Brain disease risk prediction method and device based on adaptive sub-graph contrast distillation, equipment and storage medium

Abstract

The application discloses a brain disease risk prediction method and device based on adaptive subgraph comparison distillation, equipment and a storage medium. The method comprises the following steps: collecting multi-modal brain image data and preprocessing, segmenting the multi-modal brain image data into multiple regions of interest and extracting blood oxygen level dependent signal time series, analyzing connection weights to generate a weighted brain network graph, inputting the weighted brain network graph into a target graph neural network, outputting key lesion subgraph features, constructing a multi-objective loss term based on the key lesion subgraph features, training a model based on the multi-objective loss term, predicting the brain disease risk through the trained model, and realizing the mining of deep pathological features from limited and unbalanced neuroimaging data and the improvement of the brain disease risk prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method, apparatus, device, and storage medium for predicting the risk of brain diseases based on adaptive subgraph comparative distillation. Background Technology

[0002] With the accelerating aging of the global population, chronic neurodegenerative diseases, such as Parkinson's Disease (PD) and Alzheimer's Disease (AD), have become a public health threat to human health. These diseases have long courses, insidious early symptoms, and complex pathological mechanisms. Modern neuroscience research shows that the brain is a complex and dynamic network system, and the occurrence of neurodegenerative diseases is often accompanied by abnormalities in the topology and functional remodeling of the brain's functional connectivity network (FCN). Functional magnetic resonance imaging (fMRI), as a non-invasive brain imaging technique, can indirectly reflect neuronal activity through blood oxygen level dependent (BOLD) signals and has become a core tool for exploring the pathological mechanisms of brain diseases.

[0003] In recent years, deep learning technologies, represented by Graph Neural Networks (GNNs), have been widely used in brain network classification tasks due to their excellent ability to process non-Euclidean graph structures. However, existing GNN-based brain network analysis methods still face three major technical bottlenecks in practical clinical applications: First, the semantic loss problem in brain network modeling. Traditional methods typically use Pearson correlation coefficients to construct functional connectivity matrices and apply a fixed hard threshold for binarization. While this method simplifies computation, it often erases fine-grained semantic information about the strength of connections between brain regions and ignores the overall topological similarity of connection distribution, making it difficult to capture deep pathological signals. Second, the difficulty in locating key pathogenic subgraphs. Existing models often use global pooling (such as Mean Pooling or Sum Pooling) to aggregate the features of nodes in all brain regions into a graph-level representation. This "averaging" operation easily masks small local lesion abnormalities, making it impossible for the model to adaptively extract the key brain regions (subgraphs) most discriminative for disease diagnosis. This results in a lack of interpretability, making it difficult for clinicians to identify specific diseased brain regions from the model results. Finally, there is the risk of overfitting due to small sample sizes and high noise levels. Acquiring medical imaging data is costly, and labeled data is extremely scarce (typically only a few hundred samples). Overfitting is highly likely to occur in training deep graphical neural networks with large-scale parameters. Furthermore, fMRI signals themselves have a low signal-to-noise ratio and are greatly affected by physiological noise such as head movements and respiration, further increasing the difficulty of model training. Summary of the Invention

[0004] The main objective of this invention is to provide a method, apparatus, device, and storage medium for predicting brain disease risk based on adaptive subgraph comparative distillation. This invention aims to solve the problems of semantic gaps in brain network modeling, difficulty in locating key pathogenic subgraphs, and overfitting caused by small samples and high noise in existing technologies. These problems make it difficult to extract deep pathological features from limited and unbalanced neuroimaging data and to predict brain risk in a timely and accurate manner.

[0005] To achieve the above objectives, the present invention provides a method for predicting the risk of brain diseases based on adaptive subgraph contrastive distillation, the method comprising the following steps:

[0006] Multimodal brain imaging data was acquired and preprocessed. The preprocessed brain imaging data was segmented into multiple regions of interest, and the blood oxygenation level-dependent signal time series of each region of interest was extracted. The preprocessing included time correction, head movement correction, and spatial normalization.

[0007] Based on the blood oxygen level-dependent signal time series, a connection weight analysis is performed on each region of interest to generate a weighted brain network diagram.

[0008] Construct a target graph neural network model, input the weighted brain network map into the target graph neural network model, and output key lesion sub-graph features;

[0009] A multi-objective loss term is constructed based on the key lesion subgraph features, and the target graph neural network model is trained based on the multi-objective loss term. The trained target graph neural network model is used to predict the risk of brain diseases. The multi-objective loss term includes a classification cross-entropy loss term, a knowledge distillation loss term, and a contrastive learning loss term.

[0010] Optionally, the step of performing connectivity weight analysis on each region of interest based on the blood oxygen level-dependent signal time series to generate a weighted brain network map includes:

[0011] Pearson correlation coefficient was used to perform connectivity weight analysis on each region of interest based on the blood oxygenation level-dependent signal time series, and the original functional connectivity weights between each region of interest were obtained, as shown in the following formula:

[0012]

[0013] in, This represents the original functional connectivity weights between region of interest i and region of interest j. The signal amplitude represents the blood oxygen level of region of interest i at time point l, dependent on the signal amplitude. and Let represent the average signal amplitudes of regions of interest i and j on the blood oxygen level dependent signal time series, respectively. Indicates the total number of sampling time points;

[0014] Construct a connection distribution feature vector for each region of interest based on the original functional connection weights between each region of interest;

[0015] Calculate the intersection and union of each region of interest based on the connection distribution feature vector, referring to the following formula:

[0016]

[0017]

[0018] in, This represents the intersection between region of interest i and region of interest j, used to measure the strength of the shared connectivity patterns between the two brain regions. This represents the union between regions of interest i and j, used to measure the strength range of the overall connectivity patterns between the two brain regions. and These are regions of interest i and j, and a third-party region of interest. Connection weights;

[0019] Multi-semantic soft Jaccard similarity connection weights are constructed based on intersection and union, referring to the following formula:

[0020]

[0021] in, This represents the connection weight between region of interest i and region of interest j. This represents the intersection between region of interest i and region of interest j. This represents the union of regions of interest i and regions of interest j. Represents a positive number, used to prevent the denominator from being zero;

[0022] A reconstructed weight matrix is ​​constructed based on the multi-semantic soft Jaccard similarity connection weights, and a weighted brain network graph is generated based on the reconstructed weight matrix.

[0023] Optionally, the target graph neural network model includes a teacher model and a student model;

[0024] The teacher model is configured to use a multi-layer graph convolutional network to extract features from the weighted brain network graph to obtain global topological features of the whole brain.

[0025] The student model is configured to calculate the importance score of each brain region node relative to the global topological features of the whole brain based on an adaptive attention mechanism, wherein the brain region node is the mapping unit of the region of interest in the graph structure.

[0026] The student model is further configured to prune the brain region nodes based on the importance score to obtain key brain region nodes, perform graph convolution operation based on the subgraph structure formed by the key brain region nodes to obtain subgraph node features, and aggregate the subgraph node features into key lesion subgraph features.

[0027] Optionally, the update process of the hidden states of each brain region node in the multilayer graph convolutional network of the teacher model refers to the following formula:

[0028]

[0029]

[0030] in, and They represent the first The brain region node in the first Layer and first The feature vector of the layer, Indicates the first Layer feature dimension Represents the set of real numbers. This indicates the relationship between nodes in the graph adjacency matrix and brain regions. The set of directly connected neighboring brain region nodes, wherein the graph adjacency matrix is ​​constructed based on the reconstructed weight matrix of the weighted brain network graph. The graph adjacency matrix represents the first... The element in row j, and These represent brain region nodes. And the degree of brain region node j, This represents a symmetric normalization term used to stabilize the learning process. Indicates the first The learnable parameter matrix of the layer, It is a non-linear activation function;

[0031] The teacher model is further configured to obtain node feature vectors output by each brain region node through iterative propagation of a multilayer graph convolutional network, and to perform a hybrid pooling operation on the node feature vectors output by each brain region node through a multilayer perceptron to obtain global topological features of the whole brain. The hybrid pooling operation refers to the following formula:

[0032]

[0033] in, Represents the global topological features of the whole brain. This represents a multilayer perceptron. This is used to concatenate the result vectors of average pooling and max pooling along the feature dimension. This indicates that the node feature vectors output by all brain region nodes are subjected to dimension-wise max pooling. This indicates that the node feature vectors output by all brain region nodes are averaged and pooled. This represents the total number of nodes in the brain region, which is the total number of nodes in the region of interest. This represents the total number of layers in the graph convolutional network.

[0034] Optionally, the student model is further configured to calculate the importance score of each brain region node based on the node feature vector output by each brain region node and the global topological features of the whole brain, referring to the following formula:

[0035]

[0036] in, This indicates the importance score of brain region node i. and Let represent the learnable weight matrices, respectively, used for linearly transforming the node feature vectors of brain region nodes and the global topological features of the whole brain. For the transpose of the learnable attention vector, This represents the node feature vector of brain region node i. Represents a nonlinear activation function;

[0037] The student model is also configured to normalize the importance scores of all brain region nodes to obtain attention weights, as shown in the following formula:

[0038]

[0039] in, This represents the attention weight of the i-th brain region node. The importance score of brain region node j is indicated. This represents a temperature coefficient used to adjust the smoothness of attention distribution. Represents an exponential function;

[0040] The student model is further configured to prune the brain region nodes based on the attention weights to obtain key brain region nodes, construct a subgraph structure based on the key brain region nodes, perform graph convolution operation on the subgraph structure to obtain subgraph node features of each key brain region node, and aggregate the subgraph node features into key lesion subgraph features, as shown in the following formula:

[0041]

[0042] in, Indicates key lesion sub-graph features, This represents the subgraph node features of node i learned by the learning model on the subgraph structure. This represents a set of nodes composed of nodes from key brain regions.

[0043] Optionally, the step of constructing a multi-objective loss term based on the key lesion subgraph features, training the target graph neural network model based on the multi-objective loss term, and performing brain disease risk prediction using the trained target graph neural network model includes:

[0044] Data augmentation is performed on the key lesion sub-image features of one sample in the sample batch output by the student model to construct positive sample pairs, and other samples in the sample batch that have not been data augmented are used as negative samples to construct multi-view sample pairs;

[0045] The multi-view sample pairs are processed using a contrastive learning loss function to calculate the contrastive learning loss term;

[0046] Calculate the classification cross-entropy loss term, the knowledge distillation loss term, and the feature layer distillation loss term respectively;

[0047] The total loss function is obtained by weighting and summing various loss terms using a preset set of hyperparameters.

[0048] The target graph neural network model is trained based on the total loss function;

[0049] The trained target graph neural network model is used to predict the risk of brain diseases, output the brain risk prediction results, and based on the brain risk prediction results, the attention weights are mapped to the brain atlas to generate a heat map of key pathogenic brain regions.

[0050] Optionally, the total loss function is defined with reference to the following formula:

[0051]

[0052] in, Represents the total loss item. Represents the classification cross-entropy loss term. This represents the true label of the sample. This represents the predicted label output by the target graph neural network model based on the features of key lesion subgraphs. This represents the knowledge distillation loss term. This represents the characteristic layer distillation loss term. This represents the contrastive learning loss term. , , and These represent preset hyperparameters used to balance the importance of various loss terms;

[0053] The knowledge distillation loss term is calculated using the following formula:

[0054]

[0055]

[0056] in, Indicates the distillation temperature. KL divergence is used to measure the difference between two probability distributions. This indicates the total number of disease categories. This represents the softening probability distribution of the teacher model. This represents the softening probability distribution of the student model. The teacher model represents the category. The softening probability output, The teacher model represents the category. The logits output, This represents the softening probability output of the student model for category c. This represents the logits output of the teacher model for category j;

[0057] The distillation loss term of the characteristic layer is calculated using the following formula:

[0058]

[0059] in, Denotes the Frobenius norm of a matrix. This represents the cosine similarity metric function;

[0060] The contrastive learning loss term is calculated using the following formula:

[0061]

[0062] in, This represents the contrastive learning loss term. Indicates the sample batch size. This represents the key lesion sub-map features of the i-th sample output by the student model. This represents the augmented view of the i-th sample. Indicates positive sample pairs. The cosine similarity between vectors is represented by... This represents the temperature coefficient used in contrastive learning to adjust the scale of similarity scores. This indicates a negative sample.

[0063] Furthermore, to achieve the above objectives, the present invention also proposes a brain disease risk prediction device based on adaptive subgraph contrastive distillation, the device being configured to implement the steps of the brain disease risk prediction method based on adaptive subgraph contrastive distillation as described above, the device comprising:

[0064] The data processing module is used to acquire multimodal brain imaging data and perform preprocessing. The preprocessed brain imaging data is segmented into multiple regions of interest, and the blood oxygen level dependent signal time series of each region of interest is extracted. The preprocessing includes time correction, head movement correction and spatial normalization.

[0065] The relationship analysis module is used to perform connection weight analysis on each region of interest based on the blood oxygen level dependent signal time series, and generate a weighted brain network diagram.

[0066] The model building module is used to build a target graph neural network model, input the weighted brain network map into the target graph neural network model, and output key lesion sub-graph features;

[0067] The risk prediction module is used to construct a multi-objective loss term based on the key lesion subgraph features, and to train the target graph neural network model based on the multi-objective loss term. The trained target graph neural network model is used to predict the risk of brain diseases. The multi-objective loss term includes a classification cross-entropy loss term, a knowledge distillation loss term, and a contrastive learning loss term.

[0068] Furthermore, to achieve the above objectives, this application also proposes a brain disease risk prediction device based on adaptive subgraph comparative distillation. The device includes: a memory, a processor, and a brain disease risk prediction program based on adaptive subgraph comparative distillation stored in the memory. The processor is used to run the brain disease risk prediction program based on adaptive subgraph comparative distillation, and the computer program is configured to implement the steps of the brain disease risk prediction method based on adaptive subgraph comparative distillation as described above.

[0069] In addition, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the brain disease risk prediction method based on adaptive subgraph contrast distillation as described above.

[0070] This invention effectively eliminates data interference and improves data quality and consistency through the acquisition and preprocessing of multimodal brain imaging data. By segmenting regions of interest, extracting time-series blood oxygenation signals, and performing connectivity weight analysis, it achieves in-depth mining of brain structural and functional characteristics, transforming abstract brain imaging data into structured brain network diagrams that clearly present the functional relationships between brain regions. Through the construction and training of a target graph neural network model, it accurately extracts key lesion sub-graph features, and by combining constraints from multi-target loss terms, significantly improves the model's feature recognition and generalization abilities. Finally, the trained model enables brain disease risk prediction, effectively improving the accuracy and efficiency of brain disease risk prediction, reducing errors caused by manual intervention, and effectively mining the potential features and inter-regional relationships of brain lesions, providing important support for brain pathology risk analysis and prediction. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a schematic diagram of the structure of a brain disease risk prediction device based on adaptive subgraph comparative distillation, which is part of the hardware operating environment of the embodiment of the present invention.

[0073] Figure 2 This is a flowchart illustrating the first embodiment of the brain disease risk prediction method based on adaptive subgraph comparative distillation of the present invention.

[0074] Figure 3 This is a flowchart illustrating the second embodiment of the brain disease risk prediction method based on adaptive subgraph comparative distillation of the present invention.

[0075] Figure 4 This is a structural block diagram of the first embodiment of the brain disease risk prediction device based on adaptive subgraph comparative distillation of the present invention.

[0076] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0077] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0078] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a brain disease risk prediction device based on adaptive subgraph comparative distillation, which is part of the hardware operating environment of the embodiment of the present invention.

[0079] like Figure 1 As shown, the brain disease risk prediction device based on adaptive subgraph contrast distillation may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; the user interface 1003 may include standard wired or wireless interfaces. The network interface 1004 may optionally include standard wired or wireless interfaces (such as Wireless-Fidelity (Wi-Fi) interfaces). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0080] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the brain disease risk prediction device based on adaptive subgraph contrastive distillation, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0081] like Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a brain disease risk prediction program based on adaptive subgraph contrastive distillation.

[0082] exist Figure 1 In the brain disease risk prediction device based on adaptive subgraph comparative distillation shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the brain disease risk prediction device based on adaptive subgraph comparative distillation of the present invention can be set in the brain disease risk prediction device based on adaptive subgraph comparative distillation. The brain disease risk prediction device based on adaptive subgraph comparative distillation calls the brain disease risk prediction program based on adaptive subgraph comparative distillation stored in the memory 1005 through the processor 1001, and executes the brain disease risk prediction method based on adaptive subgraph comparative distillation provided in the embodiments of the present invention.

[0083] This invention provides a method for predicting the risk of brain diseases based on adaptive subgraph contrastive distillation, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the brain disease risk prediction method based on adaptive subgraph comparative distillation of the present invention.

[0084] In this embodiment, the brain disease risk prediction method based on adaptive subgraph contrastive distillation includes the following steps:

[0085] Step S10: Collect multimodal brain imaging data and perform preprocessing. Segment the preprocessed brain imaging data into multiple regions of interest and extract the blood oxygenation level dependent signal time series of each region of interest. The preprocessing includes time correction, head motion correction and spatial normalization.

[0086] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a terminal electronic device capable of realizing the above functions. The following description uses a brain disease risk prediction device based on adaptive subgraph contrastive distillation (hereinafter referred to as the prediction device) as an example to illustrate this embodiment and the following embodiments.

[0087] It should be noted that the region of interest refers to a specific brain region that is relevant to the target brain disease and requires focused analysis, selected from preprocessed complete brain imaging data based on the needs of brain disease research and the characteristics of brain anatomy.

[0088] It should be noted that blood oxygen level dependent signal time series refers to a continuous signal sequence extracted from each region of interest that reflects the change of blood oxygen level in brain tissue over time within that region. It is the core data reflecting the functional activity state of brain regions.

[0089] It is understood that this embodiment enhances the original functional connectivity network to highlight fine-grained pathological semantic information related to the disease, overcoming the information loss caused by traditional Pearson correlation coefficient and hard thresholding methods.

[0090] In the specific implementation, resting functional magnetic resonance imaging (fMRI) images and structural images of the subject are acquired; the resting fMRI images and structural images are sequentially processed with layer-time correction, head motion correction, and spatial standardization; the processed brain images are divided into multiple regions of interest (ROIs) using brain atlases; time series of blood oxygen level-dependent signals are extracted from each ROI; and the time series of blood oxygen level-dependent signals from all ROIs are organized to form a standardized time series set.

[0091] In some embodiments, given a containing The fMRI dataset (functional Magnetic Resonance Imaging) of each subject was preprocessed (e.g., head movement correction, spatial registration, temporal filtering) and then segmented using AAL (Automated Anatomical Labeling, a brain anatomical atlas used to map voxels in the fMRI dataset to known anatomical brain regions) maps. Each region of interest (ROI) is identified. For each ROI, its oxygenation level-dependent (BOLD) signal time series is extracted to form a dataset. ,in Representing the Time series of brain regions This represents the total number of sampling time points.

[0092] Step S20: Based on the blood oxygen level dependent signal time series, perform connectivity weight analysis on each region of interest to generate a weighted brain network diagram.

[0093] It should be noted that a weighted brain network diagram can be a structure that uses each region of interest as a network node and the connection weight between different regions as the weight of the edge between the nodes. It can present the functional connection relationship of each region of interest in the brain in a graphical and networked form, and can intuitively reflect the distribution of the functional association strength between brain regions.

[0094] Understandably, this embodiment transforms the blood oxygen signal time series of each region of interest into the functional connectivity strength between regions through connection weight analysis, quantifies the degree of functional association between different regions of the brain, breaks the limitation of independent analysis of a single region, and realizes the mining of brain functional connectivity features; by generating a weighted brain network diagram, the abstract connection weight data is transformed into an intuitive graphical structure, which facilitates the subsequent overall analysis of the brain functional network and the identification of key connections.

[0095] In practical implementation, the prediction device can construct a multi-semantic soft Jaccard functional connectivity network. It calculates the original correlations between ROIs and constructs a connectivity distribution feature vector. Based on the intersection and union relationships of these feature vectors, a soft Jaccard similarity metric is introduced to reconstruct a functional connectivity matrix containing higher-order topological semantics, generating a weighted brain network graph.

[0096] In some embodiments, the prediction device can use correlation analysis to calculate the correlation between signal time series of regions of interest (ROIs) based on the blood oxygen level dependent signal time series of each RRI. The correlation coefficient is used as the connection weight between the two regions; the higher the correlation, the greater the connection weight, and vice versa. Simultaneously, by setting a reasonable threshold, weak connections with excessively low weights are eliminated to reduce noise interference and retain statistically significant functional connections. Each RRI is treated as an independent node in the network, and the calculated connection weights between regions are used as the weights of the edges between nodes to construct a weighted brain network graph. In the network graph, the size of the nodes can be set according to the importance of the region, and the thickness of the edges can be adjusted according to the magnitude of the connection weights, visually presenting the functional connection distribution and association strength of different brain regions.

[0097] Step S30: Construct a target graph neural network model, input the weighted brain network map into the target graph neural network model, and output key lesion sub-graph features.

[0098] It should be noted that the target graph neural network model can be a deep learning model that is based on the graph neural network (GNN) architecture, combined with the structural characteristics of the brain weighted brain network map and the needs of brain disease risk prediction, and is specifically improved and constructed. It can extract, analyze and process features from networked brain imaging data and be used for feature mining related to key lesions.

[0099] Key lesion subgraph features refer to the set of subgraph features extracted from weighted brain network maps through a target graph neural network model. These features are related to the pathogenesis of brain diseases and can reflect the characteristics of lesion regions and the relationships between lesions.

[0100] Understandably, this embodiment, by constructing a targeted target graph neural network model, fully adapts to the structured characteristics of weighted brain network graphs, solving the problem that traditional deep learning models struggle to handle networked data, and achieving in-depth mining of brain network features. Through the model's attention mechanism and feature fusion operation, it can accurately focus on key lesion regions and connections related to brain diseases, effectively extract key lesion subgraph features, eliminate irrelevant interference features, and improve the targeting and effectiveness of features. At the same time, it transforms the abstract brain network graph into quantifiable lesion feature vectors, providing core feature inputs for subsequent model training and disease risk prediction, realizing a leap from brain network analysis to lesion feature mining, and improving the accuracy and efficiency of feature mining.

[0101] In some embodiments, the prediction device may be based on a hierarchical "teacher-student" graph neural network architecture to construct the target graph neural network model. A framework is built that includes a pre-trained teacher model and a lightweight student model. The teacher model utilizes deep GNNs to extract global topological features of the whole brain as prior knowledge guidance.

[0102] Step S40: Construct a multi-objective loss term based on the key lesion sub-graph features, and train the target graph neural network model based on the multi-objective loss term. Perform brain disease risk prediction through the trained target graph neural network model. The multi-objective loss term includes a classification cross-entropy loss term, a knowledge distillation loss term, and a contrastive learning loss term.

[0103] It should be noted that the classification cross-entropy loss term is used to measure the difference between the disease risk prediction results output by the target graph neural network model and the actual labels, focusing on optimizing the model's classification prediction ability and ensuring that the model can accurately distinguish between normal and diseased samples.

[0104] The knowledge distillation loss term refers to the process of introducing a pre-trained teacher model to pass the knowledge (such as feature distribution and prediction probability) of the teacher model to the target graph neural network model (student model), thereby constraining the training process of the student model, improving the model's generalization ability and prediction stability, and reducing overfitting.

[0105] The contrastive learning loss term is used to improve the model's ability to distinguish lesion features and optimize the model's feature extraction accuracy by comparing key lesion sub-image features of different samples (such as normal samples and lesion samples, samples with different lesion degrees) to narrow the feature distance of similar samples and widen the feature distance of dissimilar samples.

[0106] In its implementation, the prediction device introduces an adaptive attention mechanism into the student model, calculating the importance score of each brain region node relative to the global state. Based on the score, non-critical nodes are dynamically pruned, the most discriminative lesion sub-graphs are extracted, and these sub-graph representations are aggregated.

[0107] Data augmentation (node ​​masking, edge perturbation) is applied to the extracted brain network subgraphs to construct multi-view sample pairs. A contrastive learning loss function is used to shorten the distance between similar views and widen the distance between dissimilar views, thereby enhancing the robustness of features in the latent space.

[0108] By combining classification cross-entropy loss, knowledge distillation loss (aligning teacher-student distributions), and contrastive learning loss, the model is jointly trained end-to-end, outputting disease prediction results and heat maps of key pathogenic brain regions.

[0109] This embodiment effectively eliminates data interference and improves data quality and consistency through the acquisition and preprocessing of multimodal brain imaging data. By segmenting regions of interest, extracting time-series blood oxygenation signals, and performing connectivity weight analysis, it achieves in-depth mining of brain structural and functional characteristics, transforming abstract brain imaging data into structured brain network maps that clearly present the functional relationships between brain regions. Through the construction and training of a target graph neural network model, key lesion sub-graph features are accurately extracted. Combined with constraints from multi-target loss terms, the model's feature recognition and generalization abilities are significantly improved. Finally, the trained model is used to predict the risk of brain diseases, effectively improving the accuracy and efficiency of brain disease risk prediction, reducing errors caused by manual intervention, and effectively mining the potential features and inter-regional relationships of brain lesions, providing important support for brain pathology risk analysis and prediction.

[0110] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the brain disease risk prediction method based on adaptive subgraph comparative distillation of the present invention.

[0111] Based on the first embodiment described above, in this embodiment, step S20 further includes:

[0112] Step S201: Using Pearson correlation coefficient, perform connectivity weight analysis on each region of interest based on the blood oxygen level dependent signal time series to obtain the original functional connectivity weights between each region of interest.

[0113] It should be noted that the Pearson correlation coefficient is used to quantify the degree of linear correlation between two variables, with a value ranging from [-1, 1]. The closer the value is to 1 or -1, the stronger the linear correlation between the two variables. A value of 0 indicates no linear correlation. In this embodiment, it is used to calculate the correlation between the blood oxygenation level-dependent signal time series of two regions of interest, thereby obtaining the original functional connectivity weights. The original functional connectivity weights reflect the initial degree of correlation between the changes in blood oxygenation signals between regions.

[0114] Understandably, first, calculate any two brain regions and Original functional connection strength between This invention uses the Pearson correlation coefficient for calculation, referring to the following formula:

[0115]

[0116] in, This represents the original functional connectivity weights between region of interest i and region of interest j. The signal amplitude represents the blood oxygen level of region of interest i at time point l, dependent on the signal amplitude. and Let represent the average signal amplitudes of regions of interest i and j on the blood oxygen level dependent signal time series, respectively. This represents the total number of sampling time points; the resulting original functional connectivity matrix. Although it can reflect the linear correlation between brain regions, it is easily affected by individual differences and background noise, and fails to capture higher-order topological coordination patterns.

[0117] Step S202: Construct the connection distribution feature vector of each region of interest based on the original functional connection weights between each region of interest.

[0118] It should be noted that the connectivity distribution feature vector can be a vector formed by arranging the original functional connectivity weights of each region of interest with all other regions of interest in a preset order. It is used to characterize the overall connectivity distribution features of the region of interest and includes the initial connectivity strength information of the region with all other regions.

[0119] Step S203: Calculate the intersection and union of each region of interest based on the connection distribution feature vector.

[0120] It should be noted that, in order to overcome the above limitations, this embodiment proposes a novel multi-semantic similarity reconstruction method. This method treats the connectivity pattern of each brain region (hereinafter referred to as a brain region) as a high-dimensional feature, and redefines the connectivity weights by measuring the structural overlap between these features. (Defining brain regions) The connection distribution feature vector is This vector fully describes the brain region. Functional connectivity spectrum with all other brain regions.

[0121] Based on this, this invention introduces the concept of soft Jaccard similarity to quantify the similarity of connectivity patterns between two brain regions. First, brain regions are defined. With brain regions Connecting the distribution feature vectors and The "intersection" between them and "union" Refer to the following formula:

[0122]

[0123]

[0124] in, This represents the intersection between region of interest i and region of interest j, used to measure the strength of the shared connectivity patterns between the two brain regions. This represents the union between regions of interest i and j, used to measure the strength range of the overall connectivity patterns between the two brain regions. and These are regions of interest i and j, and a third-party region of interest. Connection weights;

[0125] The minimum of the two connection strengths was calculated, representing the "bottleneck" strength resulting from their combined effect. The function determines whether two connections are in the same direction (both positive or both negative, resulting in +1) or in opposite directions (one positive and one negative, resulting in -1), thus preserving the directional information of connection cooperation. The total intersection strength is obtained by summing the values ​​for all third-party brain regions k. The maximum values ​​of the two connection strengths were calculated, representing the union of their respective ranges.

[0126] Step S204: Construct multi-semantic soft Jaccard similarity connection weights based on intersection and union.

[0127] It should be noted that the multi-semantic soft Jaccard similarity connection weights can be optimized by calculating the soft similarity (i.e., multi-semantic soft Jaccard similarity) of the intersection and union of the vectors based on the connection distribution feature vectors of two regions of interest. Compared with the original weights, it can more accurately represent the similarity and correlation between regions and reduce noise interference.

[0128] It is understood that, based on the above definitions, this invention calculates the reconstructed multi-semantic soft Jaccard similarity connection weights. Refer to the following formula:

[0129]

[0130] in, This represents the connection weights between regions of interest i and j, which are the reconstructed connection weights. Their range has been normalized to [value range missing]. between, This represents the intersection between region of interest i and region of interest j. This represents the union of regions of interest i and regions of interest j. Represents a very small positive number, used to prevent the denominator from being zero, and to increase the numerical stability of the calculation (e.g.) ).

[0131] Compared to the original , It considers not only the direct connection strength between brain region i and brain region j, but more importantly, it incorporates the similarity of the "niche" positions of these two brain regions within the whole-brain functional topology. If the connection patterns of two brain regions are highly consistent with those of other brain regions, then their... The values ​​will be higher. This method can effectively amplify collaborative brain regions with similar pathological changes and suppress isolated noisy connections, thereby generating an enhanced functional connectivity map that can better reveal the deep pathological mechanisms. The corresponding reconstruction weight matrix is .

[0132] Step S205: Construct a reconstructed weight matrix based on the multi-semantic soft Jaccard similarity connection weights, and generate a weighted brain network graph based on the reconstructed weight matrix.

[0133] It should be noted that the reconstructed weight matrix can be a matrix constructed based on the multi-semantic soft Jaccard similarity connection weights as elements, used to characterize the optimized connection strength between all regions of interest. The matrix dimension is consistent with the number of regions of interest, with rows and columns corresponding to each region of interest, and the element values ​​being the optimized connection weights between the corresponding two regions.

[0134] This embodiment uses the Pearson correlation coefficient to calculate the original functional connectivity weights, ensuring the scientific and standardized nature of the weight calculation and providing a reliable raw data foundation for subsequent optimization. By constructing a connectivity distribution feature vector, the connectivity relationships between a single region and all other regions are integrated into a quantized vector, achieving a holistic representation of regional connectivity features. Based on intersection and union, multi-semantic soft Jaccard similarity connectivity weights are constructed, effectively optimizing the accuracy of the original weights, reducing the interference of weak connectivity noise, and enhancing the weights' ability to represent functional associations between regions. By constructing a reconstructed weight matrix and generating a weighted brain network graph, the quality of the input data for the subsequent target graph neural network model is further improved, providing more reliable networked data support for the accurate extraction of key lesion subgraph features, indirectly improving the accuracy and effectiveness of feature mining in the entire technical solution.

[0135] Furthermore, in order to effectively improve the quality and effectiveness of key lesion sub-graph features, in one embodiment, the target graph neural network model includes a teacher model and a student model;

[0136] The teacher model is configured to use a multi-layer graph convolutional network to extract features from the weighted brain network graph to obtain global topological features of the whole brain.

[0137] The student model is configured to calculate the importance score of each brain region node relative to the global topological features of the whole brain based on an adaptive attention mechanism, wherein the brain region node is the mapping unit of the region of interest in the graph structure.

[0138] The student model is further configured to prune the brain region nodes based on the importance score to obtain key brain region nodes, perform graph convolution operation based on the subgraph structure formed by the key brain region nodes to obtain subgraph node features, and aggregate the subgraph node features into key lesion subgraph features.

[0139] In obtaining enhanced brain connectivity maps Subsequently, this invention constructs a hierarchical graph neural network framework, which includes a deep, pre-trained "teacher model" and a lightweight "student model" to learn global and local pathological discrimination features, respectively. This section describes in detail the construction and function of the teacher model.

[0140] The teacher model aims to utilize the complete, reconstructed whole-brain network. We learn a global, highly abstract representation to capture the macroscopic impact of disease on the overall functional patterns of the brain. First, based on the reconstructed weight matrix... Constructing a graph adjacency matrix By setting a threshold ,like ,but , indicating brain regions and There exists an edge between them; otherwise .

[0141] Furthermore, the teacher model employs a method that includes... Feature extraction is performed using a graph convolutional network (GCN) with layers 1-3. For the 1st layer of the GCN... layer( ), the first in the network The hidden states (feature vectors) of each node (i.e., brain region). The update process follows the formula below:

[0142]

[0143]

[0144] in, and They represent the first The brain region node in the first Layer and first The feature vector of the layer, Indicates the first Layer feature dimension Represents the set of real numbers. This indicates the relationship between nodes in the graph adjacency matrix and brain regions. The set of directly connected neighboring brain region nodes, wherein the graph adjacency matrix is ​​constructed based on the reconstructed weight matrix of the weighted brain network graph. The graph adjacency matrix represents the first... The element in row j, and These represent brain region nodes. And the degree of brain region node j, This represents a symmetric normalization term used to stabilize the learning process and prevent gradient explosion or vanishing caused by excessively high or low node values. Indicates the first The learnable parameter matrix of the layer and the multi-semantic soft Jaccard similarity connection weights replace the simple binary adjacency relationship in the traditional GCN, making the intensity of information transmission proportional to the multi-semantic similarity. It is the first The learnable parameter matrix of the layer is optimized by the teacher model during training. It is a non-linear activation function. In this invention, LeakyReLU is used to avoid the problem of neurons "dying" in the negative region of ReLU.

[0145] The teacher model is further configured to obtain the node feature vectors output by each brain region node through iterative propagation of a multilayer graph convolutional network, and to obtain the global topological features of the whole brain by performing a hybrid pooling operation on the node feature vectors output by each brain region node through a multilayer perceptron.

[0146] It should be noted that, after Iterative propagation of layer graph convolution, each node Finally, a high-dimensional feature vector that incorporates information from its higher-order neighborhood is obtained. To obtain a global representation that represents the entire mind map, this invention employs a hybrid pooling operation, which is defined by the following formula:

[0147]

[0148] in, Represents the global topological features of the whole brain. This indicates that the node feature vectors output by all brain region nodes are subjected to dimension-wise max pooling, which can capture the most significant activation features of the entire brain. This represents average pooling of the node feature vectors output by all brain region nodes, which summarizes the average functional state of the whole brain. This represents the total number of nodes in the brain region, which is the total number of nodes in the region of interest. This represents the total number of layers in the graph convolutional network. The operator concatenates the result vectors of average pooling and max pooling along the feature dimension, thus preserving both global information and local salient information. It is a multilayer perceptron that performs nonlinear transformations and dimensionality reduction on the concatenated high-dimensional vectors, ultimately outputting a compact and information-rich global representation vector of the whole brain. This vector is considered the teacher model's "understanding" of the overall brain map pathology and will serve as the knowledge source for subsequent knowledge distillation processes.

[0149] Furthermore, the student model is also configured to calculate the importance score of each brain region node based on the node feature vector output by each brain region node and the global topological features of the whole brain.

[0150] The student model is also configured to normalize the importance scores of all brain region nodes to obtain attention weights.

[0151] The student model is further configured to prune the brain region nodes based on the attention weights to obtain key brain region nodes, construct a subgraph structure based on the key brain region nodes, perform graph convolution operation on the subgraph structure to obtain subgraph node features of each key brain region node, and aggregate the subgraph node features into key lesion subgraph features.

[0152] It should be noted that the goal of the student model is to automatically and adaptively locate and learn features of the key subgraphs most relevant to disease classification from the vast whole-brain network. This process not only effectively filters out noise and redundant information but also provides interpretable pathological evidence for the model's predictions. Furthermore, to address the training challenges caused by the "small sample size" of clinical data for chronic diseases, this invention introduces a knowledge distillation mechanism, forcing the student model to learn and mimic the powerful capabilities of the teacher model.

[0153] First, this invention designs an adaptive subgraph attention pruning module. This module prunes the final node features learned by the teacher model. and global representation To dynamically calculate the importance of each brain region node. Importance rating The calculation is based on the following formula:

[0154]

[0155] in, This indicates the importance score of brain region node i. and These represent the learnable weight matrices, which are used to linearly transform the node feature vectors of brain region nodes and the global topological features of the whole brain to match their dimensions. For the transpose of the learnable attention vector, This represents the node feature vector of brain region node i. This represents a non-linear activation function; the essence of this formula is to evaluate the local features of each node through a single-layer feedforward network. In global context The importance of [the subject / method].

[0156] Subsequently, the importance scores of all nodes are normalized using the Softmax function to obtain the attention weights. Refer to the following formula:

[0157]

[0158] in, This represents the attention weight of the i-th brain region node. The importance score of brain region node j is indicated. This represents a temperature coefficient used to adjust the smoothness of attention distribution. Represents an exponential function; higher This will produce a smoother distribution, while lower This will allow you to focus your attention on a few key points.

[0159] This invention is based on attention weights Sort all brain regions from largest to smallest and select the highest-ranked one. These nodes constitute the node set of the key lesion subgraph. The student model has a structure similar to the teacher model but with fewer layers. A lightweight GCN, it is only available in this... A subgraph consisting of nodes Perform graph convolution operations on the subgraph. Feature representation of the subgraph. The result is obtained by weighted summation of the subgraph node features output by the student network, as shown in the following formula:

[0160]

[0161] in, Indicates key lesion sub-graph features, This represents the subgraph node features of node i learned by the learning model on the subgraph structure. This represents a set of nodes that constitute key brain regions; this approach ensures that nodes with greater contributions have a higher proportion in the subgraph representation.

[0162] Furthermore, to address the problem of training with few samples, this invention introduces a knowledge distillation framework. The student model not only learns to complete the classification task but also learns to mimic the behavior of the pre-trained (or co-trained) teacher model. Knowledge distillation involves two levels of loss:

[0163] 1. Prediction layer distillation: This process approximates the output probability distribution of the teacher model with that of the student model. Let the predicted probability vectors for a given sample after passing through the Softmax layer be... and Predicted layer distillation loss The KL divergence is used to measure the difference between the two distributions, and the knowledge distillation loss term is given by the following formula:

[0164]

[0165]

[0166] in, KL divergence is used to measure the difference between two probability distributions. This indicates the total number of disease categories. This represents the softening probability distribution of the teacher model. This represents the softening probability distribution of the student model. The teacher model represents the category. The softening probability output, The teacher model represents the category. The logits output, This represents the softening probability output of the student model for category c. This represents the logits output of the teacher model for category j; Indicates the distillation temperature. At this time, the probability distribution becomes smoother, enabling it to convey information about the similarity between categories, the so-called "dark knowledge." Student model Similarly.

[0167] 2. Feature Layer Distillation: To enable the student model to learn a deeper level of feature organization from the teacher model, this invention introduces feature similarity loss. This loss requires the student model to learn the subgraph representation. The relationships between samples are compared with the global representation learned by the teacher model. Maintain consistency. Specifically, for a batch of samples, calculate their respective... and Pairwise similarity matrix between samples and The feature similarity loss is the difference between these two similarity matrices. The feature layer distillation loss term is calculated using the following formula:

[0168]

[0169] in, Denotes the Frobenius norm of a matrix. denoted as the cosine similarity metric function, this loss function enables the student model to reconstruct sample-to-sample structural relationships similar to those observed by the teacher model when observing the whole brain, even when only observing lesion sub-images.

[0170] Furthermore, to improve the model's ability to discriminate subtle pathological differences and enhance the robustness of the learned features, this invention introduces a self-supervised contrastive learning mechanism in the student model branch. Step S40 above may include:

[0171] Step S401: Perform data augmentation on the key lesion sub-image features of one sample in the sample batch output by the student model to construct positive sample pairs, and use the other samples in the sample batch that have not been data augmented as negative samples to construct multi-view sample pairs;

[0172] Step S402: Process the multi-view sample pairs using the contrastive learning loss function and calculate the contrastive learning loss term;

[0173] Step S403: Calculate the classification cross-entropy loss term, knowledge distillation loss term, and feature layer distillation loss term respectively;

[0174] Step S404: Retrieve the preset hyperparameter set and perform a weighted summation on various loss terms to obtain the total loss function;

[0175] Step S405: Train the target graph neural network model based on the total loss function;

[0176] Step S406: Perform brain disease risk prediction using the trained target graph neural network model, output brain risk prediction results, and map attention weights to the brain atlas based on the brain risk prediction results to generate a heat map of key pathogenic brain regions.

[0177] Understandably, the core idea of ​​contrastive learning is to bring similar samples (positive sample pairs) closer together and push dissimilar samples (negative sample pairs) further apart within a metric space. In this invention, the subgraph representation of each sample output by the student model... We created a "positive sample" view for it using data augmentation techniques. Data augmentation can be an operation on the surface of a layer, such as randomly discarding some nodes or perturbing some edges.

[0178] For a containing A batch of samples, samples Enhanced view This constitutes a positive sample pair All other samples in the batch All are considered negative samples. This invention uses the InfoNCE loss function as the contrastive learning loss. The contrastive learning loss term is calculated using the following formula:

[0179]

[0180] in, This represents the contrastive learning loss term. Indicates the sample batch size. This represents the key lesion sub-map features of the i-th sample output by the student model. This represents the augmented view of the i-th sample. Indicates positive sample pairs. The cosine similarity between vectors is represented by... Indicates a negative sample. It is the temperature coefficient of contrastive learning, used to adjust the scale of similarity scores, thereby controlling the difficulty of distinguishing between positive and negative samples. This loss function maximizes the mutual information between positive sample pairs, forcing the model to learn more essential feature representations that are invariant to data augmentation operations.

[0181] Ultimately, this invention trains the entire model end-to-end using a multi-objective joint optimization function. Total loss function. It is a weighted sum of multiple loss terms, and the total loss function is given by the following formula:

[0182]

[0183] in, Represents the total loss item. Represents the classification cross-entropy loss term. This represents the true label of the sample. This represents the predicted label output by the target graph neural network model based on the features of key lesion subgraphs. This represents the knowledge distillation loss term. This represents the characteristic layer distillation loss term. This represents the contrastive learning loss term. , , and These represent preset hyperparameters used to balance the importance of various loss terms.

[0184] Among the loss terms mentioned above, the classification cross-entropy loss term is the main supervisory signal, driving the model to learn the classification task; the knowledge distillation loss term is used to convey the "dark knowledge" of the teacher model; the contrastive learning loss term is used to enhance the discriminative power and robustness of the features; and the feature layer distillation loss term is used to maintain the structural consistency of the feature space.

[0185] In one embodiment, the target graph neural network model minimizes the total loss using the Adam optimizer. The model of this invention can collaboratively optimize, adaptively discover lesion subgraphs, and learn robust and discriminative brain network representations. After training, attention weights... It can be visualized and mapped onto standard brain atlases, thus providing clinicians with intuitive, anatomically significant decision support, indicating which brain regions and functional connections play a key role in disease diagnosis.

[0186] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a brain disease risk prediction program based on adaptive subgraph comparative distillation. When the brain disease risk prediction program based on adaptive subgraph comparative distillation is executed by a processor, it implements the steps of the brain disease risk prediction method based on adaptive subgraph comparative distillation as described above.

[0187] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0188] The aforementioned computer-readable storage medium may be included in a brain disease risk prediction device based on adaptive subgraph contrastive distillation; or it may exist independently and not assembled into a brain disease risk prediction device based on adaptive subgraph contrastive distillation.

[0189] Furthermore, this invention also proposes a computer program product, including a brain disease risk prediction program based on adaptive subgraph comparative distillation, wherein when the brain disease risk prediction program based on adaptive subgraph comparative distillation is executed by a processor, it implements the steps of the brain disease risk prediction method based on adaptive subgraph comparative distillation as described above.

[0190] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the brain disease risk prediction method based on adaptive subgraph comparative distillation described above, and will not be repeated here.

[0191] Reference Figure 4 , Figure 4 This is a structural block diagram of the first embodiment of the brain disease risk prediction device based on adaptive subgraph comparative distillation of the present invention.

[0192] like Figure 4 As shown, the brain disease risk prediction device based on adaptive subgraph contrastive distillation proposed in this embodiment of the invention includes:

[0193] Data processing module 10 is used to acquire multimodal brain imaging data and perform preprocessing, segment the preprocessed brain imaging data into multiple regions of interest, and extract the blood oxygen level dependent signal time series of each region of interest. The preprocessing includes time correction, head movement correction and spatial normalization.

[0194] The relationship analysis module 20 is used to perform connection weight analysis on each region of interest based on the blood oxygen level dependent signal time series, and generate a weighted brain network diagram.

[0195] Model building module 30 is used to build a target graph neural network model, input the weighted brain network map into the target graph neural network model, and output key lesion sub-graph features;

[0196] The risk prediction module 40 is used to construct a multi-objective loss term based on the key lesion subgraph features, and to train the target graph neural network model based on the multi-objective loss term. The trained target graph neural network model is used to predict the risk of brain diseases. The multi-objective loss term includes a classification cross-entropy loss term, a knowledge distillation loss term, and a contrastive learning loss term.

[0197] This embodiment effectively eliminates data interference and improves data quality and consistency through the acquisition and preprocessing of multimodal brain imaging data. By segmenting regions of interest, extracting time-series blood oxygenation signals, and performing connectivity weight analysis, it achieves in-depth mining of brain structural and functional characteristics, transforming abstract brain imaging data into structured brain network maps that clearly present the functional relationships between brain regions. Through the construction and training of a target graph neural network model, key lesion sub-graph features are accurately extracted. Combined with constraints from multi-target loss terms, the model's feature recognition and generalization abilities are significantly improved. Finally, the trained model is used to predict the risk of brain diseases, effectively improving the accuracy and efficiency of brain disease risk prediction, reducing errors caused by manual intervention, and effectively mining the potential features and inter-regional relationships of brain lesions, providing important support for brain pathology risk analysis and prediction.

[0198] The brain disease risk prediction device based on adaptive subgraph comparative distillation provided in this application employs the brain disease risk prediction method based on adaptive subgraph comparative distillation in the above embodiments, and can solve the technical problem of brain disease risk prediction based on adaptive subgraph comparative distillation. Compared with the prior art, the beneficial effects of the brain disease risk prediction device based on adaptive subgraph comparative distillation provided in this application are the same as the beneficial effects of the brain disease risk prediction method based on adaptive subgraph comparative distillation provided in the above embodiments, and other technical features in the brain disease risk prediction device based on adaptive subgraph comparative distillation are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0199] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0200] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0201] In addition, for technical details not described in detail in this embodiment, please refer to the brain disease risk prediction method based on adaptive subgraph contrast distillation provided in any embodiment of the present invention, which will not be repeated here.

[0202] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0203] It should be noted that the user information (including but not limited to user device information, user personal information, user location information, user behavior information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0204] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0205] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0206] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting the risk of brain diseases based on adaptive subgraph contrastive distillation, characterized in that, The brain disease risk prediction method based on adaptive subgraph contrastive distillation includes: Multimodal brain imaging data was acquired and preprocessed. The preprocessed brain imaging data was segmented into multiple regions of interest, and the blood oxygenation level-dependent signal time series of each region of interest was extracted. The preprocessing included time correction, head movement correction, and spatial normalization. Based on the blood oxygen level-dependent signal time series, a connection weight analysis is performed on each region of interest to generate a weighted brain network diagram. Construct a target graph neural network model, input the weighted brain network map into the target graph neural network model, and output key lesion sub-graph features; A multi-objective loss term is constructed based on the key lesion sub-graph features, and the target graph neural network model is trained based on the multi-objective loss term. The trained target graph neural network model is used to predict the risk of brain diseases. The multi-objective loss term includes a classification cross-entropy loss term, a knowledge distillation loss term, and a contrastive learning loss term. The step of performing connectivity weight analysis on each region of interest based on the blood oxygen level-dependent signal time series to generate a weighted brain network map includes: Pearson correlation coefficient was used to perform connectivity weight analysis on each region of interest based on the blood oxygenation level-dependent signal time series, and the original functional connectivity weights between each region of interest were obtained, as shown in the following formula: in, This represents the original functional connectivity weights between region of interest i and region of interest j. The signal amplitude represents the blood oxygen level of region of interest i at time point l, dependent on the signal amplitude. and Let represent the average signal amplitudes of regions of interest i and j on the blood oxygen level dependent signal time series, respectively. Indicates the total number of sampling time points; Construct a connection distribution feature vector for each region of interest based on the original functional connection weights between each region of interest; Calculate the intersection and union of each region of interest based on the connection distribution feature vector, referring to the following formula: in, This represents the intersection between region of interest i and region of interest j, used to measure the strength of the shared connectivity patterns between the two brain regions. This represents the union between regions of interest i and j, used to measure the strength range of the overall connectivity patterns between the two brain regions. and These are regions of interest i and j, and a third-party region of interest. Connection weights; Multi-semantic soft Jaccard similarity connection weights are constructed based on intersection and union, referring to the following formula: in, This represents the connection weight between region of interest i and region of interest j. This represents the intersection between region of interest i and region of interest j. This represents the union of regions of interest i and regions of interest j. It represents a positive number and is used to prevent the denominator from being zero; A reconstructed weight matrix is ​​constructed based on the multi-semantic soft Jaccard similarity connection weights, and a weighted brain network graph is generated based on the reconstructed weight matrix. The target graph neural network model includes a teacher model and a student model; The teacher model is configured to use a multi-layer graph convolutional network to extract features from the weighted brain network graph to obtain global topological features of the whole brain. The student model is configured to calculate the importance score of each brain region node relative to the global topological features of the whole brain based on an adaptive attention mechanism, wherein the brain region node is the mapping unit of the region of interest in the graph structure. The student model is further configured to prune the brain region nodes based on the importance score to obtain key brain region nodes, perform graph convolution operation based on the subgraph structure formed by the key brain region nodes to obtain subgraph node features, and aggregate the subgraph node features into key lesion subgraph features. The process of constructing a multi-objective loss term based on the key lesion sub-graph features, training the target graph neural network model based on the multi-objective loss term, and using the trained target graph neural network model to predict brain disease risk includes: Data augmentation is performed on the key lesion sub-image features of one sample in the sample batch output by the student model to construct positive sample pairs, and other samples in the sample batch that have not been data augmented are used as negative samples to construct multi-view sample pairs; The multi-view sample pairs are processed using a contrastive learning loss function to calculate the contrastive learning loss term; Calculate the classification cross-entropy loss term, the knowledge distillation loss term, and the feature layer distillation loss term respectively; The total loss function is obtained by weighting and summing various loss terms using a preset set of hyperparameters. The target graph neural network model is trained based on the total loss function; The target graph neural network model is trained to predict the risk of brain diseases, output the brain risk prediction results, and based on the brain risk prediction results, the attention weights are mapped to the brain atlas to generate a heat map of key pathogenic brain regions. The total loss function is defined by the following formula: in, Represents the total loss item. Represents the classification cross-entropy loss term. This represents the true label of the sample. This represents the predicted label output by the target graph neural network model based on the features of key lesion subgraphs. This represents the knowledge distillation loss term. This represents the characteristic layer distillation loss term. This represents the contrastive learning loss term. , , and These represent preset hyperparameters used to balance the importance of various loss terms; The knowledge distillation loss term is calculated using the following formula: in, Indicates the distillation temperature. KL divergence is used to measure the difference between two probability distributions. This indicates the total number of disease categories. This represents the softening probability distribution of the teacher model. This represents the softening probability distribution of the student model. The teacher model represents the category. The softening probability output, The teacher model represents the category. The logits output, This represents the softening probability output of the student model for category c. This represents the logits output of the teacher model for category j; The distillation loss term of the characteristic layer is calculated using the following formula: in, Denotes the Frobenius norm of a matrix. This represents the cosine similarity metric function; The contrastive learning loss term is calculated using the following formula: in, This represents the contrastive learning loss term. Indicates the sample batch size. This represents the key lesion sub-map features of the i-th sample output by the student model. This represents the augmented view of the i-th sample. Indicates positive sample pairs. The cosine similarity between vectors is represented by... This represents the temperature coefficient used in contrastive learning to adjust the scale of similarity scores. This indicates a negative sample.

2. The brain disease risk prediction method based on adaptive subgraph contrastive distillation as described in claim 1, characterized in that, The update process of the hidden states of each brain region node in the multi-layer graph convolutional network of the teacher model refers to the following formula: in, and They represent the first The brain region node in the first Layer and first The feature vector of the layer, Indicates the first Layer feature dimension Represents the set of real numbers. This indicates the relationship between nodes in the graph adjacency matrix and brain regions. The set of directly connected neighboring brain region nodes, wherein the graph adjacency matrix is ​​constructed based on the reconstructed weight matrix of the weighted brain network graph. The graph adjacency matrix represents the first... The element in row j, and These represent brain region nodes. And the degree of brain region node j, This represents a symmetric normalization term used to stabilize the learning process. Indicates the first The learnable parameter matrix of the layer, It is a non-linear activation function; The teacher model is further configured to obtain node feature vectors output by each brain region node through iterative propagation of a multilayer graph convolutional network, and to perform a hybrid pooling operation on the node feature vectors output by each brain region node through a multilayer perceptron to obtain global topological features of the whole brain. The hybrid pooling operation refers to the following formula: in, Represents the global topological features of the whole brain. This represents a multilayer perceptron. This is used to concatenate the result vectors of average pooling and max pooling along the feature dimension. This indicates that the node feature vectors output by all brain region nodes are subjected to dimension-wise max pooling. This indicates that the node feature vectors output by all brain region nodes are averaged and pooled. This represents the total number of nodes in the brain region, which is the total number of nodes in the region of interest. This represents the total number of layers in the graph convolutional network.

3. The brain disease risk prediction method based on adaptive subgraph contrastive distillation as described in claim 2, characterized in that, The student model is also configured to calculate the importance score of each brain region node based on the node feature vector output by each brain region node and the global topological features of the whole brain, referring to the following formula: in, This indicates the importance score of brain region node i. and Let represent the learnable weight matrices, respectively, used for linearly transforming the node feature vectors of brain region nodes and the global topological features of the whole brain. For the transpose of the learnable attention vector, This represents the node feature vector of brain region node i. Represents a non-linear activation function; The student model is also configured to normalize the importance scores of all brain region nodes to obtain attention weights, as shown in the following formula: in, This represents the attention weight of the i-th brain region node. The importance score of brain region node j is indicated. This represents a temperature coefficient used to adjust the smoothness of attention distribution. Represents an exponential function; The student model is further configured to prune the brain region nodes based on the attention weights to obtain key brain region nodes, construct a subgraph structure based on the key brain region nodes, perform graph convolution operation on the subgraph structure to obtain subgraph node features of each key brain region node, and aggregate the subgraph node features into key lesion subgraph features, as shown in the following formula: in, Indicates key lesion sub-graph features, This represents the subgraph node features of node i learned by the learning model on the subgraph structure. This represents a set of nodes composed of nodes from key brain regions.

4. A brain disease risk prediction device based on adaptive subgraph contrastive distillation, characterized in that, The apparatus is configured to implement the brain disease risk prediction method based on adaptive subgraph contrastive distillation as described in any one of claims 1 to 3, the apparatus comprising: The data processing module is used to acquire multimodal brain imaging data and perform preprocessing. The preprocessed brain imaging data is segmented into multiple regions of interest, and the blood oxygen level dependent signal time series of each region of interest is extracted. The preprocessing includes time correction, head movement correction and spatial normalization. The relationship analysis module is used to perform connection weight analysis on each region of interest based on the blood oxygen level dependent signal time series, and generate a weighted brain network diagram. The model building module is used to build a target graph neural network model, input the weighted brain network map into the target graph neural network model, and output key lesion sub-graph features; The risk prediction module is used to construct a multi-objective loss term based on the key lesion subgraph features, and to train the target graph neural network model based on the multi-objective loss term. The trained target graph neural network model is used to predict the risk of brain diseases. The multi-objective loss term includes a classification cross-entropy loss term, a knowledge distillation loss term, and a contrastive learning loss term.

5. A brain disease risk prediction device based on adaptive subgraph contrastive distillation, characterized in that, The brain disease risk prediction device based on adaptive subgraph comparative distillation includes: a memory, a processor, and a brain disease risk prediction program based on adaptive subgraph comparative distillation stored in the memory. The processor is used to run the brain disease risk prediction program based on adaptive subgraph comparative distillation, and the brain disease risk prediction program based on adaptive subgraph comparative distillation is configured to implement the brain disease risk prediction method based on adaptive subgraph comparative distillation as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a brain disease risk prediction program based on adaptive subgraph comparative distillation, which, when executed by a processor, implements the brain disease risk prediction method based on adaptive subgraph comparative distillation as described in any one of claims 1 to 3.

Images