System and method for realizing depression subtype classification processing based on multiple fusion brain network graph technology, processor and storage medium thereof

Abstract

The application relates to a system for realizing depression subtype classification processing based on deep learning multiple fusion brain network graph technology, wherein the system comprises the following modules: a data acquisition and processing module for acquiring resting-state functional magnetic resonance imaging data of a subject; a data preprocessing module for preprocessing the acquired data; a multiple functional brain network construction module for generating three functional connection matrices from obtained functional magnetic resonance imaging (fMRI) data and constructing a graph representation for each coefficient connected matrix; a multiple brain network graph fusion module for improving performance under small sample capacity by using a regularization term based on data enhancement and mapping each graph representation to a feature space by using GAT fusion area groups and difference pool area groups; and a depression subtype classification module for classifying depression subtypes. The application also relates to a corresponding method, device, processor and storage medium thereof. The system, method, device, processor and storage medium thereof help to further optimize the target intervention scheme of neural regulation technology in depression treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of clinical medicine, in particular to the technical field of subtype identification of affective disorders, and specifically refers to a system and method for classifying subtypes of depression based on deep learning and multiple fusion brain network graph technology, a processor and a computer readable storage medium thereof. BACKGROUND

[0002] Depression is the most common mental disorder, characterized by at least two weeks of depressive episodes. According to statistics, about 20% of people have suffered from depression in their lifetime. Depression has a high disability rate and mortality rate, which not only seriously affects the quality of life of individual patients, but also brings a huge burden to the patient's family and the whole society. Despite centuries of psychiatric and neurological research, our understanding of the physiological and pathological mechanisms of depression is still very limited.

[0003] There is great heterogeneity in the clinical manifestations of depression, which is one of the difficulties in scientific research and clinical diagnosis and treatment of depression. Disease heterogeneity may be reflected in the differences in genes, metabolism, brain function, clinical symptoms and treatment response of different patients. Therefore, the study of depression subtypes can improve our understanding of the pathophysiological mechanisms of depression and help to achieve precision medicine. According to the Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition) (DSM-5), melancholic depression, atypical depression and anxious depression are the three most common subtypes of depression. Our technology aims to distinguish these three subtypes and find their corresponding imaging biomarkers.

[0004] Resting-state functional magnetic resonance imaging (fMRI) is one of the commonly used brain imaging methods in mental science, which reflects the situation of spontaneous brain function activity through the fluctuation of blood-oxygen-level-dependent (BOLD) signals. fMRI is non-invasive and has no radiation damage to the human body, and the subjects are easy to cooperate, which helps to explore the changes in brain function activity of patients with mental disorders. Through fMRI scanning and analysis, it is expected to find neuroimaging biomarkers of different depression subtypes. In recent years, graph neural networks (GNNs) in deep learning technology have become one of the hot technologies for analyzing brain function connection data. In 2017, Andrew et al. defined multiple biological subtypes of depression through clustering analysis of resting-state functional connectivity. However, there are still some limitations in using GNNs to classify subtypes of depression and find related biomarkers, such as the limited information contained in a single functional connectivity network and the large sample size required for data.

[0005] To this end, we propose a deep learning multi-fusion brain network graph technology to identify depression subtypes, in order to accurately distinguish the three main depression subtypes and identify biomarkers of these subtypes. In order to mine more information, we use Diff-Pooling to fuse multiple functional connection graphs to extract multi-view and multi-level image information. In order to meet the deep learning model's demand for a large amount of data, we use a new regularization term method for data augmentation to improve the algorithm performance under small sample size. Then we use the graph attention network (GAT) to identify key image biomarkers, which also has good performance under small sample size. SUMMARY

[0006] The purpose of the present application is to overcome the above-mentioned shortcomings of the prior art, and to provide a non-invasive, accurate and reliable system, method, processor and storage medium for classifying depression subtypes based on multi-fusion brain network graph technology.

[0007] In order to achieve the above-mentioned purpose, the system, method, processor and storage medium for classifying depression subtypes based on multi-fusion brain network graph technology are as follows:

[0008] The system for classifying depression subtypes based on deep learning multi-fusion brain network graph technology, its main feature is that the system comprises:

[0009] A data acquisition and processing module for acquiring resting-state functional magnetic resonance imaging data of subjects in a specific environment;

[0010] A data preprocessing module connected to the data acquisition and processing module for correcting, registering, segmenting, standardizing, filtering and smoothing the acquired resting-state functional magnetic resonance imaging data;

[0011] A multi-functional brain network construction module connected to the data preprocessing module, which generates three functional connection matrices from the data-preprocessed functional magnetic resonance imaging data, corresponding to Pearson correlation coefficient, Spearman correlation coefficient and partial correlation measure coefficient, and constructs a graph representation for each coefficient connectivity matrix;

[0012] A multi-brain network graph fusion module connected to the multi-functional brain network construction module, which uses a regularization term based on data augmentation to improve performance under small sample size, and uses GAT to fuse region groups and difference pool groups to map each graph representation to a feature space;

[0013] The depression subtype classification module is connected with the multiple brain network graph fusion module, and the depression subtype classification module is based on deep learning and adopts a multilayer perceptron model to classify depression subtypes.

[0014] The method for classifying depression subtypes based on the deep learning multiple fusion brain network graph technology of the above system mainly has the following steps:

[0015] (1) Collecting resting-state functional magnetic resonance imaging data of subjects under specific conditions in a specific environment;

[0016] (2) Preprocessing the collected resting-state functional magnetic resonance imaging data by correction, registration, segmentation, standardization, filtering, and smoothing to obtain preprocessed resting-state functional magnetic resonance imaging data;

[0017] (3) Generating three functional connectivity matrices from the data-preprocessed functional magnetic resonance imaging data, corresponding to Pearson correlation coefficient, Spearman correlation coefficient, and partial correlation measure coefficient, respectively, and constructing a graph representation for each coefficient connectivity matrix;

[0018] (4) Using a regularization term based on data augmentation to improve performance under small sample capacity, and using GAT to fuse the region group and difference pool group to map each graph representation to a feature space;

[0019] (5) Classifying depression subtypes based on deep learning and using a multilayer perceptron model.

[0020] Preferably, the step (1) comprises:

[0021] (1.1) Selecting subjects according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) for major depressive disorder and the depressive episode state disease diagnosis evaluation standard;

[0022] (1.2) Collecting functional magnetic resonance imaging data of subjects in a specific environment by Siemens 3.0T magnetic resonance scanning.

[0023] Preferably, the step (2) comprises:

[0024] (2.1) Preprocessing the functional magnetic resonance imaging data using the SPM12 toolkit in Matlab2017b;

[0025] (2.2) Deleting the first 10 time series of each subject to maintain signal stability;

[0026] (2.3) Performing time correction and head motion correction processing;

[0027] (2.4) Registering the T1 structural image to the head-motion corrected resting-state image;

[0028] (2.5) Segmenting the T1 structural image using Lie algebra differential isomorphism registration algorithm;

[0029] (2.6) Registering the resting-state brain image data to the Montreal Neurological Institute space using DARTAL and resampling to 3x3x3mm 3 ;

[0030] (2.7) Removing linear trend and band-pass filtering the resampled resting-state brain image data to remove artifacts;

[0031] (2.8) Spatially smoothing the data using a full-width-at-half-maximum 4mm three-dimensional Gaussian kernel;

[0032] (2.9) Regressing out confounding signals (6 head motion parameters, mean cerebrospinal fluid signal, mean white matter signal, and global signal);

[0033] (2.10) Removing subjects with head motion translation > 2.5mm or rotation > 2.5°.

[0034] Preferably, the step (3) comprises:

[0035] (3.1) Extracting the BOLD signals of 90 regions of interest in the brain based on the AAL1 template;

[0036] (3.2) Extracting the functional connectivity values between brain regions using Pearson correlation to construct a brain functional network matrix based on Pearson values;

[0037] (3.3) Extracting the functional connectivity values between brain regions using Spearman correlation to construct a brain functional network matrix based on Spearman values;

[0038] (3.4) Extracting the functional connectivity values between brain regions using partial correlation to construct a brain functional network matrix based on partial correlation values.

[0039] (3.5) Treating each of the above functional connectivity matrices as an undirected weighted graph, where the features of the nodes include the mean, variance, kurtosis and skewness of the BOLD signals, and the weights of the edges are where c ij is the correlation coefficient in the functional network matrix.

[0040] Preferably, the step (4) comprises:

[0041] (4.1) Integrating the feature nodes in each brain functional network using edge-weighted graph attention network;

[0042] (4.2) Adopting the differentiable pooling operation module to compress the information in the feature network in a hierarchical clustering manner;

[0043] (4.3) After hierarchical clustering, the graph becomes more and more rough, thereby generating the hierarchical representation of the input graph;

[0044] (4.4) After two layers of differentiable pooling operation, the feature information of 90 nodes is stacked to 32 nodes;

[0045] (4.5) Repeat the above steps until stacked to 4 nodes.

[0046] Preferably, the step (5) comprises:

[0047] (5.1) For the node features of each subject, adopt the 5-fold cross-validation strategy for classification;

[0048] (5.2) The decision module adopts 2-layer MLP, wherein each layer contains 32 and 3 neurons;

[0049] (5.3) Cross-entropy loss is used for classification, wherein the decay weight setting of Adam optimization is 0.08;

[0050] (5.4) Use the multilayer perceptron model to train the model for 100 epochs, wherein the initial learning rate is 0.001;

[0051] (5.5) Extract the accuracy, F1 score, precision and recall indicators to evaluate the model performance.

[0052] The implementation of the device for processing the classification of depression subtypes based on deep learning multiple fusion brain network graph technology, its main feature is that the device comprises:

[0053] A processor configured to execute computer executable instructions;

[0054] A memory storing one or more computer executable instructions, which, when executed by the processor, implement the various steps of the above-mentioned processing of classification of depression subtypes based on deep learning multiple fusion brain network graph technology.

[0055] The processor for processing the classification of depression subtypes based on deep learning multiple fusion brain network graph technology, its main feature is that the processor is configured to execute computer executable instructions, and the computer executable instructions, when executed by the processor, implement the various steps of the above-mentioned processing of classification of depression subtypes based on deep learning multiple fusion brain network graph technology.

[0056] The computer readable storage medium is characterized in that a computer program is stored thereon, and the computer program can be executed by a processor to implement each step of the depression subtype classification processing based on the deep learning multiple fusion brain network graph technology.

[0057] The system, method, processor and computer readable storage medium for implementing the depression subtype classification processing based on the deep learning multiple fusion brain network graph technology according to the application can first screen out objective biomarkers of three main subtypes of Chinese depression population, so as to provide objective biological indicators for the specific classification of depression subtypes, and integrate and package them into a tool set to be applied to clinical treatment, so as to provide new evidence-based basis for forming the precise differentiation of depression subtypes based on objective biological indicators and fill the blank of guidelines. Meanwhile, the technical solution can help the early identification of depression subtypes by constructing an effective depression subtype identification model and detecting the multiple fusion brain network graph of the subject, and it integrates the key brain regions and brain network functional connections of depression subtype classification, which is helpful for further optimizing the target intervention scheme of neural regulation technology in the treatment of depression, can effectively reduce the burden of patients, families and society, and can provide suggestions and evidence for the improvement of guidelines. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The schematic diagram of the depression subtype classification diagnosis model for the method for implementing the depression subtype classification processing based on the deep learning multiple fusion brain network graph technology according to the application is shown.

[0059] Figure 2 The schematic diagram of the brain region with the greatest contribution in the depression subtype classification diagnosis model according to the application is shown. DETAILED DESCRIPTION

[0060] In order to more clearly describe the technical content of the application, the following further describes in combination with specific embodiments.

[0061] Before the detailed description of the embodiments according to the application, it should be noted that in the following, the terms “include”, “contain” or any other variant are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes these elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0062] The system for implementing the depression subtype classification processing based on the deep learning multiple fusion brain network graph technology, wherein the system comprises:

[0063] The data acquisition processing module is configured to acquire the resting-state functional magnetic resonance imaging data fMRI of the subject in a specific environment.

[0064] The data preprocessing module is connected with the data acquisition and processing module, and is used for correcting, registering, segmenting, standardizing, filtering and smoothing the acquired resting-state functional magnetic resonance imaging data fMRI.

[0065] The multiple functional brain network construction module is connected with the data preprocessing module, and generates three functional connection matrices from the functional magnetic resonance imaging data fMRI after data preprocessing, which respectively correspond to Pearson correlation coefficient, Spearman correlation coefficient and partial correlation measure coefficient, and constructs a graph representation for each coefficient connection matrix.

[0066] The multiple brain network graph fusion module is connected with the multiple functional brain network construction module, uses a regularization term based on data enhancement to improve the performance under small sample capacity, and uses GAT to fuse the region group and the difference pool group to map each graph representation to a feature space.

[0067] The depression subtype classification module is connected with the multiple brain network graph fusion module, and is based on deep learning and uses a multilayer perceptron model to classify and process depression subtypes.

[0068] Please refer to Figure 1 The method for classifying and processing depression subtypes based on the deep learning multiple fusion brain network graph technology using the above system is shown, wherein the method comprises the following steps:

[0069] (1) Acquire resting-state functional magnetic resonance imaging data fMRI of a subject under a specific condition in a specific environment;

[0070] (2) Perform data preprocessing of correction, registration, segmentation, standardization, filtering and smoothing on the acquired resting-state functional magnetic resonance imaging data fMRI to obtain preprocessed resting-state functional magnetic resonance imaging data fMRI;

[0071] (3) Generate three functional connection matrices from the functional magnetic resonance imaging data fMRI after data preprocessing, which respectively correspond to Pearson correlation coefficient, Spearman correlation coefficient and partial correlation measure coefficient, and construct a graph representation for each coefficient connection matrix;

[0072] (4) Use a regularization term based on data enhancement to improve the performance under small sample capacity, and use GAT to fuse the region group and the difference pool group to map each graph representation to a feature space;

[0073] (5) Based on deep learning and using a multilayer perceptron model to classify and process depression subtypes.

[0074] As a preferred embodiment of the present application, the step (1) comprises:

[0075] (1.1) According to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) and the Diagnostic Evaluation Criteria for Major Depressive Disorder and Depressive Episode State, subjects are selected for the study;

[0076] (1.2) Image data is collected by Siemens 3.0T magnetic resonance scanning, and functional magnetic resonance imaging data (fMRI) of the subjects in a specific environment is collected.

[0077] As a preferred embodiment of the present application, the step (2) comprises:

[0078] (2.1) Using the SPM12 toolkit in Matlab2017b, the resting state functional magnetic resonance imaging data (fMRI) is preprocessed;

[0079] (2.2) The first 10 time point sequences of each subject are deleted to maintain the stability of the signal;

[0080] (2.3) Time correction and head motion correction processing are performed, time layer correction is performed to correct the time difference of image acquisition of different layers, the midpoint time acquisition layer is taken as the reference layer, head motion correction is performed, and the image is realigned to correct the influence of head motion;

[0081] (2.4) Image registration, due to the low resolution of functional images and the high resolution of structural images, the magnetic resonance T1 structural images are registered to the resting state images after head motion correction, so as to better segment the images;

[0082] (2.5) Image segmentation, the Lie algebra differential homeomorphism registration algorithm is used to segment the T1 structural images;

[0083] (2.6) Standardization, using DARTAL to register the resting state brain image data to the Montreal Neurological Institute space, i.e. the individual brain is standardized to the EPI template of the standard space, and resampled with a voxel size of 3×3×3mm3;

[0084] (2.7) The resting state brain image data obtained after resampling is subjected to linear trend removal and bandpass filtering to remove artifacts;

[0085] (2.8) Smoothing, spatial smoothing processing is applied with a half-height full-width of 4mm three-dimensional Gaussian kernel function;

[0086] (2.9) Regression of mixed signals, 6 head motion parameters, average cerebrospinal fluid signal, average white matter signal and whole brain signal are put into the autoregressive model as covariates to correct the fMRI data preprocessed by the above steps;

[0087] (2.10) According to the criteria of head motion translation > 2.5mm or rotation > 2.5°, the subjects not meeting the criteria are excluded.

[0088] As a preferred embodiment of the present application, the step (3) comprises:

[0089] (3.1) Based on the AAL1 template, the BOLD signals of 90 brain regions of interest (ROI) of the brain are extracted;

[0090] (3.2) The correlation degree of the BOLD signals between two ROIs is calculated by using Pearson correlation, and the linear correlation value is regarded as the functional connection (FC) value between brain regions, thereby constructing a brain functional network matrix based on the Pearson value;

[0091] (3.3) The correlation degree of the BOLD signals between two ROIs is calculated by using Spearman correlation, and the rank correlation value is regarded as the functional connection value between brain regions, thereby constructing a brain functional network matrix based on the Spearman value;

[0092] (3.4) The correlation degree of the BOLD signals between two ROIs is calculated by using partial correlation, and the linear correlation value obtained after controlling other influencing factors is regarded as the functional connection value between brain regions, thereby constructing a brain functional network matrix based on the partial correlation value;

[0093] (3.5) Each of the above functional connection matrices is regarded as an undirected weighted graph, wherein the characteristics of the nodes include the mean, variance, kurtosis and skewness of the BOLD signals, and the weight of the edge is wherein c ij is the correlation coefficient in the functional network matrix.

[0094] As a preferred embodiment of the present application, the step (4) comprises:

[0095] (4.1) The node characteristics in the brain functional networks based on the Pearson value, the Spearman value and the partial correlation value are integrated and updated by using the edge-weighted graph attention network: firstly, the similarity coefficient between two adjacent nodes is calculated one by one to update its adjacency matrix, and secondly, based on the updated adjacency matrix, the node characteristics are updated by weighting the 1-hop neighborhood node characteristics;

[0096] (4.2) The information in the feature network is compressed in a hierarchical clustering manner based on the clustering assignment matrix by using a differentiable pooling operation module;

[0097] (4.3) After hierarchical clustering, the graph becomes more and more rough, thereby generating the hierarchical representation of the input graph;

[0098] (4.4) Through two layers of differentiable pooling operations, the feature information of 90 nodes is stacked to 32 nodes;

[0099] (4.5) Repeat the above steps until stacked to 4 nodes.

[0100] As a preferred embodiment of the present application, the step (5) comprises:

[0101] (5.1) The node features of each subject are classified using a 5-fold cross-validation strategy, wherein to prevent possible data feature confusion, data from a single subject can only be included in the training set or the test set.

[0102] (5.2) The decision module uses a 2-layer MLP, with 32 and 3 neurons in each layer;

[0103] (5.3) Cross-entropy loss is used for classification, with an Adam optimizer decay weight setting of 0.08;

[0104] (5.4) The model is trained for 100 epochs using a multilayer perceptron model, with an initial learning rate of 0.001;

[0105] (5.5) The accuracy, F1 score, precision, and recall are extracted to evaluate the performance of the model.

[0106] As shown in the following table, it is the classification performance of the depression subtype model according to the technical scheme:

[0107]

[0108] The technical scheme will be further described below in combination with specific examples:

[0109] (1) Study of subjects

[0110] 1) High senior psychiatrists first recommend subjects who meet the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) major depressive disorder, depressive episode state disease diagnosis to enter the group;

[0111] 2) Collect functional magnetic resonance imaging data (fMRI) of the subjects in the group;

[0112] (2) Preprocessing of brain image data

[0113] 1) Using SPM12 (Statistical Parametric Mapping fil.ion.ucl.ac.uk / spm) and DPRSF (Data Processing Assistant for Resting-State fMRI) toolbox in Matlab 2017b to preprocess fMRI data;

[0114] 2) Removing the first 10 time series of each subject to maintain the stability of the signal;

[0115] 3) Time correction and head motion correction;

[0116] 4) Registering T1 structural image to the resting-state image after head motion correction;

[0117] 5) Using DARTAL (Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra) to segment T1 image;

[0118] 6) Using DARTAL to register resting-state brain image data to MNI space and resample to 3x3x3mm 3 ;

[0119] 7) Linear trend removal and band-pass filtering (0.01-0.1 Hz) to remove artifacts;

[0120] 8) Spatial smoothing using a 4mm FWHM (full-width half-maximum) 3D Gaussian kernel;

[0121] 9) Regressing nuisance signals (6 head motion parameters, mean CSF signal, mean white matter signal, and whole brain signal);

[0122] 10) Excluding subjects who do not meet the criteria of head motion translation > 2.5mm or rotation > 2.5°.

[0123] (3) Construction of multiple functional brain networks

[0124] 1) Based on AAL1 template, extracting BOLD signals of 90 ROIs (Regions of Interest) in the brain;

[0125] 2) Using Pearson correlation to extract FC (Functional Connectivity) values between brain regions and construct brain function network matrix based on Pearson values;

[0126] 3) Extract the functional connectivity values between brain regions using Spearman correlation to construct a brain functional network matrix based on Spearman values;

[0127] 4) Extract the functional connectivity values between brain regions using partial correlation to construct a brain functional network matrix based on partial correlation values.

[0128] 5) Treat each of the above functional connectivity matrices as an undirected weighted graph, where the characteristics of the nodes include the mean, variance, kurtosis and skewness of the BOLD signal, and the edge weights are where c ij is the correlation coefficient in the functional network matrix.

[0129] (4) Graph attention network fusion module

[0130] 1) Use edge-weighted graph attention networks to integrate feature nodes in each brain network;

[0131] 2) Use a differentiable pooling operation module to compress information in the feature network in a hierarchical clustering manner;

[0132] 3) After hierarchical clustering, the graph becomes more and more rough, thus generating a hierarchical representation of the input graph;

[0133] 4) After two layers of differentiable pooling operations, the feature information of 90 nodes is stacked to 32 nodes;

[0134] 5) As above, then stacked to 4 nodes.

[0135] (5) Depression subtype classification model based on multi-layer perceptron (MLP) model

[0136] 1) Use 5-fold cross-validation strategy to classify the node features of each subject;

[0137] 2) The decision module uses a 2-layer MLP, with 32 and 3 neurons in each layer;

[0138] 3) Cross-entropy loss is used for classification, with an Adam optimizer decay weight of 0.08;

[0139] 4) The model is trained for 100 epochs, with an initial learning rate of 0.001;

[0140] 5) Extract accuracy, F1 score, precision and recall, etc. to evaluate the performance of the model.

[0141] For patients with clear diagnosis of depression, the system can be used for subtype classification and diagnosis of the disease, so as to prompt and help the psychiatrist on duty to select appropriate treatment plan according to the disease subtype. The specialist or non-specialist can distinguish different subtypes of depression in a short time through the characteristic information of resting state functional magnetic resonance and the classification and diagnosis results, select a relatively specific treatment plan, avoid affecting the treatment outcome of the disease because of the trial and error of the treatment plan, and shorten the diagnosis and treatment time, improve the work efficiency and the accuracy of the treatment plan.

[0142] The device for implementing the subtype classification of depression based on the deep learning multiple fusion brain network graph technology, wherein the device comprises:

[0143] The processor is configured to execute computer executable instructions.

[0144] The memory stores one or more computer executable instructions, and the computer executable instructions are executed by the processor to implement the steps of the subtype classification of depression based on the deep learning multiple fusion brain network graph technology.

[0145] The processor for implementing the subtype classification of depression based on the deep learning multiple fusion brain network graph technology, wherein the processor is configured to execute computer executable instructions, and the computer executable instructions are executed by the processor to implement the steps of the subtype classification of depression based on the deep learning multiple fusion brain network graph technology.

[0146] The computer readable storage medium, wherein a computer program is stored thereon, and the computer program can be executed by the processor to implement the steps of the subtype classification of depression based on the deep learning multiple fusion brain network graph technology.

[0147] Any process or method descriptions or any other information described herein in the flowchart form can be understood as representing one or more steps of a set of steps performed in support of one or more functions or processes described herein, and that the various steps can be performed in the sequence shown or in other sequences, in accordance with the requirements of the particular implementation, unless otherwise specifically stated herein. Thus, unless specifically stated otherwise, the order or sequence of any process or method steps can be different unless specifically stated otherwise.

[0148] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the above embodiments, a plurality of steps or methods can be realized by software or firmware stored in a memory and executed by a suitable instruction execution device.

[0149] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment method can be completed by a program instructing the relevant hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0150] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0151] In the description of the present specification, the description of the terms "an embodiment", "some embodiments", "an example", "a specific example", or "embodiments" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0152] Although the embodiments of the present application have been shown and described above, it is understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application. Those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

[0153] The system, method, processor and computer readable storage medium thereof for implementing depression subtype classification processing based on the deep learning multiple fusion brain network graph technology of the present application first screen out objective biomarkers of three main subtypes of Chinese depression population, so as to provide objective biological indicators for the specific classification of depression subtypes, and integrate and package them into a tool set for clinical application, providing new evidence-based evidence for the precise differentiation of depression subtypes based on objective biological indicators, and filling the gap in the guidelines. At the same time, the technical solution constructs an effective depression subtype identification model, and classifies depression subtypes by detecting multiple fusion brain network graphs of subjects, which greatly helps the early identification of depression subtypes, integrates the key brain regions and brain network functional connections of depression subtype classification, and helps to further optimize the target intervention scheme of neural regulation technology in the treatment of depression, which can effectively reduce the burden on patients, families and society, and can provide suggestions and evidence for the improvement of guidelines.

[0154] In this specification, the present application has been described with reference to its specific embodiments. However, it is obvious that various modifications and changes can be made without departing from the spirit and scope of the present application. Therefore, the specification and drawings should be considered as illustrative rather than limiting.

Claims

1. A system for implementing depression subtype classification processing based on deep learning multi-fusion brain network graph technology, characterized in that, The system comprises: a data acquisition and processing module for acquiring resting-state functional magnetic resonance imaging data (fMRI) of a subject in a specific environment; a data preprocessing module connected to the data acquisition and processing module for correcting, registering, segmenting, standardizing, filtering, and smoothing the acquired resting-state functional magnetic resonance imaging data (fMRI) for data preprocessing; a multiple functional brain network construction module connected to the data preprocessing module for generating three functional connectivity matrices from the data-preprocessed functional magnetic resonance imaging data (fMRI), corresponding to Pearson correlation coefficient, Spearman correlation coefficient, and partial correlation measure coefficient, and constructing a graph representation for each coefficient connectivity matrix; a multiple brain network graph fusion module connected to the multiple functional brain network construction module, which uses a regularization term based on data enhancement to improve performance under small sample capacity, and uses GAT to fuse the region group and difference pool group to map each graph representation to a feature space; a depression subtype classification module connected to the multiple brain network graph fusion module, which is based on deep learning and uses a multilayer perceptron model to classify depression subtypes; The multiple functional brain network construction module comprises the following processing procedures: extracting BOLD signals of 90 brain regions of interest based on the AAL1 template; extracting functional connectivity values between brain regions using Pearson correlation to construct a brain functional network matrix based on Pearson values; extracting functional connectivity values between brain regions using Spearman correlation to construct a brain functional network matrix based on Spearman values; extracting functional connectivity values between brain regions using partial correlation to construct a brain functional network matrix based on partial correlation values; Each of the above brain functional network matrices is considered as an undirected weighted graph, where the features of the nodes include the mean, variance, kurtosis and skewness of the BOLD signal, and the weights of the edges are where c ij is the correlation coefficient in the brain functional network matrix.

2. A method for implementing depression subtype classification processing based on deep learning multi-fusion brain network graph technology using the system of claim 1, characterized in that, The method comprises the following steps: (1) acquiring resting-state functional magnetic resonance imaging data (fMRI) of a subject in a specific environment under specific conditions; (2) correcting, registering, segmenting, standardizing, filtering, and smoothing the acquired resting-state functional magnetic resonance imaging data (fMRI) for data preprocessing to obtain preprocessed resting-state functional magnetic resonance imaging data (fMRI); (3) generating three functional connectivity matrices from the data-preprocessed functional magnetic resonance imaging data (fMRI), corresponding to Pearson correlation coefficient, Spearman correlation coefficient, and partial correlation measure coefficient, and constructing a graph representation for each coefficient connectivity matrix; (4) using a regularization term based on data enhancement to improve performance under small sample capacity, and using GAT to fuse the region group and difference pool group to map each graph representation to a feature space; (5) based on deep learning and using a multilayer perceptron model to classify depression subtypes.

3. The method of claim 2, wherein the method comprises: The step (1) comprises: (1.1) selecting subjects according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) regarding major depressive disorder, and the assessment criteria for depressive episode state diseases; (1.2) Collect image data by Siemens 3.0T magnetic resonance scanning, and collect functional magnetic resonance imaging data (fMRI) of the subjects in a specific environment.

4. The method of claim 3, wherein the method comprises: The step (2) comprises: (2.1) using the SPM12 toolkit in Matlab2017b, pre-processing the resting state functional magnetic resonance imaging data (fMRI); (2.2) deleting the first 10 time series of each subject to maintain the stability of the signal; (2.3) performing time correction and head motion correction processing; (2.4) registering the nuclear magnetic resonance T1 structural image to the resting state image after head motion correction; (2.5) using the Lie algebra differential homeomorphism registration algorithm to segment the T1 structural image; (2.6) Resting-state brain image data were registered to the Montreal Neurological Institute space using DARTAL and resampled to 3 x 3 x 3 mm 3 ; (2.7) performing linear trend and band-pass filtering processing on the resting state brain image data obtained after resampling to remove artifacts; (2.8) applying a half-full-width 4mm three-dimensional Gaussian kernel function for spatial smoothing processing; (2.9) regressing mixed signals, and putting 6 head motion parameters, average cerebrospinal fluid signal, average white matter signal and whole brain signal into an autoregressive model as covariates to correct the resting state functional magnetic resonance imaging data (fMRI) pre-processed through the above steps; (2.10) according to the standard of head motion translation > 2.5mm or rotation > 2.5°, eliminating the subjects who do not meet the standard.

5. The method of claim 2, wherein the method comprises: The step (4) comprises: (4.1) using edge-weighted graph attention network to integrate feature nodes in each brain function network; (4.2) using a differentiable pooling operation module to compress information in the feature network in a hierarchical clustering manner; (4.3) after hierarchical clustering, the graph becomes more and more rough, thereby generating a hierarchical representation of the input graph; (4.4) after two layers of differentiable pooling operations, the feature information of 90 nodes is stacked to 32 nodes; (4.5) repeating the above steps until stacking to 4 nodes.

6. The method of claim 5, wherein the method comprises: The step (5) comprises: (5.1) using a 5-fold cross-validation strategy to classify the node features of each subject; (5.2) the decision module uses a 2-layer MLP, each layer containing 32 and 3 neurons; (5.3) cross-entropy loss is used for classification, and the decay weight setting of Adam optimization is 0.08; (5.4) using a multilayer perceptron model for 100 epochs of model training, with an initial learning rate of 0.001; (5.5) extracting the accuracy, F1 score, precision and recall indicators to evaluate the model performance.

7. An apparatus for implementing a depression subtype classification process based on a deep learning multi-fusion brain network graph technology, characterized in that, The device comprises: a processor configured to execute computer executable instructions; a memory storing one or more computer executable instructions, which, when executed by the processor, implement each step of the depression subtype classification processing based on the deep learning multi-fusion brain network graph technology according to any one of claims 2 to 6.

8. A processor for implementing a depression subtype classification process based on a deep learning multi-fusion brain network graph technology, characterized in that, The processor is configured to execute computer executable instructions, and the computer executable instructions, when executed by the processor, implement each step of the depression subtype classification processing based on the deep learning multiple fusion brain network graph technology in any one of claims 2 to 6.

9. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and the computer program can be executed by a processor to implement each step of the depression subtype classification processing based on the deep learning multiple fusion brain network graph technology in any one of claims 2 to 6.

Images