A Neuro-typing Method for ASD Based on Synthetic Brain Map Data Augmentation

Abstract

This invention belongs to the field of deep learning clustering, specifically relating to an ASD neural typing method based on synthetic brain map data augmentation. The method includes: acquiring brain map data for training and inputting it into a pre-trained dual-decoder graph autoencoder to obtain latent embeddings; inputting the latent embeddings into a pre-trained latent space conditional diffusion model to obtain an augmented brain map; training a deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE using the augmented brain map to obtain a trained deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE; acquiring brain map data to be detected and inputting it into the trained deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE to obtain the ASD neural typing result. This invention effectively solves the gender imbalance and multi-site heterogeneity problems existing in ASD data, enhancing the robustness and generalization ability of the model.

Description

Technical Field

[0001] This invention belongs to the field of deep learning clustering, specifically involving an ASD neural typing method based on synthetic brain map data augmentation. Background Technology

[0002] Autism Spectrum Disorder (ASD) is a highly hereditary and heterogeneous neurodevelopmental disorder. Accurate neurological subtyping is crucial for early diagnosis and targeted intervention. Current technologies primarily utilize neuroimaging techniques, such as structural magnetic resonance imaging (sMRI) and resting-state functional magnetic resonance imaging (rs-fMRI), to analyze the brain structure and function of ASD patients in order to identify their neurological subtypes.

[0003] Existing methods for neural typing using ASD neural typing models typically improve model performance by providing more high-quality samples with imbalanced data. However, current neural typing methods still suffer from the following major drawbacks:

[0004] 1. Schema collapse may occur when processing heterogeneous data;

[0005] 2. ASD has lower accuracy, robustness, and generalization when the sample size is small and the gender distribution is extremely imbalanced;

[0006] 3. ASD neural subtyping models without good data support do not perform well.

[0007] In summary, there is an urgent need for a new ASD neural typing method that can avoid the pattern collapse problem that synthetic methods may encounter when dealing with heterogeneous data, provide more high-quality samples for ASD neural typing models, and solve problems such as gender imbalance in ASD samples, thereby improving the accuracy of ASD neural typing. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes an ASD neurotyping method based on synthetic brain map data augmentation, which includes:

[0009] S1: Obtain the brain map data for training and input it into the pre-trained dual-decoder graph autoencoder to obtain the latent embedding;

[0010] S2: Input the latent embedding into a pre-trained latent space conditional diffusion model to obtain an enhanced brain map;

[0011] S3: Using augmented mind mapping to train a deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE, a trained deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE is obtained.

[0012] S4: Obtain the brain map data to be detected and input it into a trained deep embedding clustering model based on MBVAE or a deep map clustering model based on HGDAE to obtain the ASD neuromorphic classification results.

[0013] Preferably, the dual-decoder graph autoencoder includes a graph encoder, a first decoder, and a second decoder, which are composed of a layered GCN network or a GAT network. The graph encoder processes the mind map to obtain the latent embeddings. The first decoder performs an inner product operation on the embedding representation to obtain the reconstructed graph structure. The second decoder processes the embedding representation to obtain the reconstructed node features.

[0014] Furthermore, the loss function for pre-training the dual-decoder graph autoencoder is expressed as:

[0015]

[0016]

[0017]

[0018]

[0019] in, This represents the total loss of the dual-decoder graph autoencoder. This represents the loss of reconstructing the graph structure. This represents the loss of reconstructed node features. Indicates prior loss. The weights represent the loss of reconstructing the graph structure. Represents the feature loss weights of the reconstructed nodes. In the mind map representing modality m, the element in the i-th row and k-th column of the original adjacency matrix is... In a mind map representing modality m, reconstruct the elements in the i-th row and k-th column of the adjacency matrix. Indicates the number of nodes. The original node features of the brain map representing modality m. Represents the reconstructed node features of the mind map for modality m. The reconstructed adjacency matrix of the mind map for modality m. This represents the adjustment coefficient. Describing the L1 norm, Represents the matrix nuclear norm.

[0020] Preferably, the loss function for pre-training the latent space conditional diffusion model is expressed as:

[0021]

[0022] in, This represents the total loss of the latent space conditional diffusion model. Representing latent space variables The distribution, This represents the noise actually added to the data. Follows a standard normal distribution. Represents phenotypic information, This represents the noise predicted by the denoising U-Net model. Let represent the hidden variable state at time step t during the diffusion process. Indicates the time step of the diffusion process. Indicates a condition encoder. This indicates the expected operation.

[0023] Preferably, the process of training a deep embedding clustering model based on MBVAE includes:

[0024] The MBVAE model was pre-trained using augmented mind mapping to obtain the ASD embedding of the MBVAE model;

[0025] The readout operation is used to process the ASD embeddings of the MBVAE model to obtain a graph-level representation;

[0026] The similarity between the graph representation and the cluster centers is calculated using the Student t-distribution, and the clustering probability of the samples corresponding to the graph representation is obtained, which is the neural typing result.

[0027] The divergence loss is calculated based on the clustering probability of the samples. The weighted sum of the divergence loss and the total loss of the MBVAE model is used as the total loss of the deep embedding clustering model. The parameters of the MBVAE-based deep embedding clustering model are adjusted according to the total loss of the deep embedding clustering model to obtain the trained deep embedding clustering model.

[0028] Furthermore, the MBVAE model consists of two encoders and a shared decoder, with the two encoders processing the enhanced brain map and the original brain map, respectively.

[0029] Preferably, the HGDAE-based deep graph clustering model includes an HGDAE model and a clustering module. The HGDAE model consists of two encoders and one decoder. The process of training the MBVAE-based deep embedding clustering model includes:

[0030] An enhanced mind map was used to pre-train the HGDAE model; the pre-trained HGDAE model was then fine-tuned together with the clustering module to obtain a trained HGDAE-based deep graph clustering model.

[0031] Furthermore, the loss function for training the HGDAE-based deep graph clustering model is expressed as:

[0032]

[0033]

[0034]

[0035] in, This represents the total loss of the HGDAE-based depth map clustering model. The input is an augmented mind map. The reconstruction loss of the HGDAE-based depth graph clustering model. This represents the KL divergence regularization constraint term between the original mind map distribution and the enhanced mind map distribution. The KL divergence regularization constraint term represents the relationship between the distribution of the original mind map and the distribution of the enhanced mind map in the clustering feature space Z. Indicates enhanced mind mapping, Represents the original mind map. This indicates the calculation of KL divergence. This represents the true data distribution of the original mind map G. Enhanced mind mapping After feature transformation, the probability distribution in the cluster feature space Z is estimated; Enhanced mind mapping The probability distribution in the original feature space is estimated after feature transformation.

[0036] The beneficial effects of this invention are as follows:

[0037] This invention leverages the high-quality and diverse sample generation capabilities of conditional diffusion models to synthesize complex brain maps with biological significance, providing more high-quality samples for ASD neural subtyping models. This addresses issues such as gender imbalance in ASD samples and improves the performance and robustness of deep clustering-based ASD neural subtyping models. The subtype-specific synthesis method employed in this invention helps resolve the pattern collapse problem encountered by most existing synthesis methods when dealing with heterogeneous data. This invention effectively solves the gender imbalance and multi-site heterogeneity issues in ASD data, enhancing the model's robustness and generalization ability. It demonstrates high accuracy in subtyping results when learning from few samples, showing promising application prospects. Attached Figure Description

[0038] Figure 1 This is a flowchart of the ASD neuromorphic classification method based on synthetic brain map data augmentation in this invention;

[0039] Figure 2 This is a flowchart of the enhanced mind map generation process in this invention;

[0040] Figure 3This is a flowchart of the optimization process for the data-augmented mind map-level deep clustering model in this invention;

[0041] Figure 4 This is a flowchart illustrating the optimization process of the data-augmented population graph-level deep clustering model in this invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] This invention proposes a neurogenotyping method for ASD based on synthetic brain map data augmentation, such as... Figure 1 As shown, the method includes the following:

[0044] S1: Obtain the brain map data for training and input it into the pre-trained dual-decoder graph autoencoder to obtain the latent embedding.

[0045] For acquiring brain mapping data for training, the ABIDE public dataset is preferred. This dataset primarily uses data provided by the ABIDE public dataset for ASD neural subtype analysis. ABIDE is an international multi-site dataset designed to reveal the neurological heterogeneity and pathophysiological mechanisms of ASD. The Phase I ABIDE dataset contains data from 539 ASD subjects and 573 NC (normal controls) subjects from 17 different scanning sites. The Phase II ABIDE dataset contains data from 521 ASD subjects and 593 NC subjects from 19 different scanning sites. ABIDE provides sMRI and rs-fMRI imaging data of these 2226 subjects (1060 ASD patients and 1166 NC patients), as well as rich clinical assessment information. In addition, a self-collected clinical dataset (planned to collect data from 100 ASD patients and 100 normal controls) can be prepared according to project needs.

[0046] First, the dual-decoder graph autoencoder (DGAE) is pre-trained using mind map data. Specifically:

[0047] DGAE includes a graph encoder. (subscript) (Identifier of mind map instance) First decoder Second decoder Mind map for each modality m ,in Let T be a set of nodes; Let be the set of edges; It is a picture The adjacency matrix, Non-zero elements in the array represent an edge between two nodes; It is the feature matrix of the nodes, and the feature vector of each node is defined as its correlation with all other nodes (which can be obtained from...). (obtained from) Indicates a mode.

[0048] Graph encoders consist of layered GCN networks (graph convolutional networks) or GAT networks (graph attention networks). Graph encoders process brain maps to obtain latent embeddings. The first decoder is an inner product decoder, which performs inner product operations on the embedded representations to obtain the reconstructed graph structure. ;in The activation function is sigmoid, and cross-entropy is used as the loss for reconstructing the graph structure.

[0049]

[0050] in, In a brain map representing modality m, the element in the i-th row and k-th column of the original adjacency matrix is ​​the "real graph structure adjacency matrix element," used to characterize the actual connections between nodes (brain regions). In the mind map representing modality m, the element in the i-th row and k-th column of the reconstructed adjacency matrix is ​​the "graph structure adjacency matrix element reconstructed (predicted) by the model," used to characterize the model's prediction of connections between nodes. This indicates the number of nodes (brain regions).

[0051] Feature Decoder With encoder The structure is symmetrical and also consists of multiple layers of GCN or GAT. The second decoder processes the embedded representations to obtain the reconstructed node features. The loss for reconstructing node features is:

[0052]

[0053] Furthermore, to make the synthesized brain maps more biologically meaningful, we propose to introduce prior loss constraints on the brain map structure:

[0054]

[0055] The first term is sparsity loss, used to control the sparsity of brain networks, and the second term is low-rank loss. The matrix nuclear norm is used to control the modularity of brain networks; the hyperparameter is the regulation coefficient. Used to adjust the weights among loss terms.

[0056] A large number of NC subjects were collected from multiple public datasets to pre-train DGAE. The total loss function is:

[0057]

[0058] in, This represents the total loss of the dual-decoder graph autoencoder. This represents the loss of reconstructing the graph structure. This represents the loss of reconstructed node features. Indicates prior loss. The weights represent the loss of reconstructing the graph structure. Represents the feature loss weights of the reconstructed nodes. and These are all hyperparameters used to control different loss contributions.

[0059] DGAE's dual-decoder design utilizes graph structure and node attributes in both encoding and decoding processes, resulting in a more compact representation that is beneficial for high-quality mind map synthesis.

[0060] S2: Input the latent embedding into a pre-trained latent space conditional diffusion model to obtain an enhanced brain map.

[0061] like Figure 2 As shown, data is generated using a conditional diffusion model in the pre-trained DGAE latent space. The diffusion model consists of two Markov chain processes: a forward diffusion process... and reverse diffusion process Here This refers to the embedding features output by the DGAE encoder. The forward diffusion process gives Gradually add Gaussian noise:

[0062]

[0063] in, The transfer kernel represents the forward diffusion process, describing the transition from noise-free data. To noisy data The noise addition rules follow a Gaussian distribution; These are predefined parameters used to control the amount of noise added at each step, ensuring... Approximately a standard Gaussian distribution The reverse diffusion process aims to eliminate noise from... Gradually recover the original data :

[0064]

[0065] in, The generative model representing the backdiffusion process, with parameters predicted by a denoised U-Net model, aims to progressively... Restore , and These are based on the current noise data. The mean and variance predicted by the denoised U-Net model at time step t.

[0066] The key to training a diffusion model lies in minimizing the difference between the model-generated data and the real data. Given a mind map and phenotypic information, the loss function of the latent space conditional diffusion model is defined as:

[0067]

[0068] in, This represents the total loss of the latent space conditional diffusion model. This represents implicit space variables (latent variable representation). Take the expected value. This represents the noise actually added to the data. It follows a standard normal distribution with a mean of 0 and a variance of 1. Indicates condition, This can include phenotypic information, such as age, gender, and data collection site. This represents the noise predicted by the denoising U-Net model. This represents the hidden variable corresponding to time step t. This represents the time step in the noise-adding phase during the diffusion process. This represents the expectation operation. This represents a conditional encoder that passes the encoded phenotypic information p to the denoising U-Net model via a concatenation or cross-attention layer [QKV]. The phenotypic information-guided cross-attention layer is implemented as follows:

[0069]

[0070] in, , , . This represents the i-th intermediate layer of U-Net. , and This is a learnable projection matrix. After training the latent space conditional diffusion model based on a graph autoencoder, it is used to generate ASD mind maps, i.e., to generate augmented mind maps to achieve data augmentation.

[0071] S3: Use augmented mind mapping to train a deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE to obtain a trained deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE.

[0072] This approach effectively integrates ASD subtype-specific brain map augmentation into a deep clustering algorithm to improve the accuracy and generalization ability of the ASD neural subtyping model. Overall, augmented samples are used for feature representation learning during the pre-training phase to obtain a smoother, more realistic manifold structure. In the fine-tuning phase after adding the clustering module, clean original samples are used to establish clustering targets, which serve as labels in supervised training, monitoring the clustering process of the augmented samples. The feature update and target update steps are iteratively applied to improve clustering performance; specifically:

[0073] This invention designs two model optimization methods: data-augmented mind map-level deep clustering model optimization and data-augmented population graph-level deep clustering model optimization.

[0074] The optimization of the data-augmented mind map-level deep clustering model is achieved by optimizing the MBVAE-based deep embedding clustering model. The MBVAE-based deep embedding clustering model consists of the MBVAE model and a clustering module; after pre-training the MBVAE module, the clustering module is added for fine-tuning. For example... Figure 3 As shown, the process of training a deep embedding clustering model based on MBVAE includes:

[0075] The MBVAE model was pre-trained using augmented mind mapping to obtain the ASD embedding of the MBVAE model.

[0076] The MBVAE model consists of two encoders and a shared decoder. The two encoders process the original "clean sample" and the augmented "enhanced sample" (i.e., the original brain map and the augmented brain map), respectively. The two encoders share weights to learn consistent feature representations. The encodings of these two types of samples in the latent space are compared (…). and ), and utilize loss functions (including , , and The loss function of the MBVAE model consists of the following parts: (The text then goes on to discuss how these components work together to optimize the model, enabling it to learn more robust representations of the data.)

[0077] Variational inference ELBO loss: It can be derived that the variational lower bounds for the brain network sets of NC and ASD subjects are respectively:

[0078]

[0079]

[0080] in, This represents a normal latent variable. Indicates shared hidden variables. This represents a specific latent variable for ASD. Kullback-Leibler divergence between two probability distributions. ELBO loss is defined as the negative sum of the two ELBOs: .

[0081] Total correlation loss: To promote the independence of the distributions among the three latent variables, a total correlation loss is introduced:

[0082]

[0083] in, Latent variables The joint conditional probability distribution is obtained by minimizing... Make The product approximates the marginal distribution. This term can be approximated using the density ratio technique.

[0084] Modality alignment loss: Aligning data from different modalities in the latent space facilitates subsequent variational inference. We propose constructing a proxy task within a self-supervised learning framework to implement multimodal alignment. This task promotes alignment between modalities by regularizing the joint space of the single-modality variational posterior and penalizing the dispersion of the latent variables. This process can be quantified by the following modality alignment loss:

[0085]

[0086] Where ch represents a channel, including three channels: NC-specific, ASD-specific, and Common. It is a set of multimodal latent variables; The joint distribution of the single-modal variational posterior; It serves as a measure of the dispersion of latent variables and acts as a regularization term to promote consistency in the representation of different modalities.

[0087] Total loss of MBVAE model Represented as:

[0088]

[0089] in, For ELBO's loss, For total correlation loss, For modal alignment loss, , To control the weights of each loss term, a clustering module is added in the second stage for fine-tuning, minimizing the total loss of the clustering model. ,in For the clustering module loss, Parameters contributing to loss control.

[0090] The process of fine-tuning by adding the clustering module includes:

[0091] Clustering analysis is performed in the ASD-specific embedding latent space. To address the graph-level clustering task, the ASD embeddings of the MBVAE model are processed using a readout operation. The process is performed to obtain a graph-level representation. Readout operations can include summation, averaging, max pooling, etc.

[0092]

[0093] After the Readout operation, the cluster centers are initialized using the k-means algorithm. For the nth ASD subject and the kth cluster, the data representation of the ASD-specific embedding space is computed using the Student's t-distribution. and cluster center Similarities between them:

[0094]

[0095] in, Initialized by k-means, These are the degrees of freedom of the student's t-distribution. This represents the probability of assigning sample n to cluster k, from which the soft-label distribution of ASD samples can be obtained. Further by Define a target distribution with higher confidence. :

[0096]

[0097] The loss of the clustering module is defined as minimizing the distribution. and distribution KL divergence between:

[0098]

[0099] Both autoencoders in the MBVAE model are derived from the target distribution. Supervised training can make their results tend to be consistent during training.

[0100] During the fine-tuning phase, a clustering module was added, using augmented samples. and fixed target distribution The MBVAE-based deep embedding clustering model is trained using the following loss function:

[0101]

[0102] in, Parameters contributing to loss control; The loss is the clustering module loss, which is based on augmented samples. Feature distribution Compared to clean samples Target distribution KL distance between:

[0103]

[0104] After updating the network parameters, use a clean sample. Update target distribution The training (feature update) and target distribution update of the deep embedding clustering model are performed iteratively until convergence, resulting in a trained MBVAE-based deep embedding clustering model.

[0105] The optimization of the data-augmented population graph-level deep clustering model is achieved by optimizing the HGDAE-based deep graph clustering model. The core idea of ​​the HGDAE-based deep graph clustering model is to learn a robust graph representation through contrastive and adversarial learning, including the HGDAE model and a clustering module. The HGDAE model consists of two encoders and one decoder. The HGDAE-based deep graph clustering model processes clean population graphs G and their enhanced versions. Features are extracted using a weight-sharing encoder. During training, multiple loss functions are jointly optimized, including reconstruction loss (…). ), countering losses ( ), structural contrast loss ( ) and contrast loss on the underlying representation ( and This multi-task learning approach enables the model to capture meaningful representations from graph data, specifically:

[0106] like Figure 4 As shown, the process of training a deep embedding graph clustering model based on HGDAE includes two stages. First, augmented mind maps are used to pre-train the dual autoencoders (DAE, a deep neural network-based autoencoder, and THGAE, a heterogeneous graph autoencoder based on Transformer).

[0107] The HGDAE loss function consists of three components: DAE reconstruction loss, THGAE reconstruction loss, and semi-supervised loss. The reconstruction loss of the DAE is defined as the reconstruction feature matrix. (Node features output by the decoder) and the original feature matrix Distance between (original node features of enhanced samples):

[0108]

[0109] The reconstruction loss of THGAE is defined as the reconstruction adjacency matrix. With the original adjacency matrix Distance between:

[0110]

[0111] in, Let represent the element in the i-th row and k-th column of the original adjacency matrix A. Represents the reconstruction of the adjacency matrix The element in the i-th row and k-th column, This represents the normalization coefficient of the loss function. Furthermore, to better incorporate NC into the neural subtyping of ASD, a semi-supervised loss based on similarity constraints is designed using the class labels of NC (normal samples) and ASD:

[0112]

[0113] in, This represents the constraint relationship between sample m and sample n. If both samples belong to the NC group or the ASD group, then... ,otherwise . also, and Let m and n represent the distances between samples m and n in the DAE and THGAE embedding spaces, respectively. Semi-supervised loss is used. This can make the learned features more discriminative.

[0114] The overall loss function is: The overall loss of the HGDAE model is:

[0115]

[0116] in, , Hyperparameters for controlling various losses.

[0117] In the second stage, the pre-trained HGDAE model is fine-tuned together with the clustering module. The loss function of the entire depth map clustering model consists of a dual autoencoder loss and a dual self-supervised loss. Augmented samples are used during pre-training. Training HGDAE, HGDAE network output The loss function for the entire depth map clustering model is expressed as:

[0118]

[0119] Among them, the deep embedding clustering loss of the DAE autoencoder Defined as based on augmented samples Feature distribution and based on clean samples Target distribution KL divergence between:

[0120]

[0121] Deep Embedding Clustering Loss of THGAE Autoencoder To enhance ASD samples soft label distribution Compared to clean samples Target distribution KL divergence between:

[0122]

[0123] During backpropagation, after the network parameters are updated, clean samples are used. Update target distribution The deep clustering model undergoes two iterative steps: training (feature update) and updating the target distribution, until convergence.

[0124] Finally, the clustering results of nodes / samples n, i.e., the neuromorphic classification results of ASD, can be derived from the distribution. Calculation yielded:

[0125]

[0126] S4: Obtain the brain map data to be detected and input it into a trained deep embedding clustering model based on MBVAE or a deep map clustering model based on HGDAE to obtain the ASD neuromorphic classification results.

[0127] By acquiring the brain map data to be detected and inputting it into a trained deep embedding clustering model based on MBVAE or a deep map clustering model based on HGDAE, ASD neuromorphic classification results can be obtained.

[0128] In summary, this invention can synthesize biologically meaningful complex brain maps by leveraging the high-quality and diverse sample generation capabilities of the conditional diffusion model. This provides more high-quality samples for ASD neural subtyping models, addressing issues such as gender imbalance in ASD samples, and promoting the performance and robustness of deep clustering-based ASD neural subtyping models. The subtype-specific synthesis method employed in this invention helps solve the pattern collapse problem encountered by most existing synthesis methods when dealing with heterogeneous data. This invention effectively addresses the gender imbalance and multi-site heterogeneity issues in ASD data, enhancing the model's robustness and generalization ability. It also demonstrates high accuracy in subtyping results when facing few-sample learning, showing promising application prospects.

[0129] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for ASD neurological classification based on synthetic brain map data augmentation, characterized in that, include: S1: Obtain the brain map data for training and input it into the pre-trained dual-decoder graph autoencoder to obtain the latent embedding; S2: Input the latent embedding into a pre-trained latent space conditional diffusion model to obtain an enhanced brain map; S3: Use augmented mind mapping to train a deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE to obtain a trained deep embedding clustering model based on MBVAE or a deep graph clustering model based on HGDAE. S4: Obtain the brain map data to be detected and input it into a trained deep embedding clustering model based on MBVAE or a deep map clustering model based on HGDAE to obtain the ASD neuromorphic classification results.

2. The ASD neural subtyping method based on synthetic brain map data augmentation according to claim 1, characterized in that, The dual-decoder graph autoencoder includes a graph encoder, a first decoder, and a second decoder. The graph encoder and the second decoder are composed of a layered GCN network or a GAT network. The graph encoder processes the mind map to obtain the hidden embedding. The first decoder performs an inner product operation on the embedded representation to obtain the reconstructed graph structure; the second decoder processes the embedded representation to obtain the reconstructed node features.

3. The ASD neural subtyping method based on synthetic brain map data augmentation according to claim 2, characterized in that, The loss function for pre-training a dual-decoder graph autoencoder is expressed as: ； ； ； ； in, This represents the total loss of the dual-decoder graph autoencoder. This represents the loss of reconstructing the graph structure. This represents the loss of reconstructed node features. Indicates prior loss. The weights represent the loss of reconstructing the graph structure. Represents the feature loss weights of the reconstructed nodes. In the mind map representing modality m, the element in the i-th row and k-th column of the original adjacency matrix is... In a mind map representing modality m, reconstruct the elements in the i-th row and k-th column of the adjacency matrix. Indicates the number of nodes. The original node features of the brain map representing modality m. Represents the reconstructed node features of the mind map for modality m. The reconstructed adjacency matrix of the mind map for modality m. This represents the adjustment coefficient. Describing the L1 norm, Represents the matrix nuclear norm.

4. The ASD neural subtyping method based on synthetic brain map data augmentation according to claim 1, characterized in that, The loss function for pre-training the latent space conditional diffusion model is expressed as: ； in, This represents the total loss of the latent space conditional diffusion model. Representing latent space variables The distribution, This represents the noise actually added to the data. Follows a standard normal distribution. Represents phenotypic information, This represents the noise predicted by the denoising U-Net model. Let represent the hidden variable state at time step t during the diffusion process. Indicates the time step of the diffusion process. Indicates a condition encoder. This indicates the expected operation.

5. The ASD neural subtyping method based on synthetic brain map data augmentation according to claim 1, characterized in that, The process of training a deep embedding clustering model based on MBVAE includes: The MBVAE model was pre-trained using augmented mind mapping to obtain the ASD embedding of the MBVAE model; The readout operation is used to process the ASD embeddings of the MBVAE model to obtain a graph-level representation; The similarity between the graph representation and the cluster centers is calculated using the Student t-distribution, and the clustering probability of the samples corresponding to the graph representation is obtained, which is the neural typing result. The divergence loss is calculated based on the clustering probability of the samples. The weighted sum of the divergence loss and the total loss of the MBVAE model is used as the total loss of the deep embedding clustering model. The parameters of the MBVAE-based deep embedding clustering model are adjusted according to the total loss of the deep embedding clustering model to obtain the trained deep embedding clustering model.

6. The ASD neural subtyping method based on synthetic brain map data augmentation according to claim 5, characterized in that, The MBVAE model consists of two encoders and a shared decoder. The two encoders process the enhanced brain map and the original brain map, respectively.

7. The ASD neural subtyping method based on synthetic brain map data augmentation according to claim 1, characterized in that, The HGDAE-based deep graph clustering model consists of an HGDAE model and a clustering module. The HGDAE model comprises two encoders and one decoder. The process of training the MBVAE-based deep embedding clustering model includes: An enhanced mind map was used to pre-train the HGDAE model; the pre-trained HGDAE model was then fine-tuned together with the clustering module to obtain a trained HGDAE-based deep graph clustering model.

8. The ASD neural subtyping method based on synthetic brain map data augmentation according to claim 7, characterized in that, The loss function for training the HGDAE-based deep graph clustering model is expressed as: ； ； ； in, This represents the total loss of the HGDAE-based depth map clustering model. The input is an augmented mind map. The reconstruction loss of the HGDAE-based depth graph clustering model. This represents the KL divergence regularization constraint term between the original mind map distribution and the enhanced mind map distribution. The KL divergence regularization constraint term represents the relationship between the distribution of the original mind map and the distribution of the enhanced mind map in the clustering feature space Z. Indicates enhanced mind mapping, Represents the original mind map. This indicates the calculation of KL divergence. This represents the true data distribution of the original mind map G. Enhanced mind mapping After feature transformation, the probability distribution in the cluster feature space Z is estimated; Enhanced mind mapping The probability distribution in the original feature space is estimated after feature transformation.

Images