An alzheimer's disease classification method and system based on hypergraph neural network

Abstract

The application provides an Alzheimer's disease classification method and system based on a hypergraph neural network, and the method comprises the following steps: acquiring fMRI image data of a brain of an Alzheimer's disease patient and performing preprocessing; using an RBF-SVR recursive feature elimination algorithm to construct a hypergraph for each fMRI image; adding noise to the preprocessed fMRI image data and fMRI hypergraph data and integrating the fMRI image data and the fMRI hypergraph data into training metadata; inputting the training metadata into an established Alzheimer's disease classification model to perform training; using the trained Alzheimer's disease classification model to classify fMRI image data of a brain of a patient to be diagnosed; and obtaining an Alzheimer's disease classification result; the RBF-SVR recursive feature elimination algorithm is used to nonlinearly construct a hypergraph for each fMRI image, and a hypergraph neural network model is combined, so that the performance and precision of early diagnosis of AD can be effectively improved, and the classification performance is better.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and neuroimaging technology, and more specifically, to a method and system for classifying Alzheimer's disease based on hypergraph neural networks. Background Technology

[0002] Alzheimer's disease (AD) is an irreversible neurodegenerative disease that impairs memory and cognitive abilities, with its prevalence increasing significantly with age. Early diagnosis of AD is crucial for drug development, prevention, and slowing disease progression.

[0003] With the development of neuroimaging technology, more and more neuroimaging techniques are being widely used in the early diagnosis of Alzheimer's disease (AD). Among them, resting-state functional magnetic resonance imaging (rs-fMRI) is a four-dimensional medical imaging technique with advantages such as high resolution and no radiation, effectively reflecting structural and functional changes in the brain, and has achieved good results in the early diagnosis of AD. At the same time, fMRI images, due to their high temporal and spatial resolution, have become an important tool for studying human brain function and are also widely used in research on human brain connectomics.

[0004] Traditional methods for processing fMRI data involve constructing brain functional connectivity networks, storing the data's features in a conventional graph structure represented by adjacency matrices. When constructing these networks, the time series of the average BOLD signal of each Region of Interest (ROI) is extracted, and then the Pearson correlation coefficients between different ROIs are calculated. Using the ROIs as vertices of the adjacency matrix and the Pearson correlation coefficients between different ROIs as edges, a conventional graph storing fMRI data features can be constructed. However, the interactions between different brain ROIs are often non-linear, involving multiple regions. A conventional graph structure that only represents binary correlations is insufficient to record such multivariate, higher-order interactions.

[0005] A hypergraph is a generalized graph, the most general discrete structure in a finite set, and has wide applications in information science, life sciences, and other fields. A single edge can connect any number of vertices. Hypergraph structures can effectively record the multivariate higher-order correlations between different regions of interest (ROIs) in fMRI data.

[0006] Existing technology discloses an electronic device based on a hypergraph structure information aggregation model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following: constructing a brain region gene hypergraph of the subject based on brain imaging data and gene data; converting the brain region gene hypergraph into a brain region gene edge graph and a point graph using the hypergraph structure information aggregation model; fusing the edge graph and point graph after edge-level feature information aggregation and point-level feature information aggregation respectively to obtain a low-order brain region gene sub-hypergraph; and performing high-order edge-level and point-level structural information aggregation on the brain region gene hypergraph to obtain... The method involves constructing a high-order brain region gene supermap; based on the key brain region gene supermap obtained by fusing the low-order and high-order brain region gene supermaps, the probability of the subject having Alzheimer's disease (AD) is output. Although this existing technology can use supermaps to assist in AD diagnosis, the supermap construction method in this method is a linear method. However, the interactions and correlations between various brain regions are difficult to record quantitatively and qualitatively using nonlinear methods. Simply using ordinary linear mapping methods cannot record the nonlinear interactions and correlations between various ROIs in fMRI images. In addition, traditional supermap construction methods are generally binary methods, which can only record the correlation between two brain regions. However, the activity between various brain regions is often a synergistic effect of multiple brain regions. Therefore, the accuracy of traditional early AD diagnosis methods is relatively low. Summary of the Invention

[0007] To overcome the shortcomings of low accuracy in the existing early diagnosis methods for Alzheimer's disease, this invention provides an Alzheimer's disease classification method and system based on a hypergraph neural network. The generated hypergraph effectively records the nonlinear multivariate higher-order correlations between multiple ROIs, resulting in higher performance, accuracy, and better effectiveness in early diagnosis of Alzheimer's disease.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0009] This invention provides an Alzheimer's disease classification method based on a hypergraph neural network, comprising the following steps:

[0010] S1: Acquire fMRI image data of the brain of Alzheimer's patients and perform preprocessing;

[0011] S2: Construct a hypermap for each fMRI image in the preprocessed fMRI image data and obtain fMRI hypermap data;

[0012] S3: Add noise to the preprocessed fMRI image data and fMRI hypermap data, and integrate them into training metadata;

[0013] S4: Use the training metadata to train the established Alzheimer's disease classification model to obtain a trained Alzheimer's disease classification model;

[0014] S5: Obtain fMRI image data of the patient's brain to be diagnosed and input it into the trained Alzheimer's disease classification model for classification, and obtain the Alzheimer's disease classification result.

[0015] Preferably, the specific preprocessing methods in step S1 include: image format conversion, removal of initial time points, time slice correction, head motion correction, spatial standardization, delinear drift, covariate regression, low-frequency filtering, and spatial smoothing.

[0016] Preferably, in step S2, the specific method for constructing a hypermap for each fMRI image in the preprocessed fMRI image data and obtaining the fMRI hypermap data is as follows:

[0017] The RBF-SVR recursive feature elimination algorithm is used to construct a hypermap for each fMRI image in the preprocessed fMRI image data to obtain fMRI hypermap data.

[0018] In the RBF-SVR recursive feature elimination algorithm, the nonlinear Gaussian kernel SVR is used as the estimator for the recursive feature elimination algorithm.

[0019] Each fMRI image in the preprocessed fMRI image data is divided into several ROIs, and the time series of the average BOLD signal of all ROIs in each fMRI image is extracted.

[0020] For each fMRI image, a ROI is selected as the target ROI. Based on the time series of the average BOLD signal of all ROIs, the RBF-SVR recursive feature elimination algorithm is used to calculate the fit between the target ROI and other ROIs. A hyperedge is constructed using the target ROI and several other ROIs with the highest fit.

[0021] Repeat the above operations, construct hyperedges by treating each ROI in each fMRI image as a target ROI, construct the corresponding fMRI hypermap by using all hyperedges in each fMRI image, and save all fMRI hypermaps as fMRI hypermap data.

[0022] Preferably, each fMRI image includes 116 ROIs and 130 time series.

[0023] Preferably, the specific method for calculating the fit between the target ROI and other ROIs using the RBF-SVR recursive feature elimination algorithm is the five-fold cross-validation method.

[0024] Preferably, the Alzheimer's disease classification model established in step S4 is specifically the FHyperGNN hypergraph convolutional neural network model.

[0025] Preferably, the FHyperGNN hypergraph convolutional neural network model includes an input layer, several HyperGCN hypergraph convolutional modules, a first BatchNorm regularization layer, a Dropout layer, a Linear layer, a Softmax layer, and an output layer connected in sequence.

[0026] Each HyperGCN hypergraph convolutional module includes a second BatchNorm regularization layer, a HyperGNN hypergraph convolutional layer, and an activation layer connected in sequence.

[0027] The activation function of the activation layer is the ReLU function.

[0028] Preferably, the number of HyperGCN hypergraph convolutional modules is at least three.

[0029] Preferably, the Dropout rate of the Dropout layer is 0.5.

[0030] This invention also provides an Alzheimer's disease classification system based on a hypergraph neural network, which applies the above-mentioned Alzheimer's disease classification method based on a hypergraph neural network, including:

[0031] Data acquisition module: used to acquire fMRI image data of the brains of Alzheimer's patients and perform preprocessing;

[0032] Hypergraph Construction Module: Used to construct a hypergraph for each fMRI image in the preprocessed fMRI image data and obtain fMRI hypergraph data;

[0033] Noise module: Used to add noise to preprocessed fMRI image data and fMRI hypermap data, and integrate them into training metadata;

[0034] Training module: Used to train the established Alzheimer's disease classification model using training metadata, and obtain a trained Alzheimer's disease classification model;

[0035] Classification module: Used to acquire fMRI image data of the brain of the patient to be diagnosed and input it into the trained Alzheimer's disease classification model for classification, and obtain the Alzheimer's disease classification result.

[0036] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0037] This invention provides a method and system for classifying Alzheimer's disease based on a hypergraph neural network. The method includes the following steps: acquiring fMRI image data of the brain of an Alzheimer's patient and preprocessing it; constructing a hypergraph for each fMRI image in the preprocessed fMRI image data to obtain fMRI hypergraph data; adding noise to the preprocessed fMRI image data and fMRI hypergraph data and integrating them into training metadata; training the established Alzheimer's disease classification model using the training metadata to obtain a trained Alzheimer's disease classification model; acquiring fMRI image data of the brain of a patient to be diagnosed and inputting it into the trained Alzheimer's disease classification model for classification to obtain the Alzheimer's disease classification result.

[0038] This invention uses a nonlinear Gaussian kernel SVR as the estimator for a recursive feature elimination algorithm, and employs the RBF-SVR recursive feature elimination algorithm to nonlinearly construct a hypermap for each fMRI image. This hypermap data structure is used to record the nonlinear multivariate higher-order correlations between multiple ROIs in the fMRI image data, which can demonstrate good classification performance. In addition, the Alzheimer's disease classification model constructed in this invention has higher performance and accuracy in the early diagnosis of AD and better results. Attached Figure Description

[0039] Figure 1 This is a flowchart of an Alzheimer's disease classification method based on a hypergraph neural network, as provided in Example 1.

[0040] Figure 2 This is a flowchart of an Alzheimer's disease classification method based on a hypergraph neural network, as provided in Example 2.

[0041] Figure 3 This is a flowchart of the fMRI hypermap data construction process provided in Example 2.

[0042] Figure 4 The diagram shows the structure of the FHyperGNN hypergraph convolutional neural network model provided in Example 2.

[0043] Figure 5 This is a structural diagram of an Alzheimer's disease classification system based on a hypergraph neural network, provided in Example 3. Detailed Implementation

[0044] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0045] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions;

[0046] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0047] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] Example 1

[0049] like Figure 1 As shown, this embodiment provides an Alzheimer's disease classification method based on a hypergraph neural network, including the following steps:

[0050] S1: Acquire fMRI image data of the brain of Alzheimer's patients and perform preprocessing;

[0051] S2: Construct a hypermap for each fMRI image in the preprocessed fMRI image data and obtain fMRI hypermap data;

[0052] S3: Add noise to the preprocessed fMRI image data and fMRI hypermap data, and integrate them into training metadata;

[0053] S4: Use the training metadata to train the established Alzheimer's disease classification model to obtain a trained Alzheimer's disease classification model;

[0054] S5: Obtain fMRI image data of the patient's brain to be diagnosed and input it into the trained Alzheimer's disease classification model for classification, and obtain the Alzheimer's disease classification result.

[0055] In the specific implementation process, firstly, fMRI image data of the brains of Alzheimer's patients are acquired and preprocessed; a hypermap is constructed for each fMRI image in the preprocessed fMRI image data to obtain fMRI hypermap data; noise is added to the preprocessed fMRI image data and fMRI hypermap data, and integrated into training metadata; the established Alzheimer's disease classification model is trained using the training metadata to obtain a trained Alzheimer's disease classification model; finally, fMRI image data of the brains of patients to be diagnosed are acquired and input into the trained Alzheimer's disease classification model for classification to obtain Alzheimer's disease classification results.

[0056] This method uses a nonlinear Gaussian kernel SVR as the estimator for the recursive feature elimination algorithm and employs the RBF-SVR recursive feature elimination algorithm to nonlinearly construct a hypermap for each fMRI image. This hypermap data structure is used to record the nonlinear multivariate higher-order correlations between multiple ROIs in the fMRI image data, which can demonstrate good classification performance. In addition, the Alzheimer's disease classification model constructed by this method has higher performance and accuracy in the early diagnosis of AD and better results.

[0057] Example 2

[0058] like Figure 2 As shown, this embodiment provides an Alzheimer's disease classification method based on a hypergraph neural network, including the following steps:

[0059] S1: Acquire fMRI image data of the brain of Alzheimer's patients and perform preprocessing;

[0060] S2: Construct a hypermap for each fMRI image in the preprocessed fMRI image data and obtain fMRI hypermap data;

[0061] S3: Add noise to the preprocessed fMRI image data and fMRI hypermap data, and integrate them into training metadata;

[0062] S4: Use the training metadata to train the established Alzheimer's disease classification model to obtain a trained Alzheimer's disease classification model;

[0063] S5: Obtain fMRI image data of the brain of the patient to be diagnosed and input it into the trained Alzheimer's disease classification model for classification, and obtain the Alzheimer's disease classification result;

[0064] The specific preprocessing methods in step S1 include: image format conversion, removal of initial time points, time slice correction, head motion correction, spatial normalization, delinear drift, covariate regression, low-frequency filtering, and spatial smoothing.

[0065] In step S2, a hypermap is constructed for each fMRI image in the preprocessed fMRI image data. The specific method for obtaining the fMRI hypermap data is as follows:

[0066] like Figure 2 As shown, the RBF-SVR recursive feature elimination algorithm is used to construct a hypermap for each fMRI image in the preprocessed fMRI image data to obtain fMRI hypermap data;

[0067] In the RBF-SVR recursive feature elimination algorithm, the nonlinear Gaussian kernel SVR is used as the estimator for the recursive feature elimination algorithm.

[0068] Each fMRI image in the preprocessed fMRI image data is divided into several ROIs, and the time series of the average BOLD signal of all ROIs in each fMRI image is extracted.

[0069] For each fMRI image, a ROI is selected as the target ROI. Based on the time series of the average BOLD signal of all ROIs, the RBF-SVR recursive feature elimination algorithm is used to calculate the fit between the target ROI and other ROIs. A hyperedge is constructed using the target ROI and several other ROIs with the highest fit.

[0070] Repeat the above operations, construct hyperedges by treating each ROI in each fMRI image as a target ROI, construct the corresponding fMRI hypermap by using all the hyperedges in each fMRI image, and save all the fMRI hypermaps as fMRI hypermap data.

[0071] Each of the fMRI images includes 116 ROIs and 130 time series;

[0072] The specific method for calculating the fit between the target ROI and other ROIs using the RBF-SVR recursive feature elimination algorithm is the five-fold cross-validation method.

[0073] The Alzheimer's disease classification model established in step S4 is specifically the FHyperGNN hypergraph convolutional neural network model;

[0074] like Figure 4 As shown, the FHyperGNN hypergraph convolutional neural network model includes an input layer, several HyperGCN hypergraph convolutional modules, a first BatchNorm regularization layer, a Dropout layer, a Linear layer, a Softmax layer, and an output layer connected in sequence.

[0075] Each HyperGCN hypergraph convolutional module includes a second BatchNorm regularization layer, a HyperGNN hypergraph convolutional layer, and an activation layer connected in sequence.

[0076] The HyperGCN hypergraph convolutional module has at least 3 modules;

[0077] The Dropout rate of the Dropout layer is 0.5;

[0078] The activation function of the activation layer is the ReLU function.

[0079] In the specific implementation process, fMRI image data of the brains of Alzheimer's patients are first obtained from the ADNI database and preprocessed;

[0080] During image acquisition, factors such as patient head movement, respiration, and heartbeat can generate noise, leading to deterioration in image quality. Therefore, preprocessing is necessary during data analysis to reduce the impact of irrelevant noise and improve the signal-to-noise ratio. This embodiment uses the DPARSF (Data Processing Assistant for Resting-State fMRI) toolbox and the spm12 software package for preprocessing. The preprocessing of fMRI image data includes the following steps:

[0081] 1) Image format conversion (DICOM to NIFTI): The fMRI image data in the ADNI database is in DICOM image format, while the package used by Python to read the data is nibabel, which reads data in NIFTI format. The data format processed in subsequent data preprocessing is also NIFTI. Therefore, the format of the fMRI image data needs to be converted from DICOM to NIFTI.

[0082] 2) Remove First Images: When fMRI images are first collected, many factors, such as the patient not yet being at rest, can cause data instability and affect the quality of the images. Therefore, it is necessary to remove the first few time points of the images. In this embodiment, the first 10 volumes are removed, and 130 volumes are retained in the end.

[0083] 3) Slice Timing: When collecting images at each time point, the instrument scans layer by layer from top to bottom to obtain two-dimensional images of different slices. Then, all the two-dimensional images of the slices are reconstructed into a three-dimensional image. Therefore, it is inevitable that the images of each layer will be out of sync when they are acquired. Slice Timing is to correct the images of different layers acquired at different times to the same time.

[0084] 4) Head movement correction (Realign): Ideally, the patient should be stationary during image acquisition. However, due to the long acquisition time, the patient's head may rotate, which can cause deviations in image positioning and affect image quality. Therefore, head movement correction is required. Through head movement correction, the same part of the brain in all fMRI images can be placed in the same position in the image.

[0085] 5) Spatial Normalization: Since the shape and size of the brain vary to varying degrees among individuals, brain images from different patients are registered to the Montreal Neurological Institute (MNI) space to eliminate errors caused by brain morphology and reduce the influence of irrelevant factors. In this preprocessing operation, a standard EPI (Eco-Planar Imaging) brain template is used, with each voxel (a point in a two-dimensional image is called a pixel, while a point in three-dimensional space is called a voxel) having a volume of 3×3×3 mm. 3 ;

[0086] 6) Temporally Detrend: fMRI images may show a regular increase or decrease in signal over time due to long-term physiological changes or instability of the scanning instrument. This preprocessing can remove the effects of such regular signal drift.

[0087] 7) Regression Out Covariates: Several interfering signals, such as head movement, cerebrospinal fluid signals, and white matter signals, are usually removed from fMRI images to reduce the influence of non-neuronal fluctuations. These signals will be regressed.

[0088] 8) Temporally Filtered Low-Frequency Filter: fMRI image data undergoes bandpass filtering to remove signals of a certain frequency, reducing the impact of low-frequency drift and high-frequency physiological noise. The filtered frequency is 0.01-0.08 Hz.

[0089] 9) Spatial Smoothing: After smoothing the fMRI image data, the value of each voxel is calculated by weighting the values ​​of adjacent voxels before smoothing. In this embodiment, a 4×4×4mm voxel will be used. 3 Gaussian kernels are used to spatially smooth all volumes in fMRI image data, eliminating artifacts and improving the signal-to-noise ratio.

[0090] When constructing a brain functional connectivity network, the time series of the average BOLD signal of the Region of Interest (ROI) is first extracted, and then the Pearson correlation coefficient between the time series of different ROIs is calculated as the functional connectivity strength between ROIs. Traditional correlation coefficients include Pearson correlation coefficient, Kendall correlation coefficient, Spearman correlation coefficient, etc., but the above correlation coefficients only describe the correlation between two ROIs and cannot record the nonlinear interaction between multiple regions of each ROI. Therefore, this embodiment uses the RBF-SVR recursive feature elimination algorithm to construct a hypermap for each fMRI image in the preprocessed fMRI image data to obtain fMRI hypermap data.

[0091] First, each fMRI image in the preprocessed fMRI image data is divided into 116 ROIs, and the time series of the average BOLD signal of all ROIs in each fMRI image is extracted; each fMRI image has 116 ROIs, and there are a total of 130 time series; the time series of the average BOLD signal of these 116 ROIs is extracted, and each ROI corresponds to a 130×1 one-dimensional matrix;

[0092] For each fMRI image, select one ROI as the target ROI, set the time series of the average BOLD signal of the target ROI as the label Y, and concatenate the time series of the average BOLD signals of the other 115 ROIs into a 130×115 matrix X.

[0093] The RBF-SVR recursive feature elimination algorithm is used to calculate the fit between the target ROI and other ROIs. A hyperedge is constructed using the target ROI and several other ROIs with the highest fit.

[0094] The RBF-SVR recursive feature elimination algorithm is as follows:

[0095] initialization:

[0096] 1) X = [X1, X2, X3, ..., Xk, ..., X115]: 115 other ROI data used to fit the target ROI;

[0097] 2) Y = [X0]: Regression label, X0 is the target ROI, and this algorithm constructs hyperedges based on X0;

[0098] 3) start_ROI

[115] = {1,2,3,...,114,115}: The initial ROI array, which stores the ROI number currently selected;

[0099] 4) best_MSE: The MSE value of the optimal RBF-SVR model;

[0100] 5) selected_best_ROI[K]: The set of ROI numbers selected by the optimal RBF-SVR model. The array stores the numbers of the ROIs selected in the best case (1-115). The value of K is the length of the array, that is, the number of ROIs selected.

[0101] 6) rank[K] = {0}: weighted sorting array, where K is the array length, i.e., the number of selected ROIs;

[0102] 7) The rbf_svr_multi_fold_cross_validation(X,Y) function: defines a Gaussian kernel SVR as a nonlinear regression model, and uses the MSE evaluation index as the goodness of fit of the SVR model trained with the current data X to the label Y; this function returns the MSE value obtained after performing five-fold cross-validation on the RBF-SVR model with the current data X and label Y.

[0103] pseudocode:

[0104]

[0105]

[0106] By running the above algorithm once for each of the 116 ROIs, 116 hyperedges can be obtained, which can form a hypergraph based on the current fMRI data.

[0107] Repeat the above operations, construct hyperedges by treating each ROI in each fMRI image as a target ROI, construct the corresponding fMRI hypermap by using all the hyperedges in each fMRI image, and save all the fMRI hypermaps as fMRI hypermap data.

[0108] An Alzheimer's disease classification model is established, wherein the Alzheimer's disease classification model is a hypergraph convolutional neural network model;

[0109] The commonly used HyperGCN framework uses the K-nearest neighbor algorithm to linearly construct a hypergraph and constructs a single hypergraph for all samples in the dataset to find the correlation between samples, thereby achieving excellent classification results. However, in complex and large-scale four-dimensional fMRI data, constructing a single hypergraph for all samples and classifying based solely on the correlation between samples may ignore the complex signal features between different brain regions within a single fMRI dataset. Experiments show that this method of constructing hypergraphs performs poorly on fMRI data.

[0110] Therefore, this method constructs a hypergraph on the basis of a single fMRI data sample, and uses the interaction relationships between various brain regions in the fMRI data to form the hyperedge of the hypergraph, hoping to find the interaction relationships between various brain regions and thus achieve better classification results.

[0111] In this embodiment, the established Alzheimer's disease classification model is specifically the FHyperGNN hypergraph convolutional neural network model;

[0112] The FHyperGNN hypergraph convolutional neural network model includes an input layer, three HyperGCN hypergraph convolutional modules, a first BatchNorm regularization layer, a Dropout layer, a Linear layer, a Softmax layer, and an output layer connected in sequence.

[0113] Each HyperGCN hypergraph convolutional module includes a second BatchNorm regularization layer, a HyperGNN hypergraph convolutional layer, and an activation layer connected in sequence.

[0114] The Dropout rate of the Dropout layer is 0.5;

[0115] The activation function of the activation layer is the ReLU function;

[0116] The FHyperGNN hypergraph convolutional neural network model in this method has the following advantages:

[0117] Unlike mainstream hypergraph algorithms that construct a hypergraph from all samples and classify samples based on their correlations, this invention constructs a hypergraph from a single fMRI sample and identifies the nonlinear multivariate correlations between brain regions within the sample as the hyperedge set of the hypergraph. Classification is then performed based on these nonlinear multivariate correlations between brain regions within the sample.

[0118] By adding a regularized BatchNorm layer and a ReLU activation function layer before and after the traditional HyperGNN-Conv hypergraph convolutional layer, and encapsulating them into a HyperGNN-block, the three HyperGNN-blocks constitute the backbone learning network of the model. Using BatchNorm regularization before the hypergraph convolutional layer further eliminates complex noise interference in fMRI data, accelerates model convergence, effectively avoids gradient vanishing, and thus improves the model's generalization ability. Using the ReLU activation function after the hypergraph convolutional layer introduces nonlinearity into the model, allowing the neural network to approximate any nonlinear function, thereby effectively learning nonlinear high-order features in fMRI data. A Dropout layer with a Dropout rate of 0.5 is added after the three HyperGNN-block layers to prevent overfitting during the learning process of fMRI data features, thus effectively improving the model's generalization ability.

[0119] Noise was added to the preprocessed fMRI image data and fMRI hypermap data and then integrated into training metadata. The training metadata was then input into the Alzheimer's disease classification model for training, and the trained Alzheimer's disease classification model was obtained.

[0120] Deep learning models typically require a large amount of training data to achieve good results. However, for fMRI images, it is often difficult to obtain a large number of fMRI image samples for model training. Therefore, this method augments the dataset by adding uniform random noise to the fMRI data before inputting it into the model. This enhances the randomness of the data, allowing the model to better learn the data distribution during training and make better predictions on new data. This prevents the model from overfitting the training data, thereby improving the model's generalization performance. Since fMRI data is four-dimensional, patients may slightly shake their heads during continuous acquisition over a period of time, leading to increased data noise and interfering with the model's reading of signals from normal brain regions. To mitigate this interference and improve the model's generalization ability, this method automatically adds uniform random noise to the time-series and hypergraph data of the fMRI data, effectively improving the accuracy of fMRI classification tasks.

[0121] Finally, fMRI image data of the patient's brain to be diagnosed is obtained and input into the trained Alzheimer's disease classification model for classification to obtain early diagnosis results of Alzheimer's disease.

[0122] To verify the effectiveness of this method, the proposed network model is compared with the following methods; all comparison experiments are conducted on the dataset of this embodiment, and the five-fold cross-validation method is used. The same training set and test set partitioning method is used for each fold to reduce the impact of dataset partitioning on the experimental results;

[0123] The experimental data comes from the ADNI public dataset, as shown in Table 1:

[0124] Subject grouping Number of images men and women Age (years) MMSE CDR AD(N=31) 96 14 / 17 75.27±6.17 20.84±4.24 0.875±0.43 NC (N=35) 129 15 / 20 75.79±6.90 29.14±1.41 0.044±0.14

[0125] Table 1. Dataset used in this embodiment.

[0126] 1) Pearson

[0127] Traditional methods for constructing brain functional connectivity networks use Pearson correlation to calculate the connection strength between Regions of Interest (ROIs), i.e., the connection strength between region of interest i and region of interest j is:

[0128]

[0129] Where x = [x1, x2, ..., x t ,...,x T ] and y = [y1, y2, ..., y t ,...,y T [x] represents the time series of the average BOLD signal values ​​for regions of interest i and j to be subjected to Pearson correlation analysis. t y t Let represent the average BOLD signal of region of interest i and region of interest j at time t, respectively, and T be the number of volumes in the fMRI image;

[0130] 2) MLP-TCN

[0131] Multilayer Perceptron (MLP) is used to transform the original vector at each time point (i.e., the input vector composed of the mean and standard deviation of the BOLD signal extracted from each ROI) into high-dimensional features. The output of the fully connected layer is used as the input of TCN, and the classification loss of MLP is added to the model loss as the time slice loss.

[0132] 3) STGCN

[0133] This method applies a spatiotemporal graph convolutional neural network to learn spatiotemporal information in the data simultaneously, and learns the importance of functional connections in the graph through an ST-GCN model;

[0134] 4) MCGRU

[0135] Spatial features of ROI are extracted by multi-scale one-dimensional convolutional neural networks, and temporal features in time series are learned by GRU recurrent networks.

[0136] 5) MVS-GCN

[0137] This method constructs multiple brain functional connectivity networks using different thresholds, establishes a shared graph embedding layer to learn the features of each brain functional connectivity, and uses a... Figure 1 Consistency methods ensure that the feature representations of different views remain consistent;

[0138] 6) BrainGNN

[0139] BrainGNN is an end-to-end fMRI prediction framework based on graph neural networks. It learns to assign different weights to different regions of interest (ROIs) and then uses pooling layers to cluster the ROIs to classify whole-brain fMRI images.

[0140] 7) HyperGCN

[0141] The HyperGCN framework uses the K-nearest neighbor algorithm to linearly construct a hypergraph and constructs a hypergraph from all samples in the dataset to find the correlation between samples, thereby improving the classification performance.

[0142] In the above methods, the same ensemble testing method as in this paper is applied to Pearson, MLP-TCN, and STGCN methods, with a sampling length of 64. For other comparative methods, all time points of each fMRI image are used as the input of the model.

[0143] As shown in Table 2, Table 2 shows the performance scores of various methods in the AD / NC classification task based on fMRI images. It can be seen that the FHyperGNN model in this embodiment performs better than other mainstream methods on fMRI data, and has achieved the expected results in the AD / NC classification task on fMRI data.

[0144] method ACC Pearson 0.62±0.07 MLP-TCN 0.65±0.14 STGCN 0.67±0.04 MCGRU 0.53±0.11 MVS-GCN 0.71±0.09 BrainGNN 0.62±0.03 HyperGCN 0.69±0.07 FHyperGNN 0.75±0.06

[0145] Table 2. Comparison of this method with mainstream fMRI-based AD diagnostic tasks.

[0146] This method uses a nonlinear Gaussian kernel SVR as the estimator for the recursive feature elimination algorithm and employs the RBF-SVR recursive feature elimination algorithm to nonlinearly construct a hypermap for each fMRI image. This hypermap data structure is used to record the nonlinear multivariate higher-order correlations between multiple ROIs in the fMRI image data, which can demonstrate good classification performance. In addition, the Alzheimer's disease classification model constructed by this method has higher performance and accuracy in the early diagnosis of AD and better results.

[0147] Example 3

[0148] like Figure 5 As shown, this embodiment provides an Alzheimer's disease classification system based on a hypergraph neural network, applying the Alzheimer's disease classification method based on a hypergraph neural network described in Embodiment 1 or 2, including:

[0149] Data acquisition module 301: used to acquire fMRI image data of the brain of Alzheimer's patients and perform preprocessing;

[0150] Hypergraph construction module 302: used to construct a hypergraph for each fMRI image in the preprocessed fMRI image data and obtain fMRI hypergraph data;

[0151] Noise module 303: Used to add noise to the preprocessed fMRI image data and fMRI hypermap data, and integrate them into training metadata;

[0152] Training module 304: Used to train the established Alzheimer's disease classification model using training metadata to obtain a trained Alzheimer's disease classification model;

[0153] Classification module 305: Used to acquire fMRI image data of the brain of the patient to be diagnosed and input it into the trained Alzheimer's disease classification model for classification, and obtain Alzheimer's disease classification results.

[0154] In the specific implementation process, firstly, the data acquisition module 301 acquires fMRI image data of the brains of Alzheimer's patients and performs preprocessing; the hypergraph construction module 302 constructs a hypergraph for each fMRI image in the preprocessed fMRI image data to acquire fMRI hypergraph data; the noise module 303 adds noise to the preprocessed fMRI image data and fMRI hypergraph data and integrates them into training metadata; the training module 304 uses the training metadata to train the established Alzheimer's disease classification model to obtain a trained Alzheimer's disease classification model; finally, the classification module 305 acquires fMRI image data of the brains of patients to be diagnosed and inputs it into the trained Alzheimer's disease classification model for classification to obtain the Alzheimer's disease classification result.

[0155] This system uses a nonlinear Gaussian kernel SVR as the estimator for the recursive feature elimination algorithm. It uses the RBF-SVR recursive feature elimination algorithm to nonlinearly construct a hypermap for each fMRI image. The hypermap, a graph data structure, is used to record the nonlinear multivariate higher-order correlations between multiple ROIs in the fMRI image data, which can show good classification results. In addition, the Alzheimer's disease classification model constructed by this system has higher performance and accuracy in the early diagnosis of AD and better results.

[0156] The same or similar labels correspond to the same or similar parts;

[0157] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0158] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for Alzheimer's disease classification based on hypergraph neural network, characterized in that, The method comprises the following steps: S1: obtaining fMRI image data of the brain of an Alzheimer's disease patient and preprocessing the fMRI image data; The specific method of preprocessing in the step S1 comprises image format conversion, removal of an initial time point, time slice correction, head motion correction, spatial standardization, linear drift removal, covariate regression, low-frequency filtering and spatial smoothing; S2: constructing a hypergraph for each fMRI image in the preprocessed fMRI image data to obtain fMRI hypergraph data, comprising: constructing a hypergraph for each fMRI image in the preprocessed fMRI image data by using an RBF-SVR recursive feature elimination algorithm to obtain fMRI hypergraph data; In the RBF-SVR recursive feature elimination algorithm, a nonlinear Gaussian kernel SVR is used as an estimator of the recursive feature elimination algorithm; dividing each fMRI image in the preprocessed fMRI image data into a plurality of ROIs and extracting a time sequence of average BOLD signals of all ROIs of each fMRI image; for each fMRI image, selecting one ROI as a target ROI, calculating a fitting degree between the target ROI and other ROIs by using the RBF-SVR recursive feature elimination algorithm according to the time sequence of average BOLD signals of all ROIs, and constructing a hyperedge by using the target ROI and a plurality of other ROIs having the highest fitting degrees with the target ROI; repeating the above operations to construct a hyperedge by taking each ROI in each fMRI image as the target ROI, constructing all hyperedges in each fMRI image into a corresponding fMRI hypergraph, and saving all fMRI hypergraphs as fMRI hypergraph data; S3: adding noise to the preprocessed fMRI image data and the fMRI hypergraph data to integrate the preprocessed fMRI image data and the fMRI hypergraph data into training metadata; S4: training an established Alzheimer's disease classification model by using the training metadata to obtain a trained Alzheimer's disease classification model; S5: obtaining fMRI image data of the brain of a patient to be diagnosed and inputting the fMRI image data into the trained Alzheimer's disease classification model to obtain an Alzheimer's disease classification result.

2. The Alzheimer's disease classification method based on hypergraph neural network according to claim 1, characterized in that, Each fMRI image comprises 116 ROIs and 130 time sequences.

3. The Alzheimer's disease classification method based on hypergraph neural network according to claim 2, characterized in that, The specific method for calculating the fitting degree between the target ROI and other ROIs by using the RBF-SVR recursive feature elimination algorithm is a five-fold cross-validation method.

4. The Alzheimer's disease classification method based on hypergraph neural network according to claim 1 or 3, characterized in that, The Alzheimer's disease classification model established in the step S4 is specifically a FHyperGNN hypergraph convolutional neural network model.

5. The Alzheimer's disease classification method based on hypergraph neural network according to claim 4, characterized in that, The FHyperGNN hypergraph convolutional neural network model comprises an input layer, a plurality of HyperGCN hypergraph convolutional modules, a first BatchNorm regularization layer, a Dropout layer, a Linear layer, a Softmax layer and an output layer connected in sequence; Each HyperGCN hypergraph convolutional module comprises a second BatchNorm regularization layer, a HyperGNN hypergraph convolutional layer and an activation layer connected in sequence; The activation function of the activation layer is a Relu function.

6. The Alzheimer's disease classification method based on hypergraph neural network according to claim 5, characterized in that, The number of the HyperGCN hypergraph convolutional modules is at least 3.

7. The Alzheimer's disease classification method based on hypergraph neural network according to claim 6, characterized in that, The dropout rate of the Dropout layer is 0.

5.

8. An Alzheimer's disease classification system based on hypergraph neural network, applying the Alzheimer's disease classification method based on hypergraph neural network in any one of claims 1-7. It comprises: a data acquisition module for acquiring fMRI image data of the brain of an Alzheimer's patient and preprocessing the same; a hypergraph construction module for constructing a hypergraph for each fMRI image in the preprocessed fMRI image data and acquiring fMRI hypergraph data; a noise module for adding noise to the preprocessed fMRI image data and the fMRI hypergraph data and integrating the same into training metadata; a training module for training an established Alzheimer's classification model using the training metadata to obtain a trained Alzheimer's classification model; a classification module for acquiring fMRI image data of the brain of a patient to be diagnosed and inputting the same into the trained Alzheimer's classification model for classification to obtain an Alzheimer's classification result.

Images