Autism diagnosis device based on magnetic resonance hypergraph contrast learning and storage medium

Abstract

An autism diagnosis device based on magnetic resonance hypergraph contrast learning is based on a multi-stage hypergraph contrast learning method (ASD-MSHCL), comprises a pre-training stage and a fine tuning stage, and is used for identifying autism spectrum disorder (ASD) by using limited functional magnetic resonance data. According to the multi-stage hypergraph contrast learning method (ASD-MSHCL), three wave bands of low-frequency fluctuation amplitude (ALFF) are used as node features, and Pearson correlation coefficients (PCC) of average functional magnetic resonance time sequences of different brain regions are used as edge weights; the pre-training stage focuses on self-supervised learning, combines hyper-edge perception contrast learning with adversarial training (SSL-HCLA), and extracts high-order relation features from each brain region of interest (RO I); in the fine tuning process, the pre-trained SSL-HCLA is refined, and a classification module is trained to accurately identify ASD.

Description

Technical Field

[0001] The present application relates to the field of medical rehabilitation, and in particular to an autism diagnosis device and storage medium based on magnetic resonance hypergraph contrast learning. Background Art

[0002] Autism spectrum disorder (ASD) is a developmental disorder characterized by severe challenges in social interaction, communication, and behavior. According to recent data from the Centers for Disease Control and Prevention (CDC), approximately one in 36 children is diagnosed with autism. Despite the prevalence of ASD, current diagnosis relies primarily on traditional behavioral assessments. These methods are often subjective, leading to potential delays in identifying early symptoms and an increased risk of misdiagnosis. As a result, there is a growing search in psychiatry and neuroscience for objective biomarkers of ASD. This shift aims to facilitate earlier diagnosis and timely intervention, which is critical to improving treatment outcomes.

[0003] Resting-state functional magnetic resonance imaging (rs-fMRI) is a technique that measures blood oxygen level-dependent (BOLD) signals in the brain without requiring subjects to perform specific tasks. This technique has been widely used to identify potential neuroimaging biomarkers for various psychiatric disorders, and functional connectivity networks (FCNs) explore pairwise relationships between brain regions of interest (ROIs) for each subject. FCNs help us understand brain organizational patterns and diagnose psychiatric disorders such as ASD, attention-deficit / hyperactivity disorder (ADHD), depression, bipolar disorder, and anxiety. However, previous studies usually extract handcrafted network features (such as node degree and clustering coefficient) from FCNs, which are then used in predictive models for ASD diagnosis. This approach limits the accuracy of diagnosis and relies heavily on expert knowledge. Recently, a large number of machine learning models have been applied to FCN analysis for the automatic diagnosis of brain diseases, such as support vector machines (SVMs) for distinguishing multiple system atrophy from Parkinson's disease, logistic regression for early detection of Alzheimer's disease, and deep neural networks (DNNs) for identifying schizophrenia. However, these models often vectorize the FCN matrix, thus ignoring the important topological structure information between ROIs. To reveal discriminative patterns in brain networks, FCNs should be viewed as irregular graph-structured data and analyzed as a whole, where each ROI represents a vertex and its associated FC profile serves as the vertex feature.

[0004] To this end, brain disease recognition can be viewed as a graph classification problem, where nodes represent different brain regions and edges represent the strength of connections between these regions. The main task of the graph classification problem is to learn representations of graph-structured data. Unlike shallow models, deep neural networks are able to learn hierarchical representations of functional brain networks, enabling them to capture complex interactions between brain regions. Ktena et al. proposed a Siamese GCN model that learns a graph similarity metric between FCNs to distinguish between autistic patients and healthy controls. Similarly, Qin et al. studied a graph embedding learning (GEL) model for diagnosing major depression. The model first learns FCN embeddings using GCN, and then applies a fully connected layer activated by a softmax function for final classification. However, functional connectivity networks (FCNs) constructed using traditional graph-based methods, such as Pearson correlation and partial correlation using graph Lasso, are limited to capturing relationships between pairs of regions of interest (ROIs), i.e., second-order relationships. This approach ignores the interactions between multiple ROIs, or high-order relationships ( Figure 1 , let V={v1,v2,v3,v4,v5,v6,v7,v8,v9,v 10} is a set of regions of interest (e.g. vertices); Figure 1 (a): FCN represented by a graph, each edge is represented by a black line segment, and each edge connects two regions of interest; Figure 1 (b) FCN represented by a hypergraph, e1 = {v7, v8, v9}, e2 = {v1, v2, v6, v7, v8, v9}, e3 = {v3, v4, v5, v6, v9}, e4 = {v4, v5, v9, v 10}).

[0005] Given the complexity of brain networks, high-order relationships may be crucial for understanding neural processes related to cognition and behavior. In addition, previous studies have not demonstrated significant improvements in brain disease diagnosis. We hypothesize that this performance gap is mainly due to two major limitations of deep neural networks: (1) lack of sufficient training data and (2) insufficient supervision.

[0006] Self-supervised learning (SSL) has gained attention as a promising approach that exploits the inherent structure of the data to provide supervision, which makes it particularly valuable in situations where labeled data is limited. Although SSL has been successful in computer vision (CV) and natural language processing (NLP), it is rarely used in brain network analysis, where self-supervised learning is more important to address the problems of limited data and insufficient supervision. Therefore, it is crucial to consider the topological structure of brain networks when developing image self-supervised learning. Two key aspects need to be addressed:

[0007] (1) How to develop a self-supervised encoder model that captures richer representations by leveraging the graph structure of brain networks?

[0008] (2) How can these captured features be used to improve the accuracy of autism diagnosis?

[0009] Application Contents

[0010] The purpose of this application is to provide an autism diagnosis device and storage medium based on magnetic resonance hypergraph contrast learning to solve the technical problems described in the background technology.

[0011] An autism diagnostic device based on magnetic resonance hypergraph contrastive learning, comprising: a self-supervised learning module, a graph classification model (fine-tuning stage); the self-supervised learning includes hyper-edge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, which classifies the extracted features, and the linear classifier distinguishes ASD and HC (Healthy Control) subjects by analyzing brain connection patterns and functional relationships.

[0012] Compared with the related art, the advantages and positive effects of the present application are: SSL-HCLA includes creating an original hypergraph and an edge-dropped version of the hypergraph for comparative learning and feature extraction. In this process, the model distinguishes the two hypergraph versions by removing edges using a trainable Bernoulli mask, which promotes the learning of invariant features. The screened features are robust to hypergraph changes and play a vital role in disease diagnosis. After training, the remaining edge weights can be analyzed as potential biomarkers. The graph classification model (in the fine-tuning stage) classifies these extracted features using a linear classifier that distinguishes ASD and HC subjects by analyzing brain connection patterns and functional relationships. The proposed model captures complex multi-node interactions through hypergraph representation and uses self-supervised learning to generate meaningful representations with limited labeled data, thereby improving diagnostic accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] These and other features and advantages thereof will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings and claims.

[0014] Brief description of the drawings:

[0015] Figure 1 Schematic diagram of the functional connectivity network constructed for traditional graph-based methods.

[0016] Figure 2 This is a structural diagram of the model of an autism diagnostic device based on magnetic resonance hypergraph contrast learning.

[0017] Figure 3 The present invention discloses an autism diagnosis device and storage medium based on magnetic resonance hypergraph contrast learning, and a distribution map thereof on the cortical surface.

[0018] Figure 4 The present invention is a schematic diagram showing the increase in connection strength of an autism diagnosis device and storage medium according to an embodiment of magnetic resonance hypergraph contrast learning. DETAILED DESCRIPTION

[0019] It should be understood at the outset that although illustrative implementations of one or more embodiments are provided below, the disclosed systems and / or methods may be implemented using any number of techniques, whether currently known or existing. The present disclosure should in no way be limited to the exemplary implementations, drawings, and techniques shown below, including the exemplary designs and implementations shown and described herein, but may be modified within the full scope of the appended claims and their equivalents.

[0020] As the instruction manual Figure 1-4 As shown:

[0021] An autism diagnostic device based on magnetic resonance hypergraph contrastive learning, comprising: a self-supervised learning module, a graph classification model (fine-tuning stage); the self-supervised learning includes hyper-edge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, which classifies the extracted features, and the linear classifier distinguishes ASD and HC subjects by analyzing brain connection patterns and functional relationships.

[0022] In this embodiment, SSL-HCLA includes creating an original hypergraph and an edge-dropped version of the hypergraph for contrastive learning and feature extraction. In this process, the model distinguishes between the two hypergraph versions by removing edges using a trainable Bernoulli mask, which promotes the learning of invariant features. The screened features are robust to hypergraph changes and play a vital role in disease diagnosis. After training, the remaining edge weights can be analyzed as potential biomarkers. The graph classification model (in the fine-tuning stage) classifies these extracted features using a linear classifier that distinguishes ASD and HC subjects by analyzing brain connectivity patterns and functional relationships. The proposed model captures complex multi-node interactions through hypergraph representation and uses self-supervised learning to generate meaningful representations with limited labeled data, thereby improving diagnostic accuracy and robustness.

[0023] Specifically, an autism diagnostic device based on magnetic resonance hypergraph contrast learning also includes a data preprocessing module, which includes several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice time, and correcting sensitivity distortion.

[0024] Specifically, the data preprocessing module involves using inter-frame displacement, global signal and mean tissue signal as covariates through confounding factor removal, which are regressed after registering the fMRI volume to the standard MNI152 space.

[0025] Specifically, an autism diagnostic device based on magnetic resonance hypergraph contrastive learning also includes: using the AAL1 atlas, the brain is divided into 116 regions. The average time series (BOLD signal) of each region is obtained by averaging the time series of all voxels in the region; then, the amplitude of the low-frequency fluctuation (ALFF) node feature is derived from the average time series, and the full functional connectivity (FC) matrix is ​​calculated using the Pearson correlation coefficient (PCC) between the paired average time series. The brain of each subject is represented as a hypergraph using the AAL116 atlas.

[0026]

[0027] Specifically, an autism diagnosis device based on magnetic resonance hypergraph contrast learning also includes constructing a brain hypergraph:

[0028] The brain hypergraph is constructed from the A116 atlas, which segments fMRI images into multiple regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes the edges of the graph; the average time series within each ROI is calculated by averaging the BOLD signals of all voxels in the region; then, the edge weight is calculated based on the Pearson correlation coefficient between the average time series of two different regions.

[0029] The node features are derived from three frequency bands of the amplitude of low-frequency fluctuations (ALFFs) in the BOLD signal: Slow-5 (0.01-0.027Hz), Slow-4 (0.027-0.073Hz), and the typical band (0.01-0.08Hz). ALFFs are defined as the total power in these low-frequency ranges and are obtained by the Fourier transform of the averaged time series. Some previous studies have demonstrated the effectiveness of using ALFFs as node features. The hypergraph constructed in this way is represented as in represents the set of nodes, ε represents the set of hyperedges, W represents the diagonal matrix of edge weights, Hypergraph g using incidence matrix Indicates that the rules for constructing this association matrix are described as follows:

[0030]

[0031] The degree of the node Define edge e j The degree is Super Edge j ∈ε is denoted as The diagonal matrices representing node degree and hypergraph degree are D v and D c Indicates. v In D, each element specifies the number of hyperedges connected to the corresponding node, and in D c In , each element represents the number of nodes associated with the corresponding hyperedge. Using the association matrix of the hypergraph The diagonal matrix D can be derived v and D c To encapsulate nodes and hyperedges.

[0032] Given a node feature X, the node embedding is then updated through the process of hyperedge convolution.

[0033]

[0034] Where X l ∈R O×d represents the node features of layer l, is a learnable parameter and σ(·) represents a nonlinear activation function.

[0035] Specifically, an autism diagnosis device based on magnetic resonance hypergraph contrast learning also includes a mask hypergraph encoder:

[0036] Masked hypergraph encoder for node embedding learning and edge embedding: constructing hypergraphs from fMRI data This hypergraph undergoes a series of transformations, including a hyperedge convolution block, feature concatenation, and an MLP layer. Each layer updates node features through a message passing process. Given The set of adjacent nodes is denoted as The message passing process is then defined as follows:

[0037]

[0038] in f k Represents a hyperedge convolution block that transforms adjacent node features and edge weights into an aggregate vector. Function g k is an MLP layer, which is a trainable function that maps the current node representation and the aggregation vector to a new representation. The message passing process is as follows:

[0039]

[0040] The above process can be written in matrix form:

[0041]

[0042] Among them, Norm u represents the normalized node features of u, the matrix H k Include The corresponding row, and is a trainable parameter, d n =3,d 1 = d. The symbol BN represents the batch normalization operation. The hypergraph module aims to learn the latent representation of node features. Then concatenate them to form the embedded edge features

[0043] Specifically, an autism diagnosis device based on magnetic resonance hypergraph contrast learning also includes a masked hypergraph autoencoder:

[0044] The edge features are input into the MLP layer to generate the parameters of the Bernoulli distribution and perform random edge sampling. The MLP layer consists of a and bias A linear layer followed by a ReLU activation function σ ReLU , and another with weight and bias The linear layer finally forms a sigmoid function that maps the output to the range (0,1). The process is mathematically defined as:

[0045]

[0046] In this way, the edge features are converted into scalars representing the parameters of the Bernoulli distribution.

[0047] Given parameter e uv , samples the edge discard indicator buv of each edge, so that b uv ~Bernoulli(b uv ), where b uv = 0 means the edge is discarded. uv :u,v∈V] forms a binary (0 or 1) Bernoulli mask B. In order to ensure the gradient back propagation, the sampling process needs to be reparameterized. The reparameterization trick of Luo et al. is used for this purpose.

[0048]

[0049] in λ acts as a temperature parameter that governs the smoothness of the reparameterized sampling function. Falling in the interval (1-e uv, 1), the expression in the sigmoid function is positive if it is in the interval (0, 1-e uv ), it is negative. Therefore, the output b uv The probability of converging to 1 is e uv As λ approaches 0, b uv Becoming more and more like a sampled binary indicator.

[0050] The Bernoulli mask is applied to the incidence matrix H by element-wise multiplication, when b uv = 0, masking specific edges. Therefore, the input hypergraph is converted to an edge-deleted hypergraph, retaining the same number of nodes. The parameter matrix e is generated from the hypergraph g = (V, E, W) uv The process is represented as e, and the entire data augmentation process can be expressed as:

[0051]

[0052] Specifically, in the pre-training stage of self-supervised adversarial contrastive learning, the input hypergraph and its enhanced version are processed by a hypergraph encoder and a projection head to extract hypergraph features. These inputs pass through a weight-sharing edge encoder and node encoder block.

[0053] To train the parameters of the feature extractor and hypergraph enhancement process, a loss function is used that forces two feature vectors to be close if they come from the same hypergraph and far away if they come from different hypergraphs. This way, the trained feature extractor can retain the most important information while removing excess information from the graph.

[0054] Since training is performed in batch mode, such a loss is also applied in a batch-wise manner. Specifically, the loss is defined by the infoMax principle, and the goal is to maximize:

[0055]

[0056] Where N g represents the set of hypergraphs in a batch, |N g | denotes its cardinality. The similarity measure, denoted as sim, is defined as the cosine of the angle between the two input vectors:

[0057]

[0058] maximize This can be done by keeping all edges. To encourage the removal of more edges, the regularization term aims to encourage the removal of more edges and minimize all e uv The average value

[0059]

[0060] Finally, the overall objective function is defined as:

[0061]

[0062] Determine e and The optimal value of .

[0063] Specifically, hypergraph classification in the fine-tuning stage, in the fine-tuning stage, the pre-trained SSL-CLAH model is further optimized to enhance feature representation, especially to identify important and abnormal areas related to ASD. The refined features from the SSL-CLAH block are then input into the classification block. The final output is a probability-based decision indicating the likelihood that the subject has autism spectrum disorder (ASD) or a healthy control (HC). This decision is intuitively represented in the form of probability, thereby assisting the diagnostic process. The loss function used in the fine-tuning stage is the cross-entropy loss, defined as:

[0064]

[0065] where y p Indicates the label of the pth graph, represents the prediction for PTH subjects, is the sample size.

[0066] Another embodiment, a method for diagnosing autism based on magnetic resonance hypergraph contrastive learning, includes: a self-supervised learning stage, a graph classification model (fine-tuning stage); the self-supervised stage includes hyper-edge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, which classifies the extracted features, and the linear classifier distinguishes ASD and HC subjects by analyzing brain connection patterns and functional relationships.

[0067] Specifically, a method for diagnosing autism based on MRI hypergraph contrastive learning also includes a data preprocessing stage, which includes several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice time, and correcting sensitivity distortion.

[0068] Specifically, a method for diagnosing autism based on contrastive learning of MRI hypergraphs was proposed. Confounding factors were removed in the data preprocessing stage, which involved using inter-frame displacement, global signal and average tissue signal as covariates, and regressing them after registering the fMRI volume to the standard MNI152 space.

[0069] Specifically, a method for diagnosing autism based on magnetic resonance hypergraph contrastive learning also includes: using the AAL1 atlas to divide the brain into 116 regions; the average time series (BOLD signal) of each region is obtained by averaging the time series of all voxels in the region; then, the amplitude of the low-frequency fluctuation (ALFF) node feature is derived from the average time series, and the full functional connectivity (FC) matrix is ​​calculated using the Pearson correlation coefficient (PCC) between paired average time series; and the brain of each subject is represented as a hypergraph using the AAL116 atlas.

[0070] Specifically, a method for diagnosing autism based on magnetic resonance hypergraph contrast learning also includes constructing a brain hypergraph:

[0071] The brain hypergraph is constructed from the A116 atlas, which segments fMRI images into multiple regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes the edge of the graph; the average time series within each ROI is calculated by averaging the BOLD signals of all voxels in the region; then, the edge weight is calculated based on the Pearson correlation coefficient between the average time series of two different regions;

[0072] The node features are derived from three frequency bands of the amplitude of low-frequency fluctuations (ALFFs) in the BOLD signal: Slow-5 (0.01-0.027 Hz), Slow-4 (0.027-0.073 Hz), and the typical band (0.01-0.08 Hz); ALFFs are defined as the total power in these low-frequency ranges, obtained by the Fourier transform of the averaged time series; some previous studies have demonstrated the effectiveness of using ALFFs as node features; the hypergraph constructed in this way is represented as in represents the set of nodes, ε represents the set of hyperedges, W represents the diagonal matrix of edge weights, Hypergraph g using incidence matrix Indicates that the rules for constructing this association matrix are described as follows:

[0073]

[0074] The degree of the node Define edge e j The degree is Super Edge j ∈ε is denoted as The diagonal matrices representing node degree and hypergraph degree are D v and D c Indicates; in D v In D, each element specifies the number of hyperedges connected to the corresponding node, and in Dc In , each element represents the number of nodes associated with the corresponding hyperedge; using the association matrix of the hypergraph The diagonal matrix D can be derived v and D c To encapsulate nodes and hyperedges; given the node feature X, the node embedding is then updated through the process of hyperedge convolution;

[0075]

[0076] Where X l ∈R O×d represents the node features of layer l, is a learnable parameter and σ(·) represents a nonlinear activation function.

[0077] Specifically, a method for diagnosing autism based on magnetic resonance hypergraph contrast learning also sets a mask hypergraph encoder:

[0078] Masked hypergraph encoder for node embedding learning and edge embedding: constructing hypergraphs from fMRI data This hypergraph undergoes a series of transformations, including a hyperedge convolution block, feature concatenation, and an MLP layer; each layer updates node features through a message passing process; given The set of adjacent nodes is denoted as The message passing process is then defined as follows:

[0079]

[0080] in f k represents a hyperedge convolution block that converts adjacent node features and edge weights into an aggregate vector; function g k is an MLP layer, which is a trainable function that maps the current node representation and the aggregation vector to a new representation; the message passing process is as follows:

[0081]

[0082] The above process can be written in matrix form:

[0083]

[0084] Among them, Norm u represents the normalized node features of u, the matrix H k Include The corresponding row, and is a trainable parameter, d n =3,d 1 = d; the symbol BN represents the batch normalization operation; the hypergraph module aims to learn the latent representation of node features Then concatenate them to form the embedded edge features

[0085] Specifically, a method for diagnosing autism based on magnetic resonance hypergraph contrastive learning also includes a masked hypergraph autoencoder:

[0086] The edge features are input into the MLP layer to generate the parameters of the Bernoulli distribution and perform random edge sampling. The MLP layer includes a and bias A linear layer followed by a ReLU activation function σ ReLU , and another with weight and bias The linear layer finally forms a sigmoid function that maps the output to the range (0,1); this process is mathematically defined as:

[0087]

[0088] In this way, the edge features are converted into scalars, representing the parameters of the Bernoulli distribution;

[0089] Given parameter e uv , samples the edge discard indicator buv of each edge, so that b uv ~Bernoulli(b uv ), where b uv = 0 means the edge is discarded; the matrix [b uv :u,v∈V] forms a binary (0 or 1) Bernoulli mask B; In order to ensure the gradient back propagation, the sampling process needs to be reparameterized; the reparameterization trick of Luo et al. is used for this purpose;

[0090]

[0091] in λ acts as a temperature parameter that governs the smoothness of the reparameterized sampling function; if Falling in the interval (1-e uv , 1), the expression in the sigmoid function is positive if it is in the interval (0, 1-e uv ), it is negative; therefore, the output b uv The probability of converging to 1 is e uv ; As λ approaches 0, b uv Becoming more and more like a sampled binary indicator;

[0092] The Bernoulli mask is applied to the incidence matrix H by element-wise multiplication, when b uv= 0, masking specific edges; therefore, the input hypergraph is converted to an edge-deleted hypergraph, retaining the same number of nodes; the parameter matrix e will be generated from the hypergraph g = (V, E, W). uv The process is represented as e, and the entire data augmentation process can be expressed as:

[0093]

[0094] Specifically, an autism diagnosis method based on contrastive learning of magnetic resonance hypergraphs, self-supervised adversarial contrastive learning in the pre-training stage, the input hypergraph and its enhanced version are processed by a hypergraph encoder and a projection head to extract hypergraph features, these inputs pass through a weight-sharing edge encoder and node encoder block;

[0095] To train the parameters in the feature extractor and hypergraph enhancement process, a loss function is used that forces two feature vectors to be close if they come from the same hypergraph and far away if they come from different hypergraphs.

[0096] Specifically, an autism diagnosis method based on contrastive learning of magnetic resonance hypergraphs is proposed, where the loss is defined by the infoMax principle and the goal is to maximize:

[0097]

[0098] Where N g represents the set of hypergraphs in a batch, |N g | indicates its cardinality; the similarity measure, denoted as sim, is defined as the cosine of the angle between the two input vectors:

[0099]

[0100] maximize This can be done by keeping all edges; to encourage the removal of more edges, the regularization term aims to encourage the removal of more edges and minimize all e uv The average value

[0101]

[0102] Finally, the overall objective function is defined as:

[0103]

[0104] Determine e and The optimal value of .

[0105] Specifically, a method for autism diagnosis based on contrastive learning of magnetic resonance hypergraphs, hypergraph classification in the fine-tuning stage, in which the pre-trained SSL-CLAH model is further optimized to enhance feature representation, especially to identify important and abnormal regions related to ASD; the refined features from the SSL-CLAH block are then input into the classification block; the final output is a probability-based decision indicating the likelihood that the subject has autism spectrum disorder (ASD) or a healthy control (HC); this decision is intuitively represented in the form of probability, thereby aiding the diagnostic process; the loss function used in the fine-tuning stage is the cross entropy loss, defined as:

[0106]

[0107] where y p Indicates the label of the pth graph, represents the prediction for PTH subjects, is the sample size.

[0108] A storage medium for autism diagnosis based on magnetic resonance hypergraph contrast learning, on which a computer program (instructions) is stored, and the computer program (instructions) is used to implement the above-mentioned autism diagnosis method based on magnetic resonance hypergraph contrast learning.

[0109] The present application relates to an autism diagnosis device, method and storage medium based on magnetic resonance hypergraph contrastive learning, using data sets from the public databases Autism Brain Imaging Data Exchange I (ABIDE I) and ABIDE II. The experimental results show (from the results in Table 2) that the ABIDE I and ABIDE II data sets demonstrate the superior performance of our proposed model in ASD classification. The technical solution involved in the present application always achieves the highest indicators in all evaluation categories. For ABIDE I, the technical solution involved in the present application achieved an accuracy of 93.2%, an AUC of 93.1%, a sensitivity of 92.9%, and a specificity of 95.1%, which is significantly better than other models such as ASD-DiagNet, GCN and AWSO-DBN; similarly, in the ABIDE II data set, the technical solution involved in the present application maintained its superior performance, with an accuracy of 88.9%, an AUC of 88.8%, a sensitivity of 88.5%, and a specificity of 91.7%. These results clearly show that compared with other existing methods, the technical solution involved in the present application has made substantial improvements in accuracy, AUC, sensitivity and specificity, proving its robustness and effectiveness in ASD classification.

[0110]

[0111] The autism diagnostic device, method and storage medium based on magnetic resonance hypergraph contrast learning involved in this application provide tools for personalized diagnosis of autism and assist clinical decision-making; extensive ablation studies have confirmed the robustness of the proposed model; in addition, the explanatory results emphasize the key functional connections and brain regions related to ASD.

[0112] Model interpretation was performed using a gradient-based saliency map approach to identify important brain regions and connections that aided in classification. To analyze edge features, the edge feature gradients of the ASD group were transposed and averaged to generate edge-symmetric saliency maps. Increased edge features in ASD compared to HC are represented by positive gradients, while decreases are represented by negative gradients. The top 66 connections with the highest absolute values ​​(top 1%) were identified from a total of 6670 connections (0.5×116×115). These connections are shown in Figure 3. Figure 3 (a) shows their distribution on the cortical surface. Figure 3 (b) as shown.

[0113] like Figure 3 As shown, connections with increased strength in ASD are represented by darker lines, while connections with decreased strength are represented by lighter lines. In the ABIDE I dataset, regions such as SFG, SFGmed, orbitofrontal colliculus, and OLF are primarily connected via positive connections to the parietal cortex, cerebellum, and temporal lobe. In contrast, the orbitofrontal cortex and MFG are primarily negatively connected to the cerebellum, temporal cortex, and limbic system (PHG and AMYG). In the ABIDE II data, the connections are more dispersed, with most showing increased strength ( Figure 4 (a,b)).

[0114] like Figure 3 Shown: Model interpretability on ABIDE I. Figure 3 (a): All 66 selected functional connectivity features are shown in this ring diagram; Figure 3 (b) 66 functional connections of the cerebral cortex, where the darker and lighter lines represent the enhanced and weakened functional connection strengths in ASH and normal control groups, respectively; Figure 3 (c) Enhanced node features (left) and weakened node features (right) in ASD and healthy controls. Different shapes of markers represent Slow-5, Slow-4, and classic ALFF features. Abbreviations: R, right; L, left; MFG, middle frontal gyrus; ORBmid, middle orbital gyrus; ROL, Rolandic operculum; SFGmed, medial superior frontal gyrus; ORBsupmed, medial supraclavicular gyrus; DCG, dorsal cingulate gyrus; ANG, angular gyrus; PCUN, precuneus; STG, superior temporal gyrus; MTG, middle temporal gyrus; ITG, inferior temporal gyrus.

[0115] like Figure 4Shown: Model interpretability on ABIDE II. Figure 4 (a): All 66 selected functional connectivity features are shown in this ring diagram; Figure 4 (b) 66 functional connections of the cerebral cortex, where the darker and lighter lines represent the enhanced and weakened functional connection strengths in ASH and normal control groups, respectively; Figure 4 (c) Enhanced node features (left) and weakened node features (right) in ASD and healthy controls. Different shapes of markers represent Slow-5, Slow-4, and classic ALFF features. Abbreviations: R, right; L, left; MFG, middle frontal gyrus; ORBmid, middle orbital gyrus; ROL, Rolandic operculum; SFGmed, medial superior frontal gyrus; ORBsupmed, medial supraclavicular gyrus; DCG, dorsal cingulate gyrus; ANG, angular gyrus; PCUN, precuneus; STG, superior temporal gyrus; MTG, middle temporal gyrus; ITG, inferior temporal gyrus.

[0116] The study also found abnormalities in limbic features of the cuneus / calcarine fissure and dorsolateral prefrontal cortex (DLPFC) in ASD patients compared with the HC group. Meta-analysis showed that individuals with autism showed less prefrontal activity during perceptual tasks and greater extracortical recruitment during visual processing. The cuneus is associated with visual processing and retrieval of autobiographical memory. The DLPFC is connected to areas such as the orbitofrontal cortex, temporal lobe, and neocortex (including posterior temporal, parietal, and occipital regions), is involved in goal-directed thinking and action, and has been associated with non-social abnormalities in ASD, such as restricted and repetitive behaviors, or reduced self-referential ideas. Therefore, the DLPFC has been used as a target area for transcranial current or magnetic stimulation to treat ASD. The cerebellum is also a key node for motor learning, language comprehension, social skills, visual-spatial representation, and memory functions, which overlap with abnormal areas observed in ASD patients. It has been noted that the cerebellar region has reduced structural or functional connectivity with the dorsolateral and medial prefrontal cortices, which has been associated with increased social deficits and repetitive behaviors in ASD mouse models and ASD patients.

[0117] The embodiments of the present application may be provided as a method, a system, or a computer program product.

[0118] In a possible implementation, an embodiment of the present application provides a computer-readable storage medium, which stores a program code. When the program code runs on the computer, the computer executes the above method embodiment.

[0119] In a possible implementation, an embodiment of the present application provides a computer program product. When the computer program product is run on a computer, the computer is enabled to execute the above method embodiment.

[0120] Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0121] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0123] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these changes and variations. In the various embodiments of the present application, if there is no special explanation and logical conflict, the terms and / or descriptions between the various embodiments are consistent and can be referenced to each other, and the technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationships.

Claims

1. An autism diagnostic device based on magnetic resonance hypergraph contrast learning, comprising: Self-supervised learning module, graph classification model (fine-tuning stage); self-supervised learning includes super edge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, which classifies the extracted features and distinguishes ASD and HC subjects by analyzing brain connection patterns and functional relationships.

2. The autism diagnostic device based on magnetic resonance hypergraph contrast learning according to claim 1 also includes a data preprocessing module, which includes several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice time, and correcting sensitivity distortion.

3. According to the autism diagnostic device based on magnetic resonance hypergraph contrast learning in claim 2, the data preprocessing module involves using inter-frame displacement, global signal and average tissue signal as covariates through confounding factor removal, and regressing them after registering the fMRI volume to the standard MNI 152 space.

4. The autism diagnostic device based on magnetic resonance hypergraph contrast learning according to any one of claims 1 to 3, further comprising: The brain was segmented into 116 regions using the AAL1 atlas; the mean time series (BOLD signal) of each region was obtained by averaging the time series of all voxels in the region; Subsequently, the amplitude of low-frequency fluctuations (ALFF) node features were derived from the averaged time series, and the full functional connectivity (FC) matrix was calculated using the Pearson correlation coefficient (PCC) between pairwise averaged time series; each subject's brain was represented as a hypergraph using the AAL116 atlas.

5. The autism diagnosis device based on magnetic resonance hypergraph contrast learning according to claim 4, further comprising constructing a brain hypergraph: The brain hypergraph is constructed from the A116 atlas, which segments fMRI images into multiple regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes the edge of the graph; the average time series within each ROI is calculated by averaging the BOLD signals of all voxels in the region; then, the edge weight is calculated based on the Pearson correlation coefficient between the average time series of two different regions; The node features are derived from three frequency bands of the amplitude of low-frequency fluctuations (ALFFs) in the BOLD signal: Slow-5 (0.01-0.027 Hz), Slow-4 (0.027-0.073 Hz), and the typical band (0.01-0.08 Hz); ALFFs are defined as the total power in these low-frequency ranges, obtained by the Fourier transform of the averaged time series; some previous studies have demonstrated the effectiveness of using ALFFs as node features; the hypergraph constructed in this way is represented as in represents the set of nodes, ε represents the set of hyperedges, W represents the diagonal matrix of edge weights, Hypergraph g using incidence matrix Indicates that the rules for constructing this association matrix are described as follows: The degree of the node Define edge e j The degree is Super Edge j ∈ε is denoted as The diagonal matrices representing node degree and hypergraph degree are D v and D c Indicates; in D v In D, each element specifies the number of hyperedges connected to the corresponding node, and in D c In , each element represents the number of nodes associated with the corresponding hyperedge; using the association matrix of the hypergraph The diagonal matrix D can be derived v and D c To encapsulate nodes and hyperedges; given the node feature X, the node embedding is then updated through the process of hyperedge convolution; Where X l ∈R O×d represents the node features of layer l, is a learnable parameter and σ(·) represents a nonlinear activation function.

6. The autism diagnosis device based on magnetic resonance hypergraph contrast learning according to claim 5, further comprising a masked hypergraph encoder: Masked hypergraph encoder for node embedding learning and edge embedding: constructing hypergraphs from fMRI data This hypergraph undergoes a series of transformations, including a hyperedge convolution block, feature concatenation, and an MLP layer; each layer updates node features through a message passing process; given The set of adjacent nodes is denoted as The message passing process is then defined as follows: in f k represents a hyperedge convolution block that converts adjacent node features and edge weights into an aggregate vector; function g k is an MLP layer, which is a trainable function that maps the current node representation and the aggregation vector to a new representation; the message passing process is as follows: The above process can be written in matrix form: Among them, Norm u represents the normalized node features of u, the matrix H k Include The corresponding row, and is a trainable parameter, d n =3,d 1 = d; the symbol BN represents the batch normalization operation; the hypergraph module aims to learn the latent representation of node features Then concatenate them to form the embedded edge features 7. The autism diagnosis device based on magnetic resonance hypergraph contrastive learning according to claim 6, further comprising a masked hypergraph autoencoder: The edge features are input into the MLP layer to generate the parameters of the Bernoulli distribution and perform random edge sampling. The MLP layer includes a and bias A linear layer followed by a ReLU activation function σ ReLU , and another with weight and bias The linear layer finally forms a sigmoid function that maps the output to the range (0,1); this process is mathematically defined as: In this way, the edge features are converted into scalars, representing the parameters of the Bernoulli distribution; Given parameter e uv , discard indicator b for each edge uv Sampling is performed so that b uv ~Bernoulli(b uv ), where b uv = 0 means the edge is discarded; the matrix [b uv :u,v∈V] forms a binary (0 or 1) Bernoulli mask B; in order to ensure gradient backpropagation, the sampling process needs to be reparameterized; the reparameterization trick of Luo et al. is used for this purpose; in λ acts as a temperature parameter that governs the smoothness of the reparameterized sampling function; if Falling in the interval (1-e uv , 1), the expression in the sigmoid function is positive if it is in the interval (0,1-e uv ), it is negative; therefore, the output b uv The probability of converging to 1 is e uv ; As λ approaches 0, b uv Becoming more and more like a sampled binary indicator; The Bernoulli mask is applied to the incidence matrix H by element-wise multiplication, when b uv = 0, masking specific edges; therefore, the input hypergraph is converted to an edge-deleted hypergraph, retaining the same number of nodes; the parameter matrix e will be generated from the hypergraph g = (V, E, W). uv The process is represented as e, and the entire data augmentation process can be expressed as:

8. The autism diagnostic device based on magnetic resonance hypergraph contrastive learning according to any one of claims 2-7, the self-supervised adversarial contrastive learning in the pre-training stage, the input hypergraph and its enhanced version are processed by a hypergraph encoder and a projection head to extract hypergraph features, these inputs pass through a weight-sharing edge encoder and node encoder block; To train the parameters in the feature extractor and hypergraph enhancement process, a loss function is used that forces two feature vectors to be close if they come from the same hypergraph and far away if they come from different hypergraphs.

9. The autism diagnostic device based on magnetic resonance hypergraph contrastive learning according to claim 8, wherein the loss is defined by the infoMax principle, and the goal is to maximize: Where N g represents the set of hypergraphs in a batch, |N g | indicates its cardinality; the similarity measure, denoted as sim, is defined as the cosine of the angle between the two input vectors: maximize This can be done by keeping all edges; to encourage the removal of more edges, the regularization term aims to encourage the removal of more edges and minimize all e uv The average value Finally, the overall objective function is defined as: Determine e and The optimal value of .

10. The autism diagnosis device based on magnetic resonance hypergraph contrastive learning according to any one of claims 1-9, the hypergraph classification in the fine-tuning stage, in which the pre-trained SSL-CLAH model is further optimized to enhance feature representation, in particular to identify important and abnormal areas related to ASD; the refined features from the SSL-CLAH block are then input into the classification block; the final output is a probability-based decision indicating the likelihood that the subject has autism spectrum disorder (ASD) or a healthy control (HC); this decision is intuitively represented in the form of probability, thereby assisting the diagnostic process; the loss function used in the fine-tuning stage is the cross entropy loss, defined as: where y p Indicates the label of the pth graph, represents the prediction for PTH subjects, is the sample size.

11. A method for diagnosing autism based on magnetic resonance hypergraph contrast learning, comprising: Self-supervised learning stage, graph classification model (fine-tuning stage); the self-supervised stage includes super edge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, which classifies the extracted features and distinguishes ASD and HC subjects by analyzing brain connection patterns and functional relationships.

12. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 11 also includes a data preprocessing stage, including several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice time, and correcting sensitivity distortion.

13. According to the autism diagnosis method based on magnetic resonance hypergraph contrast learning of claim 12, the data preprocessing stage involves removing confounding factors by using inter-frame displacement, global signal and average tissue signal as covariates, and regressing them after registering the fMRI volume to the standard MNI 152 space.

14. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to any one of claims 11 to 13, further comprising: The brain was segmented into 116 regions using the AAL1 atlas; the mean time series (BOLD signal) of each region was obtained by averaging the time series of all voxels in the region; Subsequently, the amplitude of low-frequency fluctuations (ALFF) node features were derived from the averaged time series, and the full functional connectivity (FC) matrix was calculated using the Pearson correlation coefficient (PCC) between pairwise averaged time series; each subject's brain was represented as a hypergraph using the AAL116 atlas.

15. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 14, further comprising constructing a brain hypergraph: The brain hypergraph is constructed from the A116 atlas, which segments fMRI images into multiple regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes the edge of the graph; the average time series within each ROI is calculated by averaging the BOLD signals of all voxels in the region; then, the edge weight is calculated based on the Pearson correlation coefficient between the average time series of two different regions; The node features are derived from three frequency bands of the amplitude of low-frequency fluctuations (ALFFs) in the BOLD signal: Slow-5 (0.01-0.027 Hz), Slow-4 (0.027-0.073 Hz), and the typical band (0.01-0.08 Hz); ALFFs are defined as the total power in these low-frequency ranges, obtained by the Fourier transform of the averaged time series; some previous studies have demonstrated the effectiveness of using ALFFs as node features; the hypergraph constructed in this way is represented as in represents the set of nodes, ε represents the set of hyperedges, W represents the diagonal matrix of edge weights, Hypergraph g using incidence matrix Indicates that the rules for constructing this association matrix are described as follows: The degree of the node Define edge e j The degree is Super Edge j ∈ε is denoted as The diagonal matrices representing node degree and hypergraph degree are D v and D c Indicates; in D v In D, each element specifies the number of hyperedges connected to the corresponding node, and in D c In , each element represents the number of nodes associated with the corresponding hyperedge; using the association matrix of the hypergraph The diagonal matrix D can be derived v and D c To encapsulate nodes and hyperedges; given the node feature X, the node embedding is then updated through the process of hyperedge convolution; Where X l ∈R O×d represents the node features of layer l, is a learnable parameter and σ(·) represents a nonlinear activation function.

16. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 15, further comprising a masked hypergraph encoder: Masked hypergraph encoder for node embedding learning and edge embedding: constructing hypergraphs from fMRI data This hypergraph undergoes a series of transformations, including a hyperedge convolution block, feature concatenation, and an MLP layer; each layer updates node features through a message passing process; given The set of adjacent nodes is denoted as The message passing process is then defined as follows: in f k represents a hyperedge convolution block that converts adjacent node features and edge weights into an aggregate vector; function g k is an MLP layer, which is a trainable function that maps the current node representation and the aggregation vector to a new representation; the message passing process is as follows: The above process can be written in matrix form: Among them, Norm u represents the normalized node features of u, the matrix H k Include The corresponding row, and is a trainable parameter, d n =3,d 1 = d; the symbol BN represents the batch normalization operation; the hypergraph module aims to learn the latent representation of node features Then concatenate them to form the embedded edge features 17. The autism diagnosis method based on magnetic resonance hypergraph contrastive learning according to claim 16, further comprising a masked hypergraph autoencoder: The edge features are input into the MLP layer to generate the parameters of the Bernoulli distribution and perform random edge sampling. The MLP layer includes a and bias A linear layer followed by a ReLU activation function σ ReLU , and another with weight and bias The linear layer finally forms a sigmoid function that maps the output to the range (0,1); this process is mathematically defined as: In this way, the edge features are converted into scalars, representing the parameters of the Bernoulli distribution; Given parameter e uv , samples the edge discard indicator buv of each edge, so that b uv ~Bernoulli(b uv ), where b uv = 0 means the edge is discarded; the matrix [b uv :u,v∈V] forms a binary (0 or 1) Bernoulli mask B; in order to ensure gradient backpropagation, the sampling process needs to be reparameterized; the reparameterization trick of Luo et al. is used for this purpose; in λ acts as a temperature parameter that governs the smoothness of the reparameterized sampling function; if Falling in the interval (1-e uv , 1), the expression in the sigmoid function is positive if it is in the interval (0, 1-e uv ), it is negative; therefore, the output b uv The probability of converging to 1 is e uv ; As λ approaches 0, b uv Becoming more and more like a sampled binary indicator; The Bernoulli mask is applied to the incidence matrix H by element-wise multiplication, when b uv = 0, masking specific edges; therefore, the input hypergraph is converted to an edge-deleted hypergraph, retaining the same number of nodes; the parameter matrix e will be generated from the hypergraph g = (V, E, W). uv The process is represented as e, and the entire data augmentation process can be expressed as:

18. The method for diagnosing autism based on magnetic resonance hypergraph contrastive learning according to any one of claims 12-17, wherein the self-supervised adversarial contrastive learning in the pre-training stage, the input hypergraph and its enhanced version are processed by a hypergraph encoder and a projection head to extract hypergraph features, and these inputs are passed through a weight-sharing edge encoder and node encoder block; To train the parameters in the feature extractor and hypergraph enhancement process, a loss function is used that forces two feature vectors to be close if they come from the same hypergraph and far away if they come from different hypergraphs.

19. The autism diagnosis method based on magnetic resonance hypergraph contrastive learning according to claim 8, wherein the loss is defined by the infoMax principle, and the goal is to maximize: Where N g represents the set of hypergraphs in a batch, |N g | indicates its cardinality; the similarity measure, denoted as sim, is defined as the cosine of the angle between the two input vectors: maximize This can be done by keeping all edges; to encourage the removal of more edges, the regularization term aims to encourage the removal of more edges and minimize all e uv The average value Finally, the overall objective function is defined as: Determine e and The optimal value of .

20. The method for diagnosing autism based on magnetic resonance hypergraph contrastive learning according to any one of claims 11-19, wherein the hypergraph classification in the fine-tuning stage, in which the pre-trained SSL-CLAH model is further optimized to enhance feature representation, in particular to identify important and abnormal regions associated with ASD; the refined features from the SSL-CLAH block are then input into the classification block; the final output is a probability-based decision indicating the likelihood that the subject has autism spectrum disorder (ASD) or a healthy control (HC); this decision is intuitively represented in the form of probability, thereby aiding the diagnostic process; the loss function used in the fine-tuning stage is the cross entropy loss, defined as: where y p Indicates the label of the pth graph, represents the prediction for PTH subjects, is the sample size.

21. A storage medium for autism diagnosis based on magnetic resonance hypergraph contrast learning, on which a computer program (instructions) is stored, and the computer program (instructions) is used to implement the autism diagnosis method based on magnetic resonance hypergraph contrast learning as described in any one of claims 11-20.

Images