A brain disease classification method based on the fusion of multi-template brain functional and effector networks

Abstract

This invention discloses a brain disease diagnosis method based on the fusion of multi-template brain functional and effector networks, comprising the following steps: Step 1, data acquisition; Step 2, multi-template data fusion; Step 3, FC network construction; Step 4, EC network construction; Step 5, FC-EC network fusion. The dual-branch functional-effect connectivity construction method designed in this invention can accurately identify and construct brain functional and effector connectivity networks. The functional-effect fusion method based on adaptive self-attention designed in this invention can effectively fuse two types of network data, thereby effectively improving the problem of the still low accuracy of rs-fMRI data in diagnosing ASD. This invention achieves high accuracy in brain disease diagnosis and exhibits good interpretability in ASD tasks.

Description

Technical Field

[0001] This invention relates to a method for modeling brain connectivity networks from rs-fMRI resting-state functional magnetic resonance imaging data. In order to utilize rs-fMRI for computer-aided analysis of brain disease targets, a brain disease classification method based on the fusion of multi-template brain functional and effector networks is designed. Background Technology

[0002] Autism spectrum disorder (ASD) is a lifelong neurodevelopmental disorder characterized by a multitude of symptoms and varying degrees of skill deficits. These symptoms primarily involve social interaction, repetitive behavioral patterns, and restricted interests. The observed heterogeneity of numerous behaviors and widespread neuroanatomical abnormalities in individuals with autism make accurate analysis of this disorder challenging. Resting-state functional magnetic resonance imaging (rs-fMRI), a non-invasive neuroimaging technique, uniquely allows this invention to observe spontaneously generated low-frequency fluctuations within the brain without the individual performing a specific task. By tracking temporal changes in blood oxygen level-dependent (BOLD) signals, rs-fMRI has the ability to capture anomalous interactions between regions of interest (ROIs) in the brain. This technique provides a means for studying brain activity, thereby gaining a deeper understanding of the neural basis of autism spectrum disorder (ASD) and related brain disorders.

[0003] Based on resting-state functional magnetic resonance imaging (fMRI) data, this invention can construct a brain connectivity network model to assist in the analysis of brain diseases. The brain network can be represented as a graphical structure, consisting of brain regions as nodes and interactions between brain regions as connecting edges. Considering the characteristics of functional integration and separation in the brain, brain connectivity is typically modeled in three different modes: structural connectivity (SC), functional connectivity (FC), and effector connectivity (EC). Structural connectivity refers to the physical or anatomical connectivity networks connecting clusters of neurons; these connections can be obtained from biologically or physically relevant structures extracted from diffusion-weighted imaging (DWI) data. In contrast, functional and effector connectivity networks can be directly generated from rs-fMRI data. FC describes brain activity from the perspective of statistical dependence between brain regions. Functional connectivity is defined as the statistical dependence between remote neurophysiological events, quantifying brain activity between regions through a measure of statistical dependence, forming an undirected graph. EC, on the other hand, emphasizes the directed causal relationship between one brain region and another, describing the influence of one nervous system on another and storing it in the form of a directed graph.

[0004] Currently, numerous methods have been proposed for modeling brain connectivity networks using rs-fMRI data and applied to the analysis of brain diseases, such as Support Vector Machines (SVM), Random Forests (RF), Convolutional Neural Networks (CNN), and convolution-based autoencoders. Furthermore, Graph Neural Networks (GNNs) stand out due to their ability to capture network topology information. Graph classification methods use GNNs as nonlinear functions that take a graph constructed from rs-fMRI data as input and output labels for brain disease categories. For example, BrainGNN utilizes pooling regularization with a globally shared mask to extract graph-level representation features. However, most methods only extract features from fully connected (FC) networks for brain disease classification, neglecting to reveal differences in causal dependencies between ASD patients and healthy controls by modeling EC (ECs). Studies have shown that in many brain diseases, EC networks in several resting-state networks (such as DMN, AN, and FPN) are significantly altered compared to healthy controls, suggesting that ECs can serve as biomarkers for brain disease classification. Meanwhile, the aforementioned methods all use only a single brain segmentation template data for analysis. In contrast, using multiple templates allows for the extraction of different feature sets from various segmentation templates, providing more comprehensive brain information. Furthermore, some segmentation methods can acquire information that better aligns with clinical findings, while others may introduce increased noise due to poor fit to experimental data. Therefore, combining data from multiple segmentation templates can employ various nonlinear registration methods to effectively reduce the negative impact of registration errors and provide different but complementary information to identify different disease states, thereby improving the accuracy of the analysis. Summary of the Invention

[0005] To address the low accuracy of current rs-fMRI data analysis for ASD, this invention proposes a brain disease classification method based on the fusion of multi-template brain function and effector networks (MaFECAF). This method first fuses data from two brain segmentation templates using a cosine similarity matrix as weights, building upon a GNN model. Next, the fused data is used to construct FC and EC networks respectively through a residual spatiotemporal graph convolutional network and a variational autoencoder parameterized by a graph neural network architecture. Finally, the two networks are sequentially processed through a bi-branch dynamic graph convolutional network and an adaptive self-attention (ASA) unit to extract shallow and deep features. Simultaneously, a dynamically changing adjacency matrix promotes the interaction and feature fusion between the two branches. The fused FC and EC features are then used for ASD patient classification.

[0006] The main idea behind this invention is as follows: On the one hand, the functional connectivity network (FC network) and effector connectivity network (EC network) of the brain describe complex behavioral patterns. FC networks describe statistical correlations between brain cells, while EC networks emphasize directional causal relationships. Integrating and utilizing the complementary information of these two networks can uncover further biomarkers related to brain diseases. On the other hand, using data from multiple segmentation templates allows for a more comprehensive utilization of brain information, mitigating the problems of limited data volume and data bias caused by using only a single template, thereby improving the accuracy of brain disease classification.

[0007] A brain disease classification method based on the fusion of multi-template brain functional and effector networks includes the following steps:

[0008] Step 1, Data Acquisition: To verify the effectiveness of the model proposed in this invention, experiments will be conducted on two segmentation templates of the ABIDE I dataset, which is commonly used for brain disease classification, to evaluate the model's brain disease classification performance.

[0009] Step 2, Multi-template data fusion: First, calculate the cosine similarity matrix using the time series of two templates, and then use the cosine similarity matrix as weights to fuse the multi-template data.

[0010] Step 3, FC Network Construction: A two-branch spatiotemporal graph convolutional network is used to reassemble the time series of the two fused templates. Simultaneously, after each layer of the graph convolutional network, the cosine similarity matrix is ​​used again to fuse the data from the two templates. Finally, the Pearson correlation coefficient of the reassembled time series is calculated and subjected to Fisher transform to obtain the FC network, which is stored in the adjacency matrix.

[0011] Step 4, EC Network Construction: The two fused time series obtained in Step 3 are added together as input, and the generated time series is obtained by using a variational autoencoder parameterized by a graph neural network architecture. The EC network is obtained by iterating by minimizing the residual between the input and the generated time series.

[0012] Step 5, FC and EC network fusion: The FC and EC networks are processed separately through a two-branch dynamic neural network and an adaptive self-attention unit to extract shallow and deep features, respectively. Simultaneously, the two branches influence each other through a dynamic adjacency matrix. The features obtained from the two branches are summed and then processed again through ASA to obtain the fused features of FC and EC, which are used for brain disease classification.

[0013] Compared with the prior art, the present invention has the following obvious advantages and beneficial effects;

[0014] (1) A novel brain disease classification method based on the fusion of multi-modal brain function and effect networks is proposed, which has higher brain disease classification accuracy compared with most machine and deep learning methods.

[0015] (2) The dual-branch functional effect connection construction method designed in this invention can accurately identify and construct brain functional and effect connection networks.

[0016] (3) The functional effect fusion method based on adaptive self-attention designed in this invention can effectively fuse two types of network data, thereby effectively improving the problem that the accuracy of ASD analysis of rs-fMRI data is still low.

[0017] (4) Experimental results on real rs-fMRI data show that the brain disease classification accuracy of the present invention is high and has good interpretability in the ASD task. Attached Figure Description

[0018] Figure 1 : Flowchart of the training process for the model involved in this method.

[0019] Figure 2 Important brain regions learned by MaFECAF on the ABIDE I real dataset. Detailed Implementation

[0020] The specific implementation methods and detailed steps of the present invention are described below. The specific implementation process of the present invention is as follows: Figure 1 As shown, it specifically includes:

[0021] (Step 1) Data acquisition.

[0022] To verify the effectiveness of the model proposed in this invention, it was evaluated using ABIDE I, a commonly used evaluation method.

[0023] Experiments were conducted on a publicly available dataset for the ASD brain network classification task, and the experimental data can be downloaded free of charge from ABIDE (nitrc.org). It includes functional magnetic resonance imaging (fMRI) and phenotypic data from 1112 participants from 17 international sites, including 539 patients with autism and 573 healthy controls. The raw fMRI data were preprocessed using the DPARSF procedure. This included removing the initial five time points, performing slice timing correction, processing head movements exceeding 2 degrees, eliminating data from subjects with horizontal head movements exceeding 2 mm, and performing sequence image registration, smoothing, and filtering. Finally, the brain was segmented into multiple regions of interest (ROIs) with different functions using brain segmentation templates, and the average time series of each ROI was extracted. In this invention, two brain segmentation templates were selected: AAL90 and CC200, because AAL90 is the most widely used template for brain region extraction. The AAL template has 116 ROIs available for the whole brain and 90 cerebellar ROIs. The CC200 template, obtained by Cradock using spatially constrained spectral clustering, contains more brain regions and can provide more comprehensive and complete brain information.

[0024] (Step 2) Multi-template data fusion.

[0025] Each participant had two time-series data points derived from two splitting templates.

[0026]

[0027] The weights for merging the two templates are obtained. Then, a new template is created that incorporates the weights from the second template. The data can be obtained using Formula 2. Similarly, It was also obtained through similar processing. .

[0028]

[0029] in, Used to adjust the blending ratio of two templates.

[0030] (Step 3) FC network construction.

[0031] This invention utilizes a two-branch spatiotemporal graph convolutional network to reconstruct data from two templates. Simultaneously, a cosine similarity matrix is ​​used to enhance the fusion of the two template data within the network. Subsequently, a fully connected (FC) network is constructed using Pearson correlation coefficients and Fisher transform. In the spatiotemporal graph convolutional network, the time series data of brain regions are used as nodes in the graph, and the connectivity relationships between brain regions calculated using Pearson correlation coefficients are used as edges. The computation process of the entire module is as follows:

[0032]

[0033] in, and These represent temporal graph convolution and spatial graph convolution, respectively, both involving the aggregation of features from neighboring nodes to update information for each region of the brain. For -th The brain region update process is as follows:

[0034] in Represents space, Representing convolutional networks -th layer, express -th A set of neighboring nodes in a brain region. After passing through a spatiotemporal graph convolutional network, the original time series can be reconstructed to obtain time series data with new features of the brain region. Then, an FC network is obtained by calculating the Pearson correlation coefficient and Fisher transform.

[0035] (Step 4) EC network construction.

[0036] In this invention, in order to obtain nonlinear causal relationships between data, a graph neural network architecture is added to the linear structural equation model:

[0037]

[0038] Among them, nonlinear changes , Representing a graph neural network variation, in this architecture, the weighted adjacency matrix A can be learned along with other neural network parameters, and each variable can be considered as either an input or output of a GNN. Then, a variational autoencoder framework is used, through:

[0039]

[0040] The corresponding decoder from China Reconstruction get Finally, by minimizing the generation and reality Learn from the differences between them Potential characteristics Within this framework, input It is obtained through step two ( Therefore, A is... The characteristic distribution is also the causal relationship between brain regions after multi-modal data fusion (EC network).

[0041] (Step 5) FC EC network convergence.

[0042] After steps three and four, a brain FC and EC network stored in matrix form was constructed for each subject. The two brain networks were implemented using a two-branch dynamic graph convolutional network (DGCN) and an adaptive self-attention unit (ASA), respectively, where the dynamic adjacency matrix facilitates information exchange between the two branches. Subsequently, the feature maps learned from these two branches were superimposed, and the fused features of the FC and EC networks were extracted again using the ASA unit. This part of the computation can be performed as follows:

[0043]

[0044] DGCN consists of two synchronized graph convolutional networks and a depth-encoded adjacency matrix. First, randomly initialize a matrix. Then The output is expanded and encoded through two fully connected layers, and finally reshaped into a matrix form. In two synchronous graph convolutional networks, FC and EC networks are used as graph nodes, respectively, and both are... As edges in a graph. In graph convolutional networks, each time a node aggregates neighbor information around it through an edge and updates its own data, FC and EC networks can... Influence the other party, thereby uncovering the inherent connections between FC and EC networks.

[0045] The framework of the self-attention mechanism includes three weight matrices: a query weight matrix and a query weight matrix. A key weight matrix and a value weight matrix In this invention, the query will be performed. ,key ,value Defined as: It can be and Meanwhile, the self-attention mechanism in the Transformer lacks sensitivity to positional information in the input sequence. This invention takes into account the regional characteristics of the brain and the correlation between brain regions, and introduces... In the computation of the self-attention mechanism:

[0046]

[0047] In Equation (8), the global spatial connectivity weights of the brain are evaluated to obtain the spatial features of the ROI. Furthermore, considering the significant differences in data from different sites, a multi-layer residual perceptron network is used to dynamically adjust the data processed by ASA, so that more targeted discriminative features can be obtained for different subjects. Finally, the output features are processed through two fully connected layers and a softmax layer to complete the classification of ASD brain disease. The entire model is an end-to-end architecture, optimized using the Adam optimizer.

[0048] To fully verify the superiority of this method, comparative experiments were conducted on the real ABIDE I dataset, comparing it with RandomForest (RF), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) methods on two segmentation templates. The results were evaluated using six metrics: accuracy, recall, precision, specificity, F1 score, and AUC. These six metrics are widely used in evaluating brain network classification. The final results are presented in Table 1 as the mean ± standard deviation of ten cross-validations.

[0049] Table 1: Experimental results of various methods on the ABIDE I dataset

[0050] As shown in the table above, the model of this invention significantly outperforms other comparative methods across all five evaluation metrics. Specifically, based on the AAL template, the proposed model outperforms the best-performing CNN by approximately 9%, and on the CC200 template, it outperforms CNN by approximately 8%. Furthermore, the proposed method exhibits a smaller standard deviation, indicating more stable performance and the ability to adaptively adjust to data from different sites. In the proposed MaFECAF method, the metrics after fusing data from two templates are higher than those using only a single template, demonstrating the effectiveness of the two-template fusion step in the proposed method.

[0051] To further verify the interpretability of the proposed model, this invention uses the feature map output by MaFECAF. The importance of brain regions is assessed using weights. This invention first uses the results of an average of ten cross-validations, then sorts them in reverse order to obtain the top 10 ROIs with the highest weights. The top 10 important brain regions under two brain segmentation templates are as follows: Figure 2As shown, in the AAL90 template, the top 10 important brain regions are the dorsolateral superior frontal gyrus (SFGdor.L), the pericalcaneal cortex (CAL.L, CAL.R), the lingual gyrus (LING.L, LING.R), the supraoccipital gyrus (SOG.L), the paracentral lobule (PCL.R), the lentiform putamen (PUT.R) nucleus, and the thalamus (THA.L). The CC200 template does not have individual names because it uses a clustering-based method. Therefore, this invention uses the names of the AAL brain regions with the highest percentage in the region clustering to name the important brain regions in the CC200 template, including the middle frontal gyrus (MFG.L), the supplementary motor area (SMA.L), the medial superior frontal gyrus (SFGmed.L, SFGmed.R), the insula (INS.R), the angular gyrus (ANG.L), and the superior temporal gyrus (STG.L). Among these, the middle frontal gyrus has been previously confirmed by researchers to be associated with attention deficit. Abnormal brain activity associated with ASD is found in the middle temporal gyrus and the cortex surrounding the calcarine fissure. The angular gyrus (ANG.L.), the visual-language center, the lingual gyrus (LING.L., LING.R.), the area for processing visual memory, and the lenticular putamen (PUT.R.), which influences interests and hobbies, are consistent with the three typical symptoms of ASD: social impairment, repetitive behavioral patterns, and limited interests.

[0052] The above experiments show that the model MaFECAF proposed in this invention has higher accuracy in classifying ASD patients using rs-fMRI data compared with other methods, and therefore has great application prospects in computer-aided analysis of brain diseases.

Claims

1. A brain disease classification method based on multi-template brain function and effect network fusion, characterized in that, Includes the following steps: Step 1, Data Acquisition: Experiments were conducted using two brain segmentation templates in the ABIDEI dataset for brain disease classification; Step 2, Multi-template data fusion: First, the cosine similarity matrix is ​​calculated using time series data from two brain templates, and then the cosine similarity matrix is ​​used as the weight to fuse the multi-template data; Step 3, Construction of Functionally Connected (FC) Network: A two-branch spatiotemporal graph convolutional network is used to reassemble the time series of the two fused templates respectively. At the same time, after each layer of the graph convolutional network, the cosine similarity matrix is ​​used to fuse the data of the two templates again. Finally, the Pearson correlation coefficient of the reassembled time series is calculated and subjected to Fisher transform to obtain the functionally connected (FC) network, which is stored in the adjacency matrix. Step 4, constructing the effect connection EC network: The two fused time series obtained in Step 2 are added together as input, and the generated time series is obtained by using a variational autoencoder parameterized by a graph neural network architecture. The effect connection EC network is obtained by iterating by minimizing the residual between the input and the generated time series. Step 5: Fusing the Functional Connectivity (FC) Network and the Effective Connectivity (EC) Network: The FC network and the EC network are processed by a two-branch dynamic neural network and an adaptive self-attention unit to extract shallow and deep features, respectively. At the same time, the two-branch dynamic neural network influences each other through a dynamic adjacency matrix. The features obtained from the two-branch dynamic neural network are added together and passed through the adaptive self-attention unit again to obtain the fused features of the FC network and the EC network, which are used for brain disease classification. In step two, multi-template data fusion, each participant possesses two time-series data points from two brain templates; through... ； Two template fusion weights are obtained; then a new fusion second template is obtained Data is obtained by formula (2); By analogy, By analogy, ; ； wherein, for adjusting the fusion ratio of the two templates.

2. The brain disease classification method based on multi-template brain function and effect network fusion according to claim 1, characterized in that, In step three, the construction of the fully connected (FC) network, in the spatiotemporal graph convolutional network, the time series data of brain regions are used as nodes in the graph, and the connectivity between brain regions calculated using the Pearson correlation coefficient is used as edges in the graph; the calculation process of the spatiotemporal graph convolutional network is as follows: ； where, and denote temporal and spatial graph convolution, respectively, and both involve aggregation of neighbor node features and update the information of each region of the brain with it; for -th brain region, the update process is: ; in Represents space, Representing convolutional networks -th layer, express -th A set of neighboring nodes in a brain region; after passing through a spatiotemporal graph convolutional network, the original time series is reconstructed to obtain time series data with new features of the brain region; then, the FC network is obtained by calculating the Pearson correlation coefficient and Fisher transform.

3. The brain disease classification method based on multi-template brain function and effect network fusion according to claim 1, characterized in that, In step four, during EC network construction, a graph neural network architecture is added to the linear structural equation model to obtain non-linear causal relationships between data. ； where the non-linear change is , representing the graph neural network changes, the weighted adjacency matrix A is learned together with other neural network parameters, each variable is considered as an input or output of a GNN; then using the framework of variational autoencoder, by: ； The corresponding decoder from China Reconstruction get Finally, by minimizing the generation and reality Learn from the differences between them Potential characteristics Within this framework, input It is obtained through step two ( Therefore, A is... The characteristic distribution is also the causal relationship between brain regions after the fusion of multi-modal data, i.e., the EC network.

4. The brain disease classification method based on multi-template brain function and effect network fusion according to claim 1, characterized in that, In step five, the FCEC network fusion is performed. After steps three and four, a brain FC and EC network stored in matrix form is constructed for each subject. The two brain networks are respectively processed by a dual-branch dynamic graph convolutional network (DGCN) and an adaptive self-attention unit (ASA), where the dynamic adjacency matrix facilitates information exchange between the two branches. Subsequently, the feature maps learned from the two branches are superimposed and the fused features of FC and EC are extracted again through the ASA unit. The calculation process is as follows: ； The DGCN consists of two synchronized graph convolutional networks and a depth-encoded adjacency matrix. First, randomly initialize a matrix. Then The output is expanded and encoded through two fully connected layers, and finally reshaped into a matrix form. ; In two synchronized graph convolutional networks, FC and EC networks are taken as graph nodes respectively, and both are taken as edges in the graph In the graph convolutional networks, FC and EC networks affect each other through influencing each other through when aggregating neighbor information around the nodes and updating their own data through edges each time, so as to mine the internal relationship between FC and EC networks. The framework of the self-attention mechanism includes three weight matrices: a query weight matrix and a query weight matrix. A key weight matrix and a value weight matrix ; will query ,key ,value Defined as: Meanwhile, the self-attention mechanism in Transformer lacks sensitivity to positional information in the input sequence; considering the regional characteristics of the brain and the correlation between brain regions, a... In the computation of the self-attention mechanism: ； In formula (8), the global spatial connectivity weights of the brain are evaluated to obtain the spatial features of the ROI; a multilayer residual perceptron network is used to dynamically adjust the data processed by ASA to obtain targeted discriminative features; The output features are processed through two fully connected layers and a softmax layer to classify ASD (Acute Myocardial Infarction) brain disease. The entire model is an end-to-end architecture and is optimized using the Adam optimizer.

Images