Intelligent classification and target recommendation system for non-structural brain injury diseases

Abstract

The invention discloses a non-structural brain injury disease intelligent classification and target recommendation system. The system comprises an individualized brain function network construction module which obtains a function connection matrix according to an individual resting state function image and a structure image, represents brain function connection in a graph form, and constructs an individualized brain function network; the disease judgment and classification model training module is used for constructing a disease judgment and classification model and training the disease judgment and classification model by utilizing a network topology combined graph neural network; the key brain region identification module is used for selecting K nodes with the highest SHAP value on the trained disease judgment and classification model by applying the SHAP and taking the K nodes as key brain regions for classification; and the personalized intervention target spot recommendation module is used for registering the key brain region to a standard brain space template and matching and recommending a treatment target spot. The method can more comprehensively and efficiently identify and recommend treatment targets, and provides more accurate data support for intelligent classification judgment and treatment of mental diseases.

Description

Technical Field

[0001] The present invention relates to the field of medical rehabilitation therapy, and in particular to an intelligent classification and target recommendation system for diseases without structural brain damage. Background Art

[0002] Mental illnesses without structural brain damage refer to those diseases that do not have directly observable brain structural abnormalities, but the patients show obvious mental symptoms or abnormal behaviors, including depression, anxiety, sleep disorders, motor dysfunction, etc. Due to the lack of obvious brain structural abnormalities, the classification of these diseases mainly depends on the patient's symptoms, behavioral manifestations and the doctor's clinical experience, resulting in challenges in the accuracy and consistency of disease classification. Even if doctors repeatedly compare clinical examination results with image information to make intelligent classifications, the sensitivity and specificity are not ideal, and manual classification still has many difficulties.

[0003] Treatment of mental illness without structural brain damage usually includes a variety of methods such as drug therapy, psychotherapy and social support. At present, most rehabilitation departments of hospitals use transcranial magnetic stimulation (TMS), a painless, non-invasive and green treatment method, to treat patients. Transcranial magnetic stimulation (TMS) is a magnetic stimulation technology that uses pulsed magnetic fields to act on the central nervous system of the brain, change the membrane potential of the nerve cells in the cerebral cortex, generate induced currents, affect brain metabolism and neural electrical activity, and thus cause a series of physiological and biochemical reactions. At present, the selection of transcranial magnetic stimulation targets is usually based on the functional brain areas that can be reasonably and conveniently determined, or the patient is given a positioning cap and the target area is manually marked on the positioning cap for target treatment selection. The selection of treatment targets for patients with different diseases plays a key role in their treatment. Designing a more appropriate target recommendation method is an effective means to improve the efficacy of rTMS in the treatment of mental illness.

[0004] Since the etiology of mental illness without structural brain damage is complex and involves knowledge from multiple fields, the study of its etiology faces great difficulties. The traditional method mainly involves analyzing the biological samples of the subject to detect the presence of one or more biomarkers. Doctors need to have high medical experience and relevant theoretical knowledge research and it takes a long time. The target selection during treatment depends on the doctor's subjective experience, which has certain limitations in medical applications (Longitudinal Analysis Method and System for Magnetic Resonance Imaging Data of Mild Brain Injury (CN202111501620.X) New Biomarkers and Methods for Diagnosis and Evaluation of Traumatic Brain Injury (CN201880084813.X)). Machine Learning (ML) is favored by medical workers due to its nonlinear fitting, no need to pre-restrict model complexity, and excellent classification and prediction capabilities. It is widely used in the classification of AD and the identification of MCI subtypes. By introducing machine learning methods into the field of medical rehabilitation, it can bring higher prediction accuracy, thereby facilitating medical staff to better implement treatment plans, optimize the efficacy of mental illness treatment, and provide strong support for target selection research for mental illness without structural brain damage, overcome the limitations of traditional methods, and provide patients and medical staff with accurate and optimized medical treatment plans. Summary of the invention

[0005] The system designed by the present invention aims to solve the problem of intelligent classification and judgment of mental illness without structural brain damage and recommendation of intervention targets. The system is based on the fusion of brain function network analysis and graph neural network, establishes an individualized brain function network, and realizes intelligent classification and target recommendation. The system first collects the patient's resting fMRI data to construct an individualized brain function network to capture the functional connection pattern between different brain regions in the resting state; then uses the network topology analysis method, combined with the graph neural network training to accurately distinguish the classification model between patients and normal people, and identifies the TOP-K key brain regions that contribute the most to the classification results through the graph convolution network (GCN) and SHAP (SHapley Additive exPlanations) value analysis method, and then aligns the identified key brain regions to the standard template, calculates their distance or weighted distance to the recommended target of non-invasive intervention (such as transcranial magnetic stimulation therapy), and combines the weights of these key brain regions in the classification model and the distance to the recommended target to obtain the recommended score of the treatment target. Finally, according to the calculated recommended score of the treatment target, a personalized recommended treatment plan for intervention targets is provided to doctors, thereby optimizing the treatment effect of neuropsychiatric diseases. Through these modular designs, the system provides a complete set of efficient, accurate and explainable solutions from data collection to intervention target recommendation, which comprehensively solves the problems of intelligent classification and judgment and treatment target selection for mental illnesses without structural brain damage.

[0006] The present invention provides an intelligent classification and target recommendation system for diseases without structural brain damage, which is mainly divided into the following modules: a data acquisition module, a network topology analysis module, a classification model training module, a model interpretability enhancement module, a target recommendation calculation module and a personalized intervention plan recommendation module. First, the system collects individual resting fMRI data of individuals to construct a patient's individualized brain function network, captures the functional connection pattern between different brain regions in the resting state, and provides a basis for subsequent network analysis. Then, the network topology analysis method is used in combination with the graph neural network (GNN) technology to train a classification model that can accurately distinguish patients from normal people. Through the effective feature extraction and classification learning of the brain function network by GNN, the model can accurately predict diseases based on the differences in individual brain functional connections. In order to improve the interpretability of the model, the pooling layer of the graph convolutional network (GCN) is combined with the SHAP (SHapley Additive exPlanations) value analysis method to identify the TOP-K key brain regions that contribute the most to the classification results. The TOP-k key brain regions identified in the patient classification are registered to the standard template, and their distance or weighted distance to the recommended targets for non-invasive interventions (such as transcranial magnetic stimulation) is calculated. By combining the weight of the brain region in the classification and the distance to the recommended target, the treatment target recommendation score is obtained, providing a scientific basis for doctors to formulate personalized intervention target recommendation treatment plans.

[0007] By integrating resting-state fMRI technology, graph neural network and graph convolutional network analysis methods, a personalized brain function network classification model with high accuracy and strong interpretability was established, and successfully applied to the recommendation of treatment targets, providing new ideas and methods for the intelligent classification, judgment and treatment of neuropsychiatric diseases. Specifically, using resting-state functional magnetic resonance imaging (fMRI) technology, combined with graph neural network and graph convolutional network analysis methods, personalized brain function networks for patients and normal people are established and optimized to achieve accurate disease classification and judgment and recommendation of potential treatment targets. The system realizes intelligent classification and personalized intervention target recommendation of mental illness without structural brain damage through modular design. The modules work together to provide a complete set of efficient, accurate and interpretable solutions from data collection to intervention plan recommendation, bringing innovative methods and technical support to the diagnosis and treatment of neuropsychiatric diseases.

[0008] The purpose of the present invention is achieved by at least one of the following technical solutions.

[0009] An intelligent classification and target recommendation system for diseases without structural brain damage, including:

[0010] Individualized brain function network construction module: Based on individual resting-state functional images and structural images, the functional connection matrix is ​​obtained and the brain functional connection is represented in the form of a graph to construct an individualized brain function network;

[0011] Disease judgment classification model training module: build a disease judgment classification model, and use network topology combined with graph neural network to train the disease judgment classification model;

[0012] Identify key brain regions: Apply SHAP to the trained disease judgment classification model to select the K nodes with the highest SHAP values ​​as the key brain regions for classification;

[0013] Recommended personalized intervention target module: align key brain areas to standard brain space templates and match recommended treatment targets.

[0014] Furthermore, in the individualized brain function network construction module, the patient's resting-state functional magnetic resonance imaging (rs-fMRI) and T1-weighted structural imaging data are obtained based on the patient's nuclear magnetic resonance imaging; the functional and structural images of the individual space are obtained to construct the scalp model and the cerebral cortex model;

[0015] The collected patient MRI data are preprocessed as follows:

[0016] Remove artifacts from raw MRI data, including head motion correction and magnetic field inhomogeneity correction;

[0017] The acquired rs-fMRI data were time-corrected to ensure signal synchronization of all voxels;

[0018] The patient's functional and structural images were registered to a standard space.

[0019] Furthermore, in the individualized brain function network construction module, regions of interest are defined and functional connectivity is calculated: a brain atlas or a group of regions of interest ROIs are selected, the correlation of blood oxygen level-dependent (BOLD) signals of different brain regions at rest is calculated, multiple regions of interest ROIs are defined according to standard brain atlases or segmented based on individual anatomical images, and multiple regions of interest ROIs are used as nodes in the brain function network; representative fMRI BOLD sequences of each node are obtained by averaging or singular value decomposition (SVD) the time series of all voxels in the node;

[0020] One of the simplest and most commonly used methods in the neuroimaging community is through the pairwise correlation (i.e., temporal correlation) between the BOLD time courses of two ROIs. Other methods include partial correlation, mutual information, consistency, Granger causality, etc. After selecting the functional connectivity (FC) metric, evaluate the connectivity strength between each pair of ROIs of interest; calculate the functional connectivity between ROIs of interest (i.e., the temporal correlation between different BOLD signals), and for each group of ROIs of interest, extract its BOLD time series during the entire rs-fMRI scan; calculate the Pearson correlation coefficient of the BOLD signal between each pair of ROIs of interest, that is, calculate the correlation information between different regions to obtain the functional connectivity matrix, and each element in the functional connectivity matrix represents the functional connectivity strength between the two ROIs of interest; according to the set threshold, judge whether the functional connectivity strength between each pair of ROIs of interest meets the requirements, so as to indicate the existence of connection in the functional connectivity matrix;

[0021] Finally, based on the analysis results of functional connectivity, a brain functional network was constructed: each region of interest (ROI) in the functional connectivity matrix was taken as a node to represent different areas in the brain, and the edge represented the functional connectivity between these areas. When the functional connectivity strength between two areas exceeded the set threshold, it was considered that there was an edge. The constructed brain functional network was further analyzed, and the characteristics of the brain functional network were calculated, including the degree value, clustering coefficient and path length of each node, to obtain the degree value and edge weight matrix of each node, so as to reveal the connection pattern between brain areas.

[0022] Furthermore, in the individualized brain function network construction module, brain function networks are constructed for different groups, and individualized analysis is performed. Visualization tools are used to present the analysis results and compare the differences in brain function networks between different individuals.

[0023] Furthermore, in the disease judgment classification model training module, a disease judgment classification model is constructed, including a node feature construction module, a message passing module, an attention enhancement message passing module, and a pool strategy module;

[0024] Among them, the node feature construction module constructs a node feature vector set by receiving the functional connection matrix, uses the graph convolution layer to extract node features, and then the message passing module aggregates the feature representations of neighboring nodes, and learns the weight information of each pair of nodes through the attention enhanced message passing module. Finally, the pooling strategy module obtains the global feature representation.

[0025] Furthermore, the node feature construction module inputs the functional connection matrix generated by the individualized brain function network construction module, the degree value of each node and the edge weight matrix, performs feature decomposition on the functional connection matrix, and generates a node feature vector of each node through singular value decomposition (SVD); the feature vector of the functional connection matrix, the degree value of the node and the connection strength information of each edge in the edge weight matrix are sequentially spliced ​​into a first node feature vector, and finally outputs a first node feature vector set;

[0026] In the message transmission module, the first node feature vector output by the node feature construction module is input, and each node receives information from its neighboring nodes; the message transmission process aggregates the first node feature vectors of the neighboring nodes, takes the average value of the first node feature vectors of the neighboring nodes, gradually summarizes the feature information of the neighboring nodes, splices the aggregated neighboring node information with the features of the current node into a new node feature vector, updates the node features using a non-activated function, and gradually enriches the representation of the node through multi-level message transmission, capturing more complex graph structure information; after a set number of message transmissions, a second node feature vector is output, and the second node feature vector includes information about the node and its neighborhood network;

[0027] In the attention-enhanced message transmission module, the second node feature vector of each node output by the message transmission module and the second node feature vector of each node's neighbor node are input; in the message transmission process, the features of all neighbor nodes contribute equally to the target node; after the attention mechanism is introduced, different weights are assigned to different neighbor nodes of the node, and the neighbors that contribute most to the node feature update are focused on; the attention mechanism enhances the influence of key connections by learning the weights between each pair of nodes; for the target node and the neighbor nodes of the target node, the attention score is calculated by a linear transformation function, and the attention score is converted into an attention weight to ensure that the sum of the attention weights of all neighbor nodes of the target node is 1, and the calculated attention weight is used to perform weighted aggregation on the second node feature vector of the neighbor nodes, and the aggregated second node feature vector of the neighbor nodes is combined with the second node feature vector of the target node, and a linear transformation is used to update it, and a third node feature vector is output, which contains comprehensive information of the target node and its neighborhood, and focuses on those neighbor nodes that contribute most to the feature update;

[0028] In the pooling strategy module, the third node feature vector of each node output by the attention enhancement message passing module is input, and the third node feature vector is aggregated using the pooling strategy. The pooling strategy operates on the third node feature vectors of all nodes, averages each feature dimension and pools them to obtain a global feature vector, thereby extracting the global feature representation of the brain network, which represents the characteristics of the entire brain network.

[0029] Furthermore, in the disease judgment and classification model training module, during the training process of the disease judgment and classification model, the difference between the predicted result and the true label is calculated through the cross entropy loss function, and the Adam optimizer is used to update the parameters of the disease judgment and classification model. Based on the network topology characteristics, the graph neural network is combined to gradually capture local and global topology information to achieve accurate classification of patients and normal people.

[0030] Furthermore, the interpretability of applying machine learning to disease classification is not strong, especially the graph neural network model is analyzed in the form of a graph, so it is necessary to combine other methods to analyze the key brain areas for interpretability. SHAP (SHapley Additive exPlanations) is a method to explain the output of machine learning models. It is an additive interpretation framework. Its idea comes from game theory: for any individual, the prediction model will output a prediction value, and the SHAP framework assumes that each feature is a "contributor" to the target prediction and assigns them an important value (Shapley value). The cumulative sum of the shapley values ​​of all features of a given individual and the average prediction value is the target prediction. The core idea of ​​SHAP is to regard each feature as a "contributor" and calculate its contribution to the prediction result.

[0031] The Shapley value comes from game theory and is used to distribute the benefits in cooperative games. It defines how to fairly distribute the contributions of participants. The Shapley value of a specified feature xj is the average marginal contribution of the feature to the prediction - obtained by calculating the difference in predictions with and without feature xj in all possible combinations, that is, by weighted summing all possible feature combinations. To explain, consider all possible feature subsets. For a model containing p features, divide it into multiple feature subsets. For each feature subset S, calculate the difference in model output when it contains and does not contain a certain feature xj (marginal contribution calculation). Take a weighted average of the marginal contributions of all feature subsets. The weight of subset S means that there are p possibilities for p features in any order. After determining subset S, there are p possibilities for p features in a specific order.

[0032] |S|! (p-|S|-1)! possibilities, |S|! (p-|S|-1)! / p! is the possible proportion of feature combinations in subset S. The shapley value calculated in this way satisfies the efficiency, symmetry, virtuality and additivity. The characteristics ensure that SHAP regards the shapley value as an additive attribution method and can interpret the model's prediction as the sum of the attribution values ​​of each input feature.

[0033]

[0034] While complex machine learning algorithms bring higher prediction accuracy, they also increase the difficulty of learning the algorithms. Their internal mechanisms become difficult to understand, and there is no balance between high model accuracy and interpretability. Understanding why the model makes a certain prediction is as important as the accuracy of the prediction. Explainable results are helpful for doctors' final decisions. SHAP provides a unified method to explain the output of various machine learning models by calculating the contribution value of each feature. The importance of each feature is evaluated by comparing the SHAP values ​​of the features, and the feature contribution of a single sample is analyzed to understand the model's prediction reason for the sample. The overall behavior and decision rules of the model are understood by accumulating the SHAP values ​​of multiple samples.

[0035] Combining the SHAP classification model of patients trained with graph neural networks (GNN) with that of normal people and selecting TOP-K nodes to identify key brain regions can help understand the decision-making process of the model and identify the brain regions that have the greatest impact on classification, which will help further medical research and classification judgment for recommendation of treatment targets.

[0036] In the module for identifying key brain regions, the specific process is as follows:

[0037] SHAP is applied to the trained disease judgment classification model to calculate the contribution value of each node in the brain function network to the classification decision of each sample. For each sample, the SHAP values ​​of all nodes are calculated to obtain the contribution of each node to the model prediction result; all nodes are sorted according to the average absolute SHAP value of the node. The higher the SHAP value, the greater the contribution of the node to the classification result. The top K nodes are selected as key brain regions, which play an important role in the classification decision of the model; by combining the disease judgment classification model with SHAP, the most critical brain regions (nodes) for classification can be effectively selected, and the model decision-making process can be explained to improve the transparency and interpretability of the model.

[0038] Furthermore, in the personalized intervention target module, the MNI (Montreal Neurological Institute) standard brain space template is selected as the target of brain region registration. The key brain regions selected in the key brain region identification module are registered to the standard brain space template, including correcting rotation and translation differences. The registered K key brain regions are marked on the standard brain space template, and their coordinates and spatial positions are recorded.

[0039] Furthermore, in the module for recommending personalized intervention targets, the recommended targets for non-invasive intervention corresponding to the relevant diseases are determined. The targets are usually determined based on neuroscience research and have fixed coordinates on the standard template;

[0040] Calculate the distance between each recommended intervention target and all K key brain regions, and use the weighted distance method to combine the weights of each recommended intervention target and all K key brain regions with the corresponding key brain regions in the classification;

[0041] The importance of key brain regions is determined by their SHAP values. Key brain regions with higher SHAP values ​​contribute more to the recommended targets. The recommended score of therapeutic targets is calculated by the inverse of the weighted distance. The inverse operation ensures that the closer the distance and the larger the weight, the greater the contribution of the key brain regions to the score.

[0042] All intervention targets are ranked according to the recommendation scores, and targets with higher scores are recommended as treatment targets.

[0043] Compared with the prior art solutions, the advantages of the technical solution of the present invention are as follows:

[0044] The causes of mental illness are complex and involve multiple fields, and the optimization of research and treatment plans faces challenges. At present, because mental illness without structural brain damage does not have obvious brain structural changes, the traditional classification and judgment methods that rely on imaging examinations are invalid, and the classification and judgment of the disease are mainly based on the patient's symptoms, behavioral manifestations and the doctor's clinical experience, resulting in insufficient accuracy and consistency. On the other hand, the selection of transcranial magnetic stimulation (TMS) treatment targets depends on the rough positioning or artificial marking of functional brain areas, lacking personalization and precision, which affects the treatment effect.

[0045] In order to solve the above-mentioned problems, the present invention relates to an intelligent classification and target recommendation system for non-structural brain injury diseases, which collects resting-state fMRI data and biomarkers, constructs an individualized brain function network, captures the functional connection patterns between different brain regions and improves the accuracy. Through advanced graph neural network machine learning methods, a high-accuracy disease classification model is established to overcome the limitations of traditional methods. The SHAP value analysis method is used to identify the key brain areas that contribute most to the classification results, improve the interpretability of the model, help doctors understand and trust the classification judgment results, combine the characteristics of the patient's brain function network and the disease classification results, calculate the distance or weighted distance to the TMS recommended target, provide scientific personalized intervention target recommendations, and improve the effect of TMS treatment through more accurate target recommendations, thereby providing patients with more effective non-invasive intervention plans.

[0046] The present invention proposes an intelligent classification and target recommendation system for diseases without structural brain damage, which solves many difficult problems in the classification and treatment of mental illnesses without structural brain damage, realizes an intelligent classification and judgment method with high accuracy and strong interpretability and a personalized treatment target recommendation system, improves the accuracy and credibility of disease classification judgment and treatment target selection, and recommends personalized treatment targets through scientific calculations, thereby optimizing the effect of transcranial magnetic stimulation treatment.

[0047] 1. Compared with traditional treatment plans, the intelligent classification and target recommendation system for non-structural brain damage diseases mentioned in the present invention has the advantages of higher accuracy and personalization, and can identify and recommend treatment targets more comprehensively and efficiently, providing more accurate data support for the intelligent classification, judgment and treatment of mental illnesses.

[0048] 2. The individualized brain function network construction scheme mentioned in the present invention achieves highly accurate disease classification and target recommendation by combining resting-state fMRI and graph neural network technology, providing an innovative method for the classification and judgment of neuropsychiatric diseases and target selection for treatment, which helps to improve the treatment effect and accuracy.

[0049] 3. This solution can maximize the classification accuracy while ensuring the interpretability of the model, allowing doctors and patients to obtain better treatment options and provide strong support for clinical decision-making.

[0050] 4. In the specific processing of brain functional network feature extraction, the present invention proposes a disease judgment and classification model training module that combines network topology analysis and graph neural network, which achieves accurate capture of brain region functional connections and provides an effective solution for the classification and judgment of mental illnesses without structural brain damage.

[0051] 5. The graph convolutional network and SHAP value analysis method used in the present invention are more in line with the actual needs of clinical treatment, can more accurately identify key brain areas, and provide a stronger scientific basis for personalized treatment target recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is an execution flow chart of an intelligent classification and target recommendation system for diseases without structural brain damage in an embodiment of the present invention;

[0053] Figure 2 is an execution flow chart of the individualized brain function network construction module in an embodiment of the present invention;

[0054] Figure 3 1 is an execution flow chart of a disease judgment classification model training module in an embodiment of the present invention;

[0055] Figure 4 is an execution flow chart of a key brain region identification module in an embodiment of the present invention;

[0056] Figure 5 This is an execution flow chart of the personalized intervention target recommendation module in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the specific implementation of the present invention is described in detail below with reference to the accompanying drawings and examples.

[0058] Example:

[0059] An intelligent classification and target recommendation system for diseases without structural brain damage, such as Figure 1 As shown, including:

[0060] Individualized brain function network construction module: Based on individual resting-state functional images and structural images, the functional connection matrix is ​​obtained and the brain functional connection is represented in the form of a graph to construct an individualized brain function network;

[0061] Disease judgment classification model training module: build a disease judgment classification model, and use network topology combined with graph neural network to train the disease judgment classification model;

[0062] Identify key brain regions: Apply SHAP to the trained disease judgment classification model to select the K nodes with the highest SHAP values ​​as the key brain regions for classification;

[0063] Recommended personalized intervention target module: align key brain areas to standard brain space templates and match recommended treatment targets.

[0064] like Figure 2 As shown, in the individualized brain function network construction module, the patient's resting-state functional magnetic resonance imaging (rs-fMRI) and T1-weighted structural imaging data are obtained according to the patient's nuclear magnetic resonance imaging; in one embodiment, the resolution of T1-weighted structural imaging is not less than 1mm 3 To ensure that the individual's cerebral cortex model and scalp model can be accurately constructed, functional and structural images of the individual space can be obtained, and scalp and cerebral cortex models can be constructed;

[0065] The collected patient MRI data are preprocessed as follows:

[0066] Performing artifact removal on the raw MRI image data, including head motion correction and magnetic field inhomogeneity correction; in one embodiment, these operations are implemented using a preprocessing pipeline in tools such as AFNI, FSL, and SPM;

[0067] Performing time correction on the acquired rs-fMRI data to ensure signal synchronization of all voxels; in one embodiment, a commonly used method includes slice timing correction based on time interpolation;

[0068] The functional and structural images of the patient are registered to a standard space, such as MNI space in one embodiment, to facilitate subsequent brain functional network analysis. In one embodiment, a nonlinear registration algorithm in SPM or a FLIRT tool in FSL is used.

[0069] In the individualized brain function network building module, regions of interest are defined and functional connectivity is calculated: a brain atlas or a set of regions of interest ROIs are selected, and the correlation of blood oxygen level dependent (BOLD) signals of different brain regions at rest is calculated. According to a standard brain atlas, in one embodiment, such as AAL or HO Atlas, or segmented based on individual anatomical images, multiple regions of interest ROIs are defined, and the multiple regions of interest ROIs are used as nodes in the brain function network; a representative fMRI BOLD sequence of each node is obtained by averaging or singular value decomposition (SVD) the time series of all voxels in the node;

[0070] One of the simplest and most commonly used methods in the neuroimaging community is through the pairwise correlation (i.e., temporal correlation) between the BOLD time courses of two ROIs. Other methods include partial correlation, mutual information, consistency, Granger causality, etc. After selecting the functional connectivity (FC) metric, evaluate the connectivity strength between each pair of ROIs of interest; calculate the functional connectivity between the ROIs of interest (i.e., the temporal correlation between different BOLD signals), and for each group of ROIs of interest, extract its BOLD time series during the entire rs-fMRI scan; calculate the Pearson correlation coefficient of the BOLD signal between each pair of ROIs of interest, that is, calculate the correlation information between different regions to obtain a functional connectivity matrix, and each element in the functional connectivity matrix represents the functional connectivity strength between the two ROIs of interest; according to the set threshold, in one embodiment, the Pearson correlation coefficient is greater than 0.5, and it is judged whether the functional connectivity strength between each pair of ROIs of interest meets the requirements, so as to be represented as a connection in the functional connectivity matrix;

[0071] Finally, based on the analysis results of functional connectivity, a brain functional network was constructed: each region of interest (ROI) in the functional connectivity matrix was taken as a node to represent different areas in the brain, and the edge represented the functional connectivity between these areas. When the functional connectivity strength between two areas exceeded the set threshold, it was considered that there was an edge. The constructed brain functional network was further analyzed, and the characteristics of the brain functional network were calculated, including the degree value, clustering coefficient and path length of each node, to obtain the degree value and edge weight matrix of each node, so as to reveal the connection pattern between brain areas.

[0072] In the individualized brain function network construction module, brain function networks are constructed for different groups and individualized analysis is performed. Visualization tools are used to present the analysis results and compare the differences in brain function networks between different individuals.

[0073] like Figure 3 As shown, in the disease judgment classification model training module, a disease judgment classification model is constructed, including a node feature construction module, a message passing module, an attention enhancement message passing module and a pool strategy module;

[0074] Among them, the node feature construction module constructs a node feature vector set by receiving the functional connection matrix, uses the graph convolution layer to extract node features, and then the message passing module aggregates the feature representations of neighboring nodes, and learns the weight information of each pair of nodes through the attention enhanced message passing module. Finally, the pooling strategy module obtains the global feature representation.

[0075] Furthermore, the node feature construction module inputs the functional connection matrix generated by the individualized brain function network construction module, the degree value of each node and the edge weight matrix, performs feature decomposition on the functional connection matrix, and generates a node feature vector of each node through singular value decomposition (SVD); the feature vector of the functional connection matrix, the degree value of the node and the connection strength information of each edge in the edge weight matrix are sequentially spliced ​​into a first node feature vector, and finally outputs a first node feature vector set;

[0076] In the message transmission module, the first node feature vector output by the node feature construction module is input, and each node receives information from its neighboring nodes; the message transmission process aggregates the first node feature vectors of the neighboring nodes, takes the average value of the first node feature vectors of the neighboring nodes, gradually summarizes the feature information of the neighboring nodes, splices the aggregated neighboring node information with the features of the current node into a new node feature vector, updates the node features using a non-activated function, and gradually enriches the representation of the node through multi-level message transmission, capturing more complex graph structure information; after a set number of message transmissions, a second node feature vector is output, and the second node feature vector includes information about the node and its neighborhood network;

[0077] In the attention-enhanced message transmission module, the second node feature vector of each node output by the message transmission module and the second node feature vector of each node's neighbor node are input; in the message transmission process, the features of all neighbor nodes contribute equally to the target node; after the attention mechanism is introduced, different weights are assigned to different neighbor nodes of the node, and the neighbors that contribute most to the node feature update are focused on; the attention mechanism enhances the influence of key connections by learning the weights between each pair of nodes; for the target node and the neighbor nodes of the target node, the attention score is calculated by a linear transformation function, and the attention score is converted into an attention weight to ensure that the sum of the attention weights of all neighbor nodes of the target node is 1, and the calculated attention weight is used to perform weighted aggregation on the second node feature vector of the neighbor nodes, and the aggregated second node feature vector of the neighbor nodes is combined with the second node feature vector of the target node, and a linear transformation is used to update it, and a third node feature vector is output, which contains comprehensive information of the target node and its neighborhood, and focuses on those neighbor nodes that contribute most to the feature update;

[0078] In the pooling strategy module, the third node feature vector of each node output by the attention enhancement message passing module is input, and the third node feature vector is aggregated using the pooling strategy. The pooling strategy operates on the third node feature vectors of all nodes, averages each feature dimension and pools them to obtain a global feature vector, thereby extracting the global feature representation of the brain network, which represents the characteristics of the entire brain network.

[0079] In the disease judgment and classification model training module, during the training process of the disease judgment and classification model, the difference between the predicted result and the true label is calculated through the cross entropy loss function, and the Adam optimizer is used to update the parameters of the disease judgment and classification model. Based on the network topology characteristics, the graph neural network is combined to gradually capture local and global topology information to achieve accurate classification of patients and normal people.

[0080] The interpretability of applying machine learning to disease classification is not strong, especially the graph neural network model is analyzed in the form of a graph, so it is necessary to combine other methods to analyze the key brain areas for interpretability. SHAP (SHapley Additive ex Planations) is a method to explain the output of machine learning models. It is an additive interpretation framework. Its idea comes from game theory: for any individual, the prediction model will output a prediction value, and the SHAP framework assumes that each feature is a "contributor" to the target prediction and assigns them an important value (Shapley value). The cumulative shapley value of all features of a given individual and the sum of the average prediction value is the target prediction. The core idea of ​​SHAP is to regard each feature as a "contributor" and calculate its contribution to the prediction result.

[0081] The Shapley value comes from game theory and is used to distribute the benefits in cooperative games. It defines how to fairly distribute the contributions of participants. The Shapley value of a specified feature xj is the average marginal contribution of the feature to the prediction - obtained by calculating the difference in predictions with and without feature xj in all possible combinations, that is, by weighted summing all possible feature combinations. To explain, consider all possible feature subsets. For a model containing p features, divide it into multiple feature subsets. For each feature subset S, calculate the difference in model output when it contains and does not contain a certain feature xj (marginal contribution calculation). Take a weighted average of the marginal contributions of all feature subsets. The weight of subset S means that there are p possibilities for p features in any order. After determining subset S, there are p possibilities for p features in a specific order.

[0082] |S|! (p-|S|-1)! possibilities, |S|! (p-|S|-1)! / p! is the possible proportion of feature combinations in subset S. The shapley value calculated in this way satisfies the efficiency, symmetry, virtuality and additivity. The characteristics ensure that SHAP regards the shapley value as an additive attribution method and can interpret the model's prediction as the sum of the attribution values ​​of each input feature.

[0083]

[0084] While complex machine learning algorithms bring higher prediction accuracy, they also increase the difficulty of learning the algorithms. Their internal mechanisms become difficult to understand, and there is no balance between high model accuracy and interpretability. Understanding why the model makes a certain prediction is as important as the accuracy of the prediction. Explainable results are helpful for doctors' final decisions. SHAP provides a unified method to explain the output of various machine learning models by calculating the contribution value of each feature. The importance of each feature is evaluated by comparing the SHAP values ​​of the features, and the feature contribution of a single sample is analyzed to understand the model's prediction reason for the sample. The overall behavior and decision rules of the model are understood by accumulating the SHAP values ​​of multiple samples.

[0085] Combining the SHAP classification model of patients trained with graph neural networks (GNN) with that of normal people and selecting TOP-K nodes to identify key brain regions can help understand the decision-making process of the model and identify the brain regions that have the greatest impact on classification, which will help further medical research and classification judgment for recommendation of treatment targets.

[0086] like Figure 4 As shown in the figure, in the module for identifying key brain regions, the specific process is as follows:

[0087] SHAP is applied to the trained disease judgment classification model to calculate the contribution value of each node in the brain function network to the classification decision of each sample. For each sample, the SHAP values ​​of all nodes are calculated to obtain the contribution of each node to the model prediction result; all nodes are sorted according to the average absolute SHAP value of the node. The higher the SHAP value, the greater the contribution of the node to the classification result. The top K nodes are selected as key brain regions, which play an important role in the classification decision of the model; by combining the disease judgment classification model with SHAP, the most critical brain regions (nodes) for classification can be effectively selected, and the model decision-making process can be explained to improve the transparency and interpretability of the model.

[0088] like Figure 5 As shown in the figure, in the personalized intervention target module, the MNI (Montreal Neurological Institute) standard brain space template is selected as the target of brain region registration. The key brain regions selected in the key brain region identification module are registered to the standard brain space template, including correcting the rotation and translation differences. The registered K key brain regions are marked on the standard brain space template, and their coordinates and spatial positions are recorded.

[0089] In the module of recommending personalized intervention targets, the recommended targets for non-invasive intervention corresponding to the relevant diseases are determined. The targets are usually determined based on neuroscience research and have fixed coordinates on the standard template.

[0090] Calculate the distance between each recommended intervention target and all K key brain regions, and use the weighted distance method to combine the weights of each recommended intervention target and all K key brain regions with the corresponding key brain regions in the classification;

[0091] The importance of key brain regions is determined by their SHAP values. Key brain regions with higher SHAP values ​​contribute more to the recommended targets. The recommended score of therapeutic targets is calculated by the inverse of the weighted distance. The inverse operation ensures that the closer the distance and the larger the weight, the greater the contribution of the key brain regions to the score.

[0092] All intervention targets are ranked according to the recommendation scores, and targets with higher scores are recommended as treatment targets.

Claims

1. An intelligent classification and target recommendation system for diseases without structural brain damage, characterized in that: include: Individualized brain function network construction module: Based on individual resting-state functional images and structural images, the functional connection matrix is ​​obtained and the brain functional connection is represented in the form of a graph to construct an individualized brain function network; Disease judgment classification model training module: build a disease judgment classification model, and use network topology combined with graph neural network to train the disease judgment classification model; Identify key brain regions: Apply SHAP to the trained disease judgment classification model to select the K nodes with the highest SHAP values ​​as the key brain regions for classification; Recommended personalized intervention target module: align key brain areas to standard brain space templates and match recommended treatment targets.

2. The intelligent classification and target recommendation system for non-structural brain damage diseases according to claim 1, characterized in that: In the individualized brain function network construction module, the patient's resting-state functional magnetic resonance imaging and T1-weighted structural imaging data are obtained based on the patient's nuclear magnetic resonance images; the functional images and structural images of the individual space are obtained to build the scalp model and the cerebral cortex model; The collected patient MRI data are preprocessed as follows: Remove artifacts from raw MRI data, including head motion correction and magnetic field inhomogeneity correction; The acquired rs-fMRI data were time-corrected to ensure signal synchronization of all voxels; The patient's functional and structural images were registered to a standard space.

3. The intelligent classification and target recommendation system for non-structural brain damage diseases according to claim 2, characterized in that: In the individualized brain function network construction module, define regions of interest and calculate functional connectivity: select a brain atlas or a group of regions of interest ROIs, calculate the correlation of blood oxygen level dependent (BOLD) signals in different brain regions at rest, define multiple regions of interest ROIs according to standard brain atlases or segmentation based on individual anatomical images, and use multiple regions of interest ROIs as nodes in the brain function network; obtain the representative fMRI BOLD sequence of each node by averaging or singular value decomposition of the time series of all voxels in the node; After selecting the functional connectivity measure, the strength of connectivity between each pair of regions of interest (ROIs) was evaluated; The functional connectivity between the regions of interest (ROIs) was calculated, and for each group of regions of interest (ROIs), the BOLD time series during the entire rs-fMRI scan were extracted. Calculate the Pearson correlation coefficient of the BOLD signal between each pair of regions of interest ROIs, that is, calculate the correlation information between different regions to obtain the functional connectivity matrix. Each element in the functional connectivity matrix represents the functional connectivity strength between two regions of interest ROIs. According to the set threshold, it is judged whether the functional connectivity strength between each pair of regions of interest ROIs meets the requirements, so as to be represented as having a connection in the functional connectivity matrix; Finally, based on the analysis results of functional connectivity, a brain functional network was constructed: each region of interest (ROI) in the functional connectivity matrix was taken as a node to represent different areas in the brain, and the edge represented the functional connectivity between these areas. When the functional connectivity strength between two areas exceeded the set threshold, it was considered that there was an edge. The constructed brain functional network was further analyzed, and the characteristics of the brain functional network were calculated, including the degree value, clustering coefficient and path length of each node. The degree value and edge weight matrix of each node were obtained to reveal the connection pattern between brain regions.

4. The intelligent classification and target recommendation system for non-structural brain damage diseases according to claim 3 is characterized in that: In the individualized brain function network construction module, brain function networks are constructed for different groups and individualized analysis is performed. Visualization tools are used to present the analysis results and compare the differences in brain function networks between different individuals.

5. The intelligent classification and target recommendation system for non-structural brain damage diseases according to claim 1, characterized in that: In the disease judgment classification model training module, a disease judgment classification model is constructed, including a node feature construction module, a message passing module, an attention enhancement message passing module, and a pool strategy module; Among them, the node feature construction module constructs a node feature vector set by receiving the functional connection matrix, uses the graph convolution layer to extract node features, and then the message passing module aggregates the feature representations of neighboring nodes, and learns the weight information of each pair of nodes through the attention enhanced message passing module. Finally, the pooling strategy module obtains the global feature representation.

6. The intelligent classification and target recommendation system for diseases without structural brain damage according to claim 5, characterized in that: In the node feature construction module, the functional connection matrix, the degree value of each node and the edge weight matrix generated by the individualized brain function network construction module are input, the functional connection matrix is ​​eigen-decomposed, and a node feature vector of each node is generated by singular value decomposition; the eigenvector of the functional connection matrix, the degree value of the node and the connection strength information of each edge in the edge weight matrix are sequentially spliced ​​into a first node feature vector, and finally a first node feature vector set is output; In the message transmission module, the first node feature vector output by the node feature construction module is input, and each node receives information from its neighboring nodes; the message transmission process aggregates the first node feature vectors of the neighboring nodes, averages the first node feature vectors of the neighboring nodes, gradually summarizes the feature information of the neighboring nodes, and splices the aggregated neighboring node information with the features of the current node into a new node feature vector; after a set number of message transmissions, a second node feature vector is output, and the second node feature vector includes information about the node and its neighborhood network; In the attention enhancement message passing module, the second node feature vector of each node output by the message passing module and the second node feature vector of each node's neighboring node are input; For the target node and its neighboring nodes, the attention scores are calculated through a linear transformation function, and the attention scores are converted into attention weights to ensure that the sum of the attention weights of all neighboring nodes of the target node is 1. The calculated attention weights are used to perform weighted aggregation on the second node feature vectors of the neighboring nodes. The aggregated second node feature vectors of the neighboring nodes are combined with the second node feature vector of the target node, and are updated using a linear transformation to output a third node feature vector. The third node feature vector contains comprehensive information about the target node and its neighborhood, focusing on those neighboring nodes that contribute most to feature updates. In the pooling strategy module, the third node feature vector of each node output by the attention enhancement message passing module is input, and the third node feature vector is aggregated using the pooling strategy. The pooling strategy operates on the third node feature vectors of all nodes, averages each feature dimension and pools them to obtain a global feature vector, thereby extracting the global feature representation of the brain network, which represents the characteristics of the entire brain network.

7. The intelligent classification and target recommendation system for non-structural brain damage diseases according to claim 5, characterized in that: In the disease judgment and classification model training module, during the training process of the disease judgment and classification model, the difference between the predicted result and the true label is calculated through the cross entropy loss function, and the Adam optimizer is used to update the parameters of the disease judgment and classification model. Based on the network topology characteristics, the graph neural network is combined to gradually capture local and global topology information to achieve accurate classification of patients and normal people.

8. The intelligent classification and target recommendation system for non-structural brain damage diseases according to claim 1, characterized in that: In the module for identifying key brain regions, the specific process is as follows: Apply SHAP to the trained disease judgment classification model to calculate the contribution value of each node in the brain function network to the classification decision of each sample. For each sample, calculate the SHAP value of all nodes to obtain the contribution of each node to the model prediction result; All nodes were sorted according to their average absolute SHAP value. The higher the SHAP value, the greater the contribution of the node to the classification result. The top K nodes were selected as key brain regions.

9. The intelligent classification and target recommendation system for non-structural brain damage diseases according to claim 1, characterized in that: In the module of recommending personalized intervention targets, the MNI standard brain space template is selected as the target of brain region registration. The key brain regions selected in the module of identifying key brain regions are registered to the standard brain space template, including correcting the rotation and translation differences. The K key brain regions after registration are marked on the standard brain space template, and their coordinates and spatial positions are recorded.

10. The intelligent classification and target recommendation system for diseases without structural brain damage according to claim 9, characterized in that: In the module of recommending personalized intervention targets, determine the recommended targets for non-invasive interventions corresponding to relevant diseases; Calculate the distance between each recommended intervention target and all K key brain regions, and use the weighted distance method to combine the weights of each recommended intervention target and all K key brain regions with the corresponding key brain regions in the classification; The importance of key brain regions is determined by their SHAP values. Key brain regions with higher SHAP values ​​contribute more to the recommended targets. The recommended score of therapeutic targets is calculated by the inverse of the weighted distance. The inverse operation ensures that the closer the distance and the larger the weight, the greater the contribution of the key brain regions to the score. All intervention targets are ranked according to the recommendation scores, and targets with higher scores are recommended as treatment targets.

Images