Brain region correlation analysis system and method for autistic children

Abstract

The invention discloses a brain region correlation analysis system and method for autistic children. The system comprises a brain region division module used for dividing the cerebral cortex into a plurality of brain regions; the training data set construction module is used for constructing a training data set, and each training sample comprises a sensor space function connection matrix and a source space function connection matrix corresponding to the sensor space function connection matrix; the signal preprocessing module is used for acquiring a real electroencephalogram signal and calculating a real value of a corresponding sensor space function connection matrix; the deep learning mapping module is used for learning a mapping relation from the sensor space function connection matrix to the source space function connection matrix and outputting a predicted value of the source space function connection matrix; and the brain region correlation analysis module is used for calculating the correlation between the brain regions. The method can be used for accurately analyzing the correlation between the brain areas of the autism children.

Description

Technical Field

[0001] This invention belongs to the field of brain electromagnetic signal processing, and more specifically, relates to a brain region correlation analysis system and method for children with autism. Background Technology

[0002] The brain, as the most important part of the central nervous system, is the core of processes such as perception, memory, learning, and behavior. It also serves as a window into the body's pathological processes. Brain imaging using extracranial electromagnetic sampling data is a crucial method for exploring the mechanisms of brain activity. It enables the localization and assessment of cognitive processes and functional impairments in the brain, and is of great value in exploring the localization of functional areas within the brain and sensing the dynamics of brain activity. It has wide applications in brain research and clinical practice. Correlation analysis of brain regions in children with autism can be applied to the auxiliary diagnosis of autism, individualized intervention assessment, and early risk screening. Summary of the Invention

[0003] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a brain region correlation analysis system and method for children with autism, which can accurately analyze the correlation between brain regions of children with autism.

[0004] To achieve the above objectives, according to one aspect of the present invention, a brain region correlation analysis system for children with autism is provided, comprising: The brain region segmentation module is used to divide the cerebral cortex into multiple brain regions; The training dataset construction module is used to obtain the lead field matrix, generate sensor noise simulation signals, generate brain source activity time series simulation signals for each brain region, calculate the source spatial functional connectivity sub-matrix for the source activity time series simulation signals for each brain region, concatenate the source spatial functional connectivity sub-matrixes of all brain regions into a source spatial functional connectivity matrix, generate sensor simulation signals based on the sensor noise simulation signals, the lead field matrix, and the source activity time series simulation signals of all brain regions, calculate the sensor spatial functional connectivity matrix corresponding to the sensor simulation signals, and each sensor spatial functional connectivity matrix and its corresponding source spatial functional connectivity matrix constitute a training sample pair, constructing a training dataset containing multiple training sample pairs; The signal preprocessing module is used to acquire real EEG sensor sampling signals and preprocess the real EEG sensor sampling signals, and calculate the corresponding sensor spatial functional connectivity matrix true value for the preprocessed real EEG sensor sampling signals. The deep learning mapping module is trained end-to-end using the training dataset to learn the mapping relationship from the sensor space functional connectivity matrix to the source space functional connectivity matrix. The deep learning mapping module is also used to receive the true value of the sensor space functional connectivity matrix and output the predicted value of the source space functional connectivity matrix. The brain region correlation analysis module is used to obtain the source spatial functional connectivity submatrix of each brain region from the predicted value of the source spatial functional connectivity matrix, and to calculate the correlation between brain regions based on the source spatial functional connectivity submatrix of each brain region.

[0005] Preferably, the step of splicing the source spatial functional connectivity submatrices of all brain regions into a source spatial functional connectivity submatrices involves splicing the source spatial functional connectivity submatrices of all brain regions along the diagonal.

[0006] Preferably, generating time-series simulated brain activity signals for each brain region includes the following steps: Let N be the number of brain regions. Each time a time-series simulated signal of brain-derived activity is generated, select... One active brain region, 1≤ ≤N, is Each active brain region generates a time-series simulated signal of brain-derived activity.

[0007] Preferably, generating time-series simulated brain activity signals for each brain region includes the following steps: The parameters of the source activity time series simulation signal are predefined. Each time the brain source activity time series simulation signal is generated, the parameter conditions of the source activity time series simulation signal are set so that the brain source activity time series simulation signal generated for each brain region meets the statistical parameter setting conditions.

[0008] Preferably, the preprocessing includes the following steps: The real EEG signal before preprocessing is denoted as , It is The matrix, where The number of time sampling points, each time point , , Indicates a point in time The sampled values, For the number of sensor channels, Baseline correction was performed, and the baseline-corrected EEG signal was recorded as follows: ; right Denoising was performed, and the denoised EEG signal was recorded as follows: ; Constructing the projection matrix Through spatial projection The signal is obtained through processing. , Indicates matrix transpose; right Filtering is performed to obtain the signal The calculation formula for the preprocessed real EEG signal is as follows: ; in, Represents the Fast Fourier Transform. H(f) represents the inverse fast Fourier transform, and H(f) is the bandpass filter function.

[0009] Preferably, the deep learning mapping module includes: The Transformer branch is used to convert the sensor spatial functional connection matrix into a grayscale image of size H×W, where H and W are the height and width of the image, respectively. The first feature map of H / 4 × W / 4, the second feature map of H / 8 × W / 8, the third feature map of H / 16 × W / 16, and the fourth feature map of H / 32 × W / 32 are extracted from the H×W grayscale image. The EfficientNet branch is used to extract the fifth feature map of H / 2 × W / 2, the sixth feature map of H / 4 × W / 4, the seventh feature map of H / 8 × W / 8, the eighth feature map of H / 16 × W / 16, and the ninth feature map of H / 32 × W / 32 from the H×W grayscale image, respectively. The decoder is used to take the fifth feature map as the first fusion feature, fuse the first feature map and the sixth feature map to obtain the second fusion feature, fuse the second feature map and the seventh feature map to obtain the third fusion feature, fuse the third feature map and the eighth feature map to obtain the fourth fusion feature, fuse the fourth feature map and the ninth feature map to obtain the fifth fusion feature, and decode based on the first fusion feature, the second fusion feature, the third fusion feature, the fourth fusion feature, and the fifth fusion feature to obtain the predicted value of the source space function connectivity matrix.

[0010] Preferably, the brain region correlation analysis module includes: The first analysis module is used to obtain the functional connectivity matrix of each brain region of the test user from the predicted value of the source space functional connectivity matrix. The calculation formula for calculating the brain region correlation parameter based on the functional connectivity matrix of each brain region is predefined. The brain region correlation parameter of the test user is calculated, and the difference between the brain region correlation parameter of the test user and the brain region correlation parameter of the normal control group of children is calculated. If the difference is greater than the preset value, the test user is determined to be a child with autism. The second analysis module is used to calculate the correlation between the time series of source activities between two brain regions. The calculation formula is as follows: ; in, Indicates the first i The mean of the time series of source activity in each brain region. Indicates the first j Mean of time series of source activity in individual brain regions; The third analysis module is used to calculate the coherence of the source activity time series between two brain regions. The calculation formula is as follows: ; in, It is frequency Next i The brain regions and the first j Coherence of source activity time series in individual brain regions It is the first i The brain regions and the first j Joint power spectral density of source activity time series of individual brain regions It is the first i The self-power spectral density of the source activity time series of each brain region It is the first j The self-power spectral density of the source activity time series of each brain region.

[0011] Preferably, the source space functional connection sub-matrix, the source space functional connection matrix, and the sensor space functional connection matrix are all covariance matrices.

[0012] According to another aspect of the present invention, a method for analyzing brain region correlations in children with autism is provided, comprising the steps of: The cerebral cortex is divided into multiple brain regions; The process involves obtaining the lead field matrix, generating sensor noise simulation signals, generating brain source activity time-series simulation signals for each brain region, calculating the source spatial functional connectivity sub-matrix for each brain region's source activity time-series simulation signals, concatenating the source spatial functional connectivity sub-matrixes of all brain regions into a single source spatial functional connectivity sub-matrix, generating sensor simulation signals based on the sensor noise simulation signals, the lead field matrix, and the source activity time-series simulation signals of all brain regions, calculating the sensor spatial functional connectivity matrix corresponding to the sensor simulation signals, and constructing a training sample pair consisting of each sensor spatial functional connectivity matrix and its corresponding source spatial functional connectivity matrix. Acquire real EEG signals and preprocess the real EEG sensor sampling signals, and calculate the real value of the corresponding sensor spatial functional connectivity matrix for the preprocessed real EEG sensor sampling signals; The deep learning mapping module is trained end-to-end using the training dataset to learn the mapping relationship from the sensor space functional connectivity matrix to the source space functional connectivity matrix. The deep learning mapping module is also used to receive the true value of the sensor space functional connectivity matrix and output the predicted value of the source space functional connectivity matrix. The source spatial functional connectivity submatrix of each brain region is obtained from the predicted value of the source spatial functional connectivity matrix, and the correlation between brain regions is calculated based on the source spatial functional connectivity submatrix of each brain region.

[0013] Overall, the technical solutions conceived in this invention have beneficial effects compared with existing technologies: they can capture and quantify abnormal brain activity in children with autism, supporting auxiliary diagnosis, individualized intervention assessment, and early risk screening of autism based on EEG / MEG data. Compared with traditional methods that rely solely on behavioral assessment, this system can provide objective and reproducible neuroimaging features, providing crucial support for clinical research and intervention strategy design in autism. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the module composition of the brain region correlation analysis system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of brain region-based activity according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the working principle of the brain region correlation analysis system according to an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0016] In the description of the embodiments of this application, the term "multiple" means at least two, such as two, three, etc., unless otherwise expressly and specifically defined. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0017] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0018] This invention provides a brain region correlation analysis system and method for children with autism, which will be described below.

[0019] like Figure 1 As shown, an embodiment of the present invention provides a brain region correlation analysis system for children with autism, comprising: The brain region segmentation module is used to divide the cerebral cortex into multiple brain regions; The training dataset construction module is used to obtain the lead field matrix, generate sensor noise simulation signals, generate brain source activity time series simulation signals for each brain region, calculate the source spatial functional connectivity sub-matrix for the source activity time series simulation signals for each brain region, concatenate the source spatial functional connectivity sub-matrixes of all brain regions into a source spatial functional connectivity matrix, generate sensor simulation signals based on the sensor noise simulation signals, the lead field matrix, and the source activity time series simulation signals of all brain regions, calculate the sensor spatial functional connectivity matrix corresponding to the sensor simulation signals, and each sensor spatial functional connectivity matrix and its corresponding source spatial functional connectivity matrix constitute a training sample pair, constructing a training dataset containing multiple training sample pairs; The signal preprocessing module is used to acquire real EEG signals and preprocess them, and to calculate the real value of the corresponding sensor spatial functional connectivity matrix for the preprocessed real EEG signals. The deep learning mapping module is trained end-to-end using the training dataset to learn the mapping relationship from the sensor space functional connectivity matrix to the source space functional connectivity matrix. The deep learning mapping module is also used to receive the true value of the sensor space functional connectivity matrix and output the predicted value of the source space functional connectivity matrix. The brain region correlation analysis module is used to obtain the source spatial functional connectivity submatrix of each brain region from the predicted value of the source spatial functional connectivity matrix, and to calculate the correlation between brain regions based on the source spatial functional connectivity submatrix of each brain region.

[0020] Each module will be explained below.

[0021] (1) Brain region division module Brain region segmentation uses the guidance field of the voxel closest to the center point of each brain region as the guidance field of that region, and the intensity of the source activity of that voxel as the overall activity of that brain region. The entire cortex is divided into... Each brain region contains one equivalent current dipole degree of freedom.

[0022] These brain regions can be divided based on anatomical structure, such as the frontal lobe, parietal lobe, occipital lobe, and temporal lobe, or based on function, such as the anterior cingulate cortex and precuneus in the default mode network.

[0023] In one embodiment, N=68.

[0024] (2) Training dataset construction module The training dataset construction module aims to provide a large number of paired spatial functional connectivity matrix samples for the deep learning mapping module. Each sample pair consists of the sensor spatial functional connectivity matrix at the observation level and the corresponding source spatial functional connectivity matrix. In actual electroencephalograms, the precise activity patterns and covariance structure of brain source signals cannot be directly obtained. Therefore, this method introduces a large-scale simulation strategy based on biophysical principles to synthesize brain source signal data with controllable statistical properties.

[0025] Each time a time-series simulated signal of brain-derived activity is generated, from Randomly selected from brain regions Active areas ( Simulate different task activation modes. (Targeting...) Multiple temporal signals are constructed in each brain region. ( In each sampling, several active brain regions are randomly selected, and a corresponding source signal covariance submatrix is ​​generated for each active brain region. By concatenating all submatrices along the diagonal, the complete source space correlation matrix is ​​obtained. ( ,…, Regardless of how many brain regions are activated in the simulation, it is necessary to maintain the consistency of the matrix structure for subsequent unified analysis. For inactive brain regions, their corresponding covariance submatrices can be set to zero, and then all submatrices can be concatenated diagonally to form the final source space correlation matrix.

[0026] Furthermore, statistical parameters of the source activity time-series simulated signals can be predefined during the generation process. When generating the simulated signals, parameter conditions can be set to ensure that the generated brain source activity time-series simulated signals for each brain region meet the statistical parameter settings. These parameters are adjustable control variables used to simulate characteristics that more closely resemble real brain signals, such as the frequency distribution of the source activity time-series simulated signals, the stability of the source activity time-series simulated signals, and whether the source activity time-series simulated signals contain noise. In the construction of the source spatial correlation matrix, the source spatial correlation matrix is ​​calculated based on the generated source activity time-series simulated signals. And by adjusting the parameters, we ensure that the topology and connection strength distribution of the covered functional networks meet the design requirements.

[0027] Sensor time series The simulation, also known as the sensor time series accuracy, is based on a known head model (such as a three-layer spherical model or a finite element model) and the solved lead field matrix. Where M is the number of sensors and D is the dimension of the source space. According to the linear observation model... ; in, It is a sensor noise simulation signal from all brain regions. It is a sensor analog signal, noise Based on the preset signal-to-noise ratio distribution, the covariance matrix of the sensor space can be calculated. For each time series sample set, calculate its sample covariance. This yields the covariance of the entire brain, eliminating the need for further splicing. This refers to the sensor time series accuracy, which is used as an input feature for deep networks.

[0028] Organization and storage of training sample pairs: Each pair To form a training sample, the sensor covariance matrix As network input, the source space correlation matrix As the desired network output, the above process is repeated, traversing various brain region activation patterns, signal-to-noise ratios, source correlations, and different head model configurations, generating sample pairs ranging from hundreds of thousands to millions. This provides rich and diverse training data for the deep learning mapping module, laying a solid foundation for the accurate deconvolution and network structure reconstruction of the deep learning mapping module in real EEG signal analysis.

[0029] (3) Signal preprocessing module This module is mainly used to acquire real EEG sensor sampling signals. Preprocessing of EEG sensor sampling signals in children with autism is a core step to ensure data quality and directly affects the reliability of subsequent analysis. Its core lies in gradually removing noise and enhancing effective signals.

[0030] The entire process begins with the input of raw data. First, it is necessary to clarify the spatiotemporal structure of the data, assuming the raw data... It is The matrix, where The number of time sampling points, each time point Corresponding to a sampled value, and The number of sensor channels, and the observation value for each channel. This data records electroencephalogram (EEG) or magnetoencephalogram (MEG) activity at specific locations. Such multidimensional data naturally contains baseline drift, physiological noise, and environmental interference; therefore, the first step in preprocessing is preliminary screening to ensure data integrity and validity. For example, it is necessary to verify whether the sampling rate conforms to the Nyquist criterion. To avoid aliasing of high-frequency signals due to insufficient sampling, it is also necessary to check whether the channel mapping is consistent with the standard lead system to prevent spatial resolution errors. If extreme values ​​appear in the data of a certain channel, outliers must be removed using statistical methods to prevent subsequent processing from failing due to data defects.

[0031] After the initial screening, the baseline correction phase begins, aiming to eliminate low-frequency drift in the signal. This drift can be caused by slow fluctuations in breathing, sweating, or the device itself, and its mathematical model can represent a linear or non-linear trend. The simplest method for baseline correction is to subtract the time mean. ; However, for non-stationary signals, high-order polynomial fitting can be used for baseline correction. For example, a baseline can be corrected by fitting a signal using the least squares method. polynomial of order Then subtract from the original signal ; in, Indicates the first The order coefficients are the parameters to be estimated in polynomial fitting. Represent the independent variable of Powers are used to construct polynomials.

[0032] This step provides a stable signal background for subsequent processing, preventing low-frequency interference from masking the effective high-frequency components. The next step is interference removal, the core of which lies in separating independent noise sources from the signal. This mainly relies on independent component analysis. . Assuming the observed signal It is a signal from multiple independent sources. A linear mixture, i.e.: ; in It is a mixed matrix. For independent component matrices, This is the observed signal matrix. The algorithm consists of three steps: first, the data is centered and whitened to eliminate correlation; then, the non-Gaussianity of the components is maximized through iterative optimization (such as the FastICA algorithm) to solve for the separation matrix. Finally, based on the spectral characteristics or topographic map distribution of the components, noise components are identified and removed, and the signal is reconstructed. ; in, denoted as the j-th column in the mixing matrix, it represents the contribution coefficient of the j-th source signal to the observed signal. Let be the j-th independent component signal, representing a potential source signal. In the independent component analysis model... In this process, signal reconstruction is achieved by removing certain artifact components (such as the m-th component), resulting in the denoised signal represented as follows: .

[0033] The advantage of ICA lies in its ability to globally separate noise, but it may retain local interference, thus requiring further refinement through spatial projection. Projection matrix The construction can be based on brain structural templates or signal statistical features, such as using principal component analysis (PCA) to select the main signal direction, or constructing a projection basis based on the region of interest (ROI) defined by MRI. The denoising formula is: ; Among them, here This indicates the matrix transpose.

[0034] This step, through spatial constraints, further suppresses residual noise or enhances signals in specific brain regions, creating a noise reduction effect that complements ICA.

[0035] After spatial noise reduction, the next stage is filtering, with the goal of preserving frequency bands relevant to brain activity (such as alpha waves 8–13 Hz and beta waves 13–30 Hz) while filtering out irrelevant noise. Common methods include Fourier transform filtering and wavelet transform filtering. Fourier filtering is achieved through frequency domain truncation: the signal is subjected to a Fast Fourier Transform (FFT), and after preserving the target frequency band, an inverse transform is performed to reconstruct the time-domain signal. ; Where H(f) is the bandpass filter function: ; Wavelet filtering, on the other hand, achieves denoising of non-stationary signals through multi-resolution analysis. It uses the Daubechies-4 wavelet basis to decompose the signal into detail coefficients at different scales. And approximation coefficients, and then perform soft thresholding on the detail coefficients: ; The final reconstructed signal is While preserving the effective frequency band, it suppresses interference from high-frequency noise.

[0036] The final step in the preprocessing process is quality control, which requires verifying the output data. The signal-to-noise ratio and spatiotemporal consistency were assessed. The energy proportion of the target frequency band was calculated using power spectral density to ensure it was significantly higher than the noise frequency band. Correlation between adjacent channels was examined to reflect the spatial continuity of brain activity. Simultaneously, temporal waveforms, spectrograms, and topographic maps before and after preprocessing were plotted for comparison, visually verifying the processing effect. Parameter optimization was also a crucial step. The number of components needs to be determined based on the eigenvalue attenuation, the filtering frequency band needs to be combined with the research objectives, and real-time processing also needs to consider causal constraints. Finite impulse response filters are used to avoid future information leakage.

[0037] The entire preprocessing workflow forms a clearly defined noise reduction chain: ; Baseline correction provides a stable background for subsequent processing. Global noise is separated, spatial projection refines local interference, and filtering further purifies the signal in the frequency domain. This multi-stage processing not only progressively improves the signal-to-noise ratio but also lays a reliable data foundation for subsequent brain network analysis and deep learning models.

[0038] (4) Deep learning mapping module The working principle of deep learning modules is as follows: Figure 3 As shown.

[0039] Furthermore, the deep learning module employs the TransEfficientNet model.

[0040] Traditional machine learning models face the challenge of solving the source space correlation matrix. Traditional machine learning models suffer from low accuracy and difficulty in solving problems. To overcome these limitations, deep learning models are chosen to address this issue. Deep learning models are capable of mining the accuracy matrix of sensor time series data. Source space correlation matrix The deep connections between them. However, deep network training becomes difficult due to the vanishing gradient problem. When gradients are backpropagated to earlier layers, repeated multiplication can cause the gradients to become very small, leading to network performance saturation or even a rapid decline as the number of layers increases.

[0041] To overcome these problems, this invention employs a novel deep learning model: a segmentation and fusion network architecture combining Transformer and EfficientNet in parallel: TransEfficientNet. The overall framework of TransEfficientNet follows an encoder-decoder architecture. In the encoder architecture of the TransEfficientNet network, a dual-branch parallel design is used, innovatively fusing two heterogeneous network structures: Transformer and EfficientNet.

[0042] First, let's assess the sensor's time series accuracy. The data is normalized and mapped to a 271×271 resolution single-channel grayscale image, with each data point corresponding to the grayscale value of a pixel in the image. This two-dimensional grayscale matrix is ​​then expanded into a three-channel input tensor of dimension H×W×1, where H and W are the height and width of the image (both 271), and the single-channel data constitutes the third dimension of the tensor. The entire network follows a classic encoder-decoder architecture, where the encoder extracts features from the input image. This model innovatively uses a dual-branch parallel encoder to capture local and global features separately. This embodiment of the invention, through dual-channel synchronous input, constructs a multimodal feature extraction paradigm: firstly, it leverages the hierarchical feature extraction advantages of deep convolutional neural networks to obtain deep local semantic information; secondly, it utilizes the Transformer architecture to achieve global semantic modeling. This parallel computing mode maximizes the preservation of image data integrity and multi-dimensional feature expression. The Transformer branch is based on a block attention mechanism. The Swin Transformer performs better in image segmentation, and this paper introduces it as the training model for the Transformer branch. First, the image is input into the PatchPartition module for block segmentation, where each 4x4 adjacent pixel group forms a patch. Then, the image is flattened along the channel direction. In this embodiment, a single-channel image is input, so each patch has 4x4=16 pixels. Since each pixel has a value, the flattened shape is 16x1=16. Therefore, after Patch Partition, the image shape changes from [H, W, 1] to [H / 4, W / 4, 16]. At this point, the feature dimension of each patch is 4. 4 1. Then, a linear transformation is performed on the channel data of each pixel through the Linear Embedding layer, mapping it from 16 to C, that is, the image shape changes from [H / 4, W / 4, 16] to [H / 4, W / 4, C].

[0043] The specific operation first involves sending the input image to the Patch Partition module, where it is then segmented into... 4 A 4-size tile, where each tile has a feature dimension of 4. 4 1. After flattening on the channel, a sequence of tiles is obtained. Then, it needs to go through four stages of operation. After entering Stage 1, the sequence undergoes a linear transformation through the Linear Embedding layer. The dimension is mapped to C dimensions, where C is set to 16. After passing through two stacked Swing Transformer Blocks, the sequence is obtained. In Stage 2, after a patch merging layer, adjacent 2x2 patches are merged together, and pixels at the same location within each patch are stitched together at depth. This is then followed by a layer normalization (LN) module to obtain... Each patch has a channel dimension of 4C. After passing through a linear layer, the channel dimension finally becomes 2C. Then, after undergoing two more Swin Transformer Block feature transformations, the final result is... Stage 3 and Stage 4 operate similarly to Stage 2, while Stage 3 yields... Stage 4 then yields Therefore, the final features extracted by the Swin Transformer encoder are: In fact, Patch Partition and Linear Embedding are implemented directly through a single convolutional layer. Then, four stages are used to construct feature maps of different sizes. Except for Stage 1, which first uses a Linear Embedding layer, the remaining three stages first use a Patch Merging layer for downsampling. Then, the SwinTransformer Block is repeatedly stacked. Finally, the decoder receives: a first feature map of H / 4 × W / 4, a second feature map of H / 8 × W / 8, a third feature map of H / 16 × W / 16, and a fourth feature map of H / 32 × W / 32.

[0044] The EfficientNet branch employs a composite scaling strategy, jointly optimizing network depth, width, and input resolution to construct an efficient feature extraction architecture. Specifically, the input image is first fed into an initial convolutional layer (Stem) for processing. This convolutional layer typically has a stride of 2, reducing the image size from H×W to H / 2 × W / 2 and establishing an initial channel dimension. Afterward, the main body of the network consists of multiple stages, each composed of a series of core MBConv modules stacked together. The MBConv module is the basic building block of EfficientNet, and its internal operations include: first, channel expansion using a 1x1 convolution; then, efficient 3x3 or 5x5 depthwise separable convolution to extract features; next, attention weighting of the channel features using an (SE) module; and finally, compression of the channel number back to the original dimension using a 1x1 convolution, which is then added to the input via a residual connection. Different stages of the network achieve size reduction and channel expansion by controlling the stride of the first MBConv module. For example, after Stage 2, the feature map size is H / 4 × W / 4, and the number of channels is C. Entering Stage 3, the stride of the first MBConv module is set to 2, downsampling the feature map to H / 8 × W / 8, while the number of output channels increases to approximately 1.5C or 2C. Subsequent Stages 4, 5, etc., follow this pattern, gradually halving the feature map size while increasing the channel width. For example, Stage 4 outputs a feature map of H / 16 × W / 16, and Stage 5 outputs a feature map of H / 32 × W / 32, with the number of channels increasing accordingly. The core innovation of EfficientNet lies in its composite scaling strategy. It doesn't adjust the network depth (the number of repetitions of the MBConv module), width (number of channels), or input resolution in isolation, but rather uses a unified composite coefficient Φ and three scaling factors α, β, and γ to collaboratively balance these three dimensions. The final inputs to the decoder are: the fifth feature map of H / 2 × W / 2, the sixth feature map of H / 4 × W / 4, the seventh feature map of H / 8 × W / 8, the eighth feature map of H / 16 × W / 16, and the ninth feature map of H / 32 × W / 32. Its core module integrates depthwise separable convolution and channel attention mechanisms, enhancing feature representation capabilities while reducing computational complexity. This design effectively alleviates the "depth-resolution" trade-off in traditional convolutional neural networks, achieving progressive fusion of multi-scale features. It can accurately capture subtle morphological differences and boundary features in EEG signals, providing high-resolution, high-semantic-density local feature representations. Traditional convolutional neural networks suffer from spatial resolution decay in deep feature extraction, leading to loss of detail information and blurred segmentation edges; while the pure Transformer architecture, lacking a local feature enhancement mechanism, is prone to semantic confusion when processing complex textures and microstructures. This fusion architecture, through the complementary advantages of heterogeneous networks, constructs a "global-local" collaborative feature extraction system, achieving deep fusion of multi-scale contextual information, significantly improving the accuracy and reliability of image segmentation, and providing a new technical paradigm for solving the problems of the embodiments of this invention.

[0045] In the decoder design of the TransEfficientNet network, a Hybrid Feature Fusion Decoder (HFFD) is employed. This decoder is based on the classic U-Net architecture and incorporates the ideas of Feature Pyramid Network (FPN) and the Atrous Spatial Pyramid Pooling (ASPP) module to enhance context awareness. It receives feature maps fused from EfficientNet and Swin Transformer branches at five different resolutions as input. In this embodiment, these five fused feature maps are first defined as the source of skip connections for the decoder. The first fused feature, F_fused_1 (H / 2 × W / 2), comes from the fifth feature map of the EfficientNet Stage 2 layer (Swin Transformer does not have this resolution). The second fused feature, F_fused_2 (H / 4 × W / 4), is a feature fusion of the sixth feature map from EfficientNet Stage 2 and the first feature map from Swin Transformer Stage 1. The third fusion feature, F_fused_3 (H / 8 × W / 8), is a feature fusion of the seventh feature map from EfficientNet Stage 3 and the second feature map from Swin Transformer Stage 2. The fourth fusion feature, F_fused_4 (H / 16 × W / 16), is a feature fusion of the eighth feature map from EfficientNet Stage 4 and the third feature map from Swin Transformer Stage 3. The fifth fusion feature, F_fused_5 (H / 32 × W / 32), is a feature fusion of the ninth feature map from EfficientNet Stage 5 and the fourth feature map from Swin Transformer Stage 4. Before feeding the two branch features from each scale into the decoder, fusion is required. First, channel alignment is performed using a 1x1 convolutional layer to adjust the feature maps of EfficientNet and Swin Transformer at the same resolution, ensuring they have the same number of channels. Then, feature concatenation is performed, concatenating the two channel-aligned feature maps along the channel dimension. Finally, information extraction is performed. The concatenated feature maps are then processed through a 3x3 convolutional layer for information extraction and dimensionality reduction to obtain the final fused feature map F_fused_i.

[0046] The decoder's specific structure: The decoder starts with the deepest feature map, F_fused_5. To capture multi-scale contextual information to the maximum extent before decoding begins, this embodiment does not directly upsample it, but instead feeds it into an ASPP (Atrous Spatial Pyramid Pooling) module. This module uses multiple atrous convolutions with different dilation rates (e.g., rates = [6, 12, 18]) in parallel, along with a global average pooling branch. This allows the model to simultaneously perceive receptive fields of different sizes from the single feature map F_fused_5 without reducing resolution, effectively capturing details of small objects and contours of large objects. The results of all parallel branches are concatenated along the channel dimension and fused through a 1x1 convolution to generate a feature map with extremely rich semantic information, which this embodiment calls D_5 (H / 32 x W / 32).

[0047] The decoder consists of four main decoding stages, each of which performs the "upsampling -> fusion -> refinement" operation.

[0048] Decoding Stage 1: First, upsampling is performed. The feature map D_5 output from the bottleneck layer is passed through a transposed convolution to double its spatial resolution, becoming H / 16 × W / 16. Simultaneously, the number of channels is halved to control computational load. Next, feature fusion is performed. The upsampled result is concatenated with F_fused_4 (H / 16 × W / 16) from the encoder fusion layer along the channel dimension. Finally, the concatenated feature map is fed into the decoding block for information integration and refinement. This decoding block typically consists of two cascaded 3x3 convolutional layers, each followed by a batch normalization (BatchNorm) and ReLU activation function. The output feature map in this embodiment is referred to as D_4 (H / 16 x W / 16).

[0049] Decoding Stage 2: First, upsampling is performed, doubling the output D_4 from the previous stage to obtain a feature map of H / 8 × W / 8. Next, feature fusion is performed, concatenating the feature map with the encoder's fusion layer's F_fused_3 (H / 8 × W / 8). Finally, the concatenated result is fed into another decoding block with the same structure, outputting feature map D_3 (H / 8 x W / 8).

[0050] Decoding stage 3 is similar to decoding stage 4. Finally, the output D_1(H / 2 x W / 2) of the last stage of the decoder is upsampled by a factor of two (usually using bilinear interpolation to preserve details). The upsampled feature map is then fed into a final 1x1 convolutional layer. The output of this convolutional layer is the result required in this embodiment of the invention, namely the predicted value of the source space functional connectivity matrix.

[0051] (5) Brain region correlation analysis module Multiple EEG and magnetoencephalography studies have shown that children with autism exhibit low-frequency (e.g., Wide-area connectivity attenuation exists within the range of (wavelength) and high-frequency (e.g.) frequencies. The study identified a pattern of enhanced local connectivity, which was observed in default mode networks, social brain networks, and limbic systems. These anomalous connectivity patterns provide a quantifiable pathway for early screening, biomarker extraction, and intervention assessment.

[0052] The brain region correlation analysis module aims to quantify the interaction strength between brain regions, covering linear, nonlinear, and time-frequency characteristics, quantitatively analyzing the functional connectivity characteristics between different brain regions, and revealing their interaction relationships.

[0053] The brain region correlation analysis module includes a first analysis module, which is used to obtain the functional connectivity matrix of each brain region of the test user from the predicted value of the source space functional connectivity matrix. The first analysis module predefines the calculation formula for calculating the brain region correlation parameters based on the functional connectivity matrix of each brain region, calculates the brain region correlation parameters of the test user, and calculates the difference between the brain region correlation parameters of the test user and the brain region correlation parameters of the normal control group of children. If the difference is greater than the preset value, the test user is determined to be a child with autism.

[0054] The brain region correlation analysis module may also include a second analysis module, which calculates correlations using correlation analysis. Pearson correlation is a statistical method that measures the strength and direction of the linear relationship between two variables, with values ​​ranging from -1 to 1. For time series data of two brain regions, their functional connectivity is typically calculated using the Pearson correlation coefficient. ; in, Indicates the first i The mean of the time series of source activity in each brain region. Indicates the first j Mean of the time series of source activity in each brain region.

[0055] The Pearson correlation coefficient can be used to identify children with autism. Specifically, the Pearson correlation coefficient of the control group is calculated, and the difference between the test user's Pearson correlation coefficient and that of the control group is calculated. If the difference is greater than a preset value, the test user is determined to have autism.

[0056] Taking the default mode network of children with autism as an example, the time series X of the anterior cingulate cortex (ROI A) and precuneus (ROI B) is calculated. a The Pearson correlation coefficients of Xb(t) and Xb(t) were calculated using a formula to determine the strength of the correlation between the two brain regions. Then, the Pearson correlation coefficients of the autistic children's group and the normal control group were statistically compared. If the Pearson correlation coefficient of the autistic children's group was found to be significantly lower than that of the normal control group, this may indicate a weakened functional connection between these two brain regions in the default mode network in autistic children.

[0057] The brain region correlation analysis module may also include a third analysis module, which estimates the correlation by calculating the coherence between different brain regions. Coherence describes the synchronicity between two signals in the frequency domain, as shown in the following formula: ; in, It is frequency Next i The brain regions and the first j Coherence of source activity time series in individual brain regions It is the first i The brain regions and the first j Joint power spectral density of source activity time series of individual brain regions It is the first i The self-power spectral density of the source activity time series of each brain region It is the first j Source activity in each brain region.

[0058] The coherence values ​​of the signals from two brain regions at various frequencies are calculated using the coherence formula. The coherence values ​​of the test user and normal children in the same brain regions and the same frequency band are compared. If the interpolated coherence values ​​of the same brain regions and the same frequency band are greater than the preset value, it indicates that the test user is a child with autism.

[0059] Specifically, coherence was calculated in different frequency bands, such as the alpha band (8-13 Hz) and the theta band (4-8 Hz). Coherence in different frequency bands reflects different neural activity processes. In the alpha band, the coherence between the frontal and parietal lobes of autistic children may differ from that of typically developing children. A decreased coherence value indicates a reduced synchronicity of neural activity in these two brain regions within the alpha band. This reduced synchronicity may be related to sensory processing abnormalities in autistic children, as the parietal lobe is involved in the processing and integration of sensory information, while the frontal lobe is involved in higher-level cognitive control of sensory information. Abnormal connections between the two may lead to phenomena such as hypersensitivity or hyposensitivity. The study also investigated patterns of coherence combinations in different frequency bands. Autistic children exhibit increased coherence in low-frequency bands (such as the delta band, 0.5-4 Hz) and decreased coherence in high-frequency bands (such as the beta band, 13-30 Hz). This specific coherence pattern may provide a basis for the diagnosis and classification of autism.

[0060] An embodiment of the present invention provides a brain region correlation analysis method for children with autism, comprising the following steps: The cerebral cortex is divided into multiple brain regions; The process involves obtaining the lead field matrix, generating sensor noise simulation signals, generating brain source activity time-series simulation signals for each brain region, calculating the source spatial functional connectivity sub-matrix for each brain region's source activity time-series simulation signals, concatenating the source spatial functional connectivity sub-matrixes of all brain regions into a single source spatial functional connectivity sub-matrix, generating sensor simulation signals based on the sensor noise simulation signals, the lead field matrix, and the source activity time-series simulation signals of all brain regions, calculating the sensor spatial functional connectivity matrix corresponding to the sensor simulation signals, and constructing a training sample pair consisting of each sensor spatial functional connectivity matrix and its corresponding source spatial functional connectivity matrix. Acquire real EEG signals and preprocess the real EEG sensor sampling signals, and calculate the real value of the corresponding sensor spatial functional connectivity matrix for the preprocessed real EEG sensor sampling signals; The deep learning mapping module is trained end-to-end using the training dataset to learn the mapping relationship from the sensor space functional connectivity matrix to the source space functional connectivity matrix. The deep learning mapping module is also used to receive the true value of the sensor space functional connectivity matrix and output the predicted value of the source space functional connectivity matrix. The source spatial functional connectivity submatrix of each brain region is obtained from the predicted value of the source spatial functional connectivity matrix, and the correlation between brain regions is calculated based on the source spatial functional connectivity submatrix of each brain region.

[0061] The working principle of the brain region correlation analysis method is the same as that of the brain region correlation analysis system described above, and will not be repeated here.

[0062] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A brain region correlation analysis system for children with autism, characterized in that, include: The brain region segmentation module is used to divide the cerebral cortex into multiple brain regions; The training dataset construction module is used to obtain the lead field matrix, generate sensor noise simulation signals, generate brain source activity time series simulation signals for each brain region, calculate the source spatial functional connectivity sub-matrix for the source activity time series simulation signals for each brain region, concatenate the source spatial functional connectivity sub-matrixes of all brain regions into a source spatial functional connectivity matrix, generate sensor simulation signals based on the sensor noise simulation signals, the lead field matrix, and the source activity time series simulation signals of all brain regions, calculate the sensor spatial functional connectivity matrix corresponding to the sensor simulation signals, and each sensor spatial functional connectivity matrix and its corresponding source spatial functional connectivity matrix constitute a training sample pair, constructing a training dataset containing multiple training sample pairs; The signal preprocessing module is used to acquire real EEG sensor sampling signals and preprocess the real EEG sensor sampling signals, and calculate the corresponding sensor spatial functional connectivity matrix true value for the preprocessed real EEG sensor sampling signals. The deep learning mapping module is trained end-to-end using the training dataset to learn the mapping relationship from the sensor space functional connectivity matrix to the source space functional connectivity matrix. The deep learning mapping module is also used to receive the true value of the sensor space functional connectivity matrix and output the predicted value of the source space functional connectivity matrix. The brain region correlation analysis module is used to obtain the source spatial functional connectivity submatrix of each brain region from the predicted value of the source spatial functional connectivity matrix, and to calculate the correlation between brain regions based on the source spatial functional connectivity submatrix of each brain region.

2. The brain region correlation analysis system for children with autism as described in claim 1, characterized in that, The process of stitching together the source spatial functional connectivity submatrices of all brain regions into a source spatial functional connectivity submatrices involves stitching together the source spatial functional connectivity submatrices of all brain regions along the diagonal.

3. The brain region correlation analysis system for children with autism as described in claim 1, characterized in that, The step of generating time-series simulated brain activity signals for each brain region includes the following steps: Let N be the number of brain regions. Each time a time-series simulated signal of brain-derived activity is generated, select... One active brain region, 1≤ ≤N, is Each active brain region generates a time-series simulated signal of brain-derived activity.

4. The brain region correlation analysis system for children with autism as described in claim 1, characterized in that, The step of generating time-series simulated brain activity signals for each brain region includes the following steps: The parameters of the source activity time series simulation signal are predefined. Each time the brain source activity time series simulation signal is generated, the parameter conditions of the source activity time series simulation signal are set so that the brain source activity time series simulation signal generated for each brain region meets the statistical parameter setting conditions.

5. The brain region correlation analysis system for children with autism as described in claim 1, characterized in that, The preprocessing includes the following steps: The real EEG signal before preprocessing is denoted as , It is The matrix, where The number of time sampling points, each time point Corresponding to a sample value, , Indicates a point in time The sampled values, For the number of sensor channels, Baseline correction was performed, and the baseline-corrected EEG signal was recorded as follows: ; right Denoising was performed, and the denoised EEG signal was recorded as follows: ; Constructing the projection matrix Through spatial projection The signal is obtained through processing. , Indicates matrix transpose; right Filtering is performed to obtain the signal The calculation formula for the preprocessed real EEG signal is as follows: ； in, Represents the Fast Fourier Transform. H(f) represents the inverse fast Fourier transform, and H(f) is the bandpass filter function.

6. The brain region correlation analysis system for children with autism as described in claim 1, characterized in that, The deep learning mapping module includes: The Transformer branch is used to convert the sensor spatial functional connection matrix into a grayscale image of size H×W, where H and W are the height and width of the image, respectively. The first feature map of H / 4 × W / 4, the second feature map of H / 8 × W / 8, the third feature map of H / 16 × W / 16, and the fourth feature map of H / 32 × W / 32 are extracted from the H×W grayscale image. The EfficientNet branch is used to extract the fifth feature map of H / 2 × W / 2, the sixth feature map of H / 4 × W / 4, the seventh feature map of H / 8 × W / 8, the eighth feature map of H / 16 × W / 16, and the ninth feature map of H / 32 × W / 32 from the H×W grayscale image, respectively. The decoder is used to take the fifth feature map as the first fusion feature, fuse the first feature map and the sixth feature map to obtain the second fusion feature, fuse the second feature map and the seventh feature map to obtain the third fusion feature, fuse the third feature map and the eighth feature map to obtain the fourth fusion feature, fuse the fourth feature map and the ninth feature map to obtain the fifth fusion feature, and decode based on the first fusion feature, the second fusion feature, the third fusion feature, the fourth fusion feature, and the fifth fusion feature to obtain the predicted value of the source space function connectivity matrix.

7. A brain region correlation analysis system for children with autism as described in claim 1, characterized in that, The brain region correlation analysis module includes: The first analysis module is used to obtain the functional connectivity matrix of each brain region of the test user from the predicted value of the source space functional connectivity matrix. The calculation formula for calculating the brain region correlation parameter based on the functional connectivity matrix of each brain region is predefined. The brain region correlation parameter of the test user is calculated, and the difference between the brain region correlation parameter of the test user and the brain region correlation parameter of the normal control group of children is calculated. If the difference is greater than the preset value, the test user is determined to be a child with autism. The second analysis module is used to calculate the correlation between the time series of source activities between two brain regions. The calculation formula is as follows: ； in, Indicates the first i The mean of the time series of source activity in each brain region. Indicates the first j The mean of the time series of source activity in each brain region. Indicates the first i The brain regions and the first j Correlation of source activity time series in individual brain regions; The third analysis module is used to calculate the coherence of the source activity time series between two brain regions. The calculation formula is as follows: ； in, It is frequency Next i The brain regions and the first j Coherence of source activity time series in individual brain regions It is the first i The brain regions and the first j Joint power spectral density of source activity time series of individual brain regions It is the first i The self-power spectral density of the source activity time series of each brain region It is the first j The self-power spectral density of the source activity time series of each brain region.

8. A brain region correlation analysis system for children with autism as described in claim 1, characterized in that, The source space functional connectivity submatrix, the source space functional connectivity matrix, and the sensor space functional connectivity matrix are all covariance matrices.

9. A method for analyzing brain region correlations in children with autism, characterized in that, Including the following steps: The cerebral cortex is divided into multiple brain regions; The process involves obtaining the lead field matrix, generating sensor noise simulation signals, generating brain source activity time-series simulation signals for each brain region, calculating the source spatial functional connectivity sub-matrix for each brain region's source activity time-series simulation signals, concatenating the source spatial functional connectivity sub-matrixes of all brain regions into a single source spatial functional connectivity sub-matrix, generating sensor simulation signals based on the sensor noise simulation signals, the lead field matrix, and the source activity time-series simulation signals of all brain regions, calculating the sensor spatial functional connectivity matrix corresponding to the sensor simulation signals, and constructing a training sample pair consisting of each sensor spatial functional connectivity matrix and its corresponding source spatial functional connectivity matrix. Acquire real EEG signals and preprocess the real EEG sensor sampling signals, and calculate the real value of the corresponding sensor spatial functional connectivity matrix for the preprocessed real EEG sensor sampling signals; The deep learning mapping module is trained end-to-end using the training dataset to learn the mapping relationship from the sensor space functional connectivity matrix to the source space functional connectivity matrix. The deep learning mapping module is also used to receive the true value of the sensor space functional connectivity matrix and output the predicted value of the source space functional connectivity matrix. The source spatial functional connectivity submatrix of each brain region is obtained from the predicted value of the source spatial functional connectivity matrix, and the correlation between brain regions is calculated based on the source spatial functional connectivity submatrix of each brain region.

Images