classification method based on feature difference enhancement according to claim 1, characterized in that, The preprocessing described in step S1 is performed using the DPARSF tool, specifically including slice time correction, motion correction, global intensity normalization, and spatial smoothing. The preprocessing also includes spatial alignment of the rs-fMRI data using registered T1 structural images to ensure spatial consistency of data from different batches.

3. The brain disease imbalance data classification method based on feature difference enhancement according to claim 1, characterized in that, The preset brain atlas mentioned in step S1 is the AAL brain atlas, and the functional connectivity matrix is ​​constructed by calculating the Pearson correlation coefficient between the average time series of each brain region.

4. The brain disease imbalance data classification method based on feature difference enhancement according to claim 1, characterized in that, The feature reconstruction of the pre-trained autoencoder in step S2 specifically includes: introducing random mask perturbation into the input functional connectivity matrix in each training round. The random mask perturbation is to randomly mask some brain region connectivity elements in the functional connectivity matrix. The perturbation-reconstruction learning process is constructed through the encoder and decoder of the autoencoder, so that the autoencoder of each category learns the category-specific robust hidden representation of the sample in the functional connectivity matrix.

5. The brain disease imbalance data classification method based on feature difference enhancement according to claim 1, characterized in that, The analysis of the contribution of each brain region based on the layer-by-layer correlation propagation mechanism in step S2 specifically includes: sequentially masking the connections between each brain region of interest and other brain regions, and inputting the masked functional connectivity matrix into the pre-trained autoencoder of the corresponding category for reconstruction; During forward propagation, the activation values ​​of neurons in each layer of the autoencoder are recorded. During backward propagation, the reconstruction correlation of the output layer is traced back to the input layer layer by layer using the layer-by-layer correlation propagation mechanism, and the contribution of each brain region to the reconstruction of the global functional connectivity matrix is ​​calculated. The contribution calculation formula for the layer-by-layer correlation propagation mechanism is as follows: in, This represents the activation value of neuron j in layer l. , Let represent the positive and negative weights from neuron j to neuron k, respectively. It is the stability constant. Table 1 The correlation score of layer neuron k.

6. The brain disease imbalance data classification method based on feature difference enhancement according to claim 1, characterized in that, The contribution matrix of subject i mentioned in step S2 , where R is the number of brain regions, and the contribution matrix is ​​obtained by concatenating the contribution vectors of each brain region to all other brain regions. The r-th row of the matrix corresponds to the correlation contribution of the r-th brain region to all other brain regions.

7. The brain disease imbalance data classification method based on feature difference enhancement according to claim 1, characterized in that, Step S3, which involves calculating the statistical differences in brain functional connectivity patterns among different categories of samples and constructing a difference matrix, specifically includes: for each brain region connection in the contribution matrix, i.e., the jk-th brain region connection, calculating the mean contribution of each category of samples to that connection. and standard deviation , where m is the category identifier; Define the degree of difference The degree of difference in connectivity between categories in this brain region is quantified using the following formula: Where c is the total number of sample categories. , These represent the mean and standard deviation of the contribution of category n to the connectivity of brain regions j and k, respectively. The difference matrix is ​​formed by the degree of difference in connectivity among all brain regions.

8. The brain disease imbalance data classification method based on feature difference enhancement according to claim 1, characterized in that, The weighting of the feature learning process of the graph neural network using the difference matrix in step S3 specifically includes: constructing a mapping function, and weighting and amplifying the corresponding features based on the difference degree of the connections between brain regions in the difference matrix, with the brain region connections with higher difference degrees receiving greater weights; The feature learning process also includes: performing preliminary graph convolution and pooling on the weighted features, introducing Chebyshev convolution to capture deeper local and cross-hop neighborhood connection patterns of nodes, using SAGPooling for local pooling and combining the pooling features from global aggregation with node selection information to obtain the optimized feature representation.

9. A method for classifying imbalanced brain disease data based on feature difference enhancement according to claim 1, characterized in that, The graph neural network classification model described in step S4 includes a population graph modeling module, a multi-layer graph convolution module, and a multi-layer perceptron. The construction process of the group graph modeling module is as follows: stack the optimized feature representations of all subjects to form the node feature matrix of the group graph; The similarity between participants was calculated using the participants' gender, age, education level, and site information as edge weights in the group graph. The adjacency matrix was then obtained by summarizing the data. The adjacency matrix was normalized and thresholded to retain valid connections. The multilayer graph convolution module is used to fuse node features and edge weight information of the group graph and to pass messages. The multilayer perceptron is used to receive the fused features and output preliminary classification results.

10. A method for classifying imbalanced brain disease data based on feature difference enhancement according to claim 1, characterized in that, The process of calculating the dispersion of samples within a class in step S5 is as follows: Based on the preliminary classification results in step S4, extract the logits representation set of samples from each class. , where c is the category identifier; Calculate the mean vector of category c ,in The number of samples in category c; Calculate the covariance matrix of category c: Intra-class dispersion: in, The logtis representation of a node. The trace of the covariance matrix, which reflects the degree of dispersion of samples within a class, is formed by the intra-class dispersion matrix of all classes. .

11. A method for classifying imbalanced brain disease data based on feature difference enhancement according to claim 10, characterized in that, The calculation process for the inter-class separation degree of samples in step S5 is as follows: For any two classes i and j, if i ≠ j, then the inter-class separation degree is... = , Where i represents the Euclidean distance; if i = j, then =1; the separation degree between all categories constitutes the inter-class separation matrix. .

12. The brain disease imbalance data classification method based on feature difference enhancement according to claim 11, characterized in that, The adaptive cost matrix algorithm described in step S5 specifically includes: Combined with numerical smoothing constant Construct the target matrix ; Initialize the cost matrix using the sample frequencies of each category in the training data. The cost matrix is ​​dynamically updated using gradient descent. ,make With the target matrix Maintain consistency; the classification loss function is a double loss function consisting of cross-entropy loss and cost matrix loss, wherein the formula for calculating the cost matrix loss is: in, Denotes the Frobenius norm. For coefficients, For the target matrix The element in the i-th row and j-th column, Cost matrix The element in the i-th row and j-th column.

Images