Neural development disorder co-disease identification system based on residual image neural network

Abstract

The invention discloses a neural development disorder co-disease identification system based on a residual image neural network. Belongs to the technical field of neurodevelopment disorder co-disease recognition, and particularly relates to the technical field of co-disease recognition based on a neural network. The system comprises a data acquisition module for acquiring a resting-state fMRI image; the data preprocessing module is used for extracting standardized time sequences of 116 brain regions of the set resting state fMRI image; the multi-band division and feature extraction module is used for performing frequency band filtering on the standardized time sequence of each brain region to obtain a time sequence after each frequency band filtering; in each frequency band, constructing a binary topological matrix based on the PLV matrix and calculating a PLV feature vector based on the PLV matrix; the neurodevelopmental disorder co-disease recognition module is used for analyzing the brain function connection diagram through a residual image neural network model to obtain diagnostic information of neurodevelopmental disorder co-diseases; and the interpretable output module is used for outputting an abnormal brain region by taking the amplitude low-frequency fluctuation as a reference index.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of neurodevelopmental disorder comorbidity recognition, and particularly relates to the technical field of comorbidity recognition based on a neural network. BACKGROUND

[0002] Current automatic recognition and mechanism modeling for neurodevelopmental disorders (such as autism spectrum disorder ASD and attention deficit hyperactivity disorder ADHD) mostly adopt brain function connection analysis methods based on resting-state functional magnetic resonance imaging (rs-fMRI), including the following several mainstream technical paths: Traditional classification models (such as SVM and RF): usually based on full connection matrices or statistical features to construct feature vectors, and then input into traditional machine learning classifiers. Although this kind of method is simple to implement, it lacks the modeling ability of complex interaction structure between brain regions, and is difficult to capture dynamic network changes and frequency sensitivity.

[0003] Convolutional neural networks (such as BrainNetCNN): use specific topological convolution to model connection matrices, but their structure is usually fixed, lack the ability of node semantic modeling, and it is difficult to integrate local brain region features and frequency band characteristics, which has the problem of "black box", and lacks biological interpretability.

[0004] Graph neural network methods (such as BrainGNN, MHNet and HAGCN): In recent years, brain graph modeling methods based on GNN have appeared, which regard ROIs as nodes of the graph and functional connections as edges, and can capture structural dependency relationships. However, most of the existing methods have the following shortcomings: Frequency insensitivity: without introducing frequency domain features such as slow wave frequency bands, it cannot reflect the changes in brain function state under different frequencies; Weak dynamic mechanism modeling: lack of modeling of nonlinear dynamic features of original BOLD time series, such as LLE, etc. Connection mechanism is not clear: most of the graph construction methods use static linear indicators such as Pearson correlation, ignoring more physiologically meaningful connection methods such as neural activity phase synchronization; Weak interpretability: most models do not provide explanations of the contribution of frequency bands / brain regions / connections to the classification results, making it difficult to analyze the mechanism and clinical transformation. SUMMARY

[0005] In order to solve the key problems of frequency insensitivity, insufficient modeling of brain region dynamic features, lack of biological explanation of functional connections, and strong black box of the model in the existing technology, the application provides a neurodevelopmental disorder comorbidity recognition system based on a residual graph neural network.

[0006] The system comprises: A data acquisition module acquires resting-state fMRI images. The data preprocessing module: based on the AAL brain region template, the standardized time series of 116 brain regions of the resting-state fMRI image is extracted: ; and represents any one of the time series; The multi-band division and feature extraction module: the standardized time series of each brain region is filtered in four frequency bands respectively, to obtain the time series filtered in each frequency band, and the node feature matrix of the time series is extracted; in each frequency band, PLV calculation is performed to obtain the PLV matrix, and a binary topological matrix is constructed based on the PLV matrix and the PLV feature vector is calculated based on the PLV matrix; the node feature matrix, the binary topological matrix and the PLV feature vector are combined to obtain the brain functional connectivity graph; The neurodevelopmental disorder comorbidity recognition module: through the residual graph neural network model, the brain functional connectivity graph is analyzed to obtain the diagnosis information of the neurodevelopmental disorder comorbidity; The interpretable output module: taking the amplitude low-frequency fluctuation as a reference index, the abnormal brain region is output.

[0007] Further, in the data preprocessing module, before generating the standardized time series, data preprocessing is also performed, specifically: head stripping operation, head motion correction operation, spatial smoothing operation through Gaussian filtering, linear trend removal operation through the band-pass filtering tool of FSL, mean functional image calculation operation and standard space alignment operation using the mean functional image are performed in sequence.

[0008] Further, the four frequency bands are: SLOW2 (0.198-0.25Hz), SLOW3 (0.073-0.198Hz), SLOW4 (0.027-0.073Hz) and SLOW5 (0.01-0.027Hz); The time series filtered in each frequency band is , wherein represents the index of the brain region, represents the index of the frequency band, and the node feature matrix includes statistical features and nonlinear features.

[0009] Further, the statistical features include mean , standard deviation , skewness and kurtosis : ; wherein, represents the number of time points contained in the time series of each brain region in a specific frequency band, represents the index of the time point; ; ; .

[0010] Furthermore, the method for obtaining the nonlinear characteristics is: S51. Local neighborhood construction: In time series The signal value at each time point is , using the K nearest neighbor algorithm, for each time point Find the five most similar neighboring points, and measure the similarity using the Euclidean distance: ,in, Representation and time point Similar time point index, Indicates a time point and time point The Euclidean distance of ; select the 5 smallest points as the nearest neighbors, denoted as , forming a time point The local neighborhood of S52, reconstruction weight solution: for each time point , assuming that its signal value is reconstructed by a linear combination of neighboring points, that is ,in Indicates a time point Time point To minimize the reconstruction error, solve the following optimization problem: ; The constraints are ,in Indicates a time point The set of reconstruction weights of all topological points in the solution process is to first centralize the local neighborhood signal, that is, to subtract , and get the matrix , and then solve the weights by the least squares method : ;in To avoid matrix singularity, the regularization term 1 is a vector of all 1s. is the identity matrix; S53, Feature Quantization: Through Get each Nonlinear characteristics in each frequency band, where represents the variance function.

[0011] Furthermore, the PLV calculation is as follows: S61. Hilbert transform: Convert to analytical signal ,in represents the Hilbert transform; S62, Phase Extraction: Pass obtaining the phase angle of the resolved signal , denotes the operation of solving the phase angle; S63, PLV calculation: quantify the synchronization by the average cosine value of the phase difference, the formula is: , wherein is the imaginary unit, the greater the value, the stronger the phase synchronization between the two , and respectively represent the instantaneous phase angle of the two at a certain time point, and the PLV matrix is obtained by PLV calculation.

[0012] Further, the binary topological matrix is constructed based on the PLV matrix, specifically: when the value in the PLV matrix is , it is considered that there is an edge between the two , and all edge indexes form a binary topological matrix .

[0013] Further, the PLV eigenvector is calculated based on the PLV matrix, specifically: extract the upper triangular elements of each frequency band PLV matrix to form a one-dimensional eigenvector, and after splicing the eigenvectors of the four frequency band PLV matrices, a global connection feature , is the PLV eigenvector.

[0014] Further, the workflow in the residual graph neural network model is: S91, local node dynamic capture: through the first layer of graph convolution operation, the features of each node and the neighbor node features are weighted and aggregated to obtain the feature matrix : , wherein denotes the activation function, denotes the normalized degree matrix, denotes the binary topological matrix with self-connection , denotes the node feature matrix, denotes the learnable weight; S92, regional network feature aggregation: through the second layer of graph convolution operation, the feature extraction range is expanded to a larger brain network unit to obtain the feature matrix : , wherein denotes the learnable weight; S93, whole brain topological pattern mining: integrate the cross-network connection features in the whole brain range to form a global topological pattern representation, and obtain the feature matrix : ,in Represents learnable weights; using feature matrix Through global mean pooling, all The node features are aggregated into a single vector ; S94, Focus on PLV phase synchronization pattern mining: Through multi-layer nonlinear transformation, PLV feature vectors are mapped to low-dimensional space while retaining the core pattern : ,in express activation function, is the PLV feature vector, is the learnable weight; S95, cross-modal feature fusion: Will and Obtain fused features through adaptive weight fusion : ,in is the fusion weight of network learning; S96. Classification result output: The fused features First, it is processed by two layers of fully connected networks: , Among them, the first layer is through the weight matrix Compress the feature dimension to 128 dimensions, retaining the core discriminant information, To prevent overfitting function; the second layer passes the weight matrix Further refine the key correlations between features; pass softmax The function maps the processed features into three types of probability distributions: , is the weight matrix, output ,in The probabilities of having only ASD, having both ADHD and ASD, and having ASD alone are respectively taken, and the category corresponding to the maximum probability is taken as the final prediction result.

[0015] Furthermore, the amplitude of low-frequency fluctuations is calculated by The amplitudes of low-frequency oscillation signals in the four frequency bands were obtained. Independent sample tests were used to compare the differences in the low-frequency fluctuation values ​​of the amplitudes in each frequency band between the comorbidity group and the single disease group, and statistically significant abnormal brain areas were screened and output.

[0016] The system has the advantages that the system can be used for processing resting state fMRI brain network data, realizing recognition of ADHD, ASD and comorbid individuals, and having cross-frequency and cross-structure neural connection abnormality interpretation capability.

[0017] The system has the following technical advantages: Strong frequency sensitivity: SLOW-2 to SLOW-5 multiple slow wave frequency bands are comprehensively introduced, and frequency-dependent brain function activity features are captured through filtering and feature fusion, thereby enhancing the sensitivity of the model to ASD / ADHD and other frequency-specific abnormalities.

[0018] Accurate dynamic feature modeling: Local linear embedding (LLE) features are fused to accurately depict the nonlinear dynamic structure of the brain region and make up for the expression limitations of traditional statistical features.

[0019] Physiologically more meaningful connection mode: The phase locking value (PLV) is used as the brain region connection index to reflect the phase synchronization relationship between neural activities, which is more in line with the actual mechanism of neurophysiology.

[0020] Structural-dynamic fusion modeling: The residual collaborative modeling of graph convolution network and global PLV features is performed, and the node features (local) and connection features (global) are integrated to establish an end-to-end structural-functional coupling recognition model.

[0021] Multi-dimensional interpretable mechanism analysis: The frequency band contribution visualization, PLV connection difference analysis and brain region significance statistics are introduced into the model to build a complete neural network abnormal mechanism visualization system, thereby improving the clinical interpretability and transformation value of the results. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The system workflow diagram of the embodiment of the present application is shown in the figure. Figure 2 The abnormal brain region difference diagram of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0023] The technical solutions of the present application will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] Embodiment 1, The embodiment provides a neurodevelopmental disorder comorbidity recognition system based on a residual graph neural network, and the working process of the system is shown in the figure. Figure 1 The system comprises: A data acquisition module acquires resting state fMRI images; The data preprocessing module extracts the standardized time series of 116 brain regions of the resting-state fMRI image based on the AAL brain region template: Any one of the time series is represented as: The multi-band division and feature extraction module filters the standardized time series of each brain region in four frequency bands respectively to obtain the time series filtered in each frequency band, extracts the node feature matrix of the time series, calculates the PLV matrix in each frequency band to obtain the PLV matrix, and constructs the binary topological matrix based on the PLV matrix and calculates the PLV feature vector based on the PLV matrix; the node feature matrix, the binary topological matrix and the PLV feature vector are combined to obtain the brain functional connectivity graph; The neurodevelopmental disorder comorbidity recognition module analyzes the brain functional connectivity graph through a residual graph neural network model to obtain the diagnosis information of the neurodevelopmental disorder comorbidity; The explainability output module takes the amplitude low-frequency fluctuation as a reference index to output the abnormal brain region.

[0025] Embodiment 2, This embodiment is a further limitation of embodiment 1, and further describes the data acquisition module.

[0026] The data acquisition module acquires the resting-state fMRI image, and the resolution is recommended to be not less than 3mm³, and the time length is recommended to be greater than or equal to 5 minutes.

[0027] Embodiment 3, This embodiment is a further limitation of embodiment 1, and further describes the data preprocessing module.

[0028] The data preprocessing step in the data preprocessing module is: The following preprocessing steps are performed on the original rs-fMRI data using the FSL (FMRIB Software Library) tool, and the specific operations are as follows: ​1. Skull stripping: Apply the bet command of FSL to the original structural or functional image to remove non-brain tissues and preserve the pure brain region for accurate registration and analysis later. 2. Motion correction: Use the mcflirt tool of FSL to rigidly register all echo planar imaging (EPI) volumes to the first volume to correct the head motion of the subject during scanning and reduce motion artifacts. 3. Spatial smoothing: Apply Gaussian filtering to the motion-corrected data with a kernel width corresponding to a FWHM (full width at half maxima) of about 6 mm to improve the signal-to-noise ratio and smooth spatial noise. 4. Linear trend removal: Use the bandpass filter tool of FSL to remove linear trends and high-frequency noise in the time series to reduce slow drift and interference from non-neural signals. 5. Calculate mean functional image: Calculate the mean functional image after trend removal as the basis for subsequent registration. 6. Spatial registration to standard space: Use the flirt of FSL to perform affine registration to register the mean functional image to the MNI152 standard template (4x4x4 mm resolution) and apply the transformation to the entire 4D de-trended data to achieve standard space alignment. 7. Extract ROI time series: Based on the AAL brain region template, use Nilearn's NiftiLabelsMasker to extract the standardized time series of 116 brain regions from the resting-state fMRI image: ; use to represent any one of the time series.

[0029] Example 4, This embodiment is a further limitation of Example 1, and further describes the multi-band division and feature extraction module.

[0030] Brain function activities exhibit different characteristics at different frequencies, especially slow wave activities are closely related to neurodevelopmental disorders. Therefore, the present application focuses on four key slow wave frequency bands: SLOW2 (0.198-0.25Hz): related to rapid cognitive processing; SLOW3 (0.073-0.198Hz): associated with attention regulation; SLOW4 (0.027-0.073Hz): involved in memory and emotional processing; SLOW5 (0.01-0.027Hz): related to default network function. The division of these frequency bands is based on previous neuroimaging research, which can specifically capture the abnormal patterns of brain function in ASD, ADHD and comorbid patients.

[0031] Within each frequency band, the following two types of features (statistical features and nonlinear features) are extracted to form a node feature matrix.

[0032] 1. Statistical feature extraction Time series after filtering each frequency band , compute 4 basic statistics to characterize the signal from different perspectives: Mean Reflects the overall activity level in the frequency band, , where denotes the number of time points contained in the time series of each brain region in a specific frequency band, denotes the index of the time point; Standard Deviation Reflects the stability of the activity, ; Skewness Reflects the asymmetry of the signal distribution, ; Kurtosis : Reflects the steepness of the signal distribution, ; These features can effectively distinguish the differences in activity intensity and stability of specific brain regions.

[0033] 2. Nonlinear feature extraction (LLE Local Linear Embedding) LLE features aim to capture the local dynamic patterns (such as the correlation and stability of short-term fluctuations) of individual time series. The core idea is that the time series can be reconstructed by linear combination of neighboring points within a local range, and the dispersion of local patterns is quantified by solving the variance of reconstruction weights.

[0034] (1) Construction of local neighborhood For the time series after band-pass filtering , first determine the neighbors of each time point. Let the length of the time series be , and the signal value of each time point be .

[0035] Use the K-Nearest Neighbors (K-Nearest Neighbors) algorithm to find the most similar nearest neighbors for each time point . The distance measures the similarity through the Euclidean distance: ; where denotes the index of the similar time point to time point , and denotes the Euclidean distance between time point and time point ; Select the smallest 5 points as neighbors, denoted as , to form the time point local neighborhood.

[0036] (2) Reconstruction weight solution For each time point , assuming that its signal value can be reconstructed by a linear combination of neighboring points, that is ; in Indicates a time point Time point The contribution weight of (the weight of non-neighbor points is 0), in order to minimize the reconstruction error, solve the following optimization problem: ; The constraints are (make sure weights are normalized to avoid scaling bias), where Indicates a time point The set of reconstruction weights of all topological points in the solution process is to first centralize the local neighborhood signal, that is, to subtract , and get the matrix , and then solve the weights by the least squares method :

[0037] in 1 is a vector of all 1s, is the identity matrix, the main diagonal is all 1, and the rest are all 0; Trace() means (Physical meaning: equivalent to The sum of the squared Euclidean distances of all column vectors), and finally the weights Normalize to satisfy the constraints.

[0038] (3) LLE feature quantization Reconstruction weights It reflects the contribution strength of the neighboring points to the time point k, and its variance can quantify the stability of the local reconstruction pattern: if the weight variance is small, it means that the contribution distribution of the neighboring points is stable, and if the weight variance is large, it means that the contribution of the neighboring points fluctuates violently. Therefore, the LLE feature is defined as the variance of the reconstruction weights at all time points:

[0039] in is the variance function, and finally each The LLE feature at each frequency band is a scalar value.

[0040] Functional connectivity maps were constructed based on PLV, and connectivity features were designed to quantify different The phase locking value (PLV) is used as the core indicator to measure the functional synergy between brain regions, which can reflect the phase synchronization (coordination of neural activities) between brain regions.

[0041] The PLV calculation is as follows: (1) Hilbert transform: transforms the time series Convert to analytical signal , its imaginary part reflects the instantaneous phase of the signal; (2) Phase extraction: The phase of the analytical signal is , represents the phase angle of the signal at each time point; (3) PLV calculation: The synchronization is quantified by the average cosine value of the phase difference. The formula is: ;in is the imaginary unit, The larger the value, the The stronger the phase synchronization, and Respectively represent two The instantaneous phase angle at a specific time point is used to obtain the PLV matrix through PLV calculation.

[0042] Constructing a binary topological matrix based on the PLV matrix: Constructing a brain functional connection map based on the PLV matrix, with 116 "nodes" , “edge” means Connections between: when When, think and There is an edge; there are only two cases of an edge: existence or non-existence, that is, 0 / 1; All edge indices form a binary topological matrix , records the topological structure of the graph.

[0043] PLV feature vector extraction: To retain the intensity information of PLV, the upper triangular elements of the PLV matrix of each frequency band are extracted to form a one-dimensional feature vector. After the feature vectors are concatenated, a global connection feature is formed , as the global connectivity property of the graph. Among them, 26680 is composed of the upper triangular elements of 4 bands (116 115 / 2) obtained.

[0044] After the above steps, each sample (patient) is converted into a brain functional connectivity map.

[0045] Example 5 This embodiment further limits the embodiment 1 and further illustrates the neurodevelopmental disorder comorbidity identification module.

[0046] MBF-ResGCN deeply mines disease-specific patterns hidden in brain functional connectivity maps by integrating multi-stage feature extraction with neurodynamic features. This model closely integrates the dynamic characteristics of nodes in brain networks with neurodynamic connectivity patterns centered on PLVs, constructing a feature mining framework tailored to the characteristics of brain functional data. The multi-stage graph feature extraction module employs a progressive feature distillation strategy, leveraging multi-level graph convolution operations to model dynamic brain patterns from local to global scales. This allows for the efficient transfer and expression of deep graph features without relying on traditional residual structures.

[0047] Local node dynamic capture: Node feature matrix for input (Contains 116 The multi-band statistical features and nonlinear features of the proposed method are 116 × 20 = 116 brain regions × (4 statistical features + 1 nonlinear feature) × 4 frequency bands.

[0048] Through the first layer of graph convolution operation, focusing on a single The core of this is to use the graph topology structure (binary topology matrix) to ), perform weighted aggregation of each node’s features and neighbor node features to obtain the feature matrix : ; in To incorporate the binary topology matrix of self-connection (Retain the node's own dynamic characteristics), are learnable weights, is the activation function, Represents the normalized degree matrix (mathematical form: diagonal matrix (non-diagonal elements are 0)). This stage focuses on capturing the functional synergy pattern between a single brain region and its adjacent regions (such as the local interaction between the prefrontal cortex and the cingulate gyrus), laying the foundation for subsequent analysis.

[0049] It is a binary topological matrix Extensions based on this (incorporating self-joins), exist On the basis of adding "node itself and its own connection": for each (node), The position corresponding to "own row, own column" in will be set to 1 (or other weights used to retain its own features) to retain the dynamic features of the node itself; Only describe the difference between The connection between different regions (no self-connection, i.e. the elements on the diagonal of the matrix are all 0).

[0050] Regional network feature aggregation: After the first stage of processing, the feature matrix has already contained local correlation information. The second stage extends the feature extraction range to larger brain network units through deeper graph convolution: ; This stage strengthens the feature capture of disease-related brain networks and filters irrelevant connection information through the learning of the weight matrix , highlighting the functional abnormality pattern at the regional level.

[0051] Whole brain topology pattern mining: Further expand the receptive field and integrate cross-network connection features in the whole brain range to form a global topology pattern representation: ; Finally, the features of the 116 nodes are aggregated into a single vector through global average pooling, which condenses the dynamic characteristics and topology structure information of the whole brain nodes.

[0052] Focus on PLV phase synchronization pattern mining: As a key indicator reflecting the neural dynamic synchronization between brain regions, the phase-locked value (PLV) contains phase coordination patterns that are key biomarkers for distinguishing the three types of diseases. MBF-ResGCN designs a dedicated processing path to deeply mine the dynamic information in PLV: High-dimensional PLV feature dimension reduction and enhancement: the original PLV feature (the upper triangular elements of 4 frequency bands spliced) has a very high dimension and contains a lot of redundant information. Through multiple nonlinear transformations, it is mapped to a low-dimensional space while retaining the core patterns: where is the learnable weight, and through the screening effect of the ReLU activation function, the most critical phase synchronization patterns for classification are highlighted.

[0053] Frequency-specific PLV fusion: during the processing, the network automatically learns the weights of different frequency band PLVs, achieving frequency adaptive fusion of neural dynamic features and avoiding the subjectivity of manual selection of frequency bands.

[0054] The graph structure features and the neural dynamic features are fused through adaptive weights to capture complex correlations that cannot be reflected by a single feature: ; wherein is the fusion weight of network learning, and the fused feature The brain region function characteristics and neural dynamic connection mode are integrated, providing a discriminability far exceeding a single feature for subsequent classification.

[0055] The classification result output is the decision terminal of the model, taking the fused feature as input, and realizing accurate classification of ASD, ASD and ADHD comorbidity, and ADHD three types of samples through simple and efficient transformation. The core is to strengthen the feature discriminability and output an interpretable class probability.

[0056] Fused feature First, the integrated graph feature and the PLV neural dynamic feature are processed through two fully connected networks: ; Among them, the first layer compresses the feature dimension to 128 dimensions through the weight matrix , retaining the core discriminative information; The probability of 30% prevents overfitting; the second layer further refines the key association between features through .

[0057] Finally, the processed features are mapped to a three-class probability distribution through the softmax function: ; the output wherein are the probabilities of only having ASD, having ADHD and ASD together, and having ASD alone respectively, and the class corresponding to the maximum probability is taken as the final prediction result.

[0058] Embodiment 6, This embodiment is a further limitation of embodiment 1, and further describes the interpretable output module.

[0059] The interpretable results of the model are output by the following method: The amplitude low-frequency fluctuation (ALFF) index can effectively reflect the intensity of local neuron spontaneous activity, and it presents a high sensitivity to cognitive function (including attention, emotion regulation, etc.) in different slow wave frequency bands (such as SLOW-2 to SLOW-5). ALFF is obtained by calculating the amplitude mean of the low-frequency oscillation signal of each ROI in the SLOW2-SLOW5 frequency band, and the independent sample test is used to compare the differences in ALFF values of each frequency band between the comorbidity group and the single disease group, and the abnormal brain regions with statistical significance are screened out to reveal the unique neural markers of comorbidity individuals in local brain neural activity. For example Figure 2As shown, the ALFF differences between the comorbid ADHD-ASD and the only ADHD patients in the four frequency bands and the ALFF differences between the comorbid ADHD-ASD and the only ASD patients in the four frequency bands are shown.

[0060] Table 1 is a specific brain function abnormality marker of neurodevelopmental disorder comorbidity, which systematically reveals the brain function differences between the neurodevelopmental disorder comorbidity group (ASD-ADHD) and the single disease group (ADHD or ASD), by quantifying the abnormal activity intensity of different brain regions ) in key slow wave frequency bands, providing interpretable biomarkers for comorbidity mechanisms, where "+" represents enhancement, and "-" represents significant weakening.

[0061] Table 1:

[0062] Example 7, This embodiment is a further illustration of Examples 1-6. The system in Examples 1-6 is applied specifically.

[0063] Step one: rs-fMRI data acquisition In this embodiment, ASD patients, ADHD patients, ASD-ADHD comorbid patients and healthy control subjects who meet the inclusion criteria are recruited, and resting state functional image data is collected using a 3.0T magnetic resonance imaging system. The scanning parameters are set as follows: the repetition time (TR) is 2000ms, the echo time (TE) is 30ms, the flip angle is 90°, the number of scanning layers is 32, the spatial resolution is 3x3x3mm³, and the single scan duration is 8 to 10 minutes. The subjects maintain a closed-eye resting state during scanning without engaging in specific thinking activities. At the same time, the demographic information and clinical diagnosis results of all subjects are recorded for subsequent grouping and label annotation.

[0064] Step two: rs-fMRI data preprocessing After acquisition, the original functional image is subjected to standard preprocessing procedures using the FSL toolkit, which includes: Skull stripping: removing non-brain tissue to extract pure brain signals; Head motion correction: correcting artifacts caused by subject head motion using a rigid registration algorithm, and removing samples with head motion greater than 1.5mm or rotation angle greater than 2°; Spatial standardization: registering the functional image to the standard MNI152 template space to ensure spatial consistency; Spatial smoothing: using a Gaussian kernel with a full width at half maximum (FWHM) of 6mm for smoothing to improve signal-to-noise ratio; Detrending and filtering: Linear drift removal and bandpass filtering (frequency range 0.01 Hz to 0.25 Hz) were performed on the time series to retain low-frequency components related to neural activity. Noise regression: regress white matter signals, cerebrospinal fluid signals, and head motion parameters to further remove non-neural signal interference.

[0065] Step 3: Brain region division and feature extraction The brain was divided into 116 functional areas using the AAL3 brain region template ( ), extract each The BOLD signal time series is constructed at the following two levels: Node feature extraction: In SLOW2 to SLOW5 multi-band, for each The following features are calculated: Low-frequency amplitude characteristics: Calculation of ALFF (Amplitude of Low-Frequency Fluctuation) reflects the level of neural activity in the low-frequency band of the brain area; Statistical characteristics: including mean, standard deviation, skewness, and kurtosis, which describe the distribution characteristics of time series; LLE feature: Based on the Locally Linear Embedding method, K nearest neighbors are used to construct each The local reconstruction neighborhood is obtained and the variance of the reconstruction weight is calculated to characterize the nonlinear local structural characteristics.

[0066] Connection feature extraction: for all Combined, phase information is extracted based on Hilbert transform, and the phase locking value (PLV) is calculated to construct a 116×116 connection matrix reflecting the phase synchronization between brain regions; The PLVs within each sample were averaged to obtain stable connectivity patterns within the group.

[0067] Step 4: Classification through Multi-band Residual Graph Neural Network (MBF-ResGCN) The extracted node features and the connection matrix constructed by PLV are input into the MBF-ResGCN model as graph structure data. The model structure includes: Three-layer residual graph convolution module: extracts local functional connectivity, regional network features and whole-brain topological structure information layer by layer; Frequency band adaptive fusion module: Adaptively weighted integration of multi-band PLV features to reduce redundancy and enhance frequency band discrimination capabilities; Cross-modal feature fusion module: fuse node attributes and structural connection information to improve the model's perception of multi-modal brain network features; Classification layer: output three-class diagnosis probabilities (ASD, ADHD, comorbidity) through a fully connected network, and take the class corresponding to the maximum probability as the final diagnosis result.

[0068] Step five: model interpretability analysis To improve the model's interpretability in terms of neural mechanism analysis, the following analysis methods are implemented: Brain region node significance analysis: independent sample test is used to compare the ROI ALFF values between the comorbidity group and the ASD / ADHD single group, and the brain regions with p<0.05 are selected as potential neural markers, and their frequency band dependence is marked; Functional connection difference analysis: for the PLV values between groups, the differences between ADHD vs comorbidity, comorbidity vs ASD, and ADHD vs ASD are compared respectively. Select the significant increase or decrease in the connection relationship in the visualization diagram to reveal the abnormal connection patterns of different disease groups in the functional network structure.

Claims

1. A neurodevelopmental disorder comorbidity identification system based on residual graph neural network, characterized by: The system comprises: Data acquisition module: collects resting-state fMRI images; Data preprocessing module: Based on the AAL brain region template, the standardized time series of 116 brain regions in the resting-state fMRI images are extracted: ;use represents any one of these time series; Multi-band division and feature extraction module: The standardized time series of each brain region is filtered in four frequency bands to obtain the filtered time series of each frequency band, and the node feature matrix of the time series is extracted; within each frequency band, PLV calculation is performed to obtain the PLV matrix, and a binary topological matrix is ​​constructed based on the PLV matrix, and the PLV feature vector is calculated based on the PLV matrix; the node feature matrix, binary topological matrix and PLV feature vector are combined to obtain a brain functional connectivity map; Neurodevelopmental disorder comorbidity identification module: This module analyzes brain functional connectivity maps using a residual graph neural network model to obtain diagnostic information on neurodevelopmental disorder comorbidities. Interpretable output module: uses low-frequency fluctuations in amplitude as a reference indicator to output abnormal brain areas.

2. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 1 is characterized in that: In the data preprocessing module, data preprocessing is also performed before generating the standardized time series. Specifically, the following steps are performed: skull stripping, head motion correction, spatial smoothing using Gaussian filtering, linear trend removal using FSL's bandpass filtering tool, calculation of the mean functional image, and standard spatial alignment using the mean functional image.

3. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 2, characterized in that: The four frequency bands are: SLOW2 (0.198-0.25 Hz), SLOW3 (0.073-0.198 Hz), SLOW4 (0.027-0.073 Hz) and SLOW5 (0.01-0.027 Hz); The time series after filtering for each frequency band is in represents the index of the brain region, represents the index of the frequency band, and the node feature matrix includes statistical features and nonlinear features.

4. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 3 is characterized in that: The statistical characteristics include mean , standard deviation , skewness and kurtosis : ;in, Indicates the number of time points contained in the time series of each brain region in a specific frequency band, An index representing a time point; ； ； 。 5. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 4 is characterized in that: The method for obtaining the nonlinear characteristics is: S51. Local neighborhood construction: in time series The signal value at each time point is , using the K nearest neighbor algorithm, for each time point Find the five most similar neighboring points, and measure the similarity using the Euclidean distance: ,in, Representation and time point Similar time point index, Indicates a time point and time point The Euclidean distance of ; select the 5 smallest points as the nearest neighbors, denoted as , forming a time point The local neighborhood of S52, reconstruction weight solution: for each time point , assuming that its signal value is reconstructed by a linear combination of neighboring points, that is ,in Indicates a time point Time point To minimize the reconstruction error, solve the following optimization problem: ； The constraints are ,in Indicates a time point The set of reconstruction weights of all topological points in the solution process is to first centralize the local neighborhood signal, that is, to subtract , and get the matrix , and then solve the weights by the least squares method : ;in To avoid matrix singularity, the regularization term 1 is a vector of all 1s. is the identity matrix; S53, Feature Quantization: Through Get each Nonlinear characteristics in each frequency band, where represents the variance function.

6. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 5, characterized in that: The PLV calculation is as follows: S61. Hilbert transform: Convert to analytical signal ,in represents the Hilbert transform; S62, Phase Extraction: Pass Get the phase angle of the analytical signal , Indicates the operation of solving the phase angle; S63, PLV calculation: quantify the synchronization by the average cosine value of the phase difference, the formula is: ,in is the imaginary unit, The larger the value, the The stronger the phase synchronization, and Represents two The instantaneous phase angle at a specific time point is used to obtain the PLV matrix through PLV calculation.

7. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 6, characterized in that: The binary topology matrix constructed based on the PLV matrix is ​​as follows: When the PLV matrix When thinking about the two There is an edge, all edge indices form a binary topological matrix .

8. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 7, characterized in that: The calculation of PLV eigenvector based on PLV matrix is ​​as follows: extract the upper triangular elements of PLV matrix of each frequency band to form a one-dimensional eigenvector, and concatenate the eigenvectors of PLV matrices of four frequency bands to form a global connection feature vector. , That is the PLV feature vector.

9. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 8, characterized in that: The workflow in the residual graph neural network model is: S91, local node dynamic capture: Through the first layer of graph convolution operation, the features of each node are weighted and aggregated with the features of neighboring nodes to obtain the feature matrix : ,in represents the activation function, represents the normalized degree matrix, Represents a binary topological matrix that incorporates self-connection , represents the node feature matrix, represents the learnable weight; S92. Regional network feature aggregation: Through the second layer of graph convolution operation, the feature extraction range is expanded to larger brain network units to obtain the feature matrix : ,in represents the learnable weight; S93. Whole-brain topological pattern mining: Integrate cross-network connection features across the whole brain to form a global topological pattern representation and obtain a feature matrix : ,in represents the learnable weight; Using the feature matrix Through global mean pooling, all The node features are aggregated into a single vector ; S94, Focus on PLV phase synchronization pattern mining: Through multi-layer nonlinear transformation, PLV feature vectors are mapped to low-dimensional space while retaining the core pattern : ,in express activation function, is the PLV feature vector, is the learnable weight; S95, cross-modal feature fusion: Will and Obtain fused features through adaptive weight fusion : ,in is the fusion weight of network learning; S96. Classification result output: The fused features First, it is processed by two layers of fully connected networks: ， Among them, the first layer is through the weight matrix Compress the feature dimension to 128 dimensions, retaining the core discriminant information, To prevent overfitting function; the second layer passes the weight matrix Further refine the key correlations between features; pass softmax The function maps the processed features into three types of probability distributions: , is the weight matrix, output ,in The probabilities of having only ASD, having both ADHD and ASD, and having ASD alone are respectively taken, and the category corresponding to the maximum probability is taken as the final prediction result.

10. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 9, characterized in that: The amplitude of low-frequency fluctuations is calculated by each The amplitudes of low-frequency oscillation signals in the four frequency bands were obtained. Independent sample tests were used to compare the differences in the low-frequency fluctuation values ​​of the amplitudes in each frequency band between the comorbidity group and the single disease group, and statistically significant abnormal brain areas were screened and output.

Images