An autism subtype detection method and system based on multi-hypergraph collaborative optimization

Abstract

The application discloses an autism subtype detection method and system based on multi-supernetwork collaborative optimization. The method can obtain image data including a first modality and a second modality, and extract a low-frequency fluctuation amplitude graph of the first modality image data and a brain tissue density graph of the second modality image data. According to an automatic annotation atlas, brain regions are divided, and the average feature sequence of the low-frequency fluctuation amplitude and the brain tissue density of each brain region is extracted. Then, the first data matrix and the second data matrix are generated through the average feature sequence, and the first supernetwork and the second supernetwork are constructed according to the above data matrix. A community detection method for multi-supernetwork collaborative optimization is proposed for the first supernetwork and the second supernetwork to generate clustering results as the detection results of the autism subtypes. The method constructs the associated supernetwork based on the multi-modality image data, and then fuses the multi-modality supernetwork and performs collaborative optimization to detect the autism subtypes, so that the accuracy of the autism subtype detection can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for detecting autism subtypes based on multi-hypergraph collaborative optimization. Background Technology

[0002] Functional magnetic resonance imaging (fMRI) and structural magnetic resonance imaging (sMRI) are widely used in medical imaging. fMRI works by stimulating specific sensory organs to induce neural activity (functional area activation) in corresponding areas of the cerebral cortex. Its principle is to use magnetic resonance imaging to measure changes in hemodynamics caused by neuronal activity. From the perspective of functional imaging analysis, researchers have used random forests to build sample similarity matrices and employed spectral clustering algorithms to identify three biological subtypes of autism. SMRI is a non-invasive medical imaging technique that uses magnetic resonance imaging to scan the human body. From the perspective of structural imaging analysis, researchers have used unsupervised clustering of the brain's gray matter anatomical structures to classify autism into three biological subtypes.

[0003] Traditional graphs can be used to describe the relationships between samples, where vertices represent samples and edges represent relationships. However, traditional networks based on linear correlations can only capture second-order relationships between pairs of samples and cannot characterize the collaborative changes between multiple samples. Because hyperedges in a hypergraph can connect any number of samples, not just two, they can effectively characterize higher-order relationships between multiple samples. Previous studies have only used traditional unsupervised algorithms to cluster data from a single modality, which cannot model the complex higher-order relationships between samples and ignores the simultaneous changes in brain function and structure caused by autism. This results in poor detection of biological subtypes, and these subtypes cannot comprehensively reflect the heterogeneity of autism brain lesion patterns. Using hypergraphs to characterize higher-order similarity relationships between samples and constructing a hypernetwork by fusing multimodal brain imaging data can detect more objective and reliable autism subtypes. Summary of the Invention

[0004] This application provides a method and system for autism subtype detection based on multi-hypergraph collaborative optimization, in order to solve the problem of difficulty in analyzing the biological heterogeneity of autism.

[0005] In a first aspect, this invention discloses a method for detecting autism subtypes based on multi-hypergraph collaborative optimization, comprising:

[0006] Acquire image data, which includes first modal image data and second modal image data, wherein the first modal image data is functional magnetic resonance imaging and the second modal image data is structural magnetic resonance imaging;

[0007] Extract the low-frequency fluctuation amplitude feature map from the first modality image data, and extract the brain tissue density feature map from the second modality image data;

[0008] The brain regions are divided according to the low-frequency fluctuation amplitude feature map and the brain tissue density feature map based on the automatically labeled map;

[0009] Extract the average feature sequence of the brain region;

[0010] A data matrix is ​​obtained through the average feature sequence. The data matrix includes a first data matrix and a second data matrix. The first data matrix is ​​obtained based on the average feature sequence of the low-frequency fluctuation amplitude feature map, and the second data matrix is ​​obtained based on the average feature sequence of the brain tissue density feature map.

[0011] Construct a first hypernetwork and a second hypernetwork based on the data matrix;

[0012] Community detection is performed on the first hypernetwork and the second hypernetwork using multi-hypergraph collaborative optimization to generate cluster labels;

[0013] When the community detection of the multi-hypergraph collaborative optimization is completed, the clustering label is output.

[0014] Optionally, the method further includes: removing the time points of the target quantity in the first modal image data; performing temporal layer correction and head motion correction on the image data; registering the first modal image data to a standard space; and performing temporal filtering and spatial smoothing on the first modal image data, wherein the temporal filtering includes delinear trending and bandpass filtering.

[0015] Optionally, the method further includes: performing pre-joint correction on the second modality image data; performing feature removal and segmentation processing on the second modality image data to generate a gray matter density map and a white matter density map of the second modality image data; and registering the gray matter density map and the white matter density map to a standard space.

[0016] Optionally, the step of extracting the density feature map of the second modality image data includes: obtaining the gray matter density map and the white matter density map; and stitching the gray matter density map and the white matter density map together.

[0017] Optionally, the method further includes: detecting head movement in the first modal image data; removing first modal image data with head movement greater than 1 mm; obtaining the preprocessing level of the second modal image data; removing second modal image data with the preprocessing level lower than a target threshold; and finally ensuring that the samples in the first modal image data and the second modal image data are consistent.

[0018] Optionally, constructing the first hypernetwork and the second hypernetwork based on the data matrix includes: obtaining sub-samples, the sub-samples being elements of the first data matrix and the second data matrix; using the sub-samples as centroids; solving a sparse matrix according to the centroids, the column dimension of the sparse matrix being equal to the number of sub-samples, and the sparsity of the sparse matrix being represented based on a regularization parameter; concatenating the sparse matrix and performing binarization processing to obtain a hypergraph correlation matrix; labeling the columns of the hypergraph correlation matrix as hyperedges; and obtaining the first hypernetwork and the second hypernetwork based on the hyperedges.

[0019] Optionally, the regularization parameter is determined by ten-fold cross-validation.

[0020] Optionally, community detection using multi-hypergraph collaborative optimization is performed on the first hypernetwork and the second hypernetwork, including: clustering the non-zero value matrix of the hypergraph association matrix before binarization based on the Leuven multilayer community detection algorithm to generate initial cluster labels; optimizing the modularity objective of the first hypernetwork according to the initial cluster labels to generate a first clustering result; optimizing the modularity objective of the second hypernetwork according to the first clustering result to generate a second clustering result; optimizing the modularity objective of the first hypernetwork and the second hypernetwork according to the second clustering result; and stopping the community detection using multi-hypergraph collaborative optimization of the first and second hypernetworks when the modularity objectives of the first and second hypernetworks converge.

[0021] Optionally, optimizing the modularity objective of the first or second hypernetwork includes: moving nodes in the first or second network into clusters of neighboring nodes; calculating the change in modularity, the change including changes in partitioning and volume terms; detecting the gain of modularity using the change; if there is a gain in modularity, moving the node into a cluster of neighboring nodes and calculating the change in modularity; if there is no gain in modularity, marking the modularity objective as converged.

[0022] Secondly, this application also discloses an autism subtype detection system based on multi-hypergraph collaborative optimization, comprising:

[0023] The acquisition module is configured to acquire image data, which includes first modal image data and second modal image data, wherein the first modal image data is functional magnetic resonance imaging and the second modal image data is structural magnetic resonance imaging.

[0024] The processing module is configured to extract low-frequency fluctuation amplitude feature maps from the first modality image data and brain tissue density feature maps from the second modality image data; to divide the brain regions of the low-frequency fluctuation amplitude feature maps and the brain tissue density feature maps according to the automatically labeled atlas; and to extract the average feature sequence of the brain regions.

[0025] A multi-task hypergraph construction module is configured to obtain a data matrix through the average feature sequence, the data matrix including a first data matrix and a second data matrix, the first data matrix being obtained based on the average feature sequence of the low-frequency fluctuation amplitude feature map, and the second data matrix being obtained based on the average feature sequence of the brain tissue density feature map; and to construct a first hypernetwork and a second hypernetwork based on the data matrix.

[0026] The multi-hypergraph collaborative community detection module is configured to perform multi-hypergraph collaborative optimization community detection on the first hypernetwork and the second hypernetwork to generate cluster labels;

[0027] The output module is configured to output the clustering label at the end of the community detection in the multi-hypergraph collaborative optimization.

[0028] As can be seen from the above technical solutions, some embodiments of this application provide a method and system for autism subtype detection based on multi-hypergraph collaborative optimization. The method acquires image data including a first modality and a second modality, and extracts the low-frequency fluctuation amplitude map of the first modality image data and the brain tissue density map of the second modality image data. Brain regions are divided according to automatically labeled atlases, and the average feature sequences of low-frequency fluctuation amplitude and brain tissue density of each brain region are extracted. A first data matrix and a second data matrix are then generated using the average feature sequences, and a first hypernetwork and a second hypernetwork are constructed based on the aforementioned data matrices. A community detection method oriented towards multi-hypergraph collaborative optimization is proposed for the first and second hypernetworks to generate clustering results as the detection results for autism subtypes. The method constructs associated hypernetworks based on multi-modal image data, then fuses the multi-modal hypernetworks and performs collaborative optimization to detect autism subtypes, which can improve the accuracy of autism subtype detection. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart illustrating a method for detecting autism subtypes based on multi-hypergraph collaborative optimization provided in some embodiments of this application;

[0031] Figure 2 A conceptual comparison diagram of hypergraph and graph provided for some embodiments of this application;

[0032] Figure 3 This is a schematic diagram of a community detection process for multi-hypergraph collaborative optimization provided in some embodiments of this application;

[0033] Figure 4 A flowchart illustrating the optimized hypergraph modular target provided in some embodiments of this application;

[0034] Figure 5 This is an architecture diagram of an autism subtype detection system based on multi-hypergraph collaborative optimization, provided for some embodiments of this application. Detailed Implementation

[0035] To make the objectives and implementation methods of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments.

[0036] The brain exists in two states: the task state and the resting state. The task state mainly refers to the state of the brain when performing specific tasks such as memory, recognition, and movement. The resting state refers to the state of the brain when it is not performing specific cognitive tasks, remaining quiet, relaxed, and awake. It is the most basic and essential state among the various complex states of the brain. Therefore, abnormal states of brain neurons can be detected through resting-state brain imaging.

[0037] Functional magnetic resonance imaging (fMRI) and structural magnetic resonance imaging (sMRI) are widely used in medical imaging. fMRI works by stimulating specific sensory organs to induce neural activity (functional area activation) in corresponding areas of the cerebral cortex. Its principle is to use magnetic resonance imaging to measure changes in hemodynamics caused by neuronal activity. From the perspective of functional imaging analysis, researchers have used random forests to build sample similarity matrices and employed spectral clustering algorithms to identify three biological subtypes of autism. SMRI is a non-invasive medical imaging technique that uses magnetic resonance imaging to scan the human body. From the perspective of structural imaging analysis, researchers have used unsupervised clustering of the brain's gray matter anatomical structures to classify autism into three biological subtypes.

[0038] Traditional graphs can be used to describe the relationships between samples, where vertices represent samples and edges represent relationships. However, traditional networks based on linear correlations can only capture second-order relationships between pairs of samples and cannot characterize the collaborative changes between multiple samples. Because hyperedges in a hypergraph can connect any number of samples, not just two, they can effectively characterize higher-order relationships between multiple samples. Previous studies have only used traditional unsupervised algorithms to cluster data from a single modality, which cannot model the complex higher-order relationships between samples and ignores the simultaneous changes in brain function and structure caused by autism. This results in poor detection of biological subtypes, and these subtypes cannot comprehensively reflect the heterogeneity of autism brain lesion patterns. Using hypergraphs to characterize higher-order similarity relationships between samples and constructing a hypernetwork by fusing multimodal brain imaging data can detect more objective and reliable autism subtypes.

[0039] Based on the above application scenarios, and to address the difficulty in resolving the biological heterogeneity of autism, some embodiments of this application provide a method for autism subtype detection based on multi-hypergraph collaborative optimization, such as... Figure 1 As shown, the procedure includes the following steps:

[0040] S1: Acquire image data.

[0041] The imaging data includes first-modality imaging data and second-modality imaging data. The first-modality imaging data is functional magnetic resonance imaging (fMRI), and the second-modality imaging data is structural magnetic resonance imaging (SMRI). Both the first-modality and second-modality imaging data represent brain imaging data from different modalities at rest. By acquiring this multimodal brain imaging data, the features represented in the brain imaging data can be analyzed.

[0042] For example, the first modality of image data is fMRI data, and the second modality of image data is T1 image data.

[0043] In some embodiments, after acquiring image data, preprocessing is performed on the first modal image data and the second modal image data respectively using the SPM12 toolkit in Matlab.

[0044] Therefore, in some embodiments, when preprocessing the first modality image data, the time points for the number of targets in the first modality image data are removed, and temporal correction and head motion correction are performed on the image data. Then, the first modality image data is registered to a standard space, and temporal filtering and spatial smoothing are performed on the first modality image data. The temporal filtering includes delinear trending and bandpass filtering.

[0045] For example, taking fMRI brain imaging data of the first modality as an example, the first 10 unstable time points were removed using SPM12 to achieve signal equilibrium. Then, temporal and head motion corrections were performed on the fMRI data. Excessive head motion data (>1 mm) were filtered out to reduce the impact of head motion on the first modality data. The fMRI data was then registered to the MNI (Montreal Neurological Institute) standard space, and regression covariate analysis, delinearization, bandpass filtering, and spatial smoothing were performed sequentially to preserve the low-frequency signals (0.01-0.1 Hz) in the fMRI data.

[0046] In some embodiments, when preprocessing the second modality image data, pre-joint correction is performed on the second modality image data. Then, feature removal and segmentation processing are performed on the second modality image data to generate gray matter density maps and white matter density maps. The gray matter density maps and white matter density maps are then registered to a standard space.

[0047] For example, taking T1 brain imaging data as the second modality, AC (Anterior Commissury) correction is performed on the T1 data using SPM12, setting the image origin of the T1 data at the AC. Then, the brain images of the T1 data undergo craniotomy and segmentation to obtain the gray and white matter images of the subjects. Next, the gray and white matter images of each subject are sequentially registered to the standard MNI space. Finally, the T1 data images are modulated to compensate for the effects of MNI space normalization.

[0048] Furthermore, to improve the quality of the image data, quality control is performed. In some embodiments, head movement in the first modality image data is detected; first modality image data with head movement greater than 1 mm is discarded. Simultaneously, the preprocessing level of the second modality image data is obtained, and second modality image data with a preprocessing level below a target threshold is discarded. For example, to limit the influence of head movement, subjects with excessive head movement (displacement > 1 mm) in the first modality image data are excluded; the second modality image data, i.e., T1 brain imaging data, is preprocessed using SPM12, and T1 brain imaging data with a preprocessing level below C is discarded. Ultimately, this ensures that the samples in the first and second modality image data are consistent.

[0049] S2: Extract the low-frequency fluctuation amplitude feature map from the first modality image data, and extract the brain tissue density feature map from the second modality image data.

[0050] After acquiring and preprocessing the image data, the low-frequency fluctuation amplitude feature map of the first modality image data and the density feature map of the second modality image data are extracted to determine the feature vectors in the image data. Among them, the brain tissue density feature map includes gray matter density map and white matter density map.

[0051] For example, taking fMRI brain imaging data as the first modality of image data, the ALFF (Amplitude of Low Frequency Fluctuation) feature of the fMRI data is extracted, and the ALFF feature is standardized to obtain mALFF.

[0052] Since the density feature map includes a gray matter density map and a white matter density map, in some embodiments, when extracting the density feature map of the second modality image data, the gray matter density map and the white matter density map are obtained and then stitched together in sequence to obtain the gray and white matter density map of the second modality image data.

[0053] For example, taking T1 brain imaging data as the second modality, the gray matter density and white matter density features of the T1 data are extracted, and the gray matter density and white matter density features are stitched together to obtain a gray-white matter density map.

[0054] S3: Based on the automatically labeled map, the brain regions are divided into low-frequency fluctuation amplitude feature maps and brain tissue density feature maps.

[0055] After extracting the low-frequency fluctuation amplitude feature map and brain tissue density feature map, they are further divided into multiple brain regions using an automatic annotation atlas to facilitate the analysis of the image data. The automatic annotation atlas can employ AAL (Anatomical Automatic Labeling) atlas to divide the low-frequency fluctuation amplitude feature map and brain tissue density feature map into a certain number of brain regions.

[0056] For example, taking brain imaging data as an example. After extracting the low-frequency fluctuation amplitude feature map and density feature map of the multimodal brain imaging data, the above feature maps are divided into 90 brain regions using AAL atlas, and then the brain imaging data is analyzed through the feature sequences of each brain region.

[0057] S4: Extract the average feature sequence of brain regions.

[0058] After dividing the brain into regions, the average feature sequence of each brain region is calculated and the average feature sequence of each brain region is extracted.

[0059] S5: Obtain the data matrix by averaging the feature sequences.

[0060] After extracting the average feature sequences of each brain region, two modal data matrices are generated based on the extracted average feature sequences. The data matrices include a first data matrix and a second data matrix. The first data matrix is ​​generated based on the average feature sequence of the low-frequency fluctuation amplitude feature map, and the second data matrix is ​​generated based on the average feature sequence of the brain tissue density feature map.

[0061] For example, if 90 brain regions are divided using AAL mapping, then the two modal feature sequences of each sample can represent the data matrix X. 1 and X 2 :

[0062]

[0063]

[0064] In the formula, n represents the number of samples, with a value of 287; d represents the length of the feature sequence, with a value of 90. X 1 and X 2 These represent the data matrices for the first and second modal image data, respectively. The data matrix is ​​abbreviated as X. j Then we can conclude that:

[0065]

[0066] S6: Construct the first and second supernetworks based on the data matrix.

[0067] After obtaining the first and second data matrices, the corresponding hypernetwork is constructed using these two data matrices.

[0068] The first hypernetwork corresponds to the first hypergraph, and the second hypernetwork corresponds to the second hypergraph. For example... Figure 2 As shown, an edge in a graph can only connect to two vertices; however, an edge in a hypergraph can connect to any number of vertices and is called a hyperedge. Formally, a hypergraph H is a pair H = (X, E), where X is a set of elements called nodes or vertices. E is a set of non-empty subsets of X, called hyperedges.

[0069] It should be noted that in the embodiments of this application, a vertex of the first hypergraph or the second hypergraph represents a subject, and a hyperedge represents a higher-order relationship between multiple subjects.

[0070] To facilitate the construction of the first and second hypernetworks from the data matrix, the hypernetworks can be constructed using a multi-task sparse representation approach. Specifically, in some embodiments, when constructing the first and second hypernetworks from the data matrix, subsamples are obtained. These subsamples are elements of the first and second data matrices. A sparse matrix is ​​solved using the subsamples as centroids. The column dimension of the sparse matrix is ​​equal to the number of subsamples, and the sparsity of the sparse matrix is ​​represented by a regularization parameter. The sparse matrices are concatenated and binarized to generate a hypergraph correlation matrix. The columns of the hypergraph correlation matrix are labeled as hyperedges, and the first and second hypernetworks are obtained based on these hyperedges.

[0071] Furthermore, in order to improve the model's fit, in some embodiments, the regularization parameter is determined by 10-fold cross-validation, which can effectively improve the model's fit.

[0072] For example, a hypernetwork can be constructed using multi-task sparse representation, with the following optimization objective:

[0073]

[0074] In the formula, W = [w 1 ,…,w j ,…,w m ], D is:

[0075]

[0076] In the formula, ||W||1 is the sparse L1 norm of the parameters, λ1 represents the regularization parameter, which is used to control the sparsity of the model. The larger λ1 is, the sparser the W matrix is; the smaller λ1 is, the denser the W matrix is. D is the relational constraint between modalities, which can be used to maintain the relative distance between the feature vectors extracted from different modalities of the same subject before and after feature projection. and These represent the feature vectors of the i-th sample in the j-th and k-th modalities, respectively. Representing the eigenvector and The relative distance before feature projection; Representing the eigenvector and The relative distance after feature projection; λ2 represents the regularization parameter, used to control the parameters w between different modes. j The greater the similarity of λ2, the greater the parameter w between different modes. j The more similar the modes are, the smaller λ2 becomes, and the more similar the parameters w between different modes become. j The less similar they are, the better. The regularization parameters λ1 and λ2 are determined using ten-fold cross-validation to find the optimal parameters.

[0077] Therefore, when constructing a hypergraph through multiple tasks, the feature vectors of two different modalities for each sample are sequentially combined. As the centroid, i.e. the label, a sparse matrix W with the same number of samples can be obtained. The obtained W is then concatenated column by column to form hypergraph association matrices H1 and H2. Each column of H1 and H2 represents a hyperedge, which contains the subject as the centroid and the subjects corresponding to the parameters in W that are greater than 0. H1 and H2 also serve as a hypernetwork for the association of two modalities.

[0078] For example, the correlation matrix H of a hypergraph is defined as follows:

[0079]

[0080] Since the sparse matrix W obtained contains both 0 and non-zero values, but the non-zero values ​​are not all 1, we perform binarization on W to obtain a hypergraph incidence matrix containing only 0 and 1. At the same time, we retain the non-zero value matrix in W.

[0081] S7: Perform community detection using multi-supergraph collaborative optimization on the first and second supernetworks to generate cluster labels.

[0082] After constructing the first and second hypernetworks, community detection is performed on them using multi-hypergraph collaborative optimization to generate the final cluster labels.

[0083] In some embodiments, when performing community detection using multi-hypergraph collaborative optimization on the first hypernetwork and the second hypernetwork, clustering is performed on the non-zero value matrix of the hypergraph association matrix before binarization based on the Leuven multilayer community detection algorithm to generate initial cluster labels. The modularity objective of the first hypernetwork is optimized according to the initial cluster labels to generate a first clustering result. The modularity objective of the second hypernetwork is then optimized according to the first clustering result to generate a second clustering result. The modularity objective of the first hypernetwork is optimized using the second clustering result, and the modularity objective of the second hypernetwork is also optimized using the first clustering result. This process is repeated until the modularity objectives of the first and second hypernetworks converge, at which point the community detection using multi-hypergraph collaborative optimization of the first and second hypernetworks is stopped.

[0084] For example, such as Figure 3 As shown, the non-zero values ​​in the retained sparse matrix are clustered using multi-layer community detection, and the clustering results are used as the initial label z for multi-hypergraph fusion clustering. Starting from the initial label z, the modularity objective of the first hypernetwork is optimized, and the clustering results are used as the initial label z' of the second hypergraph to optimize the modularity objective of the second hypernetwork; the clustering results are again used as the initial label z of the first hypernetwork to optimize the modularity objective of the first hypernetwork; this process is repeated, optimizing the modularity objectives of the two hypernetworks until the modularity objectives of the two hypernetworks converge, at which point the community detection of multi-hypergraph co-optimization ends, and the obtained clustering label z is the final autism subtype identification result.

[0085] Therefore, in some embodiments, the community detection in multi-hypergraph collaborative optimization is optimized using the maximum likelihood estimation method. The maximum likelihood function is:

[0086]

[0087] in:

[0088]

[0089]

[0090]

[0091] In the formula, z represents the clustering result, Ω is the affinity function, used to control the probability of placing a hyperedge on the node tuple R; R represents a set of nodes, if a set of nodes Ω(z) R The larger the value of θ, the higher the probability of forming a hyperedge between the nodes in that group; θ is the degree parameter of each node, used to control the degree value of the node. R This represents the number of hyperedges that can be placed on node R, and is subject to the parameter b. R π(θ R )Ω(z RThe Poisson distribution of b) R The number of different sorting methods for node R, π(θ) R )=∏ i∈R θ i The product of the node R-degree parameters. Q(z,Ω,θ) is the only part of the log-likelihood, and its value depends on the affinity function Ω and the cluster label z. The value of K(θ) depends on θ, and the value of C depends only on the dataset and can be ignored during inference.

[0092] In some embodiments, maximum likelihood estimation employs a coordinate ascent method for two-stage alternating estimation. The first stage assumes an estimate. To estimate and

[0093]

[0094] generated and It can be viewed as the current estimate of the label vector z This is the maximum likelihood estimate given the conditions. The second stage assumes the current estimate... and This is used to estimate the label vector z:

[0095]

[0096] generated This can be considered as based on the current estimate and This is the maximum likelihood estimate given the condition. Then, estimations are performed alternately in these two stages until convergence. The first stage estimation uses a simple closed-form, i.e. Estimated as d:

[0097]

[0098] Therefore, Q(z,Ω,θ) should be the focus during clustering, as it is related to clustering. The affinity function Ω uses the AON affinity function with parameters β and γ, so estimating Ω can be derived as estimating parameters β and γ. The corresponding optimization objective is the hypergraph modularity.

[0099]

[0100] In the formula, Q represents the hypergraph modularity, k represents the maximum degree of a hyperedge, β controls which hyperedges are most relevant, γ controls the size of the clusters, and cut is the splitting term. k (z) is:

[0101]

[0102] In the formula, Cut k (z) Calculate the number of superedges of size k; the superedges contain nodes from two or more different clusters. The term J(ω) calculates the degree of vertices in all clusters; the term J(ω) does not depend on the clustering result z and can be ignored during inference.

[0103] like Figure 4 As shown, in some embodiments, when optimizing the modularity objective of the first or second hypernetwork, nodes in the first or second hypernetwork are moved into clusters of neighboring nodes, and the change in modularity is calculated. This change includes changes in the partitioning term and the volume term. The gain in modularity is then detected using this change. If there is a gain in modularity, nodes are moved into clusters of neighboring nodes, and the change in modularity is calculated; if there is no gain in modularity, the modularity objective is marked as converged.

[0104] In other words, each node is moved sequentially into the cluster of its neighboring nodes, and the change in hypergraph modularity is calculated. This change in modularity includes changes in both the partitioning and volume terms. If there is a gain in modularity, the node is moved into the cluster of its neighboring nodes until the modularity no longer increases, i.e., convergence. At this point, the optimized z is the final clustering result.

[0105] S8: Output cluster labels when community detection in multi-hypergraph collaborative optimization is completed.

[0106] After performing community detection using multi-hypergraph collaborative optimization on the first and second hypernetworks, the community detection process ends when the modular objectives of the first and second hypernetworks converge. The resulting clustering is then the final clustering label, which serves as the final autism subtype identification result. This embodiment of the application, using multi-hypergraph collaborative optimization community detection on multimodal image data, can accurately classify and identify autism subtypes in image data, thereby improving the effectiveness of autism subtype identification.

[0107] Based on the above-mentioned autism subtype detection method using multi-hypergraph collaborative optimization, this invention also discloses an autism subtype detection system using multi-hypergraph collaborative optimization, such as... Figure 5 As shown, the subtype identification system includes a data acquisition module, a processing module, a multi-task hypergraph construction module, a multi-hypergraph collaborative community detection module, and an output module, wherein:

[0108] The acquisition module is configured to acquire image data, wherein the first modal image data is functional magnetic resonance imaging and the second modal image data is structural magnetic resonance imaging;

[0109] The processing module is configured to extract low-frequency fluctuation amplitude feature maps from the first modality image data and brain tissue density feature maps from the second modality image data; to divide the brain regions of the low-frequency fluctuation amplitude feature maps and the brain tissue density feature maps according to the automatically labeled atlas; and to extract the average feature sequence of the brain regions.

[0110] The multi-task hypergraph construction module is configured to obtain a data matrix through the average feature sequence, the data matrix including a first data matrix and a second data matrix, the first data matrix being obtained based on the average feature sequence of the low-frequency fluctuation amplitude feature map, and the second data matrix being obtained based on the average feature sequence of the brain tissue density feature map; and to construct a first hypernetwork and a second hypernetwork based on the data matrix.

[0111] The multi-hypergraph collaborative community detection module is configured to perform multi-hypergraph collaborative optimization community detection on the first hypernetwork and the second hypernetwork to generate cluster labels;

[0112] The output module is configured to output the clustering label at the end of the community detection in the multi-hypergraph collaborative optimization.

[0113] As can be seen from the above technical solutions, some embodiments of this application provide a method and system for autism subtype detection based on multi-hypergraph collaborative optimization. The method acquires image data including a first modality and a second modality, and extracts the low-frequency fluctuation amplitude map of the first modality image data and the brain tissue density map of the second modality image data. Brain regions are divided according to automatically labeled atlases, and the average feature sequences of low-frequency fluctuation amplitude and brain tissue density of each brain region are extracted. A first data matrix and a second data matrix are then generated using the average feature sequences, and a first hypernetwork and a second hypernetwork are constructed based on the aforementioned data matrices. A community detection method oriented towards multi-hypergraph collaborative optimization is proposed for the first and second hypernetworks to generate clustering results as the detection results for autism subtypes. The method constructs associated hypernetworks based on multi-modal image data, then fuses the multi-modal hypernetworks and performs collaborative optimization to detect autism subtypes, which can improve the accuracy of autism subtype detection.

[0114] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the exemplary discussion above is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the principles and practical applications, thereby enabling those skilled in the art to better utilize the embodiments and various different variations of embodiments suitable for specific application considerations.

Claims

1. A method for autism subtype detection based on multi-hypergraph collaborative optimization, characterized in that, include: Acquire image data, which includes first modal image data and second modal image data, wherein the first modal image data is functional magnetic resonance imaging and the second modal image data is structural magnetic resonance imaging; Extract the low-frequency fluctuation amplitude feature map from the first modality image data, and extract the brain tissue density feature map from the second modality image data; The brain regions are divided according to the low-frequency fluctuation amplitude feature map and the brain tissue density feature map based on the automatically labeled map; Extract the average feature sequence of the brain region; A data matrix is ​​obtained through the average feature sequence. The data matrix includes a first data matrix and a second data matrix. The first data matrix is ​​obtained based on the average feature sequence of the low-frequency fluctuation amplitude feature map, and the second data matrix is ​​obtained based on the average feature sequence of the density feature map. A first hypernetwork and a second hypernetwork are constructed based on the data matrix, wherein: subsamples are obtained, the subsamples being elements of the first data matrix and the second data matrix; the subsamples are used as centroids; a sparse matrix is ​​solved according to the centroids, the column dimension of the sparse matrix being equal to the number of subsamples, and the sparsity of the sparse matrix being represented based on a regularization parameter; the sparse matrix is ​​concatenated and binarized to obtain a hypergraph association matrix; the columns of the hypergraph association matrix are labeled as hyperedges, and the first hypernetwork and the second hypernetwork are obtained based on the hyperedges; Community detection using multi-hypergraph collaborative optimization is performed on the first hypernetwork and the second hypernetwork to generate cluster labels. Specifically, the non-zero value matrix of the hypergraph association matrix before binarization is clustered using the Leuven multilayer community detection algorithm to generate initial cluster labels. The modularity objective of the first hypernetwork is optimized according to the initial cluster labels to generate a first clustering result. The modularity objective of the second hypernetwork is optimized according to the first clustering result to generate a second clustering result. The modularity objective of the first hypernetwork is further optimized using the second clustering result, and the modularity objective of the second hypernetwork is also optimized using the first clustering result. When the modularity objectives of the first and second hypernetworks converge, the community detection using multi-hypergraph collaborative optimization of the first and second hypernetworks is stopped. When the community detection of the multi-hypergraph collaborative optimization is completed, the clustering label is output.

2. The autism subtype detection method based on multi-hypergraph collaborative optimization according to claim 1, characterized in that, Also includes: The time points at which the number of targets is removed from the first modal image data; Perform temporal correction and head motion correction on the image data; The first modal image data is registered to a standard space, and temporal filtering and spatial smoothing are performed on the first modal image data. The temporal filtering includes delinearization and bandpass filtering.

3. The autism subtype detection method based on multi-hypergraph collaborative optimization according to claim 1, characterized in that, Also includes: Perform pre-combination correction on the second modality image data; Feature removal and segmentation processing are performed on the second modality image data to generate gray matter density map and white matter density map of the second modality image data; The gray matter density map and the white matter density map are registered to the standard space.

4. The autism subtype detection method based on multi-hypergraph collaborative optimization according to claim 3, characterized in that, Extracting brain tissue density feature maps from the second modality image data, including: Obtain the gray matter density map and the white matter density map; The gray matter density map and the white matter density map are spliced ​​together.

5. The autism subtype detection method based on multi-hypergraph collaborative optimization according to claim 1, characterized in that, Also includes: Detect head movement in the first modality image data; The first modal image data with head movement greater than 1 mm were discarded; Obtain the preprocessing level of the second modality image data; Second modality image data whose preprocessing level is lower than the target threshold are removed; Ultimately, it is ensured that the samples in the first modality image data and the second modality image data are consistent.

6. The autism subtype detection method based on multi-hypergraph collaborative optimization according to claim 1, characterized in that, The regularization parameter is determined by ten-fold cross-validation.

7. The autism subtype detection method based on multi-hypergraph collaborative optimization according to claim 1, characterized in that, The modular objective of optimizing the first hypernetwork or the second hypernetwork includes: Move a node from the first hypernetwork or the second hypernetwork into a cluster of neighboring nodes; Calculate the change in modularity, whereby the change includes the changes in the segmentation and volume terms; The gain of the modularity is detected by the amount of change; If there is a gain in the modularity, move the node into a cluster of neighboring nodes and calculate the change in the modularity; If the modularity has no gain, the modularity objective is marked as convergent.

8. An autism subtype detection system based on multi-hypergraph collaborative optimization, characterized in that, include: The acquisition module is configured to acquire image data, which includes first modal image data and second modal image data, wherein the first modal image data is functional magnetic resonance imaging and the second modal image data is structural magnetic resonance imaging. The processing module is configured to extract low-frequency fluctuation amplitude feature maps from the first modality image data and brain tissue density feature maps from the second modality image data; to divide the brain regions of the low-frequency fluctuation amplitude feature maps and the brain tissue density feature maps according to the automatically labeled atlas; and to extract the average feature sequence of the brain regions. A multi-task hypergraph construction module is configured to obtain a data matrix through the average feature sequence, the data matrix including a first data matrix and a second data matrix, the first data matrix being obtained based on the average feature sequence of the low-frequency fluctuation amplitude feature map, and the second data matrix being obtained based on the average feature sequence of the brain tissue density feature map; constructing a first hypernetwork and a second hypernetwork based on the data matrix, wherein: obtaining subsamples, the subsamples being elements of the first data matrix and the second data matrix; using the subsamples as centroids; solving a sparse matrix according to the centroids, the column dimension of the sparse matrix being equal to the number of subsamples, the sparsity of the sparse matrix being represented based on a regularization parameter; concatenating the sparse matrix and performing binarization processing to obtain a hypergraph association matrix; labeling the columns of the hypergraph association matrix as hyperedges, and obtaining the first hypernetwork and the second hypernetwork based on the hyperedges; A multi-hypergraph collaborative community detection module is configured to perform community detection using multi-hypergraph collaborative optimization on the first hypernetwork and the second hypernetwork to generate clustering labels. Specifically, it performs clustering on the non-zero value matrix of the hypergraph association matrix before binarization based on the Leuven multilayer community detection algorithm to generate initial clustering labels; optimizes the modularity objective of the first hypernetwork according to the initial clustering labels to generate a first clustering result; optimizes the modularity objective of the second hypernetwork according to the first clustering result to generate a second clustering result; optimizes the modularity objective of the first hypernetwork using the second clustering result, and optimizes the modularity objective of the second hypernetwork using the first clustering result; and stops the multi-hypergraph collaborative optimization community detection of the first and second hypernetworks when the modularity objectives of the first and second hypernetworks converge. The output module is configured to output the clustering label at the end of the community detection in the multi-hypergraph collaborative optimization.

Images