Parkinson's disease early warning system and method based on graph neural network

Abstract

The invention discloses a Parkinson's disease early warning system and method based on a graph neural network, and the system comprises a data preprocessing module which is used for obtaining resting state functional magnetic resonance imaging data, carrying out the preprocessing of the obtained data, and carrying out the construction of graph structure data; the feature extraction module is used for performing feature extraction on the graph structure data and removing redundant features or irrelevant features in the extracted features to obtain a feature subset associated with the Parkinson's disease; the model training module is used for constructing a graph neural network and training a constructed model by using samples randomly selected from the feature subset to obtain a trained graph neural network; and the early warning module is used for predicting input to-be-detected data based on the trained graph neural network, outputting a prediction label of the Parkinson's disease and carrying out early warning on the Parkinson's disease. According to the invention, the image neural network is adopted to fully extract the Parkinson's disease medical image features, and the Parkinson's disease can be accurately early warned.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided detection technology, and in particular to an early warning system and method for Parkinson's disease based on graph neural network. Background Art

[0002] Parkinson's disease (PD) is a progressive neurological disease characterized by tremor, rigidity, and bradykinesia. It is associated with progressive neuronal loss in the substantia nigra and other brain structures, with a variety of motor and non-motor clinical symptoms that seriously affect patients' daily lives. In practice, traditional PD detection methods can be subjective because they rely on the detection of occasional weak movements by the human eye, which makes them difficult to classify and may lead to misdiagnosis. In the traditional research paradigm, due to the complexity of PD clinical manifestations, neuroimaging, gait, and finger movement are all included in the mainstream methods of PD detection and pathogenic factor analysis. Among them, gait analysis and finger movement analysis are different. Neuroimaging methods such as electroencephalogram (EEG), computed tomography (CT), and magnetic resonance imaging (MRI) have achieved direct measurement of brain function and structure. With the improvement of intuition and reliability, neuroimaging technology has played a more important role in exploring the pathogenic factors of PD. In addition, compared with EEG and CT, MRI has obvious advantages in spatial resolution and no radiation, so it has broader application prospects in the study of PD.

[0003] Recent evidence shows that the functional connectivity structure of the brain in PD is changed, and these changes can be used to help identify PD. Resting-state functional magnetic resonance imaging (rs-fMRI) captures the spontaneous activity of neurons by observing changes in blood flow in the brain, which accelerates the process of detecting brain diseases and dysfunctions. Compared with structural MRI, these features can more sensitively reveal the physiological and metabolic dysfunction of the cerebral cortex and changes in various circuits, and functional abnormalities often take precedence over structural changes. A study measured the amplitude of low frequency fluctuation (ALFF) and fractional amplitude of low frequency fluctuation (fALFF) in PD patients, and used support vector machine (SVM) to distinguish PD patients. It was also found that compared with healthy controls, PD patients had changes in ALFF in the bilateral lingual gyrus and left cerebellum, and changes in fALFF in the right posterior cerebellum. Other studies have shown that global graph theory indicators such as characteristic path length and efficiency, clustering coefficient and small-world degree, as well as local indicators such as integration, isolation, centrality and node degree can be extracted from functional connections to classify PD patients. At the same time, a floating automatic feature selection method was adopted, and it was found that these features belonged to the right cuneate area, left precuneus area, right superior frontal area and middle frontal area.

[0004] Traditional machine learning methods are limited by handcrafted features and cannot fully utilize the brain's connection information. The brain may be one of the most complex networks, and different forms of connections between different brain regions constitute a very complex brain network; although graph neural networks have powerful graph structure data representation capabilities and can extract features from complex brain networks. However, graph neural networks, like most neural networks, are regarded as a black box. It is difficult to intuitively understand the working principles behind the network, and the auxiliary detection results obtained are difficult for medical workers to convince.

[0005] Therefore, the prior art needs to be improved. Summary of the invention

[0006] The technical problem to be solved by the present invention is that, in view of the defects of the prior art, the present invention provides an early warning system and method for Parkinson's disease based on graph neural network, so as to solve the problem that the existing Parkinson's disease detection method cannot make full use of the brain's connection information, resulting in low accuracy of auxiliary detection.

[0007] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0008] In a first aspect, the present invention provides an early warning system for Parkinson's disease based on a graph neural network, comprising:

[0009] A data preprocessing module is used to obtain resting state functional magnetic resonance imaging data, and preprocess the obtained data to construct graph structure data;

[0010] A feature extraction module, used to extract features from the graph structure data and remove redundant features or irrelevant features from the extracted features to obtain a feature subset associated with Parkinson's disease;

[0011] A model training module, used to construct a graph neural network, and train the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network;

[0012] The early warning module is used to predict the input test data based on the trained graph neural network, output the prediction label of Parkinson's disease, and issue a Parkinson's disease early warning.

[0013] In one implementation, the data preprocessing module includes:

[0014] A data acquisition unit, configured to acquire the resting state functional magnetic resonance imaging data from a Parkinson's disease progression marker program database; wherein the resting state functional magnetic resonance imaging data includes: T1 modal images, diffusion tensor modal images, and functional magnetic resonance modal images corresponding to Parkinson's disease patients and healthy control patients, respectively;

[0015] A T1 image preprocessing unit, used to perform tissue segmentation on the T1 modal image using a probability map, align the segmented tissues according to template pairing and standardization operations, and divide the volumes corresponding to gray matter, white matter, and cerebrospinal fluid into a plurality of regions using a preset template;

[0016] A diffusion tensor image preprocessing unit, used to resample the diffusion tensor modal image according to a preset resampling resolution, remove scalp and blank space from the resampled image, and construct and extract the structural connection area of ​​the brain network using a deterministic network according to preset diffusion matrix parameters and fiber tracking and network parameters;

[0017] The functional magnetic resonance image preprocessing unit is used to remove time points, perform layer time correction and head motion correction on the functional magnetic resonance modal image, and perform image matching, image denoising and image standardization on the corrected image, and extract the functional connection area according to the determined brain area of ​​interest.

[0018] In one implementation, the feature extraction module includes:

[0019] A Fisher score calculation unit, used to calculate the score of each feature in the graph structure data based on the Fisher criterion, and select the features with the top m scores;

[0020] A mutual information selection unit, used to analyze the relevance of the selected features in predicting the target variable and the redundancy relative to other variables based on the mutual information selection method, and remove the redundant features or the irrelevant features;

[0021] The regularization and recursive selection unit is used to perform feature selection on the remaining features based on the L1 regularization method and the recursive feature elimination method to obtain the feature subset associated with Parkinson's disease.

[0022] In one implementation, the model training module includes:

[0023] A graph neural network construction unit is used to reduce the dimension of the feature matrix through a parameterized function to obtain a new feature matrix after dimension reduction, and calculate a probability map; use the probability map to construct an adjacency matrix through a k-nearest neighbor algorithm; perform graph convolution, and obtain a final output matrix through multi-layer graph convolution operations; normalize the output matrix through a softmax function to obtain a label prediction Y for each node;

[0024] The graph neural network training unit is used to create a data graph of a preset size by random sampling; wherein the data graph of the preset size includes an unknown sample a and n known samples; the constructed model is trained according to the data graph of the preset size, and the unknown sample a is predicted to obtain a predicted label y; the predicted label y is compared with the actual label of the unknown sample a A comparison is performed, and based on the error generated by the comparison, the network parameters of the model are adjusted and updated using a back propagation algorithm; the model is optimized using a five-fold cross-validation method and an optimization algorithm to obtain the trained graph neural network.

[0025] In a second aspect, the present invention provides a method for detecting and processing Parkinson's disease data in a Parkinson's disease early warning system based on a graph neural network as described in the first aspect, comprising:

[0026] Acquire resting state functional magnetic resonance imaging data, preprocess the acquired data, and construct graph structure data;

[0027] Extracting features from the graph structure data, and removing redundant features or irrelevant features from the extracted features to obtain a feature subset associated with Parkinson's disease;

[0028] Constructing a graph neural network, and training the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network;

[0029] Based on the trained graph neural network, the input data to be tested is predicted, a prediction label of Parkinson's disease is output, and an early warning of Parkinson's disease is performed.

[0030] In one implementation, the step of acquiring resting state functional magnetic resonance imaging data and preprocessing the acquired data to construct graph structure data includes:

[0031] Acquire the resting state functional magnetic resonance imaging data from the Parkinson's disease progression marker program database; wherein the resting state functional magnetic resonance imaging data includes: T1 modal images, diffusion tensor modal images, and functional magnetic resonance modal images corresponding to Parkinson's disease patients and healthy control patients, respectively;

[0032] The T1 modality image, the diffusion tensor modality image and the functional magnetic resonance modality image are preprocessed respectively to construct corresponding graph structure data.

[0033] In one implementation, the preprocessing of the T1 modality image, the diffusion tensor modality image, and the functional magnetic resonance modality image to construct graph structure data includes:

[0034] Performing tissue segmentation on the T1 modality image using the probability map, aligning the segmented tissues according to template pairing and standardization operations, and dividing the volumes corresponding to the gray matter, white matter, and cerebrospinal fluid into a number of regions using a preset template;

[0035] Resampling the diffusion tensor modality image according to a preset resampling resolution, removing the scalp and blanking from the resampled image, and constructing and extracting the structural connection area of ​​the brain network using a deterministic network according to preset diffusion matrix parameters and fiber tracking and network parameters;

[0036] The functional magnetic resonance image is processed by removing time points, correcting layer time and correcting head motion, and the corrected image is processed by image matching, image denoising and image standardization, and the functional connection area is extracted according to the determined brain area of ​​interest.

[0037] In one implementation, the functional magnetic resonance image is subjected to time point removal, slice time correction, and head motion correction, and the corrected image is subjected to image matching, image denoising, and image standardization, and the functional connection area is extracted according to the determined brain region of interest, including:

[0038] Deleting a preset number of initial time points in the functional magnetic resonance modality image, adjusting the data of all layers to the same time point using a layer time correction method, and performing head motion correction using the head motion correction algorithm;

[0039] The corrected image is processed to remove the neck and scalp, and the T1 modality image is matched to the functional image of the removed image, multiple noise covariates are excluded, and the image is normalized according to a preset bounding box and voxel size;

[0040] The brain region of interest is determined, correlation analysis is performed between the brain region of interest and other brain regions, a functional connectivity map is generated, and the functional connectivity region is extracted based on the functional connectivity map.

[0041] In one implementation, the feature extraction is performed on the graph structure data, and redundant features or irrelevant features in the extracted features are removed to obtain a feature subset associated with Parkinson's disease, including:

[0042] Calculate the score of each feature in the graph structure data based on the Fisher criterion, and select the features with the top m scores;

[0043] Analyze the relevance of the selected features in predicting the target variable and their redundancy relative to other variables based on the mutual information selection method, and remove the redundant features or the irrelevant features;

[0044] Feature selection is performed on the remaining features based on an L1 regularization method and a recursive feature elimination method to obtain the feature subset associated with Parkinson's disease.

[0045] In one implementation, constructing a graph neural network and training the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network includes:

[0046] The feature matrix X used to construct the small graph is taken as input and passed through the parameterized function f θ Perform dimensionality reduction processing on the feature matrix to obtain a new dimensionality-reduced feature matrix And the probability map is calculated according to the following formula:

[0047]

[0048] Among them, t represents the learnable parameters in the graph network, Δ represents the Euclidean distance, and x i represents the i-th sample;

[0049] Using the probability graph P ij (X; θ, t), construct the adjacency matrix A through the k-nearest neighbor algorithm;

[0050] Perform graph convolution:

[0051]

[0052] in, represents the feature matrix of the lth layer, W (l) represents the learnable weight matrix corresponding to the lth layer, represents a diagonal matrix, σ represents a nonlinear activation function;

[0053] Through multi-layer graph convolution operations, the final output matrix H is obtained (l) ;

[0054] The output matrix H is transformed into (l) Perform normalization to obtain the label prediction Y of each node:

[0055]

[0056] Among them, h ij and ij denote the elements in the i-th row and j-th column of matrices H and Y respectively;

[0057] Creating a data graph of a preset size by random sampling; wherein the data graph of the preset size includes an unknown sample a and n known samples;

[0058] The constructed model is trained according to the data graph of the preset scale, and the unknown sample a is predicted to obtain a predicted label y;

[0059] The predicted label y is compared with the actual label of the unknown sample a. Make a comparison, and adjust and update the network parameters of the model using the back propagation algorithm based on the error generated by the comparison;

[0060] The model is optimized using a five-fold cross-validation method and an optimization algorithm to obtain the trained graph neural network.

[0061] The present invention adopts the above technical solution to achieve the following effects:

[0062] Based on functional magnetic resonance imaging data, the present invention uses graph neural networks to fully extract significant subgraphs related to Parkinson's disease and significant features of brain regions, thereby realizing auxiliary detection of Parkinson's disease. Moreover, after obtaining the auxiliary detection results, the graph neural network interpretation technology is used to classify Parkinson's disease markers, which can effectively distinguish Parkinson's disease from healthy control groups, providing interpretability for Parkinson's disease early warning. The present invention uses graph neural networks to fully extract the medical imaging features of Parkinson's disease, improves the accuracy of early warning of Parkinson's disease, and reduces the cost of early warning of Parkinson's disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0064] Figure 1 It is a schematic diagram of the Parkinson's disease early warning system based on graph neural network in the present invention.

[0065] Figure 2 It is a flow chart of the Parkinson's disease data detection and processing method based on graph neural network in the present invention.

[0066] Figure 3 This is the overall framework diagram of the Parkinson's disease auxiliary detection in the present invention.

[0067] Figure 4 It is a schematic diagram of the quality detection report of T1 modality image preprocessing in the present invention.

[0068] Figure 5 It is a visualization diagram of the structural connection diagram of the brain network in the present invention.

[0069] Figure 6 It is a visualization diagram of the functional connectivity diagram of the regional correlation analysis in the present invention.

[0070] Figure 7 This is the neural network framework diagram of the present invention.

[0071] Figure 8 It is a structural schematic diagram of the Parkinson's disease detection model in the present invention.

[0072] Fig. 9 It is a schematic diagram of the integrated learning framework of the PD detection model in the present invention.

[0073] Fig.10 It is a functional principle diagram of a terminal in one implementation of the present invention.

[0074] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0076] Exemplary Systems

[0077] Parkinson's disease (PD) is a progressive neurological disease characterized by tremors, rigidity, and bradykinesia. Traditional PD detection methods can be subjective because they rely on detecting occasional weak movements with the human eye, which makes them difficult to classify and may lead to misdiagnosis; traditional machine learning methods are limited by hand-crafted features and cannot fully utilize the brain's connection information. The brain may be one of the most complex networks, and different forms of connections between different brain regions constitute a very complex brain network; although, graph neural networks have powerful graph structure data representation capabilities and can extract features from complex brain networks. However, graph neural networks, like most neural networks, are regarded as a black box. It is difficult to intuitively understand the working principles behind the network, and the auxiliary detection results obtained are difficult for medical workers to convince.

[0078] In view of the above technical problems, an embodiment of the present invention provides an early warning system for Parkinson's disease based on a graph neural network. Based on functional magnetic resonance imaging data, the system uses a graph neural network to fully extract significant subgraphs related to Parkinson's disease and significant features of brain regions, thereby realizing auxiliary detection of Parkinson's disease. In addition, after obtaining the auxiliary detection results, the graph neural network interpretation technology is used to classify Parkinson's disease markers, which can effectively distinguish Parkinson's disease from healthy control groups, providing interpretability for Parkinson's disease warning. The embodiment of the present invention uses a graph neural network to fully extract the medical imaging features of Parkinson's disease, improves the accuracy of early warning of Parkinson's disease, and reduces the cost of early warning of Parkinson's disease.

[0079] like Figure 1 As shown, the present invention provides an early warning system for Parkinson's disease based on a graph neural network, comprising:

[0080] A data preprocessing module 10 is used to obtain resting state functional magnetic resonance imaging data, and preprocess the obtained data to construct graph structure data;

[0081] A feature extraction module 20 is used to extract features from the graph structure data and remove redundant features or irrelevant features from the extracted features to obtain a feature subset associated with Parkinson's disease;

[0082] A model training module 30 is used to construct a graph neural network and train the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network;

[0083] The early warning module 40 is used to predict the input test data based on the trained graph neural network, output a prediction label of Parkinson's disease, and perform a Parkinson's disease early warning.

[0084] Specifically, in an implementation of this embodiment, the data preprocessing module 10 includes:

[0085] A data acquisition unit, configured to acquire the resting state functional magnetic resonance imaging data from a Parkinson's disease progression marker program database; wherein the resting state functional magnetic resonance imaging data includes: T1 modal images, diffusion tensor modal images, and functional magnetic resonance modal images corresponding to Parkinson's disease patients and healthy control patients, respectively;

[0086] A T1 image preprocessing unit, used to perform tissue segmentation on the T1 modal image using a probability map, align the segmented tissues according to template pairing and standardization operations, and divide the volumes corresponding to gray matter, white matter, and cerebrospinal fluid into a plurality of regions using a preset template;

[0087] A diffusion tensor image preprocessing unit, used to resample the diffusion tensor modal image according to a preset resampling resolution, remove scalp and blank space from the resampled image, and construct and extract the structural connection area of ​​the brain network using a deterministic network according to preset diffusion matrix parameters and fiber tracking and network parameters;

[0088] The functional magnetic resonance image preprocessing unit is used to remove time points, perform layer time correction and head motion correction on the functional magnetic resonance modal image, and perform image matching, image denoising and image standardization on the corrected image, and extract the functional connection area according to the determined brain area of ​​interest.

[0089] Specifically, in an implementation of this embodiment, the feature extraction module 20 includes:

[0090] A Fisher score calculation unit, used to calculate the score of each feature in the graph structure data based on the Fisher criterion, and select the features with the top m scores;

[0091] A mutual information selection unit, used to analyze the relevance of the selected features in predicting the target variable and the redundancy relative to other variables based on the mutual information selection method, and remove the redundant features or the irrelevant features;

[0092] The regularization and recursive selection unit is used to perform feature selection on the remaining features based on the L1 regularization method and the recursive feature elimination method to obtain the feature subset associated with Parkinson's disease.

[0093] Specifically, in an implementation of this embodiment, the model training module 30 includes:

[0094] A graph neural network construction unit is used to reduce the dimension of the feature matrix through a parameterized function to obtain a new feature matrix after dimension reduction, and calculate a probability map; use the probability map to construct an adjacency matrix through a k-nearest neighbor algorithm; perform graph convolution, and obtain a final output matrix through multi-layer graph convolution operations; normalize the output matrix through a softmax function to obtain a label prediction Y for each node;

[0095] A graph neural network training unit is used to create a data graph of a preset size by random sampling; wherein the data graph of the preset size includes an unknown sample a and n known samples; the constructed model is trained according to the data graph of the preset size, and the unknown sample a is predicted to obtain a predicted label y; the predicted label y is compared with the actual label y^ of the unknown sample a, and according to the error generated by the comparison, the network parameters of the model are adjusted and updated by using a back propagation algorithm; the model is optimized by using a five-fold cross-validation method and an optimization algorithm to obtain the trained graph neural network.

[0096] This embodiment achieves the following technical effects through the above technical solution:

[0097] Based on the functional magnetic resonance imaging data, this embodiment uses a graph neural network to fully extract the significant subgraphs related to Parkinson's disease and the significant features of the brain region, thereby realizing auxiliary detection of Parkinson's disease. In addition, after obtaining the auxiliary detection results, the graph neural network interpretation technology is used to classify Parkinson's disease markers, which can effectively distinguish Parkinson's disease from the healthy control group, providing interpretability for Parkinson's disease early warning. This embodiment uses a graph neural network to fully extract the medical imaging features of Parkinson's disease, improves the accuracy of early warning of Parkinson's disease, and reduces the cost of early warning of Parkinson's disease.

[0098] Exemplary Methods

[0099] like Figure 2 As shown, based on the above embodiments, an embodiment of the present invention further provides a method for detecting and processing Parkinson's disease data based on a graph neural network, comprising the following steps:

[0100] Step S100, acquiring resting state functional magnetic resonance imaging data, and preprocessing the acquired data to construct graph structure data.

[0101] In this embodiment, the graph neural network has a powerful graph structure data representation capability and can extract features from complex brain networks. However, like most neural networks, the graph neural network is regarded as a black box, and it is difficult to intuitively penetrate the working principle behind the network, and the auxiliary detection results obtained are difficult for medical workers to convince. This embodiment mainly uses a graph neural network to fully extract the medical imaging features of Parkinson's disease based on functional magnetic resonance imaging data to achieve auxiliary detection of Parkinson's disease. Moreover, after obtaining the auxiliary detection results, this embodiment uses a graph neural network interpretation technology to classify Parkinson's disease markers.

[0102] The overall framework of Parkinson's disease auxiliary detection in this embodiment is as follows Figure 3 As shown in the figure, rs-fMRI (resting state functional magnetic resonance imaging) data is first obtained from the Parkinson's Progression Markers Initiative (PPMI) database, and then the rs-fMRI data is preprocessed to construct graph structure data so that feature extraction can be performed using a graph neural network, and finally the predicted label of the original rs-fMRI data is output. After obtaining the trained model, the graph neural network interpreter is used to implement the classification of Parkinson's disease markers.

[0103] Specifically, in an implementation of this embodiment, step S100 includes the following steps:

[0104] Step S101, acquiring the resting state functional magnetic resonance imaging data from the Parkinson's disease progression marker project database; wherein the resting state functional magnetic resonance imaging data includes: T1 modal images, diffusion tensor modal images and functional magnetic resonance modal images corresponding to Parkinson's disease patients and healthy control patients, respectively.

[0105] In this example, the Parkinson's Progression Markers Initiative (PPMI) database was used to collect patient data, which was derived from observational clinical and follow-up studies, including: PD (Parkinson's disease) data, SWEDD (molecular imaging evidence of no dopaminergic system dysfunction) data, and HC (healthy control) data. This dataset aims to identify biological markers for risk, onset, and progression of Parkinson's disease, develop better treatment tools, and provide the research community with a comprehensive, standardized, longitudinal dataset and biobank to accelerate breakthroughs and validate the clinical application of new discoveries.

[0106] This embodiment mainly collects MRI (magnetic resonance imaging) data of 123 PD patients and 41 HC (healthy controls) patients in the PPMI database, including: T1 (the time required for the longitudinal magnetization vector to recover to 63% of its original value) modal image, DTI modal image (diffusion tensor imaging) and fMRI modal image (functional magnetic resonance modal image) data; the collected data are shown in Table 1, and the collected data are randomly divided into a training data set (n=114) and an independent test data set (n=50).

[0107] Table 1 Parkinson's disease dataset

[0108]

[0109] Step S102 , preprocessing the T1 modality image, the diffusion tensor modality image, and the functional magnetic resonance modality image respectively to construct corresponding graph structure data.

[0110] In this embodiment, after acquiring the multimodal MRI data, it is also necessary to extract brain region features from the T1 modality image, the DTI modality image, and the fMRI modality image, respectively. This process involves the use of a variety of dedicated toolboxes and software to complete.

[0111] Specifically, in an implementation of this embodiment, step S102 includes the following steps:

[0112] Step S102a, performing tissue segmentation on the T1 modality image using the probability map, aligning the segmented tissues according to template pairing and standardization operations, and dividing the volumes corresponding to the gray matter, white matter, and cerebrospinal fluid into a plurality of regions using a preset template.

[0113] In this embodiment, the toolbox CAT12 (a tool for brain image analysis) and SPM software (data mining prediction analysis software) are needed to implement the T1 modality image preprocessing process; the toolbox CAT12 and SPM are widely used in the Matlab (a mathematical software) environment to perform voxel-based morphometry (VBM) analysis of various brain diseases, and they can process various forms of structural magnetic resonance imaging data. The T1 modality image is mainly preprocessed using the CAT12 toolbox.

[0114] As an example, the specific steps of preprocessing using the CAT12 toolbox in this embodiment are as follows:

[0115] 1. Tissue Segmentation (Tissue Probability Map). Using the probability map TPM.nii for tissue segmentation, the brain structure can be divided into gray matter, white matter, and cerebrospinal fluid.

[0116] 2. Affine Regularization: Due to the inherent differences in brain anatomy between different races, appropriate reference templates should be selected based on the geographic and ethnic attributes of the sample; for example, a template representing the brain structure of the European population or a template representing the East Asian population should be selected.

[0117] 3. MNI template standardization (Montreal Neurological Institute, MNI). The standardization process involves aligning the preprocessed images to a standard reference template, MNI152.nii, which is an average model of the brain calculated by analyzing the MRI scan data of many healthy subjects and is widely used in the standardization of brain images. During the standardization process, the registration algorithms used include affine and nonlinear transformations to achieve accurate alignment of the images in the standard brain space.

[0118] 4. Segment brain regions. Use the AAL-116 template to divide the extracted gray matter, white matter, and cerebrospinal fluid volumes into 116 regions, respectively, and obtain 384 features.

[0119] 5. Data Quality Check (QC). After data segmentation is completed, the data needs to be evaluated to ensure that it meets the preset quality standards. In this process, unqualified data will be identified and excluded to ensure the accuracy and reliability of the analysis results. Figure 4 shown.

[0120] Specifically, in an implementation of this embodiment, step S102 further includes the following steps:

[0121] Step S102b, resampling the diffusion tensor modality image according to a preset resampling resolution, removing the scalp and blanking from the resampled image, and constructing and extracting the structural connection area of ​​the brain network using a deterministic network according to preset diffusion matrix parameters and fiber tracking and network parameters.

[0122] In this embodiment, PANDA software (PANDA is a DTI data analysis software running under the meta-learning FSL framework) is needed to pre-process the DTI modality images, which mainly includes the following steps:

[0123] 1. Set the resampling resolution. Typically, machine-acquired data has a resolution of [2 2 2] mm, but some devices may produce data with higher resolutions such as [1 1 1] mm or [0.9 0.9 1] mm. Although high resolutions may be beneficial for capturing details, they are not necessary for analyzing the original structural images and can significantly increase the computational burden, especially when performing complex calculations such as fiber tracking. Therefore, these high-resolution images are usually resampled to a standard resolution of [2 2 2] mm.

[0124] 2. Scalp removal and blank area removal. When processing DTI data, a parameter needs to be set in advance to perform the scalp removal operation. This parameter determines the degree of scalp removal: the larger the parameter value, the more scalp tissue is removed and the less brain is retained. Usually the parameter is set to 0.25. After scalping, a large number of information-free areas will appear around the image, which appear as black areas on the periphery of the brain. If these areas are not processed, the computational burden will be unnecessarily increased. Therefore, the next step is to crop out these useless peripheral areas to improve the efficiency of data processing. The value of the cropping parameter represents the distance from the edge of the brain to the cropping boundary, and is usually set to 3mm by default.

[0125] 3. Setting of diffusion matrix parameters (Producing diffusion metrics).

[0126] a) Local Diffusion Homogeneity (LDH). The calculation of LDH involves evaluating the correlation between the diffusion rate sequence of one voxel and the diffusion rate sequence of its neighboring voxels. The PANDA toolbox provides two correlation coefficients to perform this calculation using the correlation test (Kendall's tau) and the Spearman's rank correlation coefficient (Spearman's rho).

[0127] b) Normalizing. After obtaining the diffusion parameters such as fractional anisotropy, mean diffusivity, axial diffusivity and radial diffusivity, the next step is to align these parameter images to a standardized reference template. By default, FSL provides a FMRIB58_FA template with a resolution of 1x1x1mm for this process.

[0128] c) Smooth. Before performing whole-brain voxel-level analysis, it is usually necessary to resample and smooth the diffusion parameter image in order to reduce the errors that may occur during the registration process. Smoothing usually uses the full width at half maximum (FWHM) as a parameter, and by default, a smoothing kernel with a FWHM value of 6 mm is usually sufficient.

[0129] 4. Fiber tracking and network parameter setting.

[0130] a) Fiber tracking parameter settings (Deterministic Fiber Tracking). This step aims to depict the path of the entire brain white matter fiber. Usually, as the default option, the deterministic fiber tracking algorithm is used for fiber tracking.

[0131] b) Network Node Definition. When constructing a brain structural network, the brain is first partitioned using a gray matter atlas, and each partitioned brain region is considered a node in the network. Next, the connecting edges in the network are defined using the previously traced white matter fibers. In order to build such a network, it is necessary to obtain T1 structural MRI images of each sample to determine the network nodes and FA (image contrast adjustment) images to determine the edges of the network. In addition, it is necessary to select a partition atlas of the T1 image, of which the ALL_116 atlas is a commonly used choice.

[0132] c) Parameter settings for preprocessing T1 images. Scalp removal is performed through the function f(skull_removal), and its default parameter is set to 0.5. The useless area outside the T1 structural image is removed through Cropping_Gap, which can save storage space and increase processing speed. Resample_Resolution is responsible for resampling the resolution of the T1 image to 1x1x1mm.

[0133] d) Network Construction: Directly build a deterministic network.

[0134] 5. The construction result of the brain network, using only the deterministic network (FN), FN means that the brain network constructed by using the number of fibers between each partition of the brain map obtains the structural connection area, and 13456 features are obtained. In this embodiment, the visualization diagram corresponding to the structural connection area of ​​the brain network is as follows: Figure 5 shown.

[0135] Specifically, in an implementation of this embodiment, step S102 further includes the following steps:

[0136] Step S102c, removing time points, performing slice time correction and head motion correction processing on the functional magnetic resonance modality image, and performing image matching, image denoising and image standardization processing on the corrected image, and extracting the functional connection area according to the determined brain area of ​​interest.

[0137] Specifically, in an implementation of the present embodiment, the above step S102c includes: deleting a preset number of initial time points in the functional magnetic resonance modality image, adjusting the data of all layers to the same time point by using a layer time correction method, and performing head motion correction by using the head motion correction algorithm; removing the neck and scalp from the corrected image, matching the T1 modality image to the functional image of the removed image, excluding multiple noise covariates, and performing image standardization according to a preset bounding box and voxel size; determining the brain area of ​​interest, performing correlation analysis on the brain area of ​​interest and other brain areas, generating a functional connectivity map, and extracting the functional connectivity area according to the functional connectivity map.

[0138] In this embodiment, it is necessary to use DAPBI software (a data analysis platform) to preprocess the fMRI modality images, which mainly includes the following steps:

[0139] 1. Remove time points. When performing MRI image scanning, the images at the first few time points are usually affected by the stability of the gradient magnetic field and the subject's adaptation process, resulting in high image noise. In order to improve data quality, a certain number of initial time points are usually deleted. By default, when deleting initial time points, the first 10 time points will be removed, but at least 4 will be removed to ensure data accuracy.

[0140] 2. Slice Timing. In the process of brain MRI scanning, the intermittent layer scanning technology is usually used. This scanning method causes differences in the scanning time between adjacent layers, that is, the 1st and 3rd layers will have an earlier scanning time than the 2nd layer. In order to eliminate the impact of this time difference on the scanning results of the same brain area, it is necessary to use the slice timing correction method to adjust the data of all layers to the consistency as if they were obtained at the same time point.

[0141] 3. Input basic information, including: total number of layers, scanning order and reference layer.

[0142] 4. Realignment: During the experiment, the subject’s head may move, which requires correction through a specific algorithm to reduce the impact of head movement on the data. This process is called realignment.

[0143] 5. Manually correct functional and structural images. During the processing of MRI data, due to the rotation, movement, or positioning origin errors of the participant's head, manual adjustment of the image may be required to ensure accurate alignment of the image. Manually correcting the image can improve the accuracy of subsequent steps, including image registration, segmentation, and standardization. The accuracy of these steps is crucial to obtaining high-quality brain image analysis results.

[0144] 6. Remove the neck and scalp, and match the T1 image to the functional image.

[0145] 7. Eliminate multiple noise covariates, such as global signals, noise in gray matter and white matter in the brain, etc. In addition, it is necessary to eliminate linear drift and remove noise signals from white matter and cerebrospinal fluid.

[0146] 8. Image Normalization: Since each human brain is unique, it must be normalized to a common coordinate space for easy comparison.

[0147] The bounding box is used to define the size of the brain space. Usually, the preset parameters are sufficient.

[0148] The default voxel size setting of [3 3 3] mm is usually sufficient. Even if the data was acquired at 3.5 or 3.7 mm resolution, the default setting of [3 3 3] can still be used. However, if the data was acquired using multi-band technology, which means that the data was acquired with a higher precision, the voxel size may be 2 mm, in which case the voxel size can be adjusted to [2 2 2].

[0149] 9. Calculate functional connectivity. After determining the region of interest (ROI), a functional connectivity map can be generated by performing correlation analysis between these ROIs and all other brain regions. Using the 116-region template of AAL to extract the functional connectivity regions, 13,456 features can be obtained; in this embodiment, the effect of visualizing these functional connectivity regions is as follows: Figure 6 shown.

[0150] like Figure 2 As shown, an embodiment of the present invention provides a method for detecting and processing Parkinson's disease data based on a graph neural network, comprising the following steps:

[0151] Step S200: extracting features from the graph structure data, and removing redundant features or irrelevant features from the extracted features to obtain a feature subset associated with Parkinson's disease.

[0152] In this embodiment, after preprocessing the T1 modality images, DTI modality images, and fMRI modality images, tens of thousands of feature dimensions are obtained. However, not all of these features are closely related to NDDs (neurodegenerative diseases). In order to streamline the data and highlight those key features, feature extraction technology is needed to screen out the feature subsets that are most strongly associated with the disease, while excluding those redundant or irrelevant features.

[0153] Generally, feature selection techniques can be summarized into three categories: packaging, embedding and filtering. In the embedding method, feature selection is combined with a learning algorithm, usually designed for a specific learning model. The packaging rule evaluates and selects feature subsets through a specific learning algorithm, and forms the final classifier based on this. The filtering rule screens feature subsets based on criteria that are independent of the learning model. In this embodiment, the feature subset with the strongest correlation with the disease is screened out based on the Fisher criterion, the mutual information selection method and the L1 regularization method.

[0154] Specifically, in an implementation of this embodiment, step S200 includes the following steps:

[0155] Step S201, calculating the score of each feature in the graph structure data based on the Fisher criterion, and selecting the features with the top m scores.

[0156] In this embodiment, Fisher's Criterion is one of the most widely used supervised feature selection methods in the filter method. The core concept of Fisher's Criterion is to find a feature subset so that in the selected feature space, the intervals between data points of different categories are large, while the intervals between data points of the same category are small. and is the mean and standard deviation of the k classes of the jth feature, let μ j and σ j represents the mean and standard deviation of the entire data set corresponding to the jth feature. Then the Fisher score of the jth feature is:

[0157]

[0158] in, After calculating the Fisher score for each feature, the top m features with larger scores are selected.

[0159] Step S202: analyzing the correlation of the selected features in predicting the target variable and their redundancy relative to other variables based on the mutual information selection method, and removing the redundant features or the irrelevant features.

[0160] In this embodiment, the mutual information selection method (MI) is also one of the filter methods to evaluate the relevance of the feature subset in predicting the target variable and its redundancy relative to other variables. The entropy H(X) of a random variable X, with p as the probability density function, is a common measure of uncertainty:

[0161] H(X)=E X [P(X)]=-∫p(x)logp(x)dx

[0162] Given two distributions p and q, the relative entropy (KullbackLeibler, KL) divergence is defined as:

[0163]

[0164] The mutual information (MI) between two random variables X and Y is defined as:

[0165]

[0166] Intuitively, the MI between X and Y measures the degree to which the uncertainty of one random variable Y (or X) is reduced after knowing the other random variable X (or Y). MI is symmetric, that is, I(X;Y) = I(Y;X). This definition can be directly extended by conditioning on a third random variable Z to obtain the conditional mutual information (CMI) between X and Y given Z:

[0167]

[0168] CMI implements the following useful chain rules:

[0169] I(X;Y,Z)=I(X;Z)+I(X;Y|Z)

[0170] CMI can be used to define the relevance and redundancy scores for feature selection problems, which naturally arise when limiting the desired regression / classification error. Y and X in A The CMI between is expressed as:

[0171]

[0172] This quantity v(A) intuitively represents the feature subset X A Importance in predicting target Y.

[0173] Step S203, performing feature selection on the remaining features based on the L1 regularization method and the recursive feature elimination method to obtain the feature subset associated with Parkinson's disease.

[0174] In this embodiment, L1 regularization is a method to prevent model overfitting, which is one of the embedded methods. Its specific implementation is to add a penalty term, i.e., a regularization term, to the loss function of the machine learning model. Its general form is as follows:

[0175]

[0176] In the above formula, J(w i ) function is the regularization term. The first term in the above formula is called expected loss. It can effectively prevent overfitting because it limits the overfitting of the model on the training set. Although this fitting performs well on the training data, it does not generalize well to new data. An overfitted model will try to fit each data point as if they were a series of scattered points, resulting in the model parameter w i The absolute value of becomes very large. During the training process, by continuously iterating the model parameter w i To minimize the loss function, the goal is to keep both the expected loss and the regularization term at a low level. There is a dynamic balance between the two - when the expected loss decreases, the value of the regularization term increases accordingly. Through this interaction, the model eventually reaches an optimal state, in which the corresponding parameter w for unimportant features is i will be reduced to near zero, thereby reducing the model's dependence on these features and enhancing the model's generalization ability. The L1 regularization formula is:

[0177]

[0178] L1 regularization has a different feature. When using a loss function with a regularization term, when the model reaches the optimal value, the coefficient w in front of the useless feature i It will become 0, which means it plays the role of feature selection. Note that there is a parameter λ before the regularization term in the above formula, which is a parameter that controls the degree of penalty. The larger it is, the less likely it is to overfit. When selecting features, the larger it is, the more features with coefficients of 0 there are.

[0179] In this embodiment, Recursive Feature Elimination (RFE) is a feature selection technique, which is one of the wrapper methods. It selects features by recursively considering smaller and smaller feature sets. The basic idea is to repeatedly build a model and remove the least important features (or feature groups) in each iteration until the required number of features is reached. This method uses the weights of a specific machine learning algorithm to judge the importance of each feature, thereby achieving the purpose of feature selection.

[0180] The training steps are as follows:

[0181] 1. Model training: First, RFE trains the model using all available features.

[0182] 2. Feature Importance Evaluation: It then evaluates the importance of each feature using the feature importance or coefficient provided by the machine learning model (such as support vector machine, linear regression, decision tree, etc.).

[0183] 3. Remove the least important features: Based on these importance scores, RFE removes one or more of the least important features.

[0184] 4. Repetition: The above process is repeated, using fewer features in each iteration, until a predetermined number of features is reached or some stopping condition is met.

[0185] like Figure 2 As shown, an embodiment of the present invention provides a method for detecting and processing Parkinson's disease data based on a graph neural network, comprising the following steps:

[0186] Step S300, construct a graph neural network, and use samples randomly selected from the feature subset to train the constructed model to obtain a trained graph neural network.

[0187] In this embodiment, through the several feature selection techniques in the above steps S201 to S203, the feature subset with the strongest correlation with the disease can be screened out, so that the constructed graph neural network can be trained using the feature subset.

[0188] Specifically, in an implementation of this embodiment, step S300 includes the following steps:

[0189] Step S301, taking the feature matrix X used to construct the small graph as input, and passing the parameterized function f θ Perform dimensionality reduction processing on the feature matrix to obtain a new dimensionality-reduced feature matrix And the probability map is calculated according to the following formula:

[0190]

[0191] Among them, t represents the learnable parameters in the graph network, Δ represents the Euclidean distance, and x i represents the i-th sample;

[0192] Step S302, using the probability map P ij (X; θ, t), construct the adjacency matrix A through the k-nearest neighbor algorithm;

[0193] Step S303, perform graph convolution:

[0194]

[0195] in, represents the feature matrix of the lth layer, W (l) represents the learnable weight matrix corresponding to the lth layer, represents a diagonal matrix containing the degree information of the nodes, which is used to normalize the adjacency matrix. σ represents a nonlinear activation function, and the ReLU activation function is usually used in graph convolutional neural networks.

[0196] Step S304, through multi-layer graph convolution operations, the final output matrix H is obtained (l) ;

[0197] Step S305: The output matrix H is processed by a softmax function. (l) Normalize it to get the probability distribution of each node belonging to different categories, that is, get the label prediction Y of each node:

[0198]

[0199] Among them, h ij and ij denote the elements in the i-th row and j-th column of matrices H and Y respectively.

[0200] In this embodiment, it is necessary to construct a graph convolutional neural network model, wherein the framework of the graph convolutional neural network is as follows: Figure 7 As shown; after constructing and training the graph convolutional neural network, it is used as the Parkinson's disease detection model in this embodiment. For the specific construction process of the Parkinson's disease detection model, please refer to the above steps S301 to S305.

[0201] Specifically, in an implementation of this embodiment, step S300 further includes the following steps:

[0202] Step S306, creating a data graph of a preset size by random sampling; wherein the data graph of the preset size includes an unknown sample a and n known samples;

[0203] Step S307, training the constructed model according to the data graph of the preset scale, and predicting the unknown sample a to obtain a predicted label y;

[0204] Step S308: compare the predicted label y with the actual label of the unknown sample a. Make a comparison, and adjust and update the network parameters of the model using the back propagation algorithm based on the error generated by the comparison;

[0205] Step S309, using a five-fold cross-validation method and an optimization algorithm to optimize the model to obtain the trained graph neural network.

[0206] In this embodiment, in order to deal with the problem of imbalance of positive and negative samples and insufficient sample size in the dataset, a new training strategy is proposed. The strategy involves using a small graph constructed by randomly selecting samples from the dataset as input. With this input, the graph convolutional neural network model is able to predict the labels of unknown nodes in the graph. In this process, the specific formula for constructing the feature matrix X of the small graph is as follows:

[0207]

[0208] Among them, x represents the sample, Represents unknown samples, c is the label of the sample, C is the total number of categories, and in each category, q known samples are selected and one unknown sample is added, so N = C·q+1. Taking the early detection task of PD as an example, assume that there are two categories: PD patients and HC (healthy controls). When q is set to 10, 10 samples are drawn from each category and one unlabeled sample is added, that is, the number of nodes in the selected graph is C·q+1=21.

[0209] The main training strategy is: first, create a small-scale data graph through random sampling, which contains an unknown sample a and n known samples. The model is trained with this data and predicts the unknown sample a to obtain the predicted label y. Then, the predicted label y is compared with the actual label of the unknown sample a. The comparison is performed, and the error generated by this comparison is used to adjust and update the network parameters of the model using the back propagation algorithm. This process is repeated, and each time a new small graph is randomly drawn, and back propagation is performed on this basis to further train and optimize the model. This iterative training will continue for 200 times, and each time the network parameters of the model will be fine-tuned. After 200 iterations, the network parameters of the model will be fully optimized, and the model at this time is considered to be the final model. Use this final model to predict the test set to evaluate the performance and accuracy of the model on unknown data. Through continuous iteration and updating, the model's ability to recognize and classify unknown samples is improved. The structure of the Parkinson's disease detection model in this embodiment is as follows Figure 8 shown.

[0210] At this point, the network architecture has been built, and the next step is to train the model. In order to evaluate the performance of the model, a five-fold cross validation method is used. In the experiment, the Adam optimization algorithm is used to train the network, the learning rate is set to 0.001, the number of training rounds (epochs) is set to 100, and the number of training iterations for each small graph is set to 200. The loss function uses the cross entropy loss function, which is a loss function commonly used in classification problems and can effectively measure the difference between the model's predicted probability and the actual label.

[0211] Furthermore, in this embodiment, the graph learning model with node classification as the main task can learn the probability graph in the form of a weighted adjacency matrix to obtain the best classification result. The integrated learning framework of the proposed PD detection model is as follows Fig. 9 The above method is used to extract the features of each brain region of the three modalities of T1, DTI, and fMRI, and then they are put into the PD detection model respectively, and then the ensemble learning method is used to obtain the final detection result.

[0212] In addition, the model uses the Adam optimizer and cross entropy loss function to optimize the network, the learning rate is set to 0.001, the batch size is set to 4, and the number of iterations is 100. The number of GCN layers used is 2, and the size of q is set to 10. The weights of the DTI, T1, and fMRI models are set to W1 = 0.2, W2 = 0.3, and W3 = 0.5. Using five-fold cross validation, the following indicators are used to evaluate the model performance: AUC, accuracy, sensitivity, and specificity.

[0213] like Figure 2 As shown, an embodiment of the present invention provides a method for detecting and processing Parkinson's disease data based on a graph neural network, comprising the following steps:

[0214] Step S400, predicting the input data to be tested based on the trained graph neural network, outputting a prediction label of Parkinson's disease, and performing a Parkinson's disease warning.

[0215] In this embodiment, after deploying the trained graph neural network model, the test data (T1 modality images, DTI modality images, fMRI modality images) input in the real-time scenario are predicted based on the trained graph neural network, and a graph neural network interpreter is used to explore Parkinson's disease markers, output the prediction label of Parkinson's disease, and perform Parkinson's disease early warning.

[0216] In this embodiment, a novel deep learning architecture is proposed, which uses a graph neural network to fully extract the medical imaging features of Parkinson's disease, and can accurately detect PD at an early stage. In addition, in this embodiment, significant subgraphs and significant features (brain regions) related to Parkinson's disease can be extracted to provide assistance for Parkinson's disease research.

[0217] It is worth mentioning that a Parkinson's disease marker research method based on functional magnetic resonance imaging is proposed in this embodiment. The method is based on a graph neural network and a graph neural network interpreter, which can effectively distinguish Parkinson's disease and healthy controls, and its classification accuracy reaches 94.36%, which is better than the traditional algorithm. In the marker exploration experiment, significant subgraphs and significant features (brain regions) related to Parkinson's disease were extracted, and the results were similar to those found in previous literature, which proved the effectiveness of the method and can provide help for the study of Parkinson's disease.

[0218] 1. The problem of small samples of medical data is solved. Deep learning methods often require a large number of data samples to support model training. The model proposed in this embodiment is suitable for small samples and can also solve the problem of imbalance between positive and negative samples. It is a robust model.

[0219] 2. Reduced the manpower cost of doctors. Doctors only need to collect the patient's rs-fMRI (resting state functional magnetic resonance imaging) data, first pass it through the software to pre-process the data, and then input it into the trained graph convolutional neural network, which can quickly predict PD patients and assist doctors in detection;

[0220] 3. Provides interpretability for Parkinson's disease detection. The model selects the n brain regions corresponding to the top n most important features for Parkinson's disease detection to achieve interpretability of brain regions.

[0221] This embodiment achieves the following technical effects through the above technical solution:

[0222] Based on the functional magnetic resonance imaging data, this embodiment uses a graph neural network to fully extract the significant subgraphs related to Parkinson's disease and the significant features of the brain region, thereby realizing auxiliary detection of Parkinson's disease. In addition, after obtaining the auxiliary detection results, the graph neural network interpretation technology is used to classify Parkinson's disease markers, which can effectively distinguish Parkinson's disease from the healthy control group, providing interpretability for Parkinson's disease early warning. This embodiment uses a graph neural network to fully extract the medical imaging features of Parkinson's disease, improves the accuracy of early warning of Parkinson's disease, and reduces the cost of early warning of Parkinson's disease.

[0223] Based on the above embodiment, the present invention further provides a terminal, whose principle block diagram can be as follows: Fig.10 shown.

[0224] The terminal includes: a processor, a memory, an interface, a display screen and a communication module connected through a system bus; wherein the processor of the terminal is used to provide computing and control capabilities; the memory of the terminal includes a storage medium and an internal memory; the storage medium stores an operating system and a computer program; the internal memory provides an environment for the operation of the operating system and the computer program in the storage medium; the interface is used to connect to external devices; the display screen is used to display corresponding information; and the communication module is used to communicate with a cloud server or other devices.

[0225] When the computer program is executed by a processor, it is used to implement the operation of the Parkinson's disease data detection and processing method based on graph neural network.

[0226] It can be understood by those skilled in the art that Fig.10 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the terminal to which the scheme of the present invention is applied. The specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0227] In one embodiment, a terminal is provided, which includes: a processor and a memory, wherein the memory stores a Parkinson's disease data detection and processing program based on a graph neural network, and the Parkinson's disease data detection and processing program based on a graph neural network is used to implement the operations of the above-mentioned Parkinson's disease data detection and processing method based on a graph neural network when the processor executes the Parkinson's disease data detection and processing program based on a graph neural network.

[0228] In one embodiment, a storage medium is provided, wherein the storage medium stores a Parkinson's disease data detection and processing program based on a graph neural network, and when the Parkinson's disease data detection and processing program based on a graph neural network is executed by a processor, it is used to implement the operations of the above-mentioned Parkinson's disease data detection and processing method based on a graph neural network.

[0229] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and volatile memory.

[0230] In summary, the present invention provides an early warning system and method for Parkinson's disease based on a graph neural network, including: a data preprocessing module for acquiring resting state functional magnetic resonance imaging data, and preprocessing the acquired data to construct graph structure data; a feature extraction module for extracting features from the graph structure data, and removing redundant features or irrelevant features from the extracted features to obtain a feature subset associated with Parkinson's disease; a model training module for constructing a graph neural network, and training the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network; an early warning module for predicting the input test data based on the trained graph neural network, outputting a prediction label for Parkinson's disease, and performing a Parkinson's disease early warning. The present invention uses a graph neural network to fully extract the medical imaging features of Parkinson's disease, and can accurately provide early warning for Parkinson's disease.

[0231] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. An early warning system for Parkinson's disease based on graph neural network, characterized in that: include: A data preprocessing module is used to obtain resting state functional magnetic resonance imaging data, and preprocess the obtained data to construct graph structure data; A feature extraction module, used to extract features from the graph structure data and remove redundant features or irrelevant features from the extracted features to obtain a feature subset associated with Parkinson's disease; A model training module, used to construct a graph neural network, and train the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network; The early warning module is used to predict the input test data based on the trained graph neural network, output the prediction label of Parkinson's disease, and issue a Parkinson's disease early warning.

2. The Parkinson's disease early warning system based on graph neural network according to claim 1 is characterized in that: The data preprocessing module comprises: A data acquisition unit, configured to acquire the resting state functional magnetic resonance imaging data from a Parkinson's disease progression marker program database; wherein the resting state functional magnetic resonance imaging data includes: T1 modal images, diffusion tensor modal images, and functional magnetic resonance modal images corresponding to Parkinson's disease patients and healthy control patients, respectively; A T1 image preprocessing unit, used to perform tissue segmentation on the T1 modal image using a probability map, align the segmented tissues according to template pairing and standardization operations, and divide the volumes corresponding to gray matter, white matter, and cerebrospinal fluid into a plurality of regions using a preset template; A diffusion tensor image preprocessing unit, used to resample the diffusion tensor modal image according to a preset resampling resolution, remove scalp and blank space from the resampled image, and construct and extract the structural connection area of ​​the brain network using a deterministic network according to preset diffusion matrix parameters and fiber tracking and network parameters; The functional magnetic resonance image preprocessing unit is used to remove time points, perform layer time correction and head motion correction on the functional magnetic resonance modal image, and perform image matching, image denoising and image standardization on the corrected image, and extract the functional connection area according to the determined brain area of ​​interest.

3. The Parkinson's disease early warning system based on graph neural network according to claim 1 is characterized in that: The feature extraction module comprises: A Fisher score calculation unit, used to calculate the score of each feature in the graph structure data based on the Fisher criterion, and select the features with the top m scores; A mutual information selection unit, used to analyze the relevance of the selected features in predicting the target variable and the redundancy relative to other variables based on the mutual information selection method, and remove the redundant features or the irrelevant features; The regularization and recursive selection unit is used to perform feature selection on the remaining features based on the L1 regularization method and the recursive feature elimination method to obtain the feature subset associated with Parkinson's disease.

4. The Parkinson's disease early warning system based on graph neural network according to claim 1 is characterized in that: The model training module includes: A graph neural network construction unit is used to reduce the dimension of the feature matrix through a parameterized function to obtain a new feature matrix after dimension reduction, and calculate a probability map; use the probability map to construct an adjacency matrix through a k-nearest neighbor algorithm; perform graph convolution, and obtain a final output matrix through multi-layer graph convolution operations; normalize the output matrix through a softmax function to obtain a label prediction Y for each node; The graph neural network training unit is used to create a data graph of a preset size by random sampling; wherein the data graph of the preset size includes an unknown sample a and n known samples; the constructed model is trained according to the data graph of the preset size, and the unknown sample a is predicted to obtain a predicted label y; the predicted label y is compared with the actual label of the unknown sample a A comparison is performed, and based on the error generated by the comparison, the network parameters of the model are adjusted and updated using a back propagation algorithm; the model is optimized using a five-fold cross-validation method and an optimization algorithm to obtain the trained graph neural network.

5. A method for detecting and processing Parkinson's disease data in a Parkinson's disease early warning system based on a graph neural network as claimed in any one of claims 1 to 4, characterized in that: include: Acquire resting state functional magnetic resonance imaging data, preprocess the acquired data, and construct graph structure data; Extracting features from the graph structure data, and removing redundant features or irrelevant features from the extracted features to obtain a feature subset associated with Parkinson's disease; Constructing a graph neural network, and training the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network; Based on the trained graph neural network, the input data to be tested is predicted, a prediction label of Parkinson's disease is output, and an early warning of Parkinson's disease is performed.

6. The method for detecting and processing Parkinson's disease data based on graph neural network according to claim 5, characterized in that: The step of acquiring resting state functional magnetic resonance imaging data and preprocessing the acquired data to construct graph structure data includes: Acquire the resting state functional magnetic resonance imaging data from the Parkinson's disease progression marker program database; wherein the resting state functional magnetic resonance imaging data includes: T1 modal images, diffusion tensor modal images, and functional magnetic resonance modal images corresponding to Parkinson's disease patients and healthy control patients, respectively; The T1 modality image, the diffusion tensor modality image and the functional magnetic resonance modality image are preprocessed respectively to construct corresponding graph structure data.

7. The method for Parkinson's disease data detection and processing based on graph neural network according to claim 6 is characterized in that: The preprocessing of the T1 modality image, the diffusion tensor modality image and the functional magnetic resonance modality image respectively to construct graph structure data includes: Performing tissue segmentation on the T1 modality image using the probability map, aligning the segmented tissues according to template pairing and standardization operations, and dividing the volumes corresponding to the gray matter, white matter, and cerebrospinal fluid into a number of regions using a preset template; Resampling the diffusion tensor modality image according to a preset resampling resolution, removing the scalp and blanking from the resampled image, and constructing and extracting the structural connection area of ​​the brain network using a deterministic network according to preset diffusion matrix parameters and fiber tracking and network parameters; The functional magnetic resonance image is processed by removing time points, correcting layer time and correcting head motion, and the corrected image is processed by image matching, image denoising and image standardization, and the functional connection area is extracted according to the determined brain area of ​​interest.

8. The method for Parkinson's disease data detection and processing based on graph neural network according to claim 7, characterized in that: The functional magnetic resonance modality image is subjected to time point removal, slice time correction and head motion correction processing, and the corrected image is subjected to image matching, image denoising and image standardization processing, and the functional connection area is extracted according to the determined brain area of ​​interest, including: Deleting a preset number of initial time points in the functional magnetic resonance modality image, adjusting the data of all layers to the same time point using a layer time correction method, and performing head motion correction using the head motion correction algorithm; The corrected image is processed to remove the neck and scalp, and the T1 modality image is matched to the functional image of the removed image, multiple noise covariates are excluded, and the image is normalized according to a preset bounding box and voxel size; The brain region of interest is determined, correlation analysis is performed between the brain region of interest and other brain regions, a functional connectivity map is generated, and the functional connectivity region is extracted based on the functional connectivity map.

9. The method for Parkinson's disease data detection and processing based on graph neural network according to claim 5, characterized in that: The feature extraction of the graph structure data and removal of redundant features or irrelevant features in the extracted features to obtain a feature subset associated with Parkinson's disease include: Calculate the score of each feature in the graph structure data based on the Fisher criterion, and select the features with the top m scores; Analyze the relevance of the selected features in predicting the target variable and their redundancy relative to other variables based on the mutual information selection method, and remove the redundant features or the irrelevant features; Feature selection is performed on the remaining features based on an L1 regularization method and a recursive feature elimination method to obtain the feature subset associated with Parkinson's disease.

10. The method for Parkinson's disease data detection and processing based on graph neural network according to claim 5, characterized in that: The constructing of the graph neural network and training the constructed model using samples randomly selected from the feature subset to obtain a trained graph neural network includes: The feature matrix X used to construct the small graph is taken as input and passed through the parameterized function f θ Perform dimensionality reduction processing on the feature matrix to obtain a new dimensionality-reduced feature matrix And the probability map is calculated according to the following formula: Among them, t represents the learnable parameters in the graph network, Δ represents the Euclidean distance, and x i represents the i-th sample; Using the probability graph P ij (X; θ, t), construct the adjacency matrix A through the k-nearest neighbor algorithm; Perform graph convolution: in, represents the feature matrix of the lth layer, W (l) represents the learnable weight matrix corresponding to the lth layer, represents a diagonal matrix, σ represents a nonlinear activation function; Through multi-layer graph convolution operations, the final output matrix H is obtained (l) ; The output matrix H is transformed into (l) Perform normalization to obtain the label prediction Y of each node: Among them, h ij and ij denote the elements in the i-th row and j-th column of matrices H and Y respectively; Creating a data graph of a preset size by random sampling; wherein the data graph of the preset size includes an unknown sample a and n known samples; The constructed model is trained according to the data graph of the preset scale, and the unknown sample a is predicted to obtain a predicted label y; The predicted label y is compared with the actual label of the unknown sample a. Make a comparison, and adjust and update the network parameters of the model using the back propagation algorithm based on the error generated by the comparison; The model is optimized using a five-fold cross-validation method and an optimization algorithm to obtain the trained graph neural network.

Images