Brain function magnetic resonance imaging data analysis method based on contrast graph neural network

Abstract

The invention discloses a brain function magnetic resonance imaging data analysis method based on a contrast graph neural network, and the method comprises the steps: firstly carrying out the data enhancement of a brain function connection graph, and simulating the heterogeneity of brain function magnetic resonance data, so as to improve the diversity of a data set; secondly, the hidden space embedding features of the brain function connection diagram are efficiently learned by fusing a double-Hough-Laplacian diagram convolutional network and a contraction incentive mechanism; the embedded features are mapped to a group of prototype vectors, and prototype allocation codes corresponding to the embedded features are calculated by adopting a Sinkhorn-Knopp algorithm; performing exchange optimization on prototype codes between different enhanced brain connection diagrams of the same subject through a contrast learning strategy, and compelling codes of homologous subjects to be aligned at the minimum cost in combination with a cross entropy loss function; and finally, applying the pre-training model to a functional magnetic resonance imaging data set of the Alzheimer's disease, and carrying out interpretability analysis on learning features to improve the classification efficiency and pathological analysis of the Alzheimer's disease under a limited tag condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical fields of deep learning network technology, data analysis technology and medical data processing technology, and particularly relates to a brain functional magnetic resonance imaging data analysis method based on a contrast graph neural network. BACKGROUND

[0002] Functional Magnetic Resonance Imaging (fMRI) is a dynamic activity imaging technology that reflects neural activity and metabolic level by measuring blood oxygen level dependent signals. It has become an effective tool for analyzing neurodegenerative diseases due to its non-invasive and high spatial resolution characteristics. The typical brain fMRI analysis method is to divide the brain into several brain regions according to the brain atlas, and then calculate the Pearson correlation coefficient between each pair of brain regions to obtain the brain functional connectivity graph for analysis. Since the pathogenesis of neurodegenerative diseases involves the complex interaction of genes, environment, central nervous system and systemic pathophysiological mechanisms, this multi-factor influence leads to significant heterogeneity in the clinical neuropathological manifestations of patients. Even individuals belonging to the same type of group may exhibit completely different brain activity patterns. This heterogeneity poses a great challenge to disease data analysis and diagnosis based on brain connectivity graphs, especially when early clinical symptoms are not obvious.

[0003] Traditional covariate regression methods regress out confounding variables such as gender, age, head motion, multi-center acquisition effects, etc. by generalized linear regression methods, but this method is usually based on single variable analysis and cannot decouple the interaction between variables. Some other studies on heterogeneity start from the pathological point of view, using clustering methods to identify disease subtypes to characterize this heterogeneity. Although this provides an important basis for evaluating and understanding brain function and disease classification, it does not extract functional connectivity features from the individual perspective. With the development of deep learning technology, graph neural networks have a natural advantage in processing connection type data and are increasingly being applied to analyze brain magnetic resonance imaging data. However, existing methods do not consider the heterogeneity of samples when applying graph neural networks to analyze brain connectivity data, and their classification models are heavily dependent on labeled data, while actual magnetic resonance imaging data often has a shortage of labeled data, making it difficult to effectively utilize a large amount of unlabeled data. SUMMARY

[0004] Therefore, the present application aims to provide a brain functional magnetic resonance imaging data analysis method based on a contrast graph neural network, which is reliable to implement, flexible to apply, and can achieve better training results with a small amount of training data.

[0005] To achieve the above technical purposes, the technical scheme adopted by the present application is as follows:

[0006] A brain functional magnetic resonance imaging data analysis method based on a contrast graph neural network, comprising:

[0007] S01, preprocessing the original fMRI brain imaging data, segmenting the brain into multiple regions of interest (ROIs) and constructing a brain functional connectivity graph;

[0008] S02, performing graph data enhancement on the brain functional connectivity graph to introduce sample heterogeneity, generating multiple variants of the brain functional connectivity graph, and setting them as enhanced graphs;

[0009] S03, inputting the enhanced graph into a heterogeneity-aware HL_Conv network, which has a dual Hodge-Laplace convolution structure, for simultaneously performing graph convolution feature extraction on the node domain and edge domain of the enhanced graph to obtain multi-level features of the enhanced graph;

[0010] S04, based on the multi-level features of the enhanced graph, obtaining a graph-level embedding representation vector corresponding to each enhanced graph through a compression-activation mechanism;

[0011] S05, mapping the graph-level embedding representation vector to a preset number of learnable prototype feature vectors, and then assigning an encoding to each enhanced graph using the Sinkhorn-Knopp algorithm;

[0012] S06, exchanging the encodings of two different enhanced graphs derived from the same object, based on the cross-entropy loss between the predicted prototype probability of each enhanced graph and the exchanged encoding, and minimizing the cross-entropy loss, jointly optimizing and training the model parameters of the HL_Conv network to obtain a pre-trained and optimized HL_Conv network;

[0013] S07, freezing the model parameters of the pre-trained and optimized HL_Conv network, and then applying it to the classification and recognition of fMRI brain imaging data.

[0014] As a possible implementation, further, in the present scheme S01, when constructing the brain functional connectivity graph, the average time series of each region of interest (ROI) is used as the node feature, and the correlation between the ROIs is used to define the connection edge.

[0015] As a preferred implementation option, preferably, in the present scheme S01, the original fMRI brain imaging data is preprocessed by the AFNI toolbox, and the Schaefer brain region division method is used to segment the brain into 200 regions of interest (ROIs) and construct a brain functional connectivity graph.

[0016] In constructing the brain functional connectivity graph, the average time series of each region of interest (ROI) is taken as the node feature, the Pearson correlation coefficient between each pair of ROIs is calculated to obtain the functional connectivity matrix, and the 15% connections with the strongest correlation are retained as the edge connections of the brain functional connectivity graph.

[0017] As a preferred implementation option, preferably, in the scheme S02, the method for performing graph data augmentation on the brain functional connectivity graph is one or more of the following:

[0018] (a) Node deletion: randomly remove some nodes and their associated edges;

[0019] (b) Edge editing: randomly add or delete some edges;

[0020] (c) Node feature editing: randomly mask node attributes;

[0021] (d) Subgraph extraction: randomly extract a subgraph from the full graph that does not destroy the core structure.

[0022] The scheme introduces sample heterogeneity using graph data augmentation, proposes an exchange contrastive learning framework based on a dual Hodge-Laplace graph neural network, trains to extract disease-related pattern features of brain functional connectivity on the entire dataset, and realizes semi-supervised learning of graph neural networks under the condition of a small amount of annotation.

[0023] As a preferred implementation option, preferably, in the scheme S03, the heterogeneity-aware HL_Conv network is designed based on a boundary operator and a Hodge-Laplace dual graph convolution structure, which represents the enhanced graph as a complex composed of nodes, edges, triangles, and higher-dimensional simplices.

[0024] Among them, the brain region node corresponds to a 0-dimensional simplex, the edge connecting two nodes corresponds to a 1-dimensional simplex, and the k-dimensional simplex is a k-dimensional polyhedron composed of (k+1) nodes; the kth order boundary operator is used to describe the relationship between the k-dimensional simplex and the (k+1) -dimensional simplex;

[0025] For a given k value, the boundary operator can be expressed in matrix form, which is defined as follows:

[0026]

[0027] where, is the (i, j) element of , which is defined to represent whether the jth k-dimensional simplex contains the ith (k-1) -dimensional simplex For a brain functional connectivity graph containing N nodes and E edges, the size of the first-order boundary operator is N x E, which is used to reflect how the nodes of the 0-dimensional simplex are connected to form the edges of the 1-dimensional simplex of the brain functional connectivity graph, and the element value is ±1, and the sign represents the direction;

[0028] The k-order Hodge-Laplace matrix can be defined by the boundary operator as follows:

[0029]

[0030] wherein, is the k-order Laplace matrix, and are the k-order boundary operator and its transpose, respectively.

[0031] As a preferred implementation option, preferably, in the scheme S03, the HL_Conv network only focuses on the interaction relationship between edges for the brain functional connectivity graph, and does not consider the interaction between edges and higher-dimensional simplices, and the Hodge-Laplace matrix is simplified to only involve the first-order boundary operator Based on this simplification, the Hodge-Laplace operator of the node and the edge can be directly calculated by the boundary operator, and the Hodge-Laplace operator corresponding to the node is defined as: The Hodge-Laplace operator corresponding to the edge is defined as: The 0-dimensional and 1-dimensional Hodge-Laplace matrices are substituted into the spectral-based graph neural network convolution formula, and the double convolution operation on the signal X can be realized, and the definition is as follows:

[0032]

[0033] wherein, X(·), are the input feature and the output feature of the double convolution operation respectively, h is the Hodge-Laplace convolution kernel, is the filter parameter to be optimized, T p is a polynomial expansion function, T p The Laguerre polynomial is selected for iterative approximation, and p is the order of the Laguerre polynomial expansion.

[0034] As a preferred implementation option, preferably, in the scheme S04, the graph-level embedding representation vector corresponding to each enhanced graph is obtained through the projection head based on the compression-excitation mechanism, which includes a compression stage, an excitation stage and a feature re-labeling; and correspondingly, the compression-excitation mechanism can be applied by constructing an attention reading model, and integrated with the compression-excitation reading module SERO.

[0035] In the compression stage, a global average operation is performed on the node feature matrix to generate a compression vector;

[0036] In the excitation stage, the node weight vector is generated by a multi-layer perception (MLP) and a sigmoid activation function

[0037] In the feature re-scaling stage, the graph features are multiplied by the weight vector and summed up

[0038] The mathematical expression of the compressed excitation readout is:

[0039]

[0040] where φ is the average operation on all node features, σ(·) is a nonlinear activation function, and W1 and W2 are trainable parameters.

[0041] Based on the multi-level features of the enhanced graph, the graph-level embedding representation vector corresponding to each enhanced graph, i.e., the graph-level feature z, is obtained through the compression-excitation mechanism.

[0042] As a preferred implementation option, in the scheme S05, the graph-level feature z is projected to the prototype feature vector to obtain the probability of the latent space clustering center;

[0043] Let the latent space prototype feature vector to be trained be where each column corresponds to a prototype feature, which can be regarded as a clustering center in the latent space; through the softmax function calculation, the probabilities that the i-th and j-th enhanced graphs belong to the k-th prototype are defined as follows:

[0044]

[0045] where are the graph-level feature values of the i-th and j-th functional brain graphs, respectively, and τ is a temperature coefficient used to adjust the discrimination strength of the positive and negative sample similarities; let the probability vector belonging to multiple prototypes be

[0046] In the mini-batch training, each enhanced graph is assigned an encoding, and the encoding is evenly distributed among all samples while maintaining the difference between samples; if the batch contains B samples, the optimization objective function of the encoding is:

[0047]

[0048] where Q ij is the i-th row and j-th column element of the matrix Q; the second term of formula (6) is an entropy regularization term, which is used to control the smoothness of the encoding calculation.

[0049] The problem of formula (6) is converted into an optimal transmission problem, and the Sinkhorn-Knopp algorithm is used to realize the assignment of samples-prototypes in the exchange contrast learning, and the solution of formula (6) is defined as follows:

[0050]

[0051] wherein, and are the renormalization vectors, and ∈ is a regularization parameter; through iterative multiplication operation, the assignment codes q i and q j of the two enhanced graphs are obtained.

[0052] As a preferred implementation option, preferably, the scheme S06 comprises: for two different enhanced graphs i and j originating from the same object, the codes q i and q j are respectively predicted through the exchanged graph-level features z i and z j , and then the cross-entropy loss between the predicted prototype probability p i and p j of each enhanced graph and the exchanged code is calculated, which is defined as follows:

[0053]

[0054] wherein, N is the number of samples, Γ is the set of graph enhancement transformations, are the codes of the i th and j th enhanced graphs belonging to the k th prototype.

[0055] With the aim of minimizing the cross-entropy loss, the graph feature extraction parameters of the HL_Conv network and the parameters in the prototype feature vector set are jointly optimized and trained according to formulas (5), (7) and (8), so as to obtain the pre-trained and optimized HL_Conv network.

[0056] As a preferred implementation option, preferably, the scheme S07 comprises: freezing the model parameters of the pre-trained and optimized HL_Conv network, then replacing the projection head with a multi-layer perceptron (MLP), training the classification network on a labeled data subset, and applying the trained classification network to the classification and recognition of fMRI brain imaging data to identify fMRI brain imaging data containing Alzheimer's disease.

[0057] In summary, the present scheme proposes a brain functional magnetic resonance imaging data analysis method based on a contrast graph neural network, and also correspondingly constructs an exchange contrast graph neural network learning framework for brain functional magnetic resonance imaging data. The framework captures the robust representation of the whole data set in a semi-supervised manner by exchange contrast learning on brain functional connectivity graphs.

[0058] In the technical implementation, the present scheme first performs data enhancement on the functional brain connection graph (brain functional connection graph), simulates the heterogeneity of brain functional magnetic resonance data to improve the diversity of the data set.

[0059] Secondly, a network model fusing a double-Hodge-Laplacian graph convolutional network and a shrinkage excitation readout module is designed to efficiently learn the embedding features of the hidden space of the brain functional connection graph. These embedding features are mapped to a set of prototype vectors, and the Sinkhorn-Knopp algorithm is used to calculate the corresponding prototype assignment code. On this basis, through a contrastive learning strategy, the prototype codes of different enhanced brain connection graphs of the same subject are exchanged and optimized, and the cross-entropy loss function is combined to force the codes of the homologous subjects to be aligned at the minimum cost.

[0060] Finally, the pre-trained model is applied to the functional magnetic resonance imaging data set of Alzheimer's disease, and the learned features are analyzed for interpretability to improve the classification performance and pathological analysis of Alzheimer's disease under the condition of limited labels.

[0061] By adopting the above technical scheme, the present scheme has the beneficial effects compared with the prior art:

[0062] 1. The present scheme processes the brain functional connection graph through graph data enhancement to introduce the heterogeneity feature description of the patient, and then exchanges the prototype codes of different enhanced graphs to realize the minimum cost alignment of the codes of the homologous subjects. The HL_Conv network framework of the present scheme can effectively capture the diversity features of the disease through the perception of the heterogeneity of the graph contrastive learning and the construction of the global data set representation, thereby ensuring the robustness of the semi-supervised learning framework.

[0063] 2. The present scheme innovatively applies the boundary operator in graph theory to expand the Hodge-Laplacian graph convolution based on nodes (brain regions) to the Hodge-Laplacian graph convolution based on edges (brain connections). The interaction features between brain regions are synchronously extracted through the double-graph convolution architecture, thereby significantly improving the utilization rate of high-order brain connection group information.

[0064] 3. The present scheme also proposes an attention readout model for the output of the graph convolutional network, which dynamically optimizes feature selection through a compression excitation mechanism to solve the information loss problem of the traditional global pooling (maximum value / average value). At the same time, the attention weights of each brain region node are calculated to quantify the differential contribution of different brain regions to disease classification.

[0065] 4. The present scheme constructs a semi-supervised disease data classification framework through the exchange contrastive learning of the graph neural network, which can achieve good disease classification effect with a small amount of labels. DETAILED DESCRIPTION

[0066] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on these drawings also belong to the protection scope of the present application.

[0067] Figure 1 is a brain function magnetic resonance imaging data analysis framework diagram based on the HL_Conv graph neural network of the present scheme.

[0068] Figure 2 is a comparative graph neural network analysis flowchart of brain function magnetic resonance imaging data of the present scheme. DETAILED DESCRIPTION

[0069] The present application will be further described in detail below in combination with the accompanying drawings and embodiments. It is particularly pointed out that the following embodiments are only used to illustrate the present application, but do not limit the scope of the present application. Similarly, the following embodiments are only some embodiments of the present application, not all embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative effort also belong to the protection scope of the present application.

[0070] In combination with Figure 1 and Figure 2 shown, the present embodiment scheme proposes a brain function magnetic resonance imaging data analysis method based on a comparative graph neural network, and also correspondingly constructs a swap comparative graph neural network learning framework for brain function magnetic resonance imaging data. The framework captures robust representations of the full data set in a semi-supervised manner by swap comparative learning of brain function connection graphs.

[0071] The proposed swap comparative graph neural network learning framework includes three core modules, specifically: (1) a heterogeneity-aware HL_Conv network (Heterogeneity-aware Hodge Laplacian convolution network, HL_Conv network) module; (2) swap code learning; (3) semi-supervised learning.

[0072] On the function of each module, in the heterogeneous perception HL_Conv network, the magnetic resonance imaging data (fMRI brain imaging data) of the brain is preprocessed to generate a brain connectivity graph (brain functional connectivity graph). On this basis, a plurality of graph augmentation techniques are applied to the brain functional connectivity graph, which can include one or more of node drop, edge edit, node feature edit, and subgraph extraction, etc. Enhancement techniques are used to construct different variants of the original graph (original fMRI brain imaging data) and set them as enhanced graphs. These enhanced graphs are then input into the HL_Conv network, which simultaneously processes the node graph and its dual edge graph, and extracts the multi-level features of the graph. Then, the projection head is used to obtain the embedding representation vector of each brain connectivity graph.

[0073] In the exchange coding learning, the embedding representation vector is mapped to the learnable prototype feature vector, and the similarity score is calculated simultaneously. The Sinkhorn-Knopp algorithm is used to assign and generate discrete codes. Then, the codes obtained under different enhanced variants are exchanged, and the cross-entropy loss between the exchange codes and the prediction probability is minimized to train the optimal feature extraction HL_Conv network of the brain connectivity graph.

[0074] In semi-supervised learning, a part of the labeled subset is extracted, and the pre-trained HL_Conv network parameters are combined to replace the projection head with a classification head for supervised learning of disease index prediction or classification.

[0075] As an example, the embodiment scheme can include the following steps:

[0076] A brain functional magnetic resonance imaging data analysis method based on a contrast graph neural network, comprising:

[0077] S01, preprocessing the original fMRI brain imaging data, segmenting the brain into a plurality of regions of interest ROI and constructing a brain functional connectivity graph; in this step, the original fMRI brain imaging data is standardized and preprocessed by the AFNI (Analysis of Functional NeuroImages) toolbox, the Schaefer brain region division method is used to segment the brain into 200 regions of interest ROI and construct a brain functional connectivity graph; when constructing the brain functional connectivity graph, the average time series of each region of interest ROI is used as the node feature, and the correlation between the ROIs is used to define the connection edge.

[0078] Specifically, in constructing the brain functional connectivity graph, the average time series of each region of interest (ROI) is taken as the node feature, the Pearson correlation coefficient between each pair of ROIs is calculated to obtain the functional connectivity matrix, and the 15% of the strongest correlations are retained as the edge connections of the brain functional connectivity graph.

[0079] S02, a domain-specific strategy is used to perform graph data augmentation on the brain functional connectivity graph to introduce sample heterogeneity, generate multiple variants of the brain functional connectivity graph, and set them as augmented graphs; in this step, the method of performing graph data augmentation on the brain functional connectivity graph is to randomly use one or more of the following methods:

[0080] (a) Node deletion: randomly remove some nodes and their associated edges;

[0081] (b) Edge editing: randomly add or delete some edges;

[0082] (c) Node feature editing: randomly mask node attributes;

[0083] (d) Subgraph extraction: randomly extract subgraphs from the full graph that do not destroy the core structure.

[0084] S03, input the augmented graph into the heterogeneity-aware HL_Conv network, which has a dual Hodge-Laplace convolution structure, which is used to simultaneously perform graph convolution feature extraction on the node domain and edge domain of the augmented graph to obtain multi-level features of the augmented graph; in this step, the heterogeneity-aware HL_Conv network is designed based on the boundary operator and the Hodge-Laplace dual graph convolution structure, which represents the augmented graph as a complex composed of nodes, edges, triangles, and higher-dimensional simplices; wherein the brain region node corresponds to a 0-dimensional simplex, the edge connecting two nodes corresponds to a 1-dimensional simplex, and the k-dimensional simplex is a k-dimensional polyhedron composed of (k+1) nodes; the kth order boundary operator is used to describe the relationship between the k-dimensional simplex and the (k+1) -dimensional simplex;

[0085] For a given k value, the boundary operator can be represented in matrix form, which is defined as follows:

[0086]

[0087] wherein, is the (i, j) element of , which is defined to represent whether the jth k-dimensional simplex contains the ith (k-1) -dimensional simplex For a brain functional connectivity graph containing N nodes and E edges, the size of the first-order boundary operator is N x E. The elements of the first-order boundary operator are ±1, and the sign represents the direction.

[0088] The k-order Hodge-Laplace matrix can be defined by the boundary operator as follows:

[0089]

[0090] wherein, is the k-order Laplace matrix, and are the k-order boundary operator and its transpose, respectively; the HL_Conv network only focuses on the interaction relationship between edges for the brain functional connectivity graph, and does not consider the interaction between edges and higher-dimensional simplices (such as triangular meshes). When the Hodge-Laplace matrix is defined as a boundary operator involving only the first order. Based on this simplification, the Hodge-Laplace operator of the node and the edge can be directly calculated by the boundary operator. The Hodge-Laplace operator of the node is defined as: The Hodge-Laplace operator of the edge is defined as: The 0-dimensional and 1-dimensional Hodge-Laplace matrices are substituted into the spectral-based graph neural network convolution formula, and the double convolution operation on the signal X can be realized. The definition is as follows:

[0091]

[0092] wherein, are the input feature and the output feature of the double convolution operation, respectively, h is the Hodge-Laplace convolution kernel, is the filter parameter to be optimized, T p is a polynomial expansion function, T p The Laguerre polynomial is selected for iterative approximation, and p is the order of the Laguerre polynomial expansion.

[0093] S04, based on the enhanced graph, the multi-level feature is obtained by the compression-excitation mechanism to obtain the graph-level embedding representation vector corresponding to each enhanced graph. In this step, the projection head based on the compression-excitation mechanism is used to obtain the graph-level embedding representation vector corresponding to each enhanced graph, which includes the compression stage, the excitation stage and the feature re-calibration. Correspondingly, the compression-excitation mechanism can be applied by constructing an attention reading model, and it is integrated into the compression-excitation reading module SERO.

[0094] For the compression-activation mechanism, in the compression phase, a global average operation is performed on the node feature matrix to generate a compressed vector; in the activation phase, a multi-layer perception (MLP) and a sigmoid activation function are used to generate a node weight vector In feature re-scaling, the graph features are multiplied by the weight vector and summed.

[0095] The mathematical representation of the compressed and activated readout is:

[0096]

[0097] where φ is an average operation on all node features, σ(·) is a nonlinear activation function, and W1 and W2 are trainable parameters.

[0098] Based on the multi-level features of the enhanced graph, a compression-activation mechanism is used to obtain a graph-level embedding representation vector corresponding to each enhanced graph, i.e., a graph-level feature z.

[0099] S05, the graph-level embedding representation vector is mapped to a preset number of learnable prototype feature vectors, and then a Sinkhorn-Knopp algorithm is used to assign a code to each enhanced graph.

[0100] In this step, the graph-level feature z is projected to the prototype feature vector to obtain the probability of the hidden space clustering center; wherein the hidden space prototype feature vector to be trained is set as where each column corresponds to a prototype feature, which can be regarded as a clustering center in the hidden space; through the softmax function calculation, the probabilities that the i-th and j-th enhanced graphs belong to the k-th prototype are defined as follows:

[0101]

[0102] wherein, are the graph-level feature values of the i-th and j-th functional brain graph, respectively, and τ is a temperature coefficient used to adjust the discrimination strength of the positive and negative sample similarities; let the probability vector belonging to multiple prototypes be:

[0103] In small batch training, each enhanced graph is assigned a code, and the code is evenly distributed among all samples while maintaining the difference between samples, i.e., the calculation of the code needs to satisfy two conditions: it must be consistent with the hidden space projection vector, and it must avoid model collapse (all enhanced graphs are assigned the same code); in this scheme, the ideal code should be evenly distributed among all samples while maintaining the difference between samples; if a batch contains B samples, the optimization objective function of the code is:

[0104]

[0105] wherein, Q ij Let be the element in the i-th row and j-th column of matrix Q; it should be a small value to prevent all samples from being uniformly distributed to trivial solutions of all prototypes.

[0106] The problem in Equation (6) is transformed into an optimal transmission problem, and the Sinkhorn-Knopp algorithm is used to implement sample-prototype allocation in exchange-contrast learning. At this time, the solution of Equation (6) is defined as follows:

[0107]

[0108] In the formula, and These are the renormalized vectors, and ∈ is the regularization parameter; through iterative multiplication, the allocation codes q of the two augmented graphs are obtained. i q j .

[0109] S06. For two different augmentation graphs originating from the same object, swap the encodings. Based on the cross-entropy loss between the predicted prototype probability of each augmentation graph and the swapped encodings, jointly optimize the model parameters of the HL_Conv network with the goal of minimizing the cross-entropy loss, to obtain a pre-trained and optimized HL_Conv network.

[0110] In this step, for two different augmented graphs i and j originating from the same object (sample graph), the graph-level features z are swapped. i z j Predict the code q respectively i q j Then calculate the predicted prototype probability p for each augmentation map. i p j The cross-entropy loss L between the transposed and encoded codes swap Its definition is as follows:

[0111]

[0112] Where N is the number of samples, and Γ is the set of graph enhancement transformations. These are the encodings of the i-th and j-th augmented graphs belonging to the k-th prototype, respectively.

[0113] With the goal of iteratively solving to minimize the cross-entropy loss, the graph feature extraction parameters and the parameters in the prototype feature vector set of the HL_Conv network are optimized by combining equations (5), (7), and (8) to obtain the pre-trained and optimized HL_Conv network.

[0114] S07. Freeze the model parameters of the pre-trained and optimized HL_Conv network, and then apply it to the classification and recognition of fMRI brain imaging data.

[0115] The scheme S07 includes: freezing the model parameters of the pre-trained and optimized HL_Conv network, then replacing the projection head with a multi-layer perceptron (MLP), training the classification network on a labeled data subset, and applying the trained classification network to the classification and recognition of fMRI brain imaging data to identify fMRI brain imaging data containing Alzheimer's disease.

[0116] Test comparison

[0117] To verify the effectiveness of the method proposed in the present scheme, the proposed method is compared with existing graph neural network-based models. The functional magnetic resonance imaging data used by these methods comes from the Alzheimer's Disease Neuroimaging Initiative project (https: / / www.loni.usc.edu / ), which contains 919 subjects, including 366 normal controls and 553 mild cognitive impairment patients. A 10-fold cross-validation strategy is used to divide the dataset into training, validation, and test sets in a ratio of 5:1:1. Then, a certain percentage (20%-100%) of labeled data sets are randomly selected to train and validate the classification model. This process is repeated 10 times, and the average performance and standard deviation are reported as the final result. The classification accuracy of each model at different label rates is shown in Table 1:

[0118] Table 1. Accuracy of models at different label rates

[0119]

[0120] Wherein: the sources of each model mentioned in Table 1 are as follows:

[0121] GCN corresponds to the method proposed by T.N. Kipf et al. (T.N. Kipf and M. Welling, "Semi-supervised classification with graph convolutional networks," arXiv preprint arXiv:1609.02907, 2017.);

[0122] GAT corresponds to the method proposed by P. et al. (P. G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y. Bengio, "Graph attention networks," in Proc. Int. Conf. Learn. Representations (ICLR), 2018.);

[0123] BrainGNN corresponds to the method proposed by X. Li et al. (X. Li, Y. Zhou, N. Dvornek, M. Zhang, S. Gao, J. Zhuang, D. Scheinost, L. H. Staib, P. Ventola, and J. S. Duncan, "BrainGNN: Interpretable brain graph neural network for fMRI analysis," Med. Image Anal., vol. 74, p. 102233, 2021.);

[0124] DGCN corresponds to the method proposed by K. Zhao et al. (K. Zhao, B. Duka, H. Xie, D. J. Oathes, V. Calhoun, and Y. Zhang, "A dynamic graph convolutional neural network framework reveals new insights into connectome dysfunctions in ADHD," NeuroImage, vol. 246, p. 118774, 2022.);

[0125] BrainNetCNN corresponds to the method proposed by J. Kawahara et al. (J. Kawahara, C. J. Brown, S. P. Miller, B. G. Booth, V. Chau, R. E. Grunau, J. G. Zwicker, and G. Hamarneh, "BrainNetCNN: Convolutional neural networks for brain networks; toward predicting neurodevelopment," NeuroImage, vol. 146, pp. 1038-1049, 2017.).

[0126] It can be seen from the results in Table 1 that the method proposed in the present solution is always superior to other graph neural network-based methods in classification results under different label rates. The advantage of the method proposed in the present solution is particularly evident when the training data size is small (<40%). At this time, the performance gap with other methods is the most significant. When the training data set size increases (>60%), the performance gap narrows, but the method proposed in the present solution is still the optimal method. It is particularly noteworthy that even with only 20% of the training data, the method proposed in the present solution can achieve an accuracy of more than 77%, which is significantly better than other algorithms.

[0127] Based on the above, it can be determined that the method proposed by the scheme is very effective even in the case of limited labeled training data, making it very suitable for the scenario of label scarcity. The strong performance of the method proposed by the scheme can be attributed to the ability of the pre-training model to extract high-order functional connection features from the brain connection graph using HL_Conv, as well as its global prototype mechanism, which can learn the optimal latent space representation from the entire dataset in an unsupervised manner. This enables the classifier to achieve satisfactory results using only a small amount of labeled data. In addition, the exchange-based enhancement strategy enhances the generalization ability of the model, further improving the robustness of the model.

[0128] The above only describes some embodiments of the present application, and does not limit the protection scope of the present application. Any equivalent device or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for analyzing functional magnetic resonance imaging (fMRI) data of the brain based on a contrastive graph neural network, characterized in that, It includes: S01. Preprocess the raw fMRI brain imaging data, divide the brain into multiple regions of interest (ROIs) and construct a brain functional connectivity map. S02. Perform graph data augmentation on the brain functional connectivity graph to introduce sample heterogeneity, generate multiple variants of the brain functional connectivity graph, and set them as augmented graphs; S03. Input the augmented graph into the heterogeneity-aware HL_Conv network, which has a dual Hodge-Laplacian convolutional structure, which is used to simultaneously perform graph convolutional feature extraction on the node domain and edge domain of the augmented graph to obtain multi-level features of the augmented graph. S04. Based on the multi-level features of the augmented graph, obtain the graph-level embedding representation vector corresponding to each augmented graph through a compression-excitation mechanism; S05. Map the graph-level embedding representation vector to a preset number of learnable prototype feature vectors, and then use the Sinkhorn-Knopp algorithm to assign an encoding to each augmented graph. S06. For two different augmentation graphs originating from the same object, swap the encodings. Based on the cross-entropy loss between the predicted prototype probability of each augmentation graph and the swapped encodings, jointly optimize the model parameters of the HL_Conv network with the goal of minimizing the cross-entropy loss, to obtain a pre-trained and optimized HL_Conv network. S07. Freeze the model parameters of the pre-trained and optimized HL_Conv network, and then apply it to the classification and recognition of fMRI brain imaging data.

2. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 1, characterized in that, In S01, when constructing the brain functional connectivity map, the average time series of each region of interest (ROI) is used as the node feature, and the correlation between each pair of ROIs is used to define the connection edges.

3. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 2, characterized in that, In S01, the raw fMRI brain imaging data were preprocessed using the AFNI toolbox, and the brain was divided into 200 regions of interest (ROIs) using the Schaefer brain region segmentation method, and a brain functional connectivity map was constructed. When constructing the brain functional connectivity map, the average time series of each region of interest (ROI) is used as the node feature. The functional connectivity matrix is ​​obtained by calculating the Pearson correlation coefficient between each pair of ROIs, and the 15% of connections with the strongest correlation are retained as the edge connections of the brain functional connectivity map.

4. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural networks as described in any one of claims 1 to 3, characterized in that, In S02, the method for performing graph data augmentation on the brain functional connectivity graph is one or more of the following: (a) Node discarding: Randomly remove some nodes and their associated edges; (b) Edge editing: Randomly add or delete some edges; (c) Node Feature Editing: Randomly occlude node attributes; (d) Subgraph extraction: Randomly extract subgraphs from the entire graph that do not disrupt the core structure.

5. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 4, characterized in that, In S03, the heterogeneity-aware HL_Conv network is designed based on boundary operators and Hodge-Laplace bi-graph convolutional structure, which represents the augmented graph as a complex composed of nodes, edges, triangles and higher-dimensional simplexes; In this system, brain region nodes correspond to 0-dimensional simplexes, edges connecting two nodes correspond to 1-dimensional simplexes, and a k-dimensional simplex is a k-dimensional polyhedron composed of (k+1) nodes; the k-th order boundary operator... Used to describe the relationship between the k-dimensional simplex and the (k+1)-dimensional simplex; For a given value of k, the boundary operator It can be represented in matrix form, and its definition is as follows: in, for The (i, j)th element is defined to represent the j-th k-dimensional simplex. Does it contain the i-th (k-1)-dimensional simplex? For a brain functional connectivity graph with N nodes and E edges, a first-order boundary operator of size N×E is used to reflect how the brain region nodes of the 0-dimensional simplex are connected to form the edges of the 1-dimensional simplex brain functional connectivity graph. Its element values ​​are ±1, and the sign indicates the direction. The k-order Hodge-Laplace matrix can be defined by the boundary operator as follows: in, It is a k-order Laplace matrix. These are the k-order boundary operators and their transposes, respectively.

6. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 5, characterized in that, In S03, the HL_Conv network, targeting brain functional connectivity maps, focuses only on the interactions between edges, without considering the interactions between edges and higher-dimensional simplexes. This simplifies the definition of the Hodge-Laplace matrix to involve only first-order boundary operators. Based on this simplification, the Hodge-Laplacian operators for nodes and edges can be directly calculated using boundary operators. The Hodge-Laplacian operator corresponding to a node is defined as follows: The Hodge-Laplace operator corresponding to an edge is defined as follows: Substituting the 0-dimensional and 1-dimensional Hodge-Laplacian matrices into the spectral graph neural network convolution formula, a double convolution operation on signal X can be achieved, defined as follows: Among them, X(·), , respectively, represent the input and output features of the double convolution operation, where h is the Hodge-Laplacian convolution kernel. For the filter parameters to be optimized, T p Let T be the polynomial expansion function. p The Laguerre polynomial is used for iterative approximation, where p is the order of the Laguerre polynomial expansion.

7. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 6, characterized in that, In S04, the graph-level embedding representation vector corresponding to each augmented graph is obtained through the projection head based on the compression-excitation mechanism, which includes the compression stage, the excitation stage and the feature recalibration. During the compression phase, a global averaging operation is performed on the node feature matrix to generate a compressed vector; During the activation phase, node weight vectors are generated using a multilayer perceptron (MLP) and a sigmoid activation function. When relabeling features, the graph features are multiplied by the weight vector and then summed. The mathematical expression read out by the compression excitation is: Where φ is the average operation on all node features, σ(·) is the non-linear activation function, and W1 and W2 are trainable parameters; Based on the multi-level features of the augmented graph, the graph-level embedding representation vector corresponding to each augmented graph is obtained through a compression-excitation mechanism, i.e., the graph-level feature z.

8. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 7, characterized in that, In S05, the graph-level feature z is projected onto the prototype feature vector to obtain the probability of the cluster center in the latent space. Let the latent space prototype feature vector to be trained be... Each column corresponds to a prototype feature, which can be regarded as a cluster center in the latent space. The probabilities of the i-th and j-th augmented images belonging to the k-th prototype are defined as follows using the softmax function: in, Let be the graph-level feature values ​​of the i-th and j-th functional brain maps, respectively, and τ be the temperature coefficient used to adjust the distinguishing strength of the similarity between positive and negative samples; let the probability vector belonging to multiple prototypes be: In mini-batch training, each augmentation map is assigned a code, ensuring that the codes are evenly distributed across all samples while preserving the differences between samples. If a batch contains B samples, the objective function for optimizing the codes is: in, Q ij Let be the element in the i-th row and j-th column of matrix Q; the second term of formula (6) is the entropy regularization term, which is used to control the smoothness of the encoding calculation; The problem in Equation (6) is transformed into an optimal transmission problem, and the Sinkhorn-Knopp algorithm is used to implement sample-prototype allocation in exchange-contrast learning. At this time, the solution of Equation (6) is defined as follows: In the formula, and These are the renormalized vectors, and ∈ is the regularization parameter; through iterative multiplication, the allocation codes q of the two augmented graphs are obtained. i q j .

9. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 8, characterized in that, S06 includes: for two different augmented graphs i and j originating from the same object, by exchanging the graph-level feature z... i z j Predict the code q respectively i q j Then calculate the predicted prototype probability p for each augmentation map. i p j The cross-entropy loss between the transposed and encoded values ​​is defined as follows: Where N is the number of samples, and Γ is the set of graph enhancement transformations. These are the codes for the i-th and j-th augmented graphs belonging to the k-th prototype, respectively. With the goal of minimizing the cross-entropy loss, the graph feature extraction parameters and the parameters in the prototype feature vector set of the HL_Conv network are optimized by combining equations (5), (7), and (8) to obtain the pre-trained and optimized HL_Conv network.

10. The brain functional magnetic resonance imaging data analysis method based on contrast graph neural network as described in claim 9, characterized in that, S07 includes: freezing the model parameters of the pre-trained and optimized HL_Conv network, then replacing the projection head with a multilayer perceptron (MLP), training a classification network on a labeled subset of data, and then applying the trained classification network to classify and identify fMRI brain imaging data containing Alzheimer's disease.

Images