Intelligent classification and target recommendation system for structural brain injury diseases

Abstract

The application discloses a structural brain injury disease intelligent classification and target recommendation system. The system comprises: an individualized brain function network construction module: according to individual resting state function images and structure images, a function connection matrix is obtained and brain function connection is represented in the form of a graph to construct an individualized brain function network; a disease judgment classification model training module: a disease judgment classification model is constructed, and the disease judgment classification model is trained by combining a network topology and a graph neural network; a key brain area identification module: the SHAP value of the K nodes with the highest SHAP value is selected by applying SHAP on the trained disease judgment classification model, and the K nodes are used as key brain areas for classification; and a personalized intervention target recommendation module: the key brain areas are registered to a standard brain space template to match recommended treatment targets. The application 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 application relates to the field of medical rehabilitation treatment, in particular to an intelligent classification and target recommendation system for structural brain injury diseases. BACKGROUND

[0002] The mental diseases of structural brain injury refer to those diseases without direct observable brain structure abnormalities, but patients show obvious mental symptoms or behavior abnormalities, covering depression, anxiety, sleep disorders, motor dysfunction, etc. Due to the lack of obvious brain structure abnormalities, the classification of these diseases mainly depends on the symptoms, behavior and clinical experience of doctors, resulting in challenges in the accuracy and consistency of disease classification. Even if the doctor repeatedly compares the clinical examination results and image information to make intelligent classification, the sensitivity and specificity are not ideal, and manual classification still has many difficulties.

[0003] The treatment of mental diseases of structural brain injury usually includes drug treatment, psychological treatment and social support, etc. At present, most hospitals use painless, non-invasive green treatment method transcranial magnetic stimulation (TMS) to treat patients. Transcranial magnetic stimulation (TMS) is a kind of magnetic stimulation technology which uses pulse magnetic field to act on the central nervous system of the brain, changes the membrane potential of the cerebral cortex nerve cells, makes them produce induced current, affects the brain metabolism and nerve electrical activity, and causes a series of physiological and biochemical reactions. At present, the selection of transcranial magnetic stimulation target is usually based on the functional brain region which can be reasonably and conveniently determined to locate the target, or the patient wears a positioning cap and directly marks the target area on the positioning cap to select the target for treatment. The selection of treatment target for patients with different diseases plays a key role in their treatment, and designing a more suitable target recommendation method is an effective means to improve the efficacy of rTMS in treating mental diseases.

[0004] Due to the complexity of the etiology of mental illness caused by non-structural brain injury, and the involvement of knowledge in multiple fields, the research on its etiology is very difficult. The traditional method mainly analyzes the biological samples of the object to detect the presence of one or more biomarkers, and the doctor needs to have high medical experience and related theoretical knowledge research and take a long time, and the target selection in treatment is based on the subjective experience of the doctor, which has certain limitations in medical application (Longitudinal analysis method and system for mild traumatic brain injury magnetic resonance image data (CN202111501620.X) New biomarkers and methods for diagnosing and evaluating traumatic brain injury (CN201880084813.X)). Machine learning (Machine Learning, ML) is favored by medical workers due to its nonlinear fitting, without pre-limiting model complexity and excellent classification and prediction ability, and has a relatively wide application in the classification of AD and the identification of MCI subtypes. By introducing machine learning methods into the field of medical rehabilitation, higher prediction accuracy can be achieved to facilitate medical personnel to better perform treatment programs, optimize the treatment effect of mental illness, and provide strong support for the target selection research of non-structural brain injury mental illness, overcome the limitations of traditional methods, and provide precise and optimized medical treatment programs for patients and medical personnel. SUMMARY

[0005] The system designed by the present application aims to solve the problems of intelligent classification and intervention target recommendation of non-structural brain injury mental illness. The system is based on brain function network analysis and graph neural network fusion to establish individualized brain function network and realize intelligent classification and target recommendation. The system first acquires the resting state fMRI data of the patient, constructs the individualized brain function network, and captures the functional connection mode between different brain regions in the resting state; then uses network topology analysis method and combines graph neural network to train a classification model that can accurately distinguish patients from normal people, and identifies the TOP-K key brain regions that contribute most to the classification result through graph convolution network (GCN) and SHAP (SHapley Additive exPlanations) value analysis method, and then registers the identified key brain regions to the standard template, calculates their distance or weighted distance from the recommended target of non-invasive intervention (such as transcranial magnetic stimulation treatment), and combines the weights of these key brain regions in the classification model and the distance from the recommended target to obtain the treatment target recommendation score. Finally, according to the calculated treatment target recommendation score, the doctor is provided with a personalized intervention target recommendation treatment program, so as to optimize the treatment effect of neuropsychiatric diseases. Through these modular designs, from data acquisition to intervention target recommendation, the system provides a complete set of efficient, accurate and interpretable solutions to comprehensively solve the problems of intelligent classification and treatment target selection of non-structural brain injury mental illness.

[0006] The application provides a structural brain injury disease intelligent classification and target recommendation system, which mainly includes the following modules: a data acquisition module, a network topology analysis module, a classification model training module, a model explainability enhancement module, a target recommendation calculation module and a personalized intervention scheme recommendation module. First, the system acquires the resting state fMRI data of an individual, constructs a personalized brain function network of the patient, captures the functional connection mode between different brain regions of the brain 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 capable of accurately distinguishing patients from normal people. Through the effective feature extraction and classification learning of the brain function network by the GNN, the model can accurately predict diseases based on the differences in individual brain function connections. In order to improve the model explainability, the pooling layer of the graph convolution network (GCN) is combined with the SHAP (SHapley Additive exPlanations) value analysis method to identify the TOP-K key brain regions with the greatest contribution to the classification result. The TOP-k key brain regions identified in the patient classification are registered to a standard template, and the distance or weighted distance between them and the recommended target of the non-invasive intervention (such as transcranial magnetic stimulation treatment) 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, which provides a scientific basis for doctors to develop a personalized intervention target recommendation treatment scheme.

[0007] By fusing the resting state fMRI technology, the graph neural network and the graph convolution network analysis method, an individualized brain function network classification model with high accuracy and strong explainability is established, and is successfully applied to the recommendation of treatment targets, providing a new idea and method for intelligent classification and treatment of neuropsychiatric diseases. Specifically, by using the resting state functional magnetic resonance imaging (fMRI) technology in combination with the graph neural network and the graph convolution network analysis method, an individualized brain function network for patients and normal people is established and optimized to realize accurate disease classification and potential treatment target recommendation. Through modular design, the system realizes intelligent classification and personalized intervention target recommendation of structural brain injury mental diseases, and each module works together to provide an efficient, accurate and explainable solution from data acquisition to intervention scheme recommendation, which brings innovative methods and technical support for the diagnosis and treatment of neuropsychiatric diseases.

[0008] The object of the application is achieved at least by one of the following technical solutions.

[0009] An intelligent classification and target recommendation system for structural brain injury diseases, comprising:

[0010] An individualized brain function network construction module: obtaining a functional connection matrix and representing brain function connection in the form of a graph according to individual resting-state functional images and structural images, and constructing an individualized brain function network;

[0011] A disease judgment classification model training module: constructing a disease judgment classification model, and training the disease judgment classification model by combining a network topology with a graph neural network;

[0012] A key brain area identification module: applying SHAP to the trained disease judgment classification model to select K nodes with the highest SHAP values as key brain areas for classification;

[0013] A recommended individualized intervention target module: registering the key brain areas to a standard brain spatial template to match recommended treatment targets.

[0014] Further, in the individualized brain function network construction module, resting-state functional magnetic resonance imaging (rs-fMRI) and T1-weighted structural imaging data of the patient are obtained from the patient's magnetic resonance imaging data; the individual spatial functional images and structural images are obtained, and a scalp model and a brain cortex model are constructed;

[0015] The collected patient magnetic resonance imaging data is preprocessed, specifically as follows:

[0016] The original magnetic resonance imaging data is subjected to artifact removal, including head motion correction and magnetic field inhomogeneity correction;

[0017] The collected rs-fMRI data is subjected to time correction to ensure synchronization of signals of all voxels;

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

[0019] Further, in the individualized brain function network construction module, a region of interest is defined and functional connectivity is calculated: a brain atlas or a set of regions of interest (ROI) is selected, the correlation of blood oxygen level dependent (BOLD) signals of different brain regions in the resting state is calculated, a plurality of regions of interest (ROI) are defined according to the standard brain atlas or based on individual anatomical images, and the plurality of regions of interest (ROI) are taken as nodes in the brain function network; by averaging or singular value decomposition (SVD) on the time series of all voxels in the node, a representative fMRI BOLD sequence of each node is obtained;

[0020] One of the simplest and most commonly used methods in neuroimaging is the pairwise correlation (i.e., temporal correlation) between the BOLD time courses of two ROIs, other methods include partial correlation, mutual information, coherence, Granger causality, etc. After selecting the functional connectivity (FC) measure, the strength of the connectivity between each pair of ROIs is evaluated; the functional connectivity between ROIs (i.e., temporal correlation between different BOLD signals) is calculated, for each set of ROIs, the BOLD time series of the ROIs during the entire rs-fMRI scan is extracted; the Pearson correlation coefficient of the BOLD signals between each pair of ROIs is calculated, i.e., the correlation information between different regions is calculated to obtain a functional connectivity matrix, each element in the functional connectivity matrix represents the strength of the functional connectivity between two ROIs; according to the set threshold, it is judged whether the strength of the functional connectivity between each pair of ROIs meets the requirements or not, and is represented as existing connection in the functional connectivity matrix;

[0021] Finally, based on the analysis results of the functional connectivity, the brain functional network is constructed: each ROI in the functional connectivity matrix is taken as a node (Node) to represent different regions in the brain, and the edge represents the functional connectivity between these regions. When the strength of the functional connectivity between two regions exceeds the set threshold, it is considered that there is an edge; the constructed brain functional network is further analyzed, and the characteristics of the brain functional network are calculated, including the degree value, clustering coefficient and path length of each node, the degree value and edge weight matrix of each node are obtained, to reveal the connection mode between brain regions.

[0022] Further, in the individualized brain functional network construction module, for different groups, the brain functional network is constructed and individualized analysis is performed, and a visualization tool is used to present the analysis results, and the differences between the brain functional networks of different individuals are compared.

[0023] Further, 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 enhanced message passing module, and a pooling strategy module.

[0024] Among them, the node feature construction module constructs a node feature vector set by receiving the functional connectivity matrix, extracts node features using a graph convolution layer, then aggregates the feature representations of neighbor nodes by the message passing module, learns the weight information of each pair of nodes by the attention enhanced message passing module, and finally obtains a global feature representation by the pooling strategy module.

[0025] Further, in the node feature construction module, the functional connectivity matrix generated by the individualized brain function network construction module, the degree value of each node and the edge weight matrix are input, the functional connectivity matrix is subjected to feature decomposition, and the node feature vector of each node is generated through singular value decomposition (SVD); the feature vector of the functional connectivity 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 the first node feature vector set is output;

[0026] In the message passing module, the first node feature vector output by the node feature construction module is input, and each node receives information from its neighbor nodes; the process of message passing is realized by aggregating the first node feature vectors of the neighbor nodes, the first node feature vectors of the neighbor nodes are averaged to aggregate, the feature information of the neighbor nodes is gradually summarized, the aggregated neighbor node information is spliced with the feature of the current node to obtain a new node feature vector, the feature of the node is updated by using a non-activation function, and through multi-level message passing, the representation of the node is gradually enriched, and more complex graph structure information is captured; after a certain number of message passing, a second node feature vector is output, and the second node feature vector includes the information of the node and its neighborhood network;

[0027] In the attention-enhanced message passing module, the second node feature vector of each node and the second node feature vector of each neighbor node of each node output by the message passing module are input; in the message passing process, the features of all neighbor nodes are equally contributed to the target node; after introducing the attention mechanism, different weights are assigned to different neighbor nodes of the node, and the neighbor nodes that most contribute to the update of the node feature are focused on; the attention mechanism learns the weight between each pair of nodes to improve the influence of the key connection; for the target node and the neighbor nodes of the target node, the attention score is calculated through a linear transformation function, the attention score is converted into an attention weight, the sum of the attention weights of all neighbor nodes of the target node is ensured to be 1, the second node feature vectors of the neighbor nodes are weighted and aggregated by using the calculated attention weight, the second node feature vectors of the aggregated neighbor nodes are combined with the second node feature vector of the target node, and the linear transformation is used for updating to output a third node feature vector; the third node feature vector contains the comprehensive information of the target node and its neighborhood, and the neighbor nodes that most contribute to the feature update are focused on.

[0028] In the pooling strategy module, the third node feature vector of each node output by the attention-enhanced message passing module is input, and the third node feature vector is aggregated by using a pooling strategy; the pooling strategy operates on all third node feature vectors of the nodes, performs average pooling on each feature dimension to obtain a global feature vector, extracts the global feature representation of the brain network, and 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 application of machine learning to disease classification is not very interpretable, especially for graph neural network models that perform analysis in a graphical form. Therefore, it is necessary to combine other methods for interpretable analysis of key brain regions. SHAP (SHapley Additive exPlanations) is a method for interpreting the output of machine learning models. It is an additive interpretation framework whose idea comes from game theory: for any individual, the prediction model will output a prediction value, while the SHAP framework assumes that each feature is a "contributor" to the target prediction and assigns them an importance value (Shapley value). The sum of the cumulative Shpley 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 a feature xj is included and not included (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, p features have p possibilities in a specific order.

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

[0033]

[0034] Complex machine learning algorithms bring higher prediction accuracy, but also increase the difficulty of algorithm learning, and the internal mechanism becomes difficult to understand. The balance between high accuracy and explainability of the model cannot be achieved. It is equally important to understand the reason why the model makes a certain prediction and the accuracy of the prediction. Explainable results assist doctors in making 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. The prediction reason of the model for a single sample is analyzed by analyzing the feature contribution of the sample. The overall behavior and decision rule of the model are understood by accumulating the SHAP values of multiple samples.

[0035] Combining SHAP with a patient and normal person classification model trained by a graph neural network (GNN), TOP-K nodes are selected to identify key brain regions, which can help understand the decision-making process of the model and identify the brain regions that have the most impact on classification, which is helpful for further medical research and classification and treatment target recommendation.

[0036] In the key brain region identification module, 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. According to the average absolute SHAP value of the nodes, all nodes are sorted. 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, improving the transparency and explainability of the model.

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

[0039] Further, in the personalized intervention target recommendation module, the target points for non-invasive intervention recommendation corresponding to the relevant diseases are determined. The target points are usually determined according to neuroscience research and have fixed coordinates on the standard template.

[0040] The distance between each recommended intervention target and all K key brain regions is calculated, and the weighted distance method is used to combine each recommended intervention target with all K key brain regions with the weight of the corresponding key brain region in the classification.

[0041] The importance of the key brain region is determined by the SHAP value, and the higher the SHAP value of the key brain region, the greater the contribution of the recommended target; the treatment target recommendation score is calculated by the reciprocal of the weighted distance, and the reciprocal operation ensures that the closer the distance, the greater the contribution of the key brain region with greater weight to the score;

[0042] According to the recommendation score, all intervention targets are sorted, and the target with a higher score is preferentially recommended as a treatment target.

[0043] Compared with the prior art, the advantages of the technical scheme of the present application are as follows:

[0044] The etiology of mental diseases is complex and involves multiple fields, and the optimization of research and treatment schemes is challenging. At present, mental diseases without structural brain damage do not have obvious brain structure changes, making the traditional classification judgment method relying on imaging examination ineffective, and the classification judgment of the disease is mainly based on the symptoms, behavior and clinical experience of the doctor, resulting in insufficient accuracy and consistency. On the other hand, the selection of transcranial magnetic stimulation (TMS) treatment targets relies on rough positioning or artificial marking of functional brain regions, lacking individualization and precision, affecting the treatment effect.

[0045] In order to solve the above-mentioned problems, the intelligent classification and target recommendation system for non-structural brain damage diseases disclosed by the present application collects resting-state fMRI data and biomarkers, constructs an individualized brain function network, captures the functional connection mode between different brain regions of the brain and improves the accuracy, establishes a high-accuracy disease classification model through an advanced graph neural network machine learning method, overcomes the limitations of traditional methods, identifies the key brain region that contributes most to the classification result through SHAP value analysis method, improves the explainability of the model, helps doctors understand and trust the classification judgment result, combines the characteristics of the patient's brain function network and the disease classification result, calculates the distance or weighted distance from the TMS recommended target, provides a scientific and individualized intervention target recommendation, improves the effect of TMS treatment through more accurate target recommendation, thereby providing a more effective non-invasive intervention scheme for the patient.

[0046] The present application provides an intelligent classification and target recommendation system for non-structural brain damage diseases, solves multiple problems in the classification and treatment of non-structural brain damage mental diseases, realizes an intelligent classification method with high accuracy and strong explainability and an individualized treatment target recommendation system, improves the accuracy and reliability of disease classification and treatment target selection, and recommends individualized treatment targets through scientific calculation, thereby optimizing the effect of transcranial magnetic stimulation treatment.

[0047] 1. The intelligent classification and target recommendation system for non-structural brain injury diseases mentioned in the application has higher precision and individualization advantages compared with traditional treatment schemes, can more comprehensively and efficiently identify and recommend treatment targets, and provides more accurate data support for intelligent classification and treatment of mental diseases.

[0048] 2. The individualized brain function network construction scheme mentioned in the application realizes high-accuracy disease classification and target recommendation by combining resting-state fMRI and graph neural network technology, provides an innovative method for classification and target selection for treatment of neuropsychiatric diseases, and helps to improve treatment effect and precision.

[0049] 3. The scheme can guarantee the classification accuracy and the interpretability of the model at the maximum, so that doctors and patients obtain better treatment schemes, and provide strong support for clinical decision-making.

[0050] 4. In the specific processing of brain function network feature extraction, the application proposes a disease judgment and classification model training module combining network topology analysis and graph neural network, realizes accurate capture of brain region functional connection, and provides an effective solution for classification and judgment of mental diseases of non-structural brain injury.

[0051] 5. The graph convolution network and SHAP value analysis method used in the application are more in line with the actual needs of clinical treatment, can more accurately identify key brain regions, and provide stronger scientific basis for personalized treatment target recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The execution flowchart of the intelligent classification and target recommendation system for non-structural brain injury diseases in the embodiment of the application is shown in the figure;

[0053] Figure 2 The execution flowchart of the individualized brain function network construction module in the embodiment of the application is shown in the figure;

[0054] Figure 3 The execution flowchart of the disease judgment and classification model training module in the embodiment of the application is shown in the figure;

[0055] Figure 4 The execution flowchart of the key brain region identification module in the embodiment of the application is shown in the figure;

[0056] Figure 5 The execution flowchart of the personalized intervention target recommendation module in the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present application clearer, more apparent and more comprehensible, the specific embodiments of the present application will be described in detail below with reference to the drawings and examples.

[0058] Embodiments:

[0059] An intelligent classification and target recommendation system for structural brain injury diseases, as shown in Figure 1 , comprises:

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

[0061] A disease judgment classification model training module: a disease judgment classification model is constructed, and the network topology is combined with a graph neural network to train the disease judgment classification model;

[0062] A key brain area recognition module: the SHAP value of the K nodes with the highest SHAP value is selected as the key brain area for classification by applying SHAP to the trained disease judgment classification model;

[0063] A personalized intervention target recommendation module: the key brain area is registered to a standard brain spatial template to match the recommended treatment target.

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

[0065] The collected patient nuclear magnetic imaging data is preprocessed, specifically as follows:

[0066] The original nuclear magnetic imaging data is artifact removed, including head motion correction and magnetic field inhomogeneity correction; in an embodiment, the preprocessing pipeline in tools such as AFNI, FSL, SPM, etc. is used to realize these operations;

[0067] The collected rs-fMRI data is time corrected to ensure that the signals of all voxels are synchronized; in an embodiment, common methods include slice timing correction based on time interpolation;

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

[0069] In the individualized brain functional network construction module, the region of interest is defined and the functional connectivity is calculated: a brain atlas or a set of regions of interest (ROI) is selected, the correlation of blood oxygen level dependent (BOLD) signals of different brain regions in the resting state is calculated, according to a standard brain atlas, in one embodiment, such as AAL or HO Atlas, or based on individual anatomical images, a plurality of regions of interest (ROI) are defined, and the plurality of regions of interest (ROI) are taken as nodes in the brain functional network; by averaging or singular value decomposition (SVD) on the time series of all voxels in the node, a representative fMRI BOLD sequence of each node is obtained;

[0070] One of the simplest and most commonly used methods in neuroimaging is 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) measure, the strength of the connectivity between each pair of regions of interest (ROI) is evaluated; the functional connectivity between regions of interest (ROI) (i.e., the temporal correlation between different BOLD signals) is calculated, for each set of regions of interest (ROI), the BOLD time series thereof during the entire rs-fMRI scan is extracted; the Pearson correlation coefficient of the BOLD signals between each pair of regions of interest (ROI) is calculated, i.e., the correlation information between different regions is calculated to obtain a functional connectivity matrix, each element in the functional connectivity matrix represents the strength of the functional connectivity between two regions of interest (ROI); according to a set threshold, in one embodiment, the Pearson correlation coefficient is greater than 0.5, to determine whether the strength of the functional connectivity between each pair of regions of interest (ROI) meets the requirements, to be represented as existing connectivity in the functional connectivity matrix;

[0071] Finally, based on the analysis results of the functional connectivity, the brain functional network is constructed: each region of interest (ROI) in the functional connectivity matrix is taken as a node (Node) representing different regions in the brain, and the edges represent the functional connectivity between these regions, when the strength of the functional connectivity between two regions exceeds the set threshold, it is considered that there is an edge; the constructed brain functional network is further analyzed, the characteristics of the brain functional network are calculated, including the degree value, clustering coefficient and path length of each node, the degree value and edge weight matrix of each node are obtained, to reveal the connection mode between brain regions.

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

[0073] As shown in Figure 3 The disease judgment classification model training module includes a node feature construction module, a message passing module, an attention enhanced message passing module, and a pooling strategy module.

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

[0075] Further, in the node feature construction module, the functional connection matrix, the degree values of each node, and the edge weight matrix generated by the individualized brain function network construction module are input, the functional connection matrix is subjected to feature decomposition, and the node feature vector of each node is generated by singular value decomposition (SVD); the feature vectors of the functional connection matrix, the degree values of the nodes, 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.

[0076] In the message passing 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 passing process aggregates the first node feature vectors of the neighboring nodes by taking the average 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 and the features of the current node into a new node feature vector, updates the features of the nodes using a non-activation function, and through multi-level message passing, the representation of the nodes is gradually enriched to capture more complex graph structure information; after a certain number of message passing, a second node feature vector is output, which includes the information of the node and its neighborhood network.

[0077] In the attention-enhanced message passing module, the second node feature vector of each node output by the input message passing module and the second node feature vector of each neighbor node of each node are input; in the message passing process, the contribution of all neighbor nodes to the target node is equal; after introducing the attention mechanism, different weights are assigned to different neighbor nodes of the target node, and the neighbor nodes that most contribute to the update of the node feature are focused on; the attention mechanism improves the influence of key connections by learning the weight between each pair of nodes; for the target node and the neighbor nodes of the target node, an attention score is calculated through a linear transformation function, the attention score is converted into an attention weight, and the sum of the attention weights of all neighbor nodes of the target node is ensured to be 1; the second node feature vectors of the neighbor nodes are aggregated by using the calculated attention weights, the second node feature vectors of the aggregated neighbor nodes are combined with the second node feature vector of the target node, and the linear transformation is used for updating to output a third node feature vector, which contains the comprehensive information of the target node and its neighborhood, and focuses on the neighbor nodes that most contribute to the feature update;

[0078] In the pool strategy module, the third node feature vector of each node output by the attention-enhanced message passing module is input, and the third node feature vector is aggregated by using a pooling strategy; the pooling strategy obtains a global feature vector by averaging each feature dimension through operation on the third node feature vectors of all nodes, extracts the global feature representation of the brain network, and represents the characteristics of the entire brain network.

[0079] In the disease judgment classification model training module, in the training process of the disease judgment classification model, the difference between the predicted result and the true label is calculated by using a cross-entropy loss function, and the parameters of the disease judgment classification model are updated by using an Adam optimizer; based on the network topology characteristics, the local and global topology information is gradually captured by combining the graph neural network, and the accurate classification of patients and normal persons is realized.

[0080] The application of machine learning to disease classification has weak explainability, especially the graph neural network model analyzes in the form of a graph, so other methods need to be combined to analyze the key brain areas. SHAP (SHapley Additive exPlanations) is a method for explaining the output of a machine learning model, which is an additive explanation framework, and its idea comes from game theory: for any individual, the prediction model outputs a prediction value, and the SHAP framework assumes that each feature is a "contributor" of the target prediction, and assigns an important value (shapley value) to it. The sum of the shpley 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.

[0081] The Shapley value is derived from game theory and is used to allocate gains in cooperative games, defining how to fairly distribute the contributions of the participants. The Shapley value of a specified feature xj is the average marginal contribution of that feature to the prediction - obtained by calculating the difference in prediction with and without the feature xj for all possible combinations, i.e. by weighted sum over all possible feature combinations. To explain, consider all possible feature subsets, for a model with p features, divide it into multiple feature subsets, for each feature subset S, calculate the difference in model output with and without a certain feature xj (marginal contribution calculation), and weight the marginal contributions of all feature subsets, the weight of the subset S is p! under any ordering, after determining the subset S, there are

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

[0083]

[0084] Complex machine learning algorithms bring higher prediction accuracy, but also increase the difficulty of algorithm learning, and the internal mechanism becomes difficult to understand. The balance between high accuracy and explainability of the model cannot be achieved, and it is equally important to understand the reason why the model makes a certain prediction and the accuracy of the prediction. Explainable results have an auxiliary role in the final decision of doctors. SHAP provides a unified method to explain the output of various machine learning models by calculating the contribution value of each feature, evaluates the importance of each feature by comparing the SHAP values of the features, analyzes the feature contribution of a single sample to understand the prediction reason of the model for the sample, and understands the overall behavior and decision rules of the model by accumulating the SHAP values of multiple samples.

[0085] Combining the SHAP with the patient and normal person classification model trained by the graph neural network (GNN), the TOP-K nodes are selected to identify the key brain areas, which can help understand the decision-making process of the model and identify the brain areas that have the most impact on classification, which is helpful for further medical research and classification and treatment target recommendation.

[0086] As shown in Figure 4 , in the key brain area identification module, the specific process is as follows:

[0087] On the trained disease judgment classification model, SHAP is applied 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. According to the average absolute SHAP value of the node, all nodes are sorted, the higher the SHAP value, the greater the contribution of the node to the classification result, and the top K nodes are selected as the key brain areas, 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 area (node) for classification can be effectively selected, and the model decision process can be explained, improving the transparency and interpretability of the model.

[0088] As shown in Figure 5 In the recommended personalized intervention target point 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, labeling the K key brain regions after registration on the standard brain space template, and recording their coordinates and spatial positions.

[0089] In the recommended personalized intervention target point module, the target points of non-invasive intervention recommendation corresponding to the related diseases are determined, and the target points are usually determined according to neuroscience research and have fixed coordinates on the standard template.

[0090] The distance between each recommended intervention target point and all K key brain regions is calculated, and the weighted distance method is used to combine each recommended intervention target point with all K key brain regions and the corresponding key brain regions in the classification.

[0091] The importance of the key brain region is determined by its SHAP value, and the higher the SHAP value of the key brain region, the greater the contribution to the recommended target point; the treatment target recommendation score is calculated by the reciprocal of the weighted distance, and the reciprocal operation ensures that the closer the distance, the greater the contribution of the key brain region with greater weight to the score.

[0092] According to the recommendation score, all intervention target points are sorted, and the target points with higher scores are preferentially recommended as treatment target points.

Claims

1. A structural brain injury disease intelligent classification and target recommendation system, characterized in that, The application comprises the following modules: An individualized brain function network construction module: a functional connection matrix is obtained according to individual resting state functional images and structural images, and the brain function connection is represented in the form of a graph to construct an individualized brain function network; A disease judgment classification model training module: a disease judgment classification model is constructed, and the disease judgment classification model is trained by combining network topology and a graph neural network; A key brain area identification module: the K nodes with the highest SHAP values are selected as key brain areas for classification by applying SHAP to the trained disease judgment classification model; A personalized intervention target recommendation module: the key brain areas are registered to a standard brain spatial template to match recommended treatment targets; In the disease judgment classification model training module, during the training process of the disease judgment classification model, the difference between the prediction result and the true label is calculated by a cross-entropy loss function, and the parameters of the disease judgment classification model are updated by an Adam optimizer, so as to capture local and global topological information gradually based on the network topological characteristics, and realize accurate classification of patients and normal persons; In the key brain area identification module, the specific process is as follows: 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, and the SHAP value of each node to the model prediction result is calculated for each sample; All nodes are sorted according to the average absolute SHAP value of the nodes, the higher the SHAP value, the greater the contribution of the node to the classification result, and the top K nodes are selected as key brain areas; In the personalized intervention target recommendation module, the MNI standard brain spatial template is selected as the target for brain area registration, the key brain areas selected in the key brain area identification module are registered to the standard brain spatial template, including correction of rotation and translation differences, the K key brain areas after registration are labeled on the standard brain spatial template, and the coordinates and spatial positions thereof are recorded; In the personalized intervention target recommendation module, the target points of non-invasive intervention recommendation corresponding to the related diseases are determined; The distance between each recommended intervention target and all K key brain areas is calculated, and the weighted distance method is used to combine each recommended intervention target with all K key brain areas and the corresponding key brain areas in the classification; The importance of the key brain areas is determined by the SHAP value thereof, the higher the SHAP value of the key brain area, the greater the contribution of the key brain area to the recommended target; the treatment target recommendation score is calculated by the reciprocal of the weighted distance, and the reciprocal operation ensures that the closer the distance, the greater the contribution of the key brain area with a greater weight to the score; All intervention targets are sorted according to the recommendation score, and the target with a higher score is preferentially recommended as a treatment target.

2. The structural brain injury disease intelligent classification and target recommendation system according to claim 1, characterized in that, In the individualized brain function network construction module, resting state functional magnetic resonance imaging and T1-weighted structural imaging data of a patient are obtained from the nuclear magnetic imaging of the patient; functional images and structural images of an individual space are obtained, and a scalp model and a brain cortex model are constructed; The collected nuclear magnetic imaging data of the patient is preprocessed, and the specific process is as follows: The original nuclear magnetic imaging data is subjected to artifact removal, including head motion correction and magnetic field inhomogeneity correction; Time correction is performed on the collected rs-fMRI data to ensure that the signals of all voxels are synchronized; The functional image and structural image of the patient are registered to the standard space.

3. The structural brain injury disease intelligent classification and target recommendation system according to claim 2, characterized in that, In the individualized brain function network construction module, the region of interest is defined and the functional connectivity is calculated: a brain atlas or a set of regions of interest (ROI) is selected, the correlation of the blood oxygen level dependent (BOLD) signals of different brain regions in the resting state is calculated, and multiple regions of interest (ROI) are defined according to the standard brain atlas or based on individual anatomical images. The multiple regions of interest (ROI) are used as nodes in the brain function network. The representative fMRI BOLD sequence of each node is obtained by averaging or singular value decomposition of the time series of all voxels in the node. After selecting the functional connectivity measure, the connectivity strength between each pair of regions of interest (ROI) is evaluated. The functional connectivity between the regions of interest (ROI) is calculated, and for each set of regions of interest (ROI), the BOLD time sequence during the entire rs-fMRI scan is extracted. The Pearson correlation coefficient of the BOLD signals between each pair of regions of interest (ROI) is calculated, i.e. the correlation information between different regions is calculated to obtain the functional connection matrix. Each element in the functional connection matrix represents the functional connectivity strength between two regions of interest (ROI). According to the set threshold, it is judged whether the functional connectivity strength between each pair of regions of interest (ROI) meets the requirements, and it is represented as existing connection in the functional connection matrix. Finally, based on the analysis results of the functional connectivity, the brain function network is constructed: each region of interest (ROI) in the functional connection matrix is taken as a node representing different regions in the brain, and the edge represents the functional connectivity between these regions. When the functional connectivity strength between two regions exceeds the set threshold, it is considered that there is an edge. The constructed brain function network is further analyzed, and the characteristics of the brain function network are calculated, including the degree value, clustering coefficient and path length of each node. The degree value and edge weight matrix of each node are obtained to reveal the connection mode between brain regions.

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

5. The structural brain injury disease intelligent classification and target recommendation system 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 enhanced message passing module, and a pooling strategy module. Among them, the node feature construction module constructs a node feature vector set by receiving the functional connection matrix, extracts node features using a graph convolution layer, then aggregates the feature representations of neighboring nodes by the message passing module, learns the weight information of each pair of nodes by the attention enhanced message passing module, and finally obtains the global feature representation by the pooling strategy module.

6. The structural brain injury disease intelligent classification and target recommendation system according to claim 5, characterized in that, In the node feature construction module, the functional connectivity 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 connectivity matrix is subjected to feature decomposition, and a node feature vector of each node is generated through singular value decomposition; the feature vector of the functional connectivity 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 passing module, the first node feature vector output by the node feature construction module is input, and each node receives information from its neighbor nodes; the message passing process is realized by aggregating the first node feature vectors of the neighbor nodes, averaging the first node feature vectors of the neighbor nodes, gradually aggregating the feature information of the neighbor nodes, splicing the aggregated neighbor node information and the feature of the current node into a new node feature vector; after a certain number of message passing, a second node feature vector is output, which includes the information of the node and its neighborhood network; In the attention-enhanced message passing module, the second node feature vector of each node and the second node feature vector of each neighbor node of each node output by the message passing module are input; For the target node and the neighbor nodes of the target node, the attention score is calculated through a linear transformation function, the attention score is converted into attention weight, the sum of the attention weights of all neighbor nodes of the target node is ensured to be 1, the second node feature vectors of the neighbor nodes are aggregated by using the calculated attention weight, the aggregated second node feature vectors of the neighbor nodes are combined with the second node feature vector of the target node, the linear transformation is used for updating, and a third node feature vector is output, which contains the comprehensive information of the target node and its neighborhood, and focuses on the neighbor nodes that contribute most to the feature update; In the pool strategy module, the third node feature vector of each node output by the attention-enhanced message passing module is input, and the third node feature vector is aggregated by using a pooling strategy; the pooling strategy is used to operate on the third node feature vectors of all nodes, average-pooling is performed on each feature dimension to obtain a global feature vector, and the global feature representation of the brain network is extracted, representing the characteristics of the whole brain network.

Images