ASD auxiliary diagnosis method fusing dynamic low-order and high-order dynamic central moment networks

Abstract

The invention relates to the technical field of medical auxiliary diagnosis, and discloses an ASD auxiliary diagnosis method fusing dynamic low-order and high-order dynamic central moment networks, and the method comprises the steps: obtaining resting state functional magnetic resonance imaging data of a subject, and carrying out the preprocessing; constructing a low-order dynamic function connection network and a high-order dynamic function connection network to respectively describe direct connection and connection collaboration of the brain region; extracting multi-order central moment characteristics of the two types of networks to obtain stable statistics; generating a low-high order dynamic central moment network by adopting a puzzle fusion strategy; inputting the fusion features into a visual Transform model for classification, and outputting an auxiliary diagnosis result of the infantile autism spectrum disorder; and the key brain region is identified by analyzing the attention weight of the model so as to enhance the interpretability. According to the method, multi-level dynamic connection information is effectively fused, the influence of time asynchronism is overcome, the method has the advantages of being high in classification accuracy, high in generalization ability and good in biological interpretation, and a reliable solution is provided for ASD auxiliary diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical auxiliary diagnostic technology, specifically to an ASD auxiliary diagnostic method that integrates dynamic low-order and high-order dynamic central moment networks. Background Technology

[0002] Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder, and early diagnosis and intervention are crucial for improving patient prognosis. Resting-state functional magnetic resonance imaging (fMRI) provides an important tool for non-invasively exploring abnormal functional connectivity in the brains of ASD patients. Currently, rs-fMRI-based auxiliary diagnostic methods for ASD mainly rely on functional connectivity analysis.

[0003] Existing functional connectivity analysis methods are mainly divided into static and dynamic functional connectivity. Static functional connectivity assumes that brain functional connectivity is constant during scanning, constructing a single network by calculating the correlation of brain region time series over the entire time period. However, this method cannot capture the instantaneous dynamic changes in brain activity. Dynamic functional connectivity, through techniques such as sliding windows, constructs a series of time-resolved networks, reflecting the time-varying characteristics of brain functional connectivity, and represents the forefront of current research. However, existing dynamic functional connectivity methods mostly focus on characterizing direct connections between brain regions (i.e., low-order connections), neglecting higher-order information about the coordinated changes between different connectivity patterns. Furthermore, due to the significant temporal asynchrony of rs-fMRI data, directly modeling and comparing dynamic connectivity time series presents challenges, leading to insufficient feature stability and discriminative power.

[0004] In summary, existing technologies suffer from the following shortcomings: First, most methods fail to effectively combine multi-level dynamic information from low-order direct connections and high-order co-connections, resulting in incomplete feature representation. Second, they lack methods for extracting stable statistical features from dynamic connection time series, making it difficult to overcome the impact of temporal asynchrony. Finally, commonly used classification models (such as shallow machine learning models or traditional convolutional neural networks) have limited global perception and feature extraction capabilities when dealing with such complex brain network structural features, and their decision-making processes lack interpretability, hindering clinicians' understanding and trust. Therefore, there is an urgent need for an ASD-assisted diagnostic method that can integrate multi-level dynamic connection information, extract stable discriminative features, and possess high accuracy and strong interpretability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an ASD-assisted diagnostic method that integrates dynamic low-order and high-order dynamic central moment networks, so as to solve the problems of incomplete feature representation, poor robustness to time asynchrony, and insufficient model classification performance and interpretability in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An ASD-assisted diagnostic method integrating dynamic low-order and high-order dynamic central moment networks includes the following steps:

[0008] Acquire resting-state functional magnetic resonance imaging data of the subjects and perform data preprocessing;

[0009] Based on the preprocessed data, a low-order dynamic functional connection network and a high-order dynamic functional connection network are constructed.

[0010] Extract the central moment features of the low-order dynamic functional connection network and the high-order dynamic functional connection network;

[0011] The low-order central moment features and the high-order central moment features are fused to generate a low-high-order dynamic central moment network;

[0012] The low- to high-order dynamic central moment network is input into the visual Transformer model for classification, and the auxiliary diagnostic results of autism spectrum disorder are output.

[0013] Preferably, the data preprocessing includes the following steps:

[0014] Remove the first 10 time points to ensure magnetization balance;

[0015] Head movement correction and spatial standardization were performed at the Montreal Neuroscience Institute.

[0016] Perform spatial smoothing;

[0017] Bandpass filtering is performed to remove low-frequency drift and high-frequency noise;

[0018] Six head motion parameters were regressed to further reduce the impact of head motion.

[0019] The brain was divided into multiple regions of interest using an automated anatomical tagging brain atlas, and the average time series of each region of interest was extracted.

[0020] Preferably, the construction of the low-order dynamic functional connectivity network adopts the sliding window technique, which sets the window length and step size, divides the time series into multiple overlapping time windows, calculates the Pearson correlation coefficient between all brain region pairs in each window, and generates the low-order dynamic functional connectivity network sequence.

[0021] Preferably, the construction of the low-order dynamic functional connectivity network adopts the sliding window technique, which sets the window length and step size, divides the time series into multiple overlapping time windows, calculates the Pearson correlation coefficient between all brain region pairs in each window, and generates the low-order dynamic functional connectivity network sequence.

[0022] Preferably, the extracted central moment features include:

[0023] For each pair of brain regions in a low-order dynamic functional connectivity network, calculate the d-th order central moments of its dynamic connectivity time series;

[0024] For each connection strength time series in a high-order dynamic functional connectivity network, calculate its d-th order central moment;

[0025] Where d is an integer from 1 to 7, generating the low-order central moment matrix set and the high-order central moment matrix set respectively.

[0026] Preferably, the fusion step adopts a jigsaw puzzle fusion strategy, which combines the low-order central moment matrix and the high-order central moment matrix according to rules to generate a hybrid matrix, and stacks the fusion matrices of all orders along the channel dimension to form a low-to-high-order dynamic central moment network.

[0027] Preferably, the input to the visual Transformer model is a low- to high-order dynamic central moment network matrix, which is divided into non-overlapping patches and linearly embedded and positionally encoded. The patches are then processed by a multi-layer Transformer encoder, and finally the classification probability is output through a fully connected layer and a Softmax function.

[0028] Preferably, the method also includes an interpretability analysis step: based on the attention weight distribution of the visual Transformer model, identify the brain regions that are most discriminative for classification, and determine a list of key brain regions by statistically analyzing the frequency of occurrence of brain regions corresponding to high attention patches.

[0029] This invention provides an ASD-assisted diagnostic method that integrates dynamic low-order and high-order dynamic central moment networks. It includes at least the following beneficial effects:

[0030] By jointly constructing low-order dynamic functional connectivity networks and high-order dynamic functional connectivity networks, we can simultaneously capture direct interactions between brain regions as well as high-order synergistic changes between connectivity patterns, thereby providing a more comprehensive and deeper dynamic representation of brain function and overcoming the limitations of existing methods with their single feature.

[0031] By introducing higher-order central moments to perform statistical analysis on dynamic functional connectivity time series, the modeling challenges caused by temporal asynchrony in resting-state functional magnetic resonance imaging data are effectively overcome. Statistical features that can stably reflect the distribution structure of connectivity strength and synergy are extracted, significantly improving the discriminative power and robustness of the features.

[0032] By employing a jigsaw puzzle fusion strategy to effectively integrate low-order and high-order central moment features, and combining the powerful global modeling capabilities of the visual Transformer model to learn the fused structured features, a high-precision, high-sensitivity, and high-specificity classification of autism spectrum disorders was achieved. Its excellent generalization performance was also verified on multi-site data.

[0033] By analyzing the attention weights of the visual Transformer model, the key brain regions that have the greatest influence on classification decisions can be located. These brain regions are highly consistent with the known neuropathological mechanisms of autism spectrum disorder, which greatly enhances the credibility and clinical interpretability of the model's diagnostic results and provides support for understanding disease mechanisms and discovering biomarkers. Detailed Implementation

[0034] The technical solution of the present invention will be clearly and completely described below. 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.

[0035] Example:

[0036] This invention provides an ASD-assisted diagnostic method that integrates dynamic low-order and high-order dynamic central moment networks. Taking the resting-state functional magnetic resonance imaging data of 92 subjects (45 patients with autism spectrum disorder and 47 normal controls) from the New York University site in the autism brain imaging data sharing platform as an example, all experiments can be completed on a workstation equipped with an NVIDIA RTX4090 graphics processor.

[0037] S1: Data Preprocessing

[0038] Resting-state functional magnetic resonance imaging (fMRI) data from the New York University site were obtained from an autism brain imaging data sharing platform. Preprocessing was performed using statistical parametric mapping software, including the following steps:

[0039] Remove the first 10 time points to ensure magnetization balance;

[0040] Head movement correction and spatial normalization to the space of the Montreal Neuroscience Institute (resolution 3×3×3mm) 3 );

[0041] Smooth spatial design (Gaussian kernel half-height full width 6mm);

[0042] Bandpass filtering (frequency range 0.005–0.1Hz) is used to remove low-frequency drift and high-frequency noise;

[0043] Six head motion parameters were regressed to further reduce the impact of head motion.

[0044] Subsequently, the brain was divided into 90 regions of interest (ROIs) using an automated anatomical tagging brain atlas. The average time series of each ROI was extracted, resulting in a 90 (brain region) × 170 (time point) data matrix for each subject. This matrix reflects the blood oxygenation level-dependent signal fluctuations in the 90 brain regions at 170 time points. Subject demographic information is detailed in Table 1.

[0045] Table 1: Demographic information of the participants:

[0046]

[0047] Note: a Chi-square test; b Two-sample, two-tailed t-test; c Two patients without autism diagnosis interview scale - revised score.

[0048] As shown in Table 1, there were no significant differences between the autism spectrum disorder group and the normal control group in terms of gender, age, and full-scale IQ, ensuring data comparability.

[0049] S2: Constructing a multi-layered dynamic functional connectivity network

[0050] To comprehensively capture the dynamic characteristics of brain functional connectivity in autistic patients, this invention jointly constructs low-order and high-order dynamic functional connectivity networks to characterize direct connections between brain regions and synergistic changes between connectivity patterns.

[0051] S2-1: Constructing a low-order dynamic functional connectivity network

[0052] A low-order dynamic functional connectivity network was constructed using the sliding window technique. With a window length W=30 and a step size S=1, 170 time points were divided into K=141 overlapping time windows. For the k-th window, the Pearson correlation coefficients between all brain region pairs were calculated, generating a symmetric correlation matrix L. k The functional connectivity between brain regions i and j in the k-th window is calculated as follows:

[0053] ;

[0054] in, and For brain regions and the Blood oxygen levels depend on the signal sequence. This represents the mean of the corresponding sequence. Coefficient. The value range is between [−1, 1], reflecting the brain region and The strength of functional connections between them. The matrix of all windows constitutes a dynamic functional connection network sequence. This forms a three-dimensional tensor L∈R141×90×90 It encodes the dynamic evolution of brain region interactions over time.

[0055] S2-2: Constructing a high-order dynamic functional connectivity network

[0056] A high-order dynamic functional connectivity network is constructed by characterizing the similarity of connectivity patterns in different brain regions (i.e., "connectivity correlation"). For the low-order functional connectivity matrix L within the k-th time window... k The i-th row can be interpreted as the short-term functional connectivity feature vector of brain region i:

[0057] ;

[0058] For any pair of brain regions i and j, calculate the Pearson correlation between their short-term functional connectivity profiles during the window period:

[0059] ;

[0060] in, This represents the "connection strength" between brain regions i and j within the k-th time window. (This is achieved by organizing all...) Form a symmetric matrix H k ∈R 90×90 This yields the higher-order functional connectivity matrix for the k-th time window. Repeating this process constructs the complete sequence of higher-order dynamic functional connectivity networks. This sequence reflects coupling variations in connectivity patterns across multiple brain regions, which may help identify potential complex functional abnormalities in patients with autism spectrum disorder.

[0061] S2-3: Time-based data augmentation

[0062] To enhance data diversity and improve the robustness of temporal feature modeling, a data augmentation strategy based on time series reconstruction is introduced. After constructing the low-order and high-order dynamic functional connection network sequences, the complete sequence is... Treating it as a cyclic timeline, the enhanced sequence is generated by randomly selecting any time point as the new starting position. This operation yields multiple recombined versions of the original dynamic connection sequence, each retaining its inherent connection content while presenting the temporal pattern in an offset-aligned manner. By observing the same connection dynamics from different temporal perspectives, this strategy expands the number of training samples to over 1000 without introducing artificial synthetic distortion, effectively improving the model's generalization ability.

[0063] S3: High-order central moment feature extraction and fusion

[0064] Because brain activity during resting-state scans is not task-specific, the time series of functional connectivity between different individuals exhibit significant temporal asynchrony, making direct comparison or modeling of the temporal connectivity matrix quite challenging. To address this issue, this invention introduces a higher-order statistic—the higher-order central moment—to extract stable and discriminative dynamic features.

[0065] S3-1: Extracting Central Moment Features

[0066] For any pair of brain regions (i, j) in a low-order dynamic functional connectivity network, its dynamic connectivity time series is defined as follows: The central moment of order d is calculated as follows:

[0067] ;

[0068] in, express The mean. Similarly, for the connection strength time series in high-order dynamic functional connectivity networks. Its central moment is defined as:

[0069] ;

[0070] This calculation process yields the set of low-order central moment matrices. and higher-order central moment matrix set Each matrix element The d-order central moments of the corresponding brain region pairs represent the distribution structure of connection strength (lower order) and connection synergy (higher order) at different levels.

[0071] S3-2: Jigsaw puzzle-style fusion construction of low- to high-order dynamic central moment networks

[0072] To preserve complementary information from low-order and high-order networks, a jigsaw puzzle-style fusion strategy was designed. For each central moment order... , lower-order matrix and higher-order matrices Merge according to the following rules:

[0073] ;

[0074] This process generates a mixing matrix. It simultaneously contains direct interaction information and higher-order coordination information between brain regions. Subsequently, a fusion matrix of all orders is then used. Stacking along the channel dimension forms a composite two-dimensional representation, namely a low-to-high-order dynamic central moment network with dimensions of 90×90×7. This fusion mechanism effectively reorganizes multi-level brain connectivity statistics into an image-like structure, enabling subsequent Transformer-based models to learn rich and structured representations.

[0075] S4: Classification and Diagnosis Based on Visual Transformer Model

[0076] To effectively model and integrate multi-level, multi-regional dynamic connectivity information embedded in feature maps, a visual Transformer is introduced as a classification module. The visual Transformer utilizes a self-attention mechanism to achieve global modeling capabilities, enabling it to capture long-range dependencies, which is particularly advantageous for feature extraction and discrimination tasks in complex brain network structures.

[0077] S4-1: Model Input and Preprocessing

[0078] Treating the low- to high-order dynamic central moment network matrix as a two-dimensional "image" input, denoted as X∈R 90×90×7 The matrix is ​​divided into non-overlapping tiles of size 15×15. Each tile is flattened and passed through a linear projection layer to generate a fixed-length vector representation, forming a tile sequence. The total number of tiles is K=36.

[0079] S4-2: Visual Transformer Model Structure

[0080] The workflow of a visual Transformer model includes the following stages:

[0081] Linear embeddings and positional encoding: All tile embeddings are concatenated into a matrix, and a learnable positional encoding E is added. pos To preserve spatial order, the input representation of the Transformer encoder is calculated as follows:

[0082] ;

[0083] Where E is a trainable embedding matrix.

[0084] Transformer encoder architecture: The backbone network of the visual Transformer consists of 6 stacked Transformer encoder blocks, each containing a multi-head self-attention mechanism (8 heads) and a feedforward neural network. Each sub-module is equipped with residual connections and layer normalization to ensure stable training and information flow.

[0085] Classification output: The feature sequence from the final Transformer encoder is compressed through a fully connected layer, and the class probabilities are generated using the Softmax function. For binary classification tasks, the output layer contains two neurons, representing the probabilities of autism spectrum disorder and normal control, respectively.

[0086] S4-3: Model Training and Diagnosis

[0087] Using the Adam optimizer (initial learning rate 1×10⁻⁶) −3 The model was trained for 100 epochs using a batch size of 16 and a cross-entropy loss function. The entire model was optimized through backpropagation, and the class with the higher probability value output by the model was considered the auxiliary diagnostic result for that subject.

[0088] S5: Interpretability analysis (the most discriminative feature in the diagnosis of autism spectrum disorder)

[0089] To further enhance the interpretability of the model and reveal the specific brain regions that the visual Transformer-Central Moment Network model focuses on during the classification process, a tile-level attention tracking method based on Transformer attention weight distribution is proposed, aiming to identify the feature regions that are most discriminative for the diagnosis of autism spectrum disorder.

[0090] Specifically, the fused feature map low-to-high-order dynamic central moment network is divided into 36 fixed-size (15×15) non-overlapping patches, each corresponding to a sub-region in the original connection matrix. The attention weights in the final Transformer layer are extracted to obtain the model's attention level to each patch when making the final classification decision. To improve the robustness of the results, the attention distribution of all subjects is averaged, and then averaged across all attention heads in the last layer to obtain the final patch-level attention score.

[0091] Subsequently, all patches were sorted according to their attention weights, and the top 20% of patches with the highest attention were selected as the most sensitive regions for classification. These patches were then mapped back to their corresponding row and column indices to identify the brain region pairs covered by each patch. Furthermore, each pair of brain regions was broken down into individual brain regions, and their frequency of occurrence in all high-attention patches was counted. The top 10 most frequently occurring brain regions were identified as the key brain regions with the highest attention in the visual Transformer-Central Moment Network model for autism spectrum disorder classification.

[0092] Attention tracking analysis using a visual Transformer-Central Moment Network model identified the top 10 most influential brain regions, including the superior frontal gyrus (left / right), precuneus (left), and hippocampus (left / right). These regions frequently appeared in patches of high model attention, with occurrence probabilities ranging from 10.02% to 2.75%, indicating their significant importance in classification tasks. Notably, most of these brain regions are closely associated with core symptoms of autism spectrum disorder. For example, the superior frontal gyrus and precuneus are key nodes in the default mode network, involved in social cognition and theory of mind; the hippocampus is involved in emotion regulation and episodic memory; and the postcentral gyrus and the trigone of the inferior frontal gyrus are associated with motor comprehension and the mirror neuron system. The consistency of the model's attention with existing neuroimaging literature demonstrates that the proposed method maintains high discriminative performance while possessing strong biological interpretability, providing reliable support for a deeper understanding of the neural mechanisms of autism spectrum disorder and the identification of potential neurobiomarkers.

[0093] Experimental results:

[0094] To verify the effectiveness of the proposed method, this embodiment conducted a systematic experiment on the New York University dataset, an autism brain imaging data exchange platform, and evaluated it from three aspects: comparison of input feature modeling methods, comparison of classification models, and cross-site generalization performance. All experiments used the same preprocessing and data partitioning strategies to ensure the comparability of results.

[0095] 1. Comparison of Input Feature Modeling Methods

[0096] To evaluate the impact of different input feature modeling strategies on the classification performance of autism spectrum disorder, a series of systematic comparative experiments were designed. All models used the same visual Transformer architecture as the classifier, differing only in their input features. The abbreviations and definitions of each input feature are as follows:

[0097] Static low-order functional connectivity: Calculation of single functional connectivity matrices based on complete resting-state functional magnetic resonance imaging time series.

[0098] Static low-order central moment features: Extracting the seventh-order central moment features of connection strength based on static low-order functional connections.

[0099] Dynamic low-order functional connectivity: Multiple functional connectivity matrices are obtained using the sliding window method, and their average value is taken.

[0100] Dynamic higher-order functional connectivity: Calculates the "correlation of connectivity" between brain regions based on dynamic lower-order functional connectivity.

[0101] Dynamic low-order central moment features and dynamic high-order central moment features: Seventh-order central moments are extracted from dynamic low-order functional connections and dynamic high-order functional connections in the time dimension, respectively.

[0102] Fusion Feature Low-High Order Dynamic Central Moment Network: The dynamic low-order central moment features and dynamic high-order central moment features are fused using a jigsaw puzzle structure.

[0103] The experimental results are shown in Table 2.

[0104] Table 2: Results of different feature modeling methods:

[0105]

[0106] As shown in Table 2, introducing dynamic information significantly improves model performance. Compared to static low-order functional connectivity (74.8% accuracy), dynamic low-order functional connectivity, by modeling the temporal evolution of brain connections, improves accuracy to 75.0%; dynamic high-order functional connectivity further improves to 76.9%, demonstrating the discriminative value of high-order information. Central moment modeling also plays a positive role in performance improvement, with dynamic low-order and dynamic high-order central moment features achieving accuracies of 77.2% and 77.5%, respectively. The low-to-high-order dynamic central moment network method proposed in this invention performs best, achieving an accuracy of 79.8%, sensitivity and specificity of 79.3% and 80.0%, respectively, and an F1 score of 79.4%, proving that low-order and high-order features have good complementarity.

[0107] 2. Classification Model Comparison Experiment

[0108] To further verify the effectiveness of the visual Transformer architecture in processing fused feature maps, a series of comparative experiments were designed, covering various classification models. All experiments used low- to high-order dynamic central moment networks as input features to ensure that performance differences stemmed solely from the classification model architecture itself. The comparison models included random forests, support vector machines, multilayer perceptrons, residual networks, convolutional neural networks, graph convolutional networks, and variants of the visual Transformer (such as no positional encoding, 3-layer encoders, and single-head attention). The experimental results are shown in Table 3.

[0109] Table 3: Results of different classification models:

[0110]

[0111] Table 3 shows that traditional shallow learning methods (such as random forests and support vector machines) generally perform poorly, while deep models typically outperform shallow methods. The Visual Transformer and its variants consistently outperform other models, with the complete Visual Transformer-Central Moment Network achieving the highest accuracy (79.8%). The performance degradation of the Visual Transformer variants validates the crucial roles of positional encoding, network depth, and multi-head attention mechanisms, indicating that these components synergistically enhance the model's ability to perceive brain connectivity map structures and extract semantic features.

[0112] 3. Cross-site generalization performance

[0113] To verify the model's generalization ability, cross-site experiments were conducted on 14 independent imaging sites in the autism brain imaging data exchange platform dataset. The results are shown in Table 4.

[0114] Table 4: Cross-site generalization performance:

[0115] As shown in Table 4, the Visual Transformer-Central Moment Network performed optimally at 7 out of 14 sites, achieving an average accuracy of 76.5%. This consistently outperformed Convolutional Neural Networks (63.2%), Support Vector Machines (68.3%), and K-Nearest Neighbors (59.0%), and slightly surpassed Multi-Head Self-Attention-Functional Connectivity Networks (73.4%). This demonstrates that the Visual Transformer-Central Moment Network possesses strong robustness and generalization ability in addressing site heterogeneity such as differences in scanning protocols and variations in data distribution, making it crucial for practical clinical applications.

[0116] In summary, this embodiment details an ASD-assisted diagnostic method that integrates dynamic low-order and high-order dynamic central moment networks. By constructing low-order and high-order dynamic functional connectivity networks, extracting multi-order central moment features, and employing a jigsaw-style fusion strategy to generate a low-to-high-order dynamic central moment network, combined with the global modeling capabilities of the visual Transformer model, a highly accurate (79.8%) and highly interpretable autism spectrum disorder classification was achieved. The key brain regions identified by the model are highly consistent with known neuropathologies of autism spectrum disorders, demonstrating the biological rationale of the method. Cross-site experiments further validated the model's generalization ability. This invention provides an efficient and reliable solution for the assisted diagnosis of autism spectrum disorders and has the potential to be extended to other neurological diseases.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An ASD-assisted diagnostic method integrating dynamic low-order and high-order dynamic central moment networks, characterized in that, Includes the following steps: Acquire resting-state functional magnetic resonance imaging data of the subjects and perform data preprocessing; Based on the preprocessed data, a low-order dynamic functional connection network and a high-order dynamic functional connection network are constructed. Extract the central moment features of the low-order dynamic functional connection network and the high-order dynamic functional connection network; The low-order central moment features and the high-order central moment features are fused to generate a low-high-order dynamic central moment network; The low- to high-order dynamic central moment network is input into the visual Transformer model for classification, and the auxiliary diagnostic results of autism spectrum disorder are output.

2. The ASD-assisted diagnostic method based on the fusion of dynamic low-order and high-order dynamic central moment networks according to claim 1, characterized in that, The data preprocessing includes the following steps: Remove the first 10 time points to ensure magnetization balance; Head movement correction and spatial standardization were performed at the Montreal Neuroscience Institute. Perform spatial smoothing; Bandpass filtering is performed to remove low-frequency drift and high-frequency noise; Six head motion parameters were regressed to further reduce the impact of head motion. The brain was divided into multiple regions of interest using an automated anatomical tagging brain atlas, and the average time series of each region of interest was extracted.

3. The ASD-assisted diagnostic method based on the fusion of dynamic low-order and high-order dynamic central moment networks according to claim 1, characterized in that, The construction of the low-order dynamic functional connectivity network adopts the sliding window technique, which sets the window length and step size, divides the time series into multiple overlapping time windows, calculates the Pearson correlation coefficient between all brain region pairs in each window, and generates the low-order dynamic functional connectivity network sequence.

4. The ASD-assisted diagnostic method based on the fusion of dynamic low-order and high-order dynamic central moment networks according to claim 1, characterized in that, The construction of the high-order dynamic functional connectivity network includes: for the low-order functional connectivity matrix within each time window, calculating the Pearson correlation between the short-term functional connectivity profiles of any two brain regions, and generating a high-order dynamic functional connectivity network sequence.

5. The ASD-assisted diagnostic method based on the fusion of dynamic low-order and high-order dynamic central moment networks according to claim 1, characterized in that, The extracted central moment features include: For each pair of brain regions in a low-order dynamic functional connectivity network, calculate the d-th order central moments of its dynamic connectivity time series; For each connection strength time series in a high-order dynamic functional connectivity network, calculate its d-th order central moment; Where d is an integer from 1 to 7, generating the low-order central moment matrix set and the high-order central moment matrix set respectively.

6. The ASD-assisted diagnostic method based on the fusion of dynamic low-order and high-order dynamic central moment networks according to claim 1, characterized in that, The fusion step adopts a jigsaw puzzle fusion strategy, which combines low-order central moment matrices and high-order central moment matrices according to rules to generate a hybrid matrix, and stacks the fusion matrices of all orders along the channel dimension to form a low-to-high-order dynamic central moment network.

7. The ASD-assisted diagnostic method based on the fusion of dynamic low-order and high-order dynamic central moment networks according to claim 1, characterized in that, The visual Transformer model takes a low- to high-order dynamic central moment network matrix as input, divides it into non-overlapping patches, performs linear embedding and position encoding, processes it through a multi-layer Transformer encoder, and finally outputs the classification probability through a fully connected layer and a Softmax function.

8. The ASD-assisted diagnostic method based on the fusion of dynamic low-order and high-order dynamic central moment networks according to claim 1, characterized in that, It also includes an interpretability analysis step: based on the attention weight distribution of the visual Transformer model, identify the brain regions with the strongest discriminative power for classification, and determine the list of key brain regions by statistically analyzing the frequency of occurrence of brain regions corresponding to high attention patches.

Images