A multi-view outlier detection system for brain disease diagnosis

Abstract

The application provides a multi-view outlier detection system for brain disease diagnosis, and relates to the technical field of computer-aided diagnosis.The application aims at the problem that outlier samples in brain network data sets can guide model learning to deviate from the real distribution, and provides a multi-view outlier detection system for brain disease diagnosis, which screens out outliers in the data while effectively denoising brain network data, and comprises a brain network construction model and a brain disease diagnosis model; the brain network construction model constructs a brain network based on brain functional magnetic resonance imaging data; the brain disease diagnosis model is used for removing outlier data in the brain functional magnetic resonance imaging data, and obtains a brain network classification result based on a multi-view joint learning method, which improves the brain disease diagnosis performance and also provides clinical interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of computer-aided diagnosis, and in particular to a multi-view outlier detection system for brain disease diagnosis. Background Art

[0002] Leveraging the development of next-generation information technology, brain imaging and computer-assisted diagnostic techniques for brain diseases are increasingly being used, providing valuable guidance for clinical decision-making. The mainstream approach uses functional magnetic resonance imaging (fMRI) to explore activity signals across brain regions. These techniques model each region as a node and the functional connections between regions as edges, constructing complex functional brain networks that reflect actual brain activity. These methods leverage the computer science principle of graph neural networks (GNNs) to construct diagnostic models for brain diseases.

[0003] In recent years, a variety of brain disease classification methods based on graph neural networks have been proposed. The paper "Hi-GCN: Ahierarchical Graph Convolution Network for Graph Embedding Learning of Brain Networks and Brain Disorders Prediction" (Jiang et al., Computers in Biology and Medicine, 2020) proposes a brain network modeling method based on hierarchical graph convolution. This method constructs the connectivity relationships between brain regions as a graph structure and designs a hierarchical graph convolution network to mine the multi-level nested structural information in the brain map. During feature extraction, Hi-GCN fully utilizes the local and global graph topology and achieves good classification performance in brain disease prediction tasks. Furthermore, Hi-GCN introduces a graph embedding learning mechanism to generate multi-scale node representations at different levels of abstraction to capture richer structural information. Experimental results show that this method demonstrates strong generalization and discrimination capabilities on multiple brain disease datasets, showing promising application prospects. Although the technical solution proposed in this paper enhances the representational capabilities of brain maps through a hierarchical structure, its method still has the following two major flaws: First, the solution does not consider the interference of outlier samples on disease diagnosis results in multi-site heterogeneous datasets such as ABIDE. Outlier samples refer to individuals that deviate significantly from the overall sample distribution. These samples may be caused by differences in scanning equipment, parameter settings, subject status, etc. Their existence may cause the trained model to deviate from the main data distribution, and then learn suboptimal or even erroneous discrimination rules, seriously affecting the accuracy of disease diagnosis. Second, functional magnetic resonance imaging signals are easily interfered with by blood oxygen fluctuations caused by the subject's physiological activities (such as heartbeat, breathing, etc.). These non-task-related signals will introduce a lot of noise.

[0004] Existing solutions do not effectively process these noise components, which affects the quality of the final constructed brain network, weakens the graph neural network model's ability to learn disease patterns, and reduces the robustness and reliability of diagnosis. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned existing technologies and provide a multi-view outlier detection system for brain disease diagnosis, which can effectively denoise brain network data while filtering out outliers in the data, including a brain network construction model and a brain disease diagnosis model;

[0006] The brain network construction model constructs a brain network based on functional magnetic resonance imaging (fMRI) data. The brain disease diagnosis model is used to remove outliers from fMRI data and obtain brain network classification results based on a multi-view joint learning method.

[0007] The multi-view joint learning method maps the brain network into multiple views, extracts features of the multiple views respectively, and performs classification based on the features of the multiple views to obtain a brain network classification result.

[0008] Furthermore, the multi-view outlier detection system for brain disease diagnosis includes a two-stage processing mechanism:

[0009] Phase 1:

[0010] Based on brain functional magnetic resonance imaging data, a brain network construction model was used to construct a brain network;

[0011] Using a brain disease diagnosis model to process the brain network, remove outlier data from brain functional magnetic resonance imaging data, and obtain brain functional magnetic resonance imaging data from which outlier data has been removed;

[0012] Phase 2:

[0013] Based on the brain functional magnetic resonance imaging data with outlier data removed, a brain network with outlier data removed was constructed using a brain network construction model;

[0014] The brain disease diagnosis model is used to process the brain network after removing outlier data to obtain the classification results of the brain network after removing outlier data.

[0015] Furthermore, based on the brain functional magnetic resonance imaging data, the specific method of constructing a brain network using a brain network construction model is as follows:

[0016] Acquiring brain functional magnetic resonance imaging data, wherein the brain functional magnetic resonance imaging data includes blood oxygen level dependent signals of each brain region;

[0017] Calculate the Pearson correlation coefficient between any two brain regions and generate an adjacency matrix. The elements in the adjacency matrix are the weights of the edges between any two nodes in the brain network. Construct a brain network based on the adjacency matrix.

[0018] The brain network includes several nodes and several edges, each node represents a brain region, and the weight of each edge represents the correlation strength between two brain regions.

[0019] Furthermore, the brain disease diagnosis model includes: a multi-threshold segmentation module, a multi-view joint learning module, a classification module, and an outlier screening module;

[0020] The brain disease diagnosis model removes outliers from functional magnetic resonance imaging data and obtains brain network classification results, including:

[0021] Use the multi-threshold segmentation module to perform threshold segmentation on the adjacency matrix of the brain network to obtain multiple views;

[0022] Use the multi-view joint learning module to generate multiple pooled views and extract features from the multiple pooled views to obtain multi-view joint features;

[0023] Use the classification module to classify the multi-view joint features and obtain the brain network classification results;

[0024] Use the outlier filtering module to identify and remove outliers in brain functional magnetic resonance imaging data based on brain network classification results.

[0025] Furthermore, a multi-threshold segmentation module is used to perform threshold segmentation on the adjacency matrix to obtain multiple views. The specific method is as follows:

[0026] Using the threshold hyperparameter The element value of the ternary adjacency matrix is nodes and The weight of the edge between nodes Greater than the threshold hyperparameter When nodes and The weight of the edge between nodes is set to 1; when the nodes and The weight of the edge between nodes Less than the threshold hyperparameter When nodes and The weight of the edge between nodes is set to -1; when the nodes and The weight of the edge between nodes Between threshold hyperparameters and threshold hyperparameters Between nodes and The weight of the edge between the nodes is set to 0; the element value of the ternary adjacency matrix is ​​used as the weight of the edge between any two nodes in the adjacency matrix to obtain the threshold hyperparameter The view below;

[0027] Set multiple different threshold hyperparameters to obtain multiple views, each corresponding to a different threshold hyperparameter.

[0028] Furthermore, the multi-view joint learning module includes a multi-image pooling submodule and a multi-task graph embedding learning submodule;

[0029] Use the multi-view joint learning module to generate multiple pooled views and extract features from multiple pooled views to obtain multi-view joint features, including:

[0030] Use the multi-graph pooling submodule to generate an indicator matrix for pooling multiple views, and pool the ternary adjacency matrix. Pool the nodes in the ternary adjacency matrix into several supernodes, and use the hyperedge between any two supernodes as the weight of the pooled edge.

[0031] Use the multi-task graph embedding learning submodule to extract features from multiple pooled views and obtain multi-view joint features;

[0032] A loss function of the multi-view joint learning module is established, and the multi-view joint learning module is trained to obtain a trained multi-view joint learning module.

[0033] Furthermore, the multi-task graph embedding learning submodule of the multi-view joint learning module includes several shared graph embedding learning layers, which include a shared graph convolutional layer and multiple dedicated graph convolutional layers, each of which corresponds to a pooled view.

[0034] The multi-task graph embedding learning submodule is used to extract the features of multiple pooled views and obtain the multi-view joint features. The specific method is as follows:

[0035] Use several shared graph embedding learning layers to perform multi-view joint learning on multiple pooled views, where A shared graph is embedded in the learning layer, and a dedicated graph convolution layer corresponding to each pooled view is used to extract dedicated features to obtain dedicated features for each pooled view;

[0036] Use the shared graph convolution layer to extract shared features from each pooled view to obtain shared features;

[0037] Fuse the dedicated features and shared features of each pooled view to obtain the fused features;

[0038] The fused features extracted from each pooled view are visually Figure 1 Consistency constraints will be Figure 1 The fusion features of each consistency constraint are spliced ​​together to obtain the multi-view joint feature.

[0039] Furthermore, the loss function of the multi-view joint learning module is established. The specific method is as follows:

[0040] Based on the binary cross entropy loss function, orthogonal loss function, balanced loss function, negative penalty loss function and L2 regularization term, the loss function of the multi-image pooling submodule is established;

[0041] Create a video Figure 1 Consistency regularization loss function is used to calculate the feature maps extracted from each pooled view. Figure 1 Consistency constraints;

[0042] Establish a priori sub-network structure regularization loss function to enhance the connection weights between nodes in the priori sub-network;

[0043] Based on the loss function of the multi-image pooling model, Figure 1 The consistency regularization loss function and the prior sub-network structure regularization loss function are used to construct the total loss function of the brain disease diagnosis model. , as shown in the following formula:

[0044] (1);

[0045] in, and are weight coefficients, Based on the visual Figure 1 Consistency regularization loss function, is the prior sub-network structure regularization loss function, is the loss function of the multi-image pooling model.

[0046] Furthermore, the classification module includes a fully connected layer; the fully connected layer is used to classify the multi-view joint features to obtain the brain network classification results.

[0047] Furthermore, an outlier filtering module is used to identify and remove outlier data based on the brain network classification results. The specific method is as follows:

[0048] Calculate the classification confidence of each category predicted by the classification module. The classification module predicts the category of the brain network as The classification confidence As shown in the following formula:

[0049] (2);

[0050] in, For category The logits score, For category The logits score;

[0051] The entropy of the probability distribution is used to quantify the uncertainty of the classification confidence of each category, and the entropy value of each category is obtained, as shown in the following formula:

[0052] (3);

[0053] in, is the total number of categories, is the entropy calculation function;

[0054] Sort the entropy values ​​of each category in descending order, select the top K% samples with the highest entropy values ​​as outlier data, and remove the outlier data.

[0055] The beneficial effects of the above technical solution are as follows: the multi-view outlier detection system for brain disease diagnosis provided by the present invention addresses the problem that outlier samples in brain network datasets can lead model learning to deviate from the true distribution. An uncertainty-based outlier screening module is designed. The uncertainty is calculated using confidence and weighted summed. The top K% samples are screened in descending order of uncertainty as outliers, and these outliers are removed from the brain data to improve diagnostic performance.

[0056] To address the problems of noisy connections and high dimensionality in single brain network data, a multi-view joint learning module was designed. This module pools multiple nodes into a single supernode, effectively removing noisy connections and reducing data dimensionality, thus preventing overfitting. Furthermore, multi-view joint learning utilizes brain structural information of varying intensities retained by multiple thresholds to learn richer complementary information, further improving the model's classification performance.

[0057] The multi-view outlier detection system for brain disease diagnosis provided by this invention improves brain disease classification performance while also providing clinical interpretability. By removing outlier samples and noisy connections, the key connections retained by the model are concentrated in functional areas such as the default mode network (DMN), the salience network (SN), and the central executive network (CEN). These networks are widely believed to be closely related to symptoms in autism research. This result is consistent with existing neuroscience findings, providing a credible functional mechanism explanation for the diagnostic results and a potential reference for subsequent clinical intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1A schematic diagram of a multi-view outlier detection system for brain disease diagnosis provided by an embodiment of the present invention;

[0059] Figure 2 A flowchart of a multi-view outlier detection system for brain disease diagnosis provided by an embodiment of the present invention;

[0060] Figure 3 Schematic diagram of a priori subnetworks related to diseases identified during the diagnosis process provided by an embodiment of the present invention, where (a) is the salience subnetwork, (b) is the default mode subnetwork, and (c) is the central executive subnetwork;

[0061] Figure 4 Schematic diagram of the interaction of prior subnetworks identified in the diagnostic process provided by an embodiment of the present invention, wherein (a) is a schematic diagram of the interaction between the saliency subnetwork and the default mode subnetwork, (b) is a schematic diagram of the interaction between the saliency subnetwork and the central executive subnetwork, (c) is a schematic diagram of the interaction between the central executive subnetwork and the visual subnetwork, (d) is a schematic diagram of the interaction between the default mode subnetwork and the sensorimotor subnetwork, (e) is a schematic diagram of the interaction between the default mode subnetwork and the visual subnetwork, and (f) is a schematic diagram of the interaction between the saliency subnetwork and the visual subnetwork. DETAILED DESCRIPTION

[0062] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0063] This embodiment is a multi-view outlier detection system for brain disease diagnosis, such as Figure 1 As shown, including: brain network construction model and brain disease diagnosis model;

[0064] The brain network construction model is used to construct a brain network based on brain functional magnetic resonance imaging data. The brain disease diagnosis model is used to remove outliers in brain functional magnetic resonance imaging data and obtain brain network classification results based on a multi-view joint learning method.

[0065] The multi-view joint learning method maps the brain network into multiple views, extracts features of the multiple views respectively, and performs classification based on the features of the multiple views to obtain a brain network classification result.

[0066] In this embodiment, the multi-view outlier detection system for brain disease diagnosis includes a two-stage processing mechanism:

[0067] Phase 1:

[0068] Based on brain functional magnetic resonance imaging data, a brain network construction model was used to construct a brain network;

[0069] Using a brain disease diagnosis model to process the brain network, remove outlier data from brain functional magnetic resonance imaging data, and obtain brain functional magnetic resonance imaging data from which outlier data has been removed;

[0070] Phase 2:

[0071] Based on the brain functional magnetic resonance imaging data with outlier data removed, a brain network with outlier data removed was constructed using a brain network construction model;

[0072] The brain disease diagnosis model is used to process the brain network after removing outlier data to obtain the classification results of the brain network after removing outlier data.

[0073] In this embodiment, the specific method of constructing a brain network using a brain network construction model based on brain functional magnetic resonance imaging data is as follows:

[0074] Acquiring brain functional magnetic resonance imaging data, wherein the brain functional magnetic resonance imaging data includes blood oxygen level dependent signals of each brain region;

[0075] Calculate the Pearson correlation coefficient between any two brain regions and generate an adjacency matrix. The elements in the adjacency matrix are the weights of the edges between any two nodes in the brain network. Construct a brain network based on the adjacency matrix.

[0076] The brain network includes several nodes and several edges, each node represents a brain region, and the weight of each edge represents the correlation strength between two brain regions.

[0077] In this embodiment, the brain disease diagnosis model includes: a multi-threshold segmentation module, a multi-view joint learning module, a classification module, and an outlier screening module;

[0078] The multi-threshold segmentation module is used to perform threshold segmentation on the adjacency matrix to obtain multiple views;

[0079] The multi-view joint learning module is used to generate multiple pooled views and extract features from the multiple pooled views to obtain multi-view joint features;

[0080] The classification module is used to classify the joint features of multiple views and obtain the brain network classification results;

[0081] The outlier screening module is used to identify and remove outlier data in brain functional magnetic resonance imaging data based on brain network classification results.

[0082] In this embodiment, the outlier data in the brain functional magnetic resonance imaging data is removed based on the multi-view outlier detection system for brain disease diagnosis, and the brain network classification results are obtained based on the multi-view joint learning method, such as Figure 2 As shown, the specific process is:

[0083] S1 constructs a brain network using a brain network construction model based on brain functional magnetic resonance imaging data;

[0084] Acquire brain functional magnetic resonance imaging data, including blood oxygen level-dependent signals in various brain regions;

[0085] Calculate the Pearson correlation coefficient between any two brain regions and generate an adjacency matrix. The elements in the adjacency matrix are the weights of the edges between any two nodes in the brain network. Construct a brain network based on the adjacency matrix.

[0086] The brain network includes a plurality of nodes and a plurality of edges, each node represents a brain region, and the weight of each edge represents the correlation strength between two brain regions;

[0087] S2 uses a multi-threshold segmentation module to perform threshold segmentation on the adjacency matrix of the brain network to obtain multiple views;

[0088] Because the adjacency matrix of the brain network contains a large amount of noise and redundant connections, the constructed brain network is highly complex. To address this issue, threshold segmentation is used to filter out weakly correlated brain region information to initially remove noise and redundant connections in the brain network.

[0089] Using the threshold hyperparameter The element value of the ternary adjacency matrix is nodes and The weight of the edge between nodes Greater than the threshold hyperparameter When nodes and The weight of the edge between nodes is set to 1; when the nodes and The weight of the edge between nodes Less than the threshold hyperparameter When nodes and The weight of the edge between nodes is set to -1; when the nodes and The weight of the edge between nodes Between threshold hyperparameters and threshold hyperparameters Between nodes and The weight of the edge between the nodes is set to 0; the element value of the ternary adjacency matrix is ​​used as the weight of the edge between any two nodes in the adjacency matrix to obtain the threshold hyperparameter The view below;

[0090] After ternary nodes and The weight of the edge between nodes As shown in the following formula:

[0091] (1);

[0092] in, is the first nodes and The weight of the edge between nodes, is the first value in the adjacency matrix after ternary nodes and The weight of the edge between nodes;

[0093] Use the element value of the ternary adjacency matrix as the weight of the edge between any two nodes in the adjacency matrix to remove the noise and redundant connections of the brain network and obtain the threshold hyperparameter The view below;

[0094] Set multiple different threshold hyperparameters to obtain multiple views, each corresponding to a different threshold hyperparameter;

[0095] S3 uses a multi-view joint learning module to generate multiple pooled views and extract features of the multiple pooled views to obtain multi-view joint features; the multi-view joint learning module includes a multi-image pooling submodule and a multi-task graph embedding learning submodule;

[0096] S3.1 uses the multi-graph pooling submodule to generate an indicator matrix for pooling multiple views, and pools the ternary adjacency matrix. The nodes in the ternary adjacency matrix are pooled into several supernodes, and the hyperedge between any two supernodes is used as the weight of the pooled edge.

[0097] After the threshold segmentation removes a certain amount of noise in a rough manner, in order to further remove noise and reduce the dimension of the adjacency matrix, reduce the learning cost of the model and prevent overfitting, the multi-graph pooling technology is used to pool multiple nodes into a super node. The edges between super nodes are used as the weights of the enhanced hyper edges to obtain a pooled graph. In addition, the supervised graph pooling method is used to directly incorporate the classification labels into the learning process.

[0098] Use the multi-image pooling submodule to generate an indicator matrix for pooling multiple views , and pool the ternary adjacency matrix , pool the nodes in the ternary adjacency matrix into several super nodes, use the hyperedge between any two super nodes as the weight of the enhanced edge, and generate a pooled graph , as shown in the following formula:

[0099] (2);

[0100] in, is the ternary adjacency matrix, is the indicator matrix, is the transposed matrix of the indicator matrix, the indicator matrix The dimension is , is the ternary adjacency matrix The number of nodes, is the number of supernodes in the pooling graph;

[0101] After pooling, the two super nodes and The weight of the hyperedge between As shown in the following formula:

[0102] (3);

[0103] in, is the ternary adjacency matrix midpoint For the importance score of the classification task, is the ternary adjacency matrix midpoint For the importance score of the classification task, For nodes and nodes The weight of the edge between them;

[0104] Indicator Matrix It reflects the learning process from the original graph to the pooled graph. Through the pooling process, noise and redundant connections are eliminated, and the dimension is reduced, resulting in a clearer and more efficient graph structure. Sharing across different brain networks ensures pooling consistency and the feasibility of comparison, effectively aggregating brain network connections, reducing dimensionality, and eliminating noise.

[0105] S3.2 uses a multi-task graph embedding learning submodule to extract features from multiple pooled views and obtain multi-view joint features. The multi-task graph embedding learning submodule of the multi-view joint learning module includes several shared graph embedding learning layers, which include a shared graph convolutional layer and multiple dedicated graph convolutional layers, each dedicated graph convolutional layer corresponding to a pooled view.

[0106] Brain networks segmented using a high threshold typically retain only strong connections, reflecting more stable and important relationships. In contrast, networks segmented using a low threshold contain more weak connections, providing a more detailed view of the network structure. However, because a single threshold segmentation can only capture structural features within a specific range of connection strengths, it can lead to information loss and bias.

[0107] To address this issue, this embodiment uses a multi-view learning approach to obtain multiple views by setting different segmentation thresholds. A high threshold focuses on capturing strong and stable connections that are critical for understanding the core brain structures, while a low threshold retains weaker connections that provide details. By integrating multiple views, key and subtle connection patterns are preserved, providing a more comprehensive and detailed representation of the brain network. The multi-view learning approach reduces information loss and potential biases by capturing information at multiple levels, enhancing the ability to identify complex and nuanced connection patterns, thereby enabling more powerful and accurate brain network analysis.

[0108] Use several shared graph embedding learning layers to perform multi-view joint learning on multiple pooled views, where A shared graph is embedded in the learning layer, and a dedicated graph convolution layer corresponding to each pooled view is used to extract dedicated features to obtain dedicated features for each pooled view;

[0109] Use the shared graph convolution layer to extract shared features from each pooled view to obtain shared features;

[0110] Fuse the dedicated features and shared features of each pooled view to obtain the fused features;

[0111] No. The pooled view is passed through the Fusion features after shared graph embedding learning layer As shown in the following formula:

[0112] (4);

[0113] in, For the Pooling map of views, For the The shared graph embedding learning layer The feature matrix output by the dedicated graph convolution layer corresponding to the pooled view, For the The shared graph embedding learning layer The weight matrix of the dedicated graph convolution layer corresponding to the pooled views, For the The shared graph embedding learning layer The views share the feature map output by the graph convolution layer;

[0114] In this embodiment, two shared graph embedding learning layers are used to capture more compact and important graph structure information while enhancing the learning of consistency between multiple views.

[0115] The feature maps extracted from each pooled view are Figure 1 Consistency constraints will be Figure 1 The feature maps with consistency constraints are spliced ​​together to obtain multi-view joint features.

[0116] S3.3 establishes a loss function for the multi-view joint learning module, trains the multi-view joint learning module, and obtains a trained multi-view joint learning module;

[0117] S3.3.1 Establish the loss function of the multi-image pooling submodule based on the binary cross entropy loss function, orthogonal loss function, balanced loss function, negative penalty loss function and L2 regularization term;

[0118] Establishing a binary cross entropy loss function , as shown in the following formula:

[0119] (5);

[0120] in, For the The true labels of samples, For the The predicted labels of samples, is the total number of samples;

[0121] Establishing a negative penalty loss function , as shown in the following formula:

[0122] (6);

[0123] in, is the ternary adjacency matrix Middle Node-pair pooling graph Middle The membership degree of a supernode, is the ReLU function;

[0124] Negative values ​​in the indicator matrix may lead to meaningless pool allocation during training, using a negative value penalty loss function Prevent the indicator matrix F from containing negative values. Specifically, in this embodiment, by using the ReLU function to penalize negative values ​​in the indicator matrix, it is ensured that the indicator matrix values ​​remain non-negative, thereby enhancing the stability of the learning process.

[0125] Establishing an orthogonal loss function , through the penalty matrix The non-diagonal elements of are used to prevent overlap between super nodes, as shown in the following formula:

[0126] (7);

[0127] in, is the diagonalization function used to extract the diagonal elements of the matrix, is the L2 norm used to calculate the sum of squares of all elements of the matrix;

[0128] In the multi-graph pooling model, if the columns of the indicator matrix F are not orthogonal, supernodes may overlap, leading to confusion between group assignments. In this embodiment, by strengthening orthogonality, the independence of supernodes is ensured, thereby ensuring separability and a more unique group structure.

[0129] Establishing a balanced loss function , as shown in the following formula:

[0130] (8);

[0131] in, is the variance;

[0132] In applications of brain disease diagnosis, it is important that supernodes have similar sizes to avoid biased results. Imbalance in supernode size may lead to meaningless insights. Balanced loss function Used to maintain the pooling map The number of original nodes contained in each super node is relatively uniform, ensuring that the super nodes have similar sizes and enhancing the interpretability of the pooling results; in this embodiment, by minimizing The variance of the diagonal elements in ,ensures a more evenly distributed supernode size.

[0133] Establishing regularization terms , as shown in the following formula:

[0134] (9);

[0135] In this embodiment, the regularization term is used Penalty indicator matrix and the weight matrix The maximum value in , constrains the size of the parameters to reduce overfitting.

[0136] Based on the binary cross entropy loss function , orthogonal loss function , balanced loss function , negative penalty loss function and regularization term , establish the loss function of the multi-image pooling model , as shown in the following formula:

[0137] (10);

[0138] in, 、 、 、 are weight parameters, To jointly optimize the indicator matrix F and the weight matrix W to find the minimum value, is the category index of positive and negative samples;

[0139] Based on the loss function of the multi-graph pooling model, the multi-graph pooling model is optimized by backpropagation under the supervised graph pooling method to generate an indicator matrix for the pooling adjacency matrix and multiple pooled views.

[0140] S3.3.2 Establishing a visual Figure 1 Consistency regularization loss function;

[0141] Since the views obtained under different segmentation thresholds essentially represent the same brain network, there is inherent consistency information between different views. Figure 1 The consistency regularization loss function ensures that features learned from different views remain consistent, helping to reduce differences between views and thus making the representation of the network structure more uniform. Considering that positive correlation and negative correlation represent different functional connectivity patterns, this embodiment optimizes the similarity between the positive correlation structure graphs and the negative correlation structure graphs of the two views, respectively, so that the feature maps output by different views maintain a certain degree of similarity.

[0142] use and Represents the first Hedi The positive correlation graph structure of the view, using and Represents the first Hedi The negative correlation graph structure of the view, using the view Figure 1 Consistency regularization loss function Optimize Hedi View positive correlation graph structure and The similarities between Hedi Negative correlation graph structure of views and The similarity between the feature maps extracted from different views is analyzed. Figure 1 Consistency constraints;

[0143] See Figure 1 Consistency regularization loss function As shown in the following formula:

[0144] (11);

[0145] in, For the The indicator matrix of the view, For the The transposed matrix of the view's indicator matrix, For a collection of views, is the Sigmoid activation function used to normalize the node representation similarity;

[0146] S3.3.3 Establish a regularized loss function for the prior sub-network structure to enhance the connection weights between nodes within the prior sub-network;

[0147] A priori subnetworks are known brain pathways established by certain fixed brain regions in the brain. Different prior subnetworks represent different brain functions. Existing studies have shown that specific prior subnetworks, such as the salience network SN and the default mode network DMN, are crucial for the diagnosis of brain diseases. Enabling brain networks to learn how to effectively extract prior subnetworks for diagnosing brain diseases can significantly improve the ability to diagnose mental illnesses. In order to integrate prior knowledge about the structure of prior subnetworks, in this embodiment, a priori subnetwork structure regularization loss function is introduced, and the prior subnetwork structure information is directly embedded in the learning process of the multi-graph pooling module to ensure that the multi-graph pooling module learns functional subnetworks that are conducive to classification.

[0148] Establishing a priori sub-network structure regularization loss function , as shown in the following formula:

[0149] (12);

[0150] in, For nodes The membership vector of node The probability of belonging to each supernode, For nodes The membership vector of node The probability of belonging to each supernode, For the The number of brain regions in each subnetwork, is a supernode mapping function used to map nodes to supernodes according to node priority, is a constant used to avoid log 0 during the optimization process;

[0151] Regularize the loss function using the prior sub-network structure Penalize the situation where brain region nodes in the same prior sub-network are pooled into different super-nodes, and retain the information of the prior sub-network;

[0152] S3.3.4 Loss function and view based on multi-image pooling model Figure 1 The consistency regularization loss function and the prior sub-network structure regularization loss function are used to construct the total loss function of the brain disease diagnosis model. , as shown in the following formula:

[0153] (13);

[0154] in, and are weight coefficients, Based on the visual Figure 1 Consistency regularization loss function, is the prior sub-network structure regularization loss function, is the loss function of the multi-image pooling model.

[0155] In this embodiment, the weight coefficient is set and 0.1 and 0.001 respectively.

[0156] S4 uses the classification module to classify the multi-view joint features and obtain the brain network classification results;

[0157] The classification module includes a fully connected layer; the fully connected layer is used to classify the joint features of multiple views to obtain the brain network classification results.

[0158] S5 uses the outlier filtering module to identify and remove outlier data based on brain network classification results;

[0159] The softmax function is used to calculate the classification confidence of each category predicted by the classification module. The classification module predicts the category of the brain network as The classification confidence As shown in the following formula:

[0160] (14);

[0161] in, For category The logits score, For category The logits score;

[0162] The entropy of the probability distribution is used to quantify the uncertainty of the classification confidence of each category, and the entropy value of each category is obtained, as shown in the following formula:

[0163] (15);

[0164] in, is the total number of categories, is the entropy calculation function;

[0165] Entropy is a standard measure of uncertainty and plays an important role in classification tasks to identify samples about which the model is uncertain. Higher entropy values ​​indicate greater uncertainty in the model’s predictions, which suggests that the corresponding brain maps may be more difficult to classify correctly and are therefore more likely to be outliers.

[0166] Sort the entropy values ​​of each category in descending order, select the top K% samples with the highest entropy values ​​as outlier data, and remove outlier data from brain functional magnetic resonance imaging data;

[0167] S6 uses a brain network construction model to construct a brain network that removes outlier data based on brain functional magnetic resonance imaging data that removes outlier data;

[0168] S7 inputs the brain network after removing the outlier data into the brain disease diagnosis model to obtain a classification result of the brain network after removing the outlier data;

[0169] The brain network without outlier data is input into the brain disease diagnosis model, and passes through the multi-threshold segmentation module, multi-view joint learning module and classification module in sequence to obtain the classification result of the brain network without outlier data.

[0170] In this embodiment, a comparative experiment was conducted on the ABIDE dataset using a multi-view outlier detection method for brain disease diagnosis and existing methods. The results of the comparative experiment are shown in Table 1. The brain disease diagnosis model BrainOSM of this embodiment outperforms traditional methods such as SVM / RF, with improvements of 5.51% and 6.42% in ACC and AUC, respectively. Compared with non-graph deep learning methods such as ASD-DiagNet and DAE, the brain disease diagnosis model BrainOSM of this embodiment further improved ACC by 3.47% and AUC by 2.89%. In addition, our method surpassed the CNN-based method, with significant gains in both indicators. Finally, by comparing with GCN-based models such as s-GCN, BrainGNN, ST-GCN, EigenGCN and GroupINN, the brain disease diagnosis model BrainOSM of this embodiment achieved significant improvements, demonstrating its effectiveness. These results show that the model has significant performance and can be used for brain disease diagnosis.

[0171] Figure 3 and Figure 4 They are respectively the disease-related brain subnetworks discovered by the multi-view outlier detection method for brain disease diagnosis provided by this embodiment and the interaction diagrams between the subnetworks, Figure 3 (a) is the saliency subnetwork, Figure 3 (b) is the default mode subnetwork, Figure 3 (c) is the central execution subnetwork, Figure 4(a) is a schematic diagram of the interaction between the saliency subnetwork and the default mode subnetwork, Figure 4 (b) is a schematic diagram of the interaction between the saliency subnetwork and the central executive subnetwork, Figure 4 (c) is a schematic diagram of the interaction between the central executive subnetwork and the visual subnetwork, Figure 4 (d) is a schematic diagram of the interaction between the default mode subnetwork and the sensorimotor subnetwork, Figure 4 (e) is a schematic diagram of the interaction between the default mode subnetwork and the visual subnetwork, Figure 4 (f) is a schematic diagram of the interaction between the saliency subnetwork and the visual subnetwork, which illustrates that the multi-view outlier detection method for brain disease diagnosis provided in the embodiment not only improves the performance of brain disease diagnosis, but also provides clinical interpretability.

[0172] In addition to significantly improving the performance of brain disease diagnosis, the embodiment also provides strong support for neural mechanism explanation. The high-score brain subnetworks and the connections therebetween (such as DMN-SN, SN-CEN, etc.) selected in the embodiment are highly consistent with existing research on functional connectivity abnormalities of autism, showing good biological interpretability. For example, the SN plays a key role in regulating attention switching and internal-external information switching, the DMN is related to self-perception and social cognition, and the CEN affects executive function and task control. The results of the embodiment show that these subnetworks and their interactions may be the core link of the pathogenesis of brain diseases, providing theoretical support and clinical reference for future early diagnosis and intervention.

[0173] Table 1 Comparison of data between the experimental group and the control group

[0174]

[0175] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features thereof; and such modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the present application.

Claims

1. A multi-view outlier detection system for brain disease diagnosis, characterized by: Including brain network construction models and brain disease diagnosis models; The brain network construction model is used to construct brain networks based on brain functional magnetic resonance imaging data; The brain disease diagnosis model is used to remove outliers from functional magnetic resonance imaging (fMRI) data and obtain brain network classification results based on a multi-view joint learning method. The multi-view joint learning method maps the brain network into multiple views, extracts features of the multiple views respectively, and classifies them based on the features of the multiple views to obtain brain network classification results; The brain disease diagnosis model includes: a multi-threshold segmentation module, a multi-view joint learning module, a classification module, and an outlier screening module; The brain disease diagnosis model removes outliers from functional magnetic resonance imaging data and obtains brain network classification results, including: Use the multi-threshold segmentation module to perform threshold segmentation on the adjacency matrix of the brain network to obtain multiple views; Use the multi-view joint learning module to generate multiple pooled views and extract features from the multiple pooled views to obtain multi-view joint features; Use the classification module to classify the multi-view joint features and obtain the brain network classification results; Use the outlier filtering module to identify and remove outliers in brain functional magnetic resonance imaging data based on brain network classification results; Use the multi-threshold segmentation module to perform threshold segmentation on the adjacency matrix to obtain multiple views. The specific method is as follows: Using the threshold hyperparameter The element value of the ternary adjacency matrix is nodes and The weight of the edge between nodes Greater than the threshold hyperparameter When nodes and The weight of the edge between nodes is set to 1; when the nodes and The weight of the edge between nodes Less than the threshold hyperparameter When nodes and The weight of the edge between nodes is set to -1; when the nodes and The weight of the edge between nodes Between threshold hyperparameters and threshold hyperparameters Between nodes and The weight of the edge between the nodes is set to 0; the element value of the ternary adjacency matrix is ​​used as the weight of the edge between any two nodes in the adjacency matrix to obtain the threshold hyperparameter The view below; Set multiple different threshold hyperparameters to obtain multiple views, each corresponding to a different threshold hyperparameter.

2. A multi-view outlier detection system for brain disease diagnosis according to claim 1, characterized in that: It includes a two-stage processing mechanism: Phase 1: Based on brain functional magnetic resonance imaging data, a brain network construction model was used to construct a brain network; Using a brain disease diagnosis model to process the brain network, remove outlier data from brain functional magnetic resonance imaging data, and obtain brain functional magnetic resonance imaging data from which outlier data has been removed; Phase 2: Based on the brain functional magnetic resonance imaging data with outlier data removed, a brain network with outlier data removed was constructed using a brain network construction model; The brain disease diagnosis model is used to process the brain network after removing outlier data to obtain the classification results of the brain network after removing outlier data.

3. The multi-view outlier detection system for brain disease diagnosis according to claim 2, characterized in that: Based on brain functional magnetic resonance imaging data, the specific method of using the brain network construction model to build a brain network is as follows: Acquiring brain functional magnetic resonance imaging data, wherein the brain functional magnetic resonance imaging data includes blood oxygen level dependent signals of each brain region; Calculate the Pearson correlation coefficient between any two brain regions and generate an adjacency matrix. The elements in the adjacency matrix are the weights of the edges between any two nodes in the brain network. Construct a brain network based on the adjacency matrix. The brain network includes several nodes and several edges, each node represents a brain region, and the weight of each edge represents the correlation strength between two brain regions.

4. The multi-view outlier detection system for brain disease diagnosis according to claim 1, characterized in that: The multi-view joint learning module includes a multi-image pooling submodule and a multi-task graph embedding learning submodule; Use the multi-view joint learning module to generate multiple pooled views and extract features from multiple pooled views to obtain multi-view joint features, including: Use the multi-graph pooling submodule to generate an indicator matrix for pooling multiple views, and pool the ternary adjacency matrix. Pool the nodes in the ternary adjacency matrix into several supernodes, and use the hyperedge between any two supernodes as the weight of the pooled edge. Use the multi-task graph embedding learning submodule to extract features from multiple pooled views and obtain multi-view joint features; A loss function of the multi-view joint learning module is established, and the multi-view joint learning module is trained to obtain a trained multi-view joint learning module.

5. The multi-view outlier detection system for brain disease diagnosis according to claim 4, characterized in that: The multi-task graph embedding learning submodule of the multi-view joint learning module includes several shared graph embedding learning layers, which include a shared graph convolutional layer and multiple dedicated graph convolutional layers, each of which corresponds to a pooled view; The multi-task graph embedding learning submodule is used to extract the features of multiple pooled views and obtain the multi-view joint features. The specific method is as follows: Use several shared graph embedding learning layers to perform multi-view joint learning on multiple pooled views, where A shared graph is embedded in the learning layer, and a dedicated graph convolution layer corresponding to each pooled view is used to extract dedicated features to obtain dedicated features for each pooled view; Use the shared graph convolution layer to extract shared features from each pooled view to obtain shared features; Fuse the dedicated features and shared features of each pooled view to obtain the fused features; The fusion features extracted from each pooled view are constrained by view consistency, and the fusion features that have passed the view consistency constraint are spliced ​​to obtain the multi-view joint features.

6. The multi-view outlier detection system for brain disease diagnosis according to claim 5, characterized in that: Establish the loss function of the multi-view joint learning module. The specific method is: Based on the binary cross entropy loss function, orthogonal loss function, balanced loss function, negative penalty loss function and L2 regularization term, the loss function of the multi-image pooling submodule is established; Establish a view consistency regularization loss function; Establish a priori sub-network structure regularization loss function to enhance the connection weights between nodes in the priori sub-network; Based on the loss function of the multi-image pooling model, the view consistency regularization loss function and the prior sub-network structure regularization loss function, the total loss function of the brain disease diagnosis model is constructed. , as shown in the following formula: (1)； in, and are weight coefficients, is the view consistency regularization loss function, is the prior sub-network structure regularization loss function, is the loss function of the multi-image pooling model.

7. The multi-view outlier detection system for brain disease diagnosis according to claim 6, characterized in that: The classification module includes a fully connected layer; the fully connected layer is used to classify the multi-view joint features to obtain a brain network classification result.

8. The multi-view outlier detection system for brain disease diagnosis according to claim 7, characterized in that: Use the outlier filtering module to identify and remove outlier data based on brain network classification results. The specific method is as follows: Calculate the classification confidence of each category predicted by the classification module. The classification module predicts the category of the brain network as The classification confidence As shown in the following formula: (2)； in, For category The logits score, For category The logits score; The entropy of the probability distribution is used to quantify the uncertainty of the classification confidence of each category, and the entropy value of each category is obtained, as shown in the following formula: (3)； in, is the total number of categories, is the entropy calculation function; Sort the entropy values ​​of each category in descending order, select the top K% samples with the highest entropy values ​​as outlier data, and remove the outlier data.

Images