A GNN-based multimodal brain imaging method for autism spectrum disorder detection

Abstract

This invention discloses a multimodal brain imaging method for autism spectrum disorder (ASD) detection based on generative neural networks (GNNs). By introducing a limb-centric functional connectivity network, distinct from traditional feature-based methods, it mines higher-order information interactions between brain regions, accurately capturing abnormal brain function patterns in autistic patients. Simultaneously, leveraging the graph structure learning capabilities of GNNs, it integrates brain imaging data with clinical text information into a unified multimodal graph, capturing complex nonlinear relationships and achieving deep integration of multimodal features. This method effectively improves the accuracy and robustness of autism detection, providing technical support for personalized diagnosis and treatment. It solves key technical challenges in multimodal data fusion and brain network analysis, enhancing the accuracy and reliability of early ASD diagnosis.

Description

Technical Field

[0001] This invention belongs to the fields of computer technology and brain science, and in particular relates to a multimodal brain imaging method for detecting autism spectrum disorder based on GNN. Background Technology

[0002] ASD (Amyotrophic Disorder) is a common neurodevelopmental disorder characterized by social difficulties, language impairments, and repetitive behaviors. With the continuous development of medical imaging technology, brain imaging techniques such as functional magnetic resonance imaging (fMRI), structural magnetic resonance imaging (sMRI), and diffusion tensor imaging (DTI) have achieved significant results in the early diagnosis of ASD. These techniques can reveal abnormal changes in brain function and structure, providing valuable clues for the early identification of ASD. However, existing brain imaging methods mainly focus on analyzing the connection strength or functional association between nodes (i.e., brain regions), neglecting the characteristics of limbic-centric brain networks, which limits their potential as biomarkers in ASD detection. With the rapid development of deep learning and graph neural networks (GNNs), more and more research is beginning to focus on the analysis of graph structure data, especially achieving significant success in processing conventional graph data such as social networks and traffic networks. Although GNNs have achieved significant applications in multiple fields, their application in the joint analysis of brain imaging and clinical text data is still in the early stages of exploration. Currently, graph neural network (GNN) technology is mainly focused on processing conventional graph data. However, its application in the framework of fusing multimodal brain imaging data with clinical text information has not been fully explored. Existing technologies have the following shortcomings:

[0003] 1. The potential of limb-centric brain networks as biomarkers has not been fully explored: Most current brain network analysis methods focus more on node-centric functional connectivity matrices, primarily used to explore abnormal changes in the connection strength between pairs of brain regions in patients, neglecting the importance of limb-centric functional connectivity networks in information flow and interaction between brain regions. The limb, as a key carrier of information transmission in brain networks, contains interactive features; however, traditional methods are relatively weak in extracting and analyzing limb information, failing to fully explore the value of the limb as a potential biomarker. This limits the accuracy and reliability of brain network-based autism detection.

[0004] 2. Limitations of Graph Structure Analysis Methods: Existing GNN technologies are mainly limited to the analysis of conventional graph data in conventional fields. Applying GNNs to the joint analysis of brain imaging and clinical text data remains a relatively new research direction, especially in the detection of ASD, where the use of GNNs is still in the exploratory stage. Current research lacks a unified framework that can effectively combine multimodal brain imaging data and clinical text information, which limits the potential of brain imaging technology in the early detection of ASD.

[0005] Therefore, a multimodal brain imaging method based on GNN for autism spectrum disorder detection needs to be proposed to solve the above problems. Summary of the Invention

[0006] The technical problem this invention aims to solve is to provide a multimodal brain imaging method for detecting autism spectrum disorder (ASD) based on gene neural networks (GNNs). By constructing a limb-centric functional connectivity network, it deeply mines high-order information interactions between brain regions, accurately captures abnormal brain function patterns in autistic patients, and integrates brain imaging data with clinical text information into a unified multimodal graph. Utilizing the graph structure learning capability of GNNs, it captures complex nonlinear relationships and achieves deep feature fusion, thereby effectively improving the accuracy and robustness of autism detection. This invention aims to improve the accuracy and reliability of early ASD diagnosis, solve key technical challenges in multimodal data fusion and brain network analysis, and thus promote the innovative development of ASD detection methods.

[0007] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows:

[0008] A multimodal brain imaging method for detecting autism spectrum disorder based on GNN includes the following steps:

[0009] S1, Acquire fMRI data of the subjects; the data comes from publicly available datasets or individual data collected in clinical trials;

[0010] S2, preprocessing the acquired fMRI data, including denoising, spatial normalization, and temporal alignment; the preprocessing transforms the data into a time series that meets the analysis requirements;

[0011] S3, based on the extracted time series, calculates the edge-centered functional connectivity matrix;

[0012] S4. Feature selection is performed on the edge-centered functional connectivity matrix to filter out key features closely related to ASD diagnosis.

[0013] S5, construct an adjacency matrix describing edge features by combining the clinical text information of the subjects; including conducting clinical diagnosis on the subjects and obtaining text data of the patients' gender and age as the basis for constructing the adjacency matrix;

[0014] S6 takes the preprocessed fMRI features and clinical text features as input and feeds them into the GNN for joint learning and classification.

[0015] Preferably, the data preprocessing method in step S2 is as follows:

[0016] The fMRI data obtained in step S1 were preprocessed using the open-source software package C-PAC, including head motion correction, time slice correction, spatial registration, spatial smoothing, and bandpass filtering.

[0017] Furthermore, head movement correction: Although auxiliary methods were used to fix the subject's head during data acquisition, the long scanning time meant that the subject's head could still move, and even slight head movements could affect the accuracy of subsequent experimental data analysis. Therefore, the subject's brain was assumed to be a rigid body, and six head movement parameters involving translation and rotation were used to correct the data and eliminate the influence of the subject's head movements.

[0018] Furthermore, head time slice correction: the acquisition time of different slices in the fMRI data is adjusted to align the data of all slices to the same time point to eliminate the influence of scan time differences on the temporal sequence of brain activity signals; through interpolation methods, the slice data are unified to the time point of the reference slice to ensure the consistency of data temporal sequence.

[0019] Furthermore, head spatial registration: registering low-resolution functional images onto high-resolution structural images of the same subject.

[0020] Furthermore, head-space smoothing: During the acquisition process, the subject and the scanning machine generate certain interference signals, resulting in image noise in the acquired images. Therefore, image smoothing is necessary to eliminate interference signals generated by the subject and the scanning machine.

[0021] Furthermore, headband filtering: During data acquisition, noise and other uncertainties can cause high-frequency signals to appear in the acquired images, affecting subsequent analysis. Therefore, filtering of the acquired data is necessary to reduce noise interference.

[0022] The functional images were registered into the standard anatomical space MNI152 to obtain a preprocessed fMRI data matrix, the dimensions of which are:

[0023] ;

[0024] in It is the size of The matrix, It refers to the number of brain regions. It refers to the number of time points; Indicates the first The brain region is a time series, with a size of [missing information]. This indicates that the brain region is in Signals at specific points in time;

[0025] In standard space, C-PAC divides brain regions according to the AAL template and extracts the average signal of all voxels within each region of interest:

[0026] ;

[0027] Indicates the first Time series of individual elements Indicates the first Areas of interest It is the first The number of all voxels in each region is counted; ultimately, a time series matrix is ​​obtained, representing the signal changes in each brain region.

[0028] Preferably, in step S3, the method for calculating the functional connectivity matrix centered on the edge is as follows:

[0029] The average time series extracted from all brain regions are normalized using z-scores to assess the deviation of sample points from the population mean, thereby achieving data normalization.

[0030] ;

[0031] in and It is the average time series of two brain regions. and yes and Two time series Fraction;

[0032] For any two brain regions, the average time series extracted are calculated, and their element-wise product is multiplied to obtain a new set of time series, called the "marginal time series".

[0033] ;

[0034] ;

[0035] Where, Indicates brain regions and brain regions The "marginal time series" between them. Indicates brain regions and brain regions The "marginal time series" between them. and Indicates brain regions and brain regions The z-score of the average time series.

[0036] Calculate the element-wise product between any two "edge time series" to obtain the final edge-centric functional connectivity matrix:

[0037] ;

[0038] In the formula, EFCN represents the resulting edge-centered functional connectivity matrix.

[0039] Preferably, in step S4, feature selection is performed on the edge-centered functional connectivity matrix to filter out key features closely related to ASD diagnosis, including:

[0040] For the high-dimensional marginal functional connectivity matrix of the subjects, a non-replacement random sampling method is used to achieve full and uniform coverage of the feature subset; assuming that each sampling... Select from features Each feature is represented as follows:

[0041] ;

[0042] in From A set of indices randomly selected from the features. ;

[0043] The random forest algorithm is used to classify and analyze the sampled feature subset to obtain the importance ranking of each diagnostic feature:

[0044] ;

[0045] in, Indicates importance score; Indicates the first On the tree, features of The amount of impurity reduction;

[0046] The importance ranking of each feature is calculated. :

[0047] ;

[0048] Where sort represents the sorting function;

[0049] The top 10%, 20%, and 30% of features were selected and included in the final feature set to provide data support for subsequent classification and brain region localization.

[0050] Preferably, in step S6, the method of using the preprocessed fMRI features and clinical text features as input to the GNN for joint learning and classification is as follows:

[0051] The extracted final feature set is used as the node features of the patient, and combined with the text features obtained through clinical diagnosis to construct an adjacency matrix, which is used as edge features to form a multimodal graph together with the node features;

[0052] Multimodal graphs are fed into a GNN for processing, and high-level features of the brain network are captured through node embedding learning of the GNN.

[0053] GNN optimizes node representations by passing information, ultimately obtaining the embedding vector for each node. :

[0054] ;

[0055] in, It is a node In the Layer representation, It is a node The set of neighboring nodes, These are the normalization coefficients, typically the degree matrix of the adjacency matrix. Or other forms of weight, It is The weight matrix of the layer, It is an activation function;

[0056] After the GNN completes its learning, the obtained shared features are input into a fully connected classifier for ASD classification prediction. The output of the fully connected layer... Represented as:

[0057] ;

[0058] in, It is the weight matrix of the output layer. It is a bias term. It is a node The system generates predictive labels; finally, it outputs the prediction results and generates personalized diagnostic reports to support clinical decision-making.

[0059] The beneficial effects of the present invention are as follows:

[0060] 1. This invention overcomes the limitation of existing brain network analysis methods that overly rely on brain region nodes, proposing a functional connectivity network analysis method centered on the edge. The edge, as the carrier of information flow and interaction between brain regions, carries the core characteristics of brain function, which existing methods have failed to fully explore. By focusing on the dynamic information flow at the edge, this invention can more accurately capture abnormal brain function patterns in autistic patients, providing more reliable biomarkers for the early diagnosis of ASD. This innovation improves the accuracy and reliability of autism detection.

[0061] 2. This invention employs a multimodal data fusion framework based on GNN, which can effectively integrate brain imaging data and clinical text information. Through the graph structure learning capability of graph neural networks, this invention can capture the complex nonlinear relationship between brain imaging data and clinical information, making up for the limitations of traditional methods that rely on only a single data source. This framework can improve the utilization efficiency of multimodal data, thereby improving the comprehensiveness and accuracy of ASD detection.

[0062] 3. While GNNs, as a powerful tool for analyzing graph-structured data, have achieved remarkable results in areas such as social networks and transportation networks, their application in the joint analysis of brain imaging and clinical text data is still in its early stages of exploration. This solution innovatively applies GNN technology to the field of ASD detection, improving the accuracy, robustness, and generalization ability of detection by adaptively learning the complex relationships between multimodal data. This innovative application breaks through the bottleneck of graph neural networks in the medical field, providing stronger technical support for ASD detection.

[0063] 4. This invention emphasizes the role of edges in brain function transmission, explores higher-order information interactions between brain regions, and breaks through the dependence on node connection strength in traditional brain network analysis, enabling more accurate identification of brain function abnormalities in autistic patients. It also addresses the limitation of traditional methods relying on a single data source, as this framework can automatically learn complex relationships between multimodal data, improving the accuracy and robustness of ASD detection. By combining graph neural networks with a multimodal data fusion framework, it provides an accurate and robust ASD detection technology. This framework can effectively handle complex high-dimensional and heterogeneous data, improving the generalization ability of ASD detection and increasing detection accuracy. Attached Figure Description

[0064] Figure 1 This is a flowchart of the present invention;

[0065] Figure 2 This is a schematic diagram of the modeling process in an embodiment of the present invention;

[0066] Figure 3 This is a schematic diagram of the abnormal brain region located based on the edge-centered functional connectivity matrix in an embodiment of the present invention. Detailed Implementation

[0067] Example 1:

[0068] like Figure 1 As shown, a multimodal brain imaging method for detecting autism spectrum disorder based on GNN includes the following steps:

[0069] S1, Acquire fMRI data of the subjects; the data comes from publicly available datasets or individual data collected in clinical trials;

[0070] S2, preprocessing the acquired fMRI data, including denoising, spatial normalization, and temporal alignment; the preprocessing transforms the data into a time series that meets the analysis requirements;

[0071] S3, based on the extracted time series, calculates the edge-centered functional connectivity matrix;

[0072] S4. Feature selection is performed on the edge-centered functional connectivity matrix to filter out key features closely related to ASD diagnosis.

[0073] S5, construct an adjacency matrix describing edge features by combining the clinical text information of the subjects; including conducting clinical diagnosis on the subjects and obtaining text data of the patients' gender and age as the basis for constructing the adjacency matrix;

[0074] S6 takes the preprocessed fMRI features and clinical text features as input and feeds them into the GNN for joint learning and classification.

[0075] Example 2:

[0076] like Figure 2 As shown in the figure, this embodiment provides a specific modeling process for the GNN-based multimodal brain imaging autism spectrum disorder detection method. The specific process is as follows:

[0077] The data preprocessing method in step S2 is as follows:

[0078] The fMRI data obtained in step S1 were preprocessed using the open-source software package C-PAC, including head motion correction, time slice correction, spatial registration, spatial smoothing, and bandpass filtering.

[0079] Furthermore, head movement correction: Although auxiliary methods were used to fix the subject's head during data acquisition, the long scanning time meant that the subject's head could still move, and even slight head movements could affect the accuracy of subsequent experimental data analysis. Therefore, the subject's brain was assumed to be a rigid body, and six head movement parameters involving translation and rotation were used to correct the data and eliminate the influence of the subject's head movements.

[0080] Furthermore, head time slice correction: the acquisition time of different slices in the fMRI data is adjusted to align the data of all slices to the same time point to eliminate the influence of scan time differences on the temporal sequence of brain activity signals; through interpolation methods, the slice data are unified to the time point of the reference slice to ensure the consistency of data temporal sequence.

[0081] Furthermore, head spatial registration: registering low-resolution functional images onto high-resolution structural images of the same subject.

[0082] Furthermore, head-space smoothing: During the acquisition process, the subject and the scanning machine generate certain interference signals, resulting in image noise in the acquired images. Therefore, image smoothing is necessary to eliminate interference signals generated by the subject and the scanning machine.

[0083] Furthermore, headband filtering: During data acquisition, noise and other uncertainties can cause high-frequency signals to appear in the acquired images, affecting subsequent analysis. Therefore, filtering of the acquired data is necessary to reduce noise interference.

[0084] The functional images were registered into the standard anatomical space MNI152 to obtain a preprocessed fMRI data matrix, the dimensions of which are:

[0085] ;

[0086] in It is the size of The matrix, It refers to the number of brain regions. It refers to the number of time points; Indicates the first The brain region is a time series, with a size of [missing information]. This indicates that the brain region is in Signals at specific points in time;

[0087] In standard space, C-PAC divides brain regions according to the AAL template and extracts the average signal of all voxels within each region of interest:

[0088] ;

[0089] Indicates the first Time series of individual elements Indicates the first Areas of interest It is the first The number of all voxels in each region is counted; ultimately, a time series matrix is ​​obtained, representing the signal changes in each brain region.

[0090] Preferably, in step S3, the method for calculating the functional connectivity matrix centered on the edge is as follows:

[0091] The average time series extracted from all brain regions are normalized using z-scores to assess the deviation of sample points from the population mean, thereby achieving data normalization.

[0092] ;

[0093] in and It is the average time series of two brain regions. and yes and Two time series Fraction;

[0094] For any two brain regions, the average time series extracted are calculated, and their element-wise product is multiplied to obtain a new set of time series, called the "marginal time series".

[0095] ;

[0096] ;

[0097] In the formula, Indicates brain regions and brain regions The "marginal time series" between them. Indicates brain regions and brain regions The "marginal time series" between them. and Indicates brain regions and brain regions The z-score of the average time series.

[0098] Calculate the element-wise product between any two "edge time series" to obtain the final edge-centric functional connectivity matrix:

[0099] ;

[0100] In the formula, EFCN represents the resulting edge-centered functional connectivity matrix.

[0101] like Figure 3 As shown, this study uses the AAL map, and participants will receive a k-dimensional edge connectivity matrix (k=6105). The edge connectivity matrix is ​​symmetric, and to improve computational efficiency, only k elements from the lower triangle of the matrix are used in the analysis.

[0102] The edge connectivity matrix can be viewed as an extension of the node-centric functional connectivity matrix, with a significantly larger dimension. This high-dimensionality provides more supplementary information for studying the relationship between brain activity and behavior. On one hand, the edge connectivity matrix omits the final averaging step in traditional computation, thus reducing information loss. The time series of each pair of nodes is preserved at a single-frame timescale, significantly improving temporal resolution and enabling a more accurate reflection of resonant activity between brain regions. On the other hand, the edge connectivity matrix can capture correlations in the interactions of multiple brain regions, forming a high-order brain network representation based on the edges. As the dimension of the edge connectivity matrix increases, it provides richer high-order brain connectivity features, thereby revealing complex interaction patterns and cross-regional correlations between multiple brain regions.

[0103] Preferably, in step S4, feature selection is performed on the edge-centered functional connectivity matrix to filter out key features closely related to ASD diagnosis, including:

[0104] For the high-dimensional marginal functional connectivity matrix of the subjects, a non-replacement random sampling method is used to achieve comprehensive and uniform coverage of the feature subset. Specifically, in the experiment, 2000 samples are performed, with 10000 features randomly selected each time. It is assumed that each time... Select from features Each feature is represented as follows:

[0105] ;

[0106] in From A set of indices randomly selected from the features. ;

[0107] The random forest algorithm is used to classify and analyze the sampled feature subset to obtain the importance ranking of each diagnostic feature:

[0108] ;

[0109] in, Indicates importance score; Indicates the first On the tree, features of The amount of impurity reduction;

[0110] The importance ranking of each feature is calculated. :

[0111] ;

[0112] Where sort represents the sorting function;

[0113] The top 10%, 20%, and 30% of features were selected and included in the final feature set to provide data support for subsequent classification and brain region localization.

[0114] Preferably, in step S6, the method of using the preprocessed fMRI features and clinical text features as input to the GNN for joint learning and classification is as follows:

[0115] The extracted final feature set is used as the node features of the patient, and combined with the text features obtained through clinical diagnosis to construct an adjacency matrix, which is used as edge features to form a multimodal graph together with the node features;

[0116] Multimodal graphs are fed into a GNN for processing, and high-level features of the brain network are captured through node embedding learning of the GNN.

[0117] GNN optimizes node representations by passing information, ultimately obtaining the embedding vector for each node. :

[0118] ;

[0119] in, It is a node In the Layer representation, It is a node The set of neighboring nodes, These are the normalization coefficients, typically the degree matrix of the adjacency matrix. Or other forms of weight, It is the first The weight matrix of the layer, It is an activation function;

[0120] After the GNN completes its learning, the obtained shared features are input into a fully connected classifier for ASD classification prediction. The output of the fully connected layer... Expressed as:

[0121] ;

[0122] in, It is the weight matrix of the output layer. It is a bias term. It is a node The system generates predictive labels; finally, it outputs the prediction results and generates personalized diagnostic reports to support clinical decision-making.

[0123] Example 3:

[0124] To objectively evaluate the effectiveness of this invention, this study used the ABIDE I public dataset for model validation, selecting data containing 863 samples that met the imaging quality standards; the data was allocated as training and validation sets in a 9:1 ratio. Compared with the traditional node-based functional connectivity matrix method, the model of this invention improved accuracy and sensitivity by 6 percentage points, as shown in Table 1 below:

[0125] Table 1: Evaluation of the advantages of limb-centered brain network analysis methods;

[0126]

[0127] Furthermore, by comparing and evaluating Support Vector Machine (SVM) and GNN, the results show that the model using GNN as the classifier improves accuracy by 14 percentage points and sensitivity by 12 percentage points compared to the model using SVM, as shown in Table 2 below:

[0128] Table 2: Classification comparison between GNN and traditional SVM;

[0129]

[0130] This invention proposes an innovative method for ASD detection, combining edge-centered brain network analysis, a multimodal data fusion framework, and GNN technology to significantly improve the diagnostic accuracy and robustness of ASD. To objectively evaluate the effectiveness of this invention, this study used the ABIDE I public dataset for model validation, selecting 863 samples that met the imaging quality standards. The data were allocated to the training and validation sets in a 9:1 ratio.

[0131] Experimental results show that, compared with the traditional node-based functional connectivity matrix method, the method of this invention improves accuracy and sensitivity by 6 percentage points, respectively. Furthermore, further evaluation and comparison show that the model using GNN as the classifier improves accuracy and sensitivity by 14 and 12 percentage points, respectively, compared with the model using SVM; as shown in Table 3 below:

[0132] Table 3: Comparison of classification performance of this model with other models;

[0133]

[0134] This method demonstrates significant technical advantages, providing stronger support for the accurate classification and early diagnosis of ASD.

Claims

1. A multimodal brain imaging method for detecting autism spectrum disorder based on GNN, characterized in that, Includes the following steps: S1, Obtain fMRI data from the subject, which may be derived from publicly available datasets or individual data collected in clinical trials; S2, preprocesses the acquired fMRI data, including denoising, spatial normalization and temporal alignment; Preprocessing transforms the data into a time series that meets the analytical requirements. S3, based on the extracted time series, calculates the edge-centered functional connectivity matrix using the following method: The average time series extracted from all brain regions are normalized using z-scores to assess the deviation of sample points from the population mean, thereby achieving data normalization. ； in and It is the average time series of two brain regions. and yes and Two time series Fraction; For any two brain regions, the average time series extracted are calculated, and their element-wise product is multiplied to obtain a new set of time series, called the "marginal time series". ； ； Where, Indicates brain regions and brain regions "Marginal time series" between; Indicates brain regions and brain regions "Marginal time series" between; and Indicates brain regions and brain regions The z-score of the average time series; Calculate the element-wise product between any two "edge time series" to obtain the final edge-centric functional connectivity matrix: ； In the formula, EFCN represents the resulting edge-centered functional connectivity matrix; S4. Feature selection is performed on the edge-centered functional connectivity matrix to filter out key features closely related to ASD diagnosis. S5, construct an adjacency matrix describing edge features by combining the clinical text information of the subjects; including conducting clinical diagnosis on the subjects and obtaining text data of the patients' gender and age as the basis for constructing the adjacency matrix; S6 takes the preprocessed fMRI features and clinical text features as input and feeds them into the GNN for joint learning and classification.

2. The method for detecting autism spectrum disorder based on GNN multimodal brain imaging according to claim 1, characterized in that, The data preprocessing method in step S2 is as follows: The fMRI data obtained in step S1 were preprocessed using the open-source software package C-PAC, including head motion correction, time slice correction, spatial registration, spatial smoothing, and bandpass filtering. The functional images were registered into the standard anatomical space MNI152 to obtain a preprocessed fMRI data matrix, the dimensions of which are: ； in It is the size of The matrix, It refers to the number of brain regions. It refers to the number of time points; Indicates the first The brain region data is a time series, with a size of [missing information]. This indicates that the brain region is in Signals at specific points in time; In standard space, C-PAC divides brain regions according to the AAL template and extracts the average signal of all voxels within each region of interest: ； Indicates the first Time series of individual elements Indicates the first Areas of interest It is the first The number of all voxels in each region is counted; ultimately, a time series matrix is ​​obtained, representing the signal changes in each brain region.

3. The method for detecting autism spectrum disorder based on GNN multimodal brain imaging according to claim 1, characterized in that, In step S4, feature selection is performed on the edge-centered functional connectivity matrix to filter out key features closely related to ASD diagnosis, including: For the high-dimensional marginal functional connectivity matrix of the subjects, a non-replacement random sampling method is used to achieve full and uniform coverage of the feature subset; assuming that each sampling... Select from features Each feature is represented as follows: ； in It is from A set of indices randomly selected from the features. ; The random forest algorithm is used to classify and analyze the sampled feature subset to obtain the importance ranking of each diagnostic feature: ； in, Indicates importance score; Indicates the first On the tree, features of The amount of impurity reduction; The importance ranking of each feature is calculated. : ； Where sort represents the sorting function; The top 10%, 20%, and 30% of features were selected and included in the final feature set to provide data support for subsequent classification and brain region localization.

4. The method for detecting autism spectrum disorder based on GNN multimodal brain imaging according to claim 3, characterized in that, In step S6, the method of feeding the preprocessed fMRI features and clinical text features as input into the GNN for joint learning and classification is as follows: The extracted final feature set is used as the node features of the patient, and combined with the text features obtained through clinical diagnosis to construct an adjacency matrix, which is used as edge features to form a multimodal graph together with the node features; Multimodal graphs are fed into a GNN for processing, and high-level features of the brain network are captured through node embedding learning of the GNN. GNN optimizes node representations by passing information, ultimately obtaining the embedding vector for each node. : ； in, It is a node In the Layer representation, It is a node The set of neighboring nodes, These are the normalization coefficients, typically the degree matrix of the adjacency matrix. Or other forms of weight, It is The weight matrix of the layer, It is an activation function; After the GNN completes its learning, the obtained shared features are input into a fully connected classifier for ASD classification prediction. The output of the fully connected layer... Represented as: ； in, It is the weight matrix of the output layer. It is a bias term. It is a node The system generates predictive labels; finally, it outputs the prediction results and generates personalized diagnostic reports to support clinical decision-making.

Images