Visual dysfunction detection method and related equipment

Abstract

The embodiment of the invention provides a visual dysfunction detection method and related equipment, and belongs to the technical field of medical signal processing and neural engineering. The method comprises the following steps: carrying out preprocessing and source space positioning on electroencephalogram signals, and reconstructing cortical neural activity; a dynamic k-nearest neighbor algorithm based on inverse distance weighting is adopted, cortex vertexes are adaptively allocated to a brain region of an AAL map, and the problem of space mismatching is solved; extracting time sequences of the visual network and the default mode network in a frequency band, and calculating dynamic function connection of the time sequences to construct a spatial-temporal feature tensor; the tensor is input into a multi-scale controllable graph convolutional network model for classification prediction, the network adopts similar Chebyshev connection and a double-gating unit to enhance the global feature extraction capability, and a residual cavity module is fused to capture multi-scale features. The accuracy, reliability and neurophysiological interpretability of visual dysfunction detection are effectively improved, and an objective basis is provided for clinical precise diagnosis and treatment.

Description

Technical Field

[0001] This application relates to the fields of medical signal processing and neural engineering technology, and in particular to a method and related equipment for detecting visual impairment. Background Technology

[0002] Visual impairments, such as amblyopia and other neurodevelopmental visual disorders, pose a significant global public health challenge. Early and accurate diagnosis is crucial for intervention and treatment. The pathological nature of these diseases involves abnormal functional connectivity in the visual cortex and neural plasticity remodeling, rather than simply optical structural defects.

[0003] Currently, traditional clinical diagnostic methods rely heavily on subjective assessments, such as best-corrected visual acuity tests and contrast sensitivity tests. These methods have limited application in non-communicative populations such as infants and young children, and patients with cognitive impairments, making it difficult to detect early neurological abnormalities and, more importantly, to reveal deeper pathological mechanisms.

[0004] Visual evoked potentials (VEPs), as a non-invasive electrophysiological technique, provide objective evidence for assessing the functional integrity of the visual pathway. However, traditional VEP analysis is often limited to local waveform parameters (such as latency and amplitude), making it difficult to comprehensively characterize the dynamic reorganization process of the whole-brain functional network. Visual dysfunction often involves cross-network interaction dysregulation between multiple brain regions (such as the visual network and the default mode network), and relying solely on coarse measurements from scalp electrodes cannot accurately pinpoint the origin and propagation pathway of abnormal neural activity.

[0005] In recent years, brain functional network analysis and graph neural networks have provided new perspectives for understanding complex brain diseases. However, combining EEG source localization, functional connectivity, and GNNs for visual impairment detection still faces a series of technical bottlenecks: 1) Spatial mismatch problem: Spatial alignment errors often exist between brain source localization results and regions of interest in anatomical atlases, leading to inaccurate functional area division and affecting the reliability of subsequent network analysis. 2) Feature redundancy and insufficient specificity: Existing methods often directly use the time series of whole-brain channels as input without distinguishing the specificity of functional networks, resulting in feature redundancy and susceptibility to interference from irrelevant brain regions, leading to poor model interpretability. 3) Structural limitations of graph neural networks: Traditional GNNs suffer from problems such as over-smoothing, difficulty in capturing long-range dependencies, and inability to effectively balance the importance of neighbor information at different levels, limiting their ability to model the dynamics of complex brain networks. Summary of the Invention

[0006] The main objective of this application is to propose a method, electronic device, storage medium, and program product for detecting visual dysfunction based on EEG source localization and functional network connectivity, in order to solve problems such as spatial mismatch, feature redundancy, poor model interpretability, and limited graph neural network modeling capabilities in the prior art.

[0007] To achieve the above objectives, one aspect of this application proposes a method for detecting visual impairment, the method comprising: The subject's electroencephalogram (EEG) signals were acquired and preprocessed. Source spatial localization of preprocessed EEG signals was performed to reconstruct neural activity sequences in the source space of the cerebral cortex; Based on standard brain atlases, the source space is functionally partitioned, the cortical vertices are divided into multiple brain regions, and the time series of each brain region in a specific functional network are extracted. Based on the time series of each brain region in the specific functional network, the brain functional connectivity matrix is ​​calculated to construct the spatiotemporal feature tensor; The spatiotemporal feature tensor is input into a trained multi-scale controllable graph convolutional network model, which outputs the detection or prediction results of visual impairment. The multi-scale controllable graph convolutional network model includes at least a graph feature extraction module. The graph feature extraction module uses Chebyshev-like polynomial connections and combines them with dual gating units to aggregate information of neighbor nodes of different orders in the graph structure.

[0008] In some embodiments, the functional partitioning of the source space based on standard brain atlas includes: Calculate the Euclidean distance from each neural source vertex in the source space to the centroid of each region of interest in the standard brain atlas; A dynamic k-nearest neighbor algorithm based on inverse distance weighting is used to adaptively assign each neural source vertex to a region of interest.

[0009] In some embodiments, the specific functional network includes a visual network and a default pattern network; the time series is derived from the EEG signals. The signal extracted from the frequency band.

[0010] In some embodiments, calculating the brain functional connectivity matrix includes: The time series is divided into multiple time segments using a sliding time window; Within each time segment, the Pearson correlation coefficients between each pair of brain regions in the visual network and the default mode network are calculated to form a dynamic functional connectivity matrix.

[0011] In some embodiments, the output of the Chebyshev-like polynomial connection in the graph feature extraction module Represented as:

[0012] in, For activation function, The maximum order of graph convolution. For the improved Chebyshev join function, For input features, for The adjacency matrix corresponding to the order, for The weights corresponding to the order; And, when At that time, through the expression Establish an invariant mapping.

[0013] In some embodiments, the dual-gating unit includes a first gating unit and a second gating unit; The first gating unit includes an update unit, a candidate hiding unit, and an output unit; the update unit adjusts the amount of information retained from previous neighboring orders and the degree to which new data from the current order is incorporated into the hidden state, and the calculation formula is as follows:

[0014] in, This represents the output representation of the update unit at order t. The weights are represented as follows, and the outputs corresponding to orders t-1 and t are respectively expressed as and ; For input features, for The adjacency matrix corresponding to the order; For activation functions; Candidate hidden states serve as intermediate representations. Information is refreshed by combining the current input with the previous hidden state, and the calculation formula is as follows:

[0015]

[0016] in, Indicates the candidate hidden state. Indicates element-wise multiplication; The output unit controls the final hidden state output and coordinates the weighted interaction between the candidate hidden state and its previous state. The calculation formula is as follows:

[0017] in, The output of the t-th order is represented by the output unit; The second gating unit is a forget gate, used to control the importance of each order of neighbor information in the final summation, and its output... Represented as:

[0018] in, For activation function, As for the weight of the Forgotten Gate, This represents the result of each order of graph convolution.

[0019] In some embodiments, the multi-scale controllable graph convolutional network model further includes a multi-channel feature extractor; The multi-channel feature extractor is based on a residual network structure and incorporates a dilated convolution module to form a residual dilated module, which is used to extract multi-scale deep spatiotemporal features from the output of the graph feature extraction module.

[0020] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0021] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0022] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0023] The beneficial effects of this application include at least the following: 1) Improved accuracy of spatial localization and functional area division: By using a dynamic k-nearest neighbor algorithm based on inverse distance weighting, cortical vertices are adaptively assigned to ROIs defined by the AAL map, effectively solving the problem of mismatch between source space reconstruction and anatomical map boundaries, and significantly improving the reliability and neurophysiological interpretability of functional connectivity analysis.

[0024] 2) Enhanced feature specificity and interpretability: By focusing on visual networks and default mode networks closely related to visual function... By analyzing neural activity across frequency bands and calculating their functional connectivity, features with clear physiological significance were constructed. This effectively reduced feature redundancy, improved the sensitivity and specificity of identifying visual pathway dysfunction, and provided an objective basis for individualized clinical assessment.

[0025] 3) Improved modeling capabilities of graph neural networks: By introducing a Chebyshev-like quick connection method, the gradient problem and the difficulty of acquiring information from distant nodes in deep GNNs are effectively solved. Through dual-gating units, the importance and flow of neighbor information of different orders are dynamically controlled, effectively balancing low-order and high-order information and mitigating the negative impact of noise and redundant information.

[0026] 4) Multi-scale deep feature extraction is achieved: By embedding the dilated convolution module into the residual structure to form a "residual dilated module", the model can fuse local details and global context information without increasing the depth and number of parameters, capture more comprehensive multi-scale spatiotemporal features, alleviate the oversmoothing problem, and improve classification and prediction performance. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of dilated convolution; Figure 2 This is a flowchart of a visual impairment detection method according to an embodiment of this application; Figure 3 This is a schematic diagram of the image feature extraction module in an embodiment of this application; Figure 4 This is a schematic diagram of the gate control unit in an embodiment of this application; Figure 5 This is a schematic diagram of the convolution module in an embodiment of this application; Figure 6 This is a schematic diagram of the multi-channel feature extraction module in an embodiment of this application; Figure 7 This is a schematic diagram of the hardware structure of the electronic device in the embodiments of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0030] Before providing a detailed description of the embodiments of this application, some terms and related knowledge involved in the embodiments of this application will be explained first.

[0031] 1) Brain power localization Brain-source localization is a technique that uses electroencephalogram (EEG) signals recorded on the scalp to infer the location of intracranial neural activity sources. Its physical basis is that the electrical current generated by neuronal activity is conducted through media such as the skull and cerebrospinal fluid to the scalp surface, forming a potential distribution. This process requires solving two core problems: the forward problem and the inverse problem.

[0032] The forward problem refers to calculating the potential distribution on the scalp given the distribution of electrical sources within the brain. Its mathematical model is based on Maxwell's equations under a quasi-static approximation, expressed as:

[0033] in To indicate individual scalp electrodes Measurement potential matrix at each time point for Current source density distribution at each candidate source location, The lead field matrix (describing the conductivity properties of the medium), also known as the forward operator, describes the electric field propagation relationship from each source location to each electrode. Its calculation depends on individualized or templated head geometry models and the conductivity distribution of brain tissue.

[0034] The inverse problem is to use the measured Inversely inferring the unknown ,because (Since the number of source points far exceeds the number of electrodes), this problem is underdetermined (there are infinitely many solutions), requiring the introduction of regularization constraints to obtain a stable solution. A commonly used method is the Minimum Norm Estimate (MNE), which searches for the source distribution with the minimum energy among all possible solutions. This solution can be expressed as:

[0035] in, It is the identity matrix. This is a regularization parameter used to control noise and the stability of the solution.

[0036] 2) Brain network partitioning Brain functional partitioning maps anatomical structures to a system of brain regions with specific cognitive or perceptual functions, forming the basis of brain network analysis. Automated Anatomical Labeling (AAL), provided by the Montreal Neurological Institute (MNI), includes 116 regions, 90 of which are brain structures. The MNI model provides a standardized coordinate space and reference images for brain data from different individuals, where each neural source has a three-dimensional coordinate in a Cartesian coordinate system. AAL is a widely used standard anatomical analysis template in brain imaging analysis, classifying each neural activity in a brain image into a specific anatomical region, also known as a Region of Interest (ROI), and is widely used in fMRI and EEG source localization. Each brain region is considered a node in the network, and the whole brain is divided into... The first functional area, then the second each region Corresponding to a set of spatial regions:

[0037]

[0038] Average time series data were extracted from each region. As a representative of the neural activity at this node:

[0039] in For the region The collection of voxels within, This represents the current density at the source location. Functional partitioning allows complex brain activity to be simplified into a dynamic system with a finite number of nodes, supporting the subsequent construction of functional connectivity networks. Two common core networks are the Default Mode Network (DMN) and the Visual Network. The DMN involves various cognitive functions such as individual thinking and memory integration, while the Visual Network is related to visual processing and visual tasks.

[0040] 3) Functional connections Functional connectivity (FC) measures the temporal synchronicity of neural activity across different brain regions, reflecting their ability to work collaboratively. Given... Time series of each functional area The full-factor (FC) of pairwise nodes can be quantified using various methods. The most commonly used is the Pearson correlation coefficient.

[0041] in, Represents a node and Functional connection strength between them For covariance, This is a time series analysis of two brain regions. Its standard deviation.

[0042] 4) Graph Convolutional Neural Network (GCN) Capturing latent spatial relationships from non-Euclidean structural graphs is crucial. Graph Neural Networks (GCNs) effectively utilize these spatial relationships between joints. Graph convolution operations can be defined as:

[0043] in, Indicates input data, This represents the normalized adjacency matrix. During the normalization process, the degree of each node needs to be calculated. Based on the distance between nodes and the center point, the adjacency matrix can be divided into three different submatrices. . The weight matrix representing the convolution operation. It is an adaptive node weight mask. It is a dot product operation.

[0044] 5) Dilated Convolution Dilated convolution is a technique in convolutional neural networks that effectively expands the receptive field without increasing the number of model parameters or computational overhead by introducing a gap (i.e., dilation rate) and padding it with zero values ​​inside the convolutional kernel. This allows it to capture a wider range of contextual information while maintaining network depth. Figure 1 As shown. This method can model long-range dependencies while keeping the kernel size constant, and is particularly suitable for processing sequence signals with complex spatiotemporal structures. Neural activity in EEG signals has complex time-frequency characteristics, and oscillation patterns at different time scales may contain key information related to visual function status. Dilated convolution helps improve the model's recognition accuracy by expanding the perceptual range and extracting multi-scale spatiotemporal features. When applied to its discrete form of one-dimensional convolution, considering parameters such as padding, stride, and kernel size, the dilated convolution operation can be expressed as:

[0045] in, Represents the input sequence. Represents the convolution kernel. Indicates the size of the convolution kernel. This is the index of the output feature map. This is the index of the position of the convolution kernel as it slides across the input. Step size, void ratio Control the sampling interval of the convolution kernel on the input sequence. This is a bias term used to implement zero padding to maintain the output dimension.

[0046] Visual dysfunction poses a significant global public health challenge. Epidemiological studies of neurodevelopmental visual disorders, with amblyopia as a prime example, show a prevalence of 1.36%-4%. Early and accurate diagnosis is crucial for intervention and treatment. The pathological nature of these diseases involves abnormal functional connectivity in the visual cortex and neural plasticity remodeling, rather than simple optical structural defects. Traditional clinical diagnosis relies heavily on subjective assessment methods, such as best-corrected visual acuity (BCVA) testing and contrast sensitivity testing, which have limited application in non-communicative populations such as infants and children with cognitive impairments. These methods not only fail to capture early neurological abnormalities but also fail to reveal deeper pathological mechanisms. For example, even after optical correction, amblyopia patients still exhibit visual cortex neurological deficits, manifesting as persistent impairment in visual information integration and spatial localization abilities. Significant blind spots exist in clinical treatment. Although approximately 25% of amblyopia patients can be cured with refractive correction alone, the remaining patients require additional treatment interventions. Even with effective treatment, the relapse rate of amblyopia remains as high as 10% to 40%. This uncertainty stems from the lack of a biomarker prediction system, which makes it impossible to identify treatment resistance subtypes in advance, posing a continuous challenge to the accurate diagnosis and treatment of visual impairment.

[0047] Visual evoked potentials (VEPs), a non-invasive electrophysiological technique, provide objective evidence for assessing the functional integrity of visual pathways by recording the cortical electrical responses of the brain to visual stimuli (such as checkerboard flipping or grating patterns). VEP signals acquired through scalp electrodes can reflect the electrophysiological activity of the striatum and extrastriate cortex. Its core lies in revealing neural conduction delays and signal integration abnormalities in visual information processing, which is particularly prominent in visual dysfunction disorders such as amblyopia. Transient visual evoked potentials assess these by measuring the latency and amplitude changes of specific spatiotemporal stimuli, while steady-state visual evoked potentials (SSVEPs) further reveal the steady-state characteristics of neural oscillations by analyzing the synchronous response of EEG to periodic stimuli. However, traditional VEP analysis is often limited to local waveform parameters, making it difficult to comprehensively characterize the dynamic reorganization process of the whole-brain functional network. Visual dysfunction often involves cross-network interaction dysregulation across multiple brain regions; relying solely on coarse measurements from scalp electrodes cannot accurately pinpoint the origin and propagation pathways of abnormal neural activity.

[0048] In recent years, brain functional network analysis has provided a new perspective for understanding complex brain diseases. The brain is not a simple collection of isolated regions, but rather a complex network of functionally specific subnetworks working collaboratively. The visual network (VN) is directly responsible for receiving and processing visual information, while the default mode network (DMN) is closely related to memory integration and attention regulation. This cross-network functional connectivity (FC) pattern is crucial for efficient visual information processing. In patients with visual impairments such as amblyopia, this cross-network synergistic and antagonistic mechanism may be disrupted, manifesting as abnormal intra-network connectivity or disruption of inter-network connectivity patterns. However, current technologies lack the ability to objectively quantify and decode this deep-level network interaction.

[0049] EEG signal source localization technology, based on biophysical modeling and mathematical inversion algorithms, can trace the origins of EEG signals recorded on the scalp back to neural activity sources within the brain, enabling spatiotemporal dynamic reconstruction of cortical and even subcortical functional areas. Compared to functional magnetic resonance imaging (fMRI), which offers high spatial resolution but insufficient temporal accuracy, and traditional EEG, which provides millisecond-level response but suffers from spatial ambiguity, source localization technology promises to combine the advantages of both, providing near-functional imaging-level brain region localization capabilities while maintaining high temporal resolution. This characteristic makes it particularly suitable for analyzing neural mechanisms involving both rapid transmission and distributed processing in the visual pathway.

[0050] Meanwhile, Graph Neural Networks (GNNs) have shown great potential in the field of medical signal processing. Graph Convolutional Networks (GCNs) effectively extract spatial and frequency domain features by modeling the spatial topological relationships between electrode channels, enabling high-order node representation learning. Numerous studies have further improved performance based on various GCN-based improved models, such as Dynamic Graph Convolutional Networks (DGCNNs), which capture time-varying dependencies by dynamically adjusting the graph structure; and Wide Graph Convolutional Networks (GCB-nets), which integrates wide networks to improve feature extraction efficiency. These methods fully utilize the graph structure characteristics of EEG, enhancing the accuracy and robustness of emotion classification. However, EEG signals possess complex multi-scale spatiotemporal characteristics, and single-scale feature extraction is insufficient to fully uncover the underlying pathological information. How to construct the optimal graph structure and how to utilize EEG functional networks for analysis remain challenges.

[0051] Therefore, there is an urgent need for an analytical framework that can deeply integrate cross-functional network connectivity features with deep learning models. This framework needs to transcend the limitations of scalp potentials and accurately extract neural activity within and between key brain networks (such as the visual network and the default mode network); secondly, it needs to effectively quantify the temporal dynamic functional connectivity patterns of these networks, constructing high-dimensional features with neurophysiological significance; and thirdly, it needs a high-performance graph neural network model that can overcome the limitations of traditional models, effectively integrate local details and global contextual information, achieve accurate modeling and classification of complex brain network dynamics, and ultimately serve the objective identification and treatment response assessment of visual dysfunction.

[0052] In view of this, this application provides a method, electronic device, storage medium, and program product for detecting visual impairment. This technical framework can simultaneously retain the millisecond-level temporal accuracy and spatial network information of EEG, effectively capturing the sources of visual function-related neural activity and realizing the analysis and decoding of the neural mechanisms of visual impairment. First, the EEG data is filtered and artifact removed to enhance the signal. Then, combining a head volume conductor model and the standardized low-resolution electromagnetic coherence tomography (sLORETA) method, the scalp signals are mapped to the cortical source space, reconstructing the three-dimensional dynamics of neural activity. Based on standard brain atlases, the source signals are functionally partitioned, and extracted... The time series of visual and default mode networks within the frequency band were analyzed. Dynamic functional connectivity was further calculated to construct multidimensional brain network features. Finally, a deep learning model was used to predict treatment response and classify patient subtypes, providing neuroimaging evidence for individualized assessment of visual dysfunction.

[0053] This application provides a method for detecting visual impairment, relating to the fields of medical signal processing and neural engineering. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing a visual impairment detection method, but is not limited to the above forms.

[0054] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0055] Figure 2 This is an optional flowchart of a visual impairment detection method provided in an embodiment of this application. Figure 2 The method may include, but is not limited to, steps S1 to S5.

[0056] S1. Acquire the subject's electroencephalogram (EEG) signals and preprocess the EEG signals; S2. Locate the source space of the preprocessed EEG signal and reconstruct the neural activity sequence of the source space of the cerebral cortex; S3. Based on standard brain atlases, the source space is functionally partitioned, the cortical vertices are divided into multiple brain regions, and the time series of each brain region in a specific functional network are extracted. S4. Based on the time series of each brain region in the specific functional network, calculate the brain functional connectivity matrix to construct the spatiotemporal feature tensor; S5. Input the spatiotemporal feature tensor into the trained multi-scale controllable graph convolutional network model and output the detection or prediction results of visual impairment; wherein, the multi-scale controllable graph convolutional network model includes at least a graph feature extraction module, which adopts Chebyshev-like polynomial connection and combines dual gating units to aggregate information of neighbor nodes of different orders in the graph structure.

[0057] The solutions of the embodiments of this application will be described in detail below with reference to the accompanying drawings and specific application examples.

[0058] This embodiment proposes a visual impairment detection method based on source localization and functional network connectivity, which specifically includes the following steps: Step 1: EEG data preprocessing The acquired EEG data underwent preprocessing to improve signal quality and reduce noise interference. First, the raw EEG signal was bandpass filtered to preserve neurophysiological activity within the target frequency band. Then, independent component analysis was used to identify and remove artifacts introduced by physiological activities such as eye movements and electromyography, achieving effective noise reduction while preserving the integrity of the neural signal.

[0059] Step 2: Source Spatial Localization After step 1, clean EEG data containing visual function information was obtained. Using the MNE-Python toolkit, the sLORETA method was employed for source localization, converting the channel-space EEG signals into source-space activity over approximately 8196 vertices. Based on the fsaverage standard brain template, a three-layer boundary element method (BEM) head model incorporating the scalp, skull, and cortical surface was constructed. Forward solutions were calculated using the Biosemi64 standard electrode locations to obtain the lead field matrix mapping from the cortical dipole source to the scalp EEG electrodes. For each EEG signal sample, depth-weighted regularization was applied within the sLORETA framework. Signal-to-noise ratio This method calculates current density imaging from scalp EEG electrodes to the source space, enabling three-dimensional spatial reconstruction of neural activity and generating high temporal resolution source space brain activity sequences. Cortical sources are distributed in an unconstrained orientation within a high-resolution source space, generated based on Octree subdivision with 4098 vertices per hemisphere. This approach converts EEG data from the channel space into intracranial current source distribution, thereby improving the spatial resolution of brain activity representation.

[0060] Step 3: Brain region division Based on the neural activity reconstructed from the source space in step 2, a standard brain atlas was introduced to functionally partition the whole brain, with each region serving as a node in the brain network. The AAL template provides specific MNI coordinates for the cortical and subcortical regions, enabling the calculation of the vectorized Euclidean distance from each of the 8196 neural source vertices to its nearest AAL centroid. To address the alignment issue between source reconstruction and atlas boundaries, a dynamic k-nearest neighbor algorithm based on inverse distance weighting was used to adaptively assign vertices to each ROI region, mitigating edge effects while ensuring robust anatomical correspondences. Following the seven-brain network partitioning method proposed by Yeo et al., brain regions distributed within the visual and default mode networks were identified from the resulting 116 brain regions. The specific brain regions and their corresponding brain networks are shown in Table 1. Table 1 Correspondence between brain regions and brain networks

[0061] Step 4: Connect the calculation function Extract Pearson correlation analysis was performed on EEG signals from 35 brain regions belonging to the visual and default mode networks within the frequency band to assess functional connectivity strength. A sliding time window method was used to segment continuous time series into non-overlapping 500ms time windows to improve statistical reliability. Within each time window, the pairwise linear correlations between time series from each brain region were calculated to construct an undirected functional connectivity matrix. Finally, a spatiotemporal feature tensor was generated for each sample, with a shape of 1 frequency band × 100 time windows × (35 × 35) connections, forming a connection feature sequence that evolves over time. This feature tensor was further normalized and vectorized to form a high-dimensional brain network feature representation. This step, through spatial aggregation and temporal segmentation of vertex-level signals, effectively reduces signal leakage artifacts while preserving the dynamic functional network structure.

[0062] Step 5: Fractal Modeling and Prediction The spatiotemporal feature tensors obtained in step 4 are input into a multi-scale controllable graph convolutional network, and finally, fully connected layers are used for label prediction. The model adopts a supervised learning framework, using clinical treatment effects or diagnostic labels as training objectives. Through model training and parameter optimization, the treatment response trends of new individuals can be inferred, thereby achieving visual impairment detection.

[0063] 5.1) First, the graph feature extraction module needs to be used to extract regularized features, such as... Figure 3 As shown, this module uses the spatiotemporal feature tensor obtained in the previous step as input.

[0064] a) This module includes a shortcut connection to enhance the transfer of feature information from near-end nodes to global nodes. In the ResNet-based shortcut connection, t-order feature extraction can be represented as:

[0065] in, This represents the output of the graph feature extraction module. For activation function, This represents the maximum order of graph convolution. Represents the Chebyshev polynomial. This is represented as a modified Chebyshev join function. This is represented as input. Represented as The adjacency matrix between nodes within the same order. Represented as The weights corresponding to the order. When At that time, through the expression Establish an invariant mapping. This structure enables feature extraction from distant nodes, ensuring robust network performance that is no worse than dedicated extraction from nearest-neighbor nodes.

[0066] b) This module also introduces a dual-gating unit to improve the flow of information. The specific implementation of the dual-gating unit is as follows: Figure 4 As shown, the gating unit selectively filters key information by constraining the data transmission and output results of each information layer shortcut connection module, which helps to extract long-term feature dependencies.

[0067] The gating unit 1 structure mainly includes an update unit, a candidate hiding unit, and an output unit. The update unit adjusts the amount of information retained from previous neighboring orders and the degree to which new data from the current order is incorporated into the hidden state. The calculation formula is as follows:

[0068] in, The weights are represented as follows, and the outputs corresponding to orders t-1 and t are respectively expressed as and And can be used express .

[0069] Candidate hidden states serve as intermediate representations. Information is refreshed by combining the current input with the previous hidden state, and the calculation formula is as follows:

[0070] in Indicates the candidate hidden state. This represents element-wise multiplication.

[0071] The output unit controls the final hidden state output and coordinates the weighted interaction between the candidate hidden state and its previous state. Based on the above two equations, the output unit can be expressed as:

[0072] Gating unit 2 is a forget gate, used to control the importance of each order of neighbor information in the final summation. The calculation process can be represented as follows:

[0073] in This is represented as the output of the forget gate. This represents the activation function. Indicates the weight of the forget gate. This represents the result of each order of graph convolution.

[0074] The dual-gated unit's gating mechanism effectively captures long-term dependencies, simplifies the parameter set, and streamlines the network structure, thus achieving a balance between performance and computational efficiency. This allows the model to reasonably balance outputs across different feature categories, improving the performance of visual impairment detection tasks. The final output of the dual-gated unit can be expressed as:

[0075] in, For activation function, Let t be the output of the output gate.

[0076] 5.2) Multi-channel feature extractors are an important component of multi-scale controllable graph convolutional network frameworks, such as... Figure 6 As shown. The features learned by the graph feature extraction module are further extracted to extract deeper features. ResNet, as the structural basis of the multi-channel feature extractor, enables the network to learn the residual mapping between input and output, rather than directly fitting the complete nonlinear transformation, effectively alleviating the gradient vanishing problem common in deep neural networks, thereby achieving network depth expansion while maintaining training efficiency. The dilated convolution module is embedded in the residual structure, as shown. Figure 5 As shown, a "residual dilated module" is formed, which combines the training stability brought by residual connections with the multi-scale perception capability of dilated convolution. This integrated structure can effectively fuse local details and global contextual information, achieve more comprehensive feature representation in multi-channel signal processing, and alleviate the problem of over-smoothing of features caused by deep stacking.

[0077] In summary, compared with the prior art, this application has at least the following advantages and beneficial effects: 1) In the existing technology, there is often a spatial mismatch between the brain power localization results and the ROI of the anatomical atlas, which leads to inaccurate functional area division and affects the reliability and physiological interpretability of subsequent network analysis.

[0078] This application proposes using a standardized low-resolution EEG method for source localization, combined with an inverse distance-weighted dynamic k-nearest neighbor algorithm, to adaptively assign cortical vertices to the ROIs defined by the AAL map, effectively solving the problem of mismatch between source space reconstruction and anatomical map boundaries, and improving the accuracy of functional area division.

[0079] Advantages: This method overcomes the boundary errors caused by traditional fixed mapping, achieves fine-grained matching of vertices to ROIs, significantly improves the spatial consistency between source localization and functional partitioning, and enhances the reliability and neurophysiological interpretability of subsequent functional connectivity analysis.

[0080] 2) Existing EEG signal analysis methods typically use time series data from the entire brain as direct input, failing to distinguish functional network specificity. This leads to feature redundancy and susceptibility to interference from unrelated brain regions. Such methods are often based on black-box models, lacking neurophysiological interpretability, making it difficult to pinpoint the origin and propagation path of abnormal activity, and resulting in low accuracy.

[0081] This application focuses on key brain networks closely related to visual function—the visual network and the default mode network—and extracts their components from the source space. The neural activity sequences in the frequency bands are used to calculate cross-network functional connectivity using Pearson correlation coefficients, constructing a spatiotemporal feature tensor with a clear anatomical-functional correspondence, thereby improving the interpretability and classification accuracy of the model.

[0082] Advantages: Visual networks and default mode networks play a crucial role in visual information integration and attention regulation. Frequency band connectivity anomalies are highly sensitive and specific, effectively distinguishing patients from healthy controls, improving the interpretability of neural mechanisms, and providing objective and quantifiable data for individualized clinical assessment, treatment response prediction, and pathological subtype identification.

[0083] 3) As the number of layers increases in deep graph neural networks, the vanishing or exploding gradient problem can lead to network degradation. This severely limits its ability to extract effective features from distant nodes, making it difficult for the model to capture global contextual information in the graph structure, thus resulting in poor performance in tasks requiring global understanding.

[0084] This application proposes a novel Chebyshev-like network connection method that employs residual learning to enhance the network's global feature extraction from remote nodes, thereby facilitating feature transmission from near-end nodes to far-end nodes.

[0085] Advantages: The shortcut connection method proposed in this application effectively solves the problem of losing low-order neighbor information when extracting high-order features, enabling the model to take into account both the local importance of low-order neighbor information and the globality of high-order neighbor features when extracting graph features.

[0086] 4) Convolutional networks often fail to balance the importance of low-order neighbor information with high-order neighbor information, and unreasonable information output often generates a lot of redundant information and noise.

[0087] This application proposes a novel dual-gating unit to improve the propagation of hierarchical feature information and dynamically update the gating, thereby selectively retaining and discarding information, which helps to extract long-term feature dependencies.

[0088] Advantages: The dual-gated unit proposed in this application can control the information flow of each order in the Chebyshev-like connection, preventing irrelevant and redundant information from flowing to higher orders, and enabling the model to better balance low-order and high-order information; by controlling the neighbor information output of each order, it can more effectively avoid the accumulation of irrelevant and redundant information, and minimize the impact of noise and irrelevant information.

[0089] 5) When processing graph-structured data such as brain functional networks, graph convolutional networks typically employ a layer-by-layer neighborhood aggregation mechanism, where the receptive field is limited by the number of network layers. Shallow structures struggle to capture long-range dependencies, while deep stacking can lead to oversmoothing, causing node features to become uniform and losing their discriminative ability. Furthermore, simple stacking cannot effectively model multi-scale spatiotemporal features, limiting its ability to express the dynamics of complex brain networks.

[0090] This application introduces a dilated convolution mechanism into graph convolution, expanding the receptive field by adjusting the dilation rate. This achieves information aggregation across multi-hop neighbors without increasing network depth or the number of parameters, effectively combining local and global functional connectivity patterns from different channels. This design captures more comprehensive multi-channel features while addressing the oversmoothing problem commonly found in graph convolutional networks. It enhances the model's ability to model collaborative activities in distant brain regions, improving classification and prediction performance.

[0091] Advantages: This model combines the expressive power of deep networks with training stability, can capture the dynamics of complex brain networks, achieve high-precision classification, support visual dysfunction screening and treatment response prediction, and has good prospects for translational applications.

[0092] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0093] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0094] Please see Figure 7 , Figure 7 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 701 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 702 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 702 and is called and executed by the processor 701 using the methods described in the embodiments of this application. The input / output interface 703 is used to implement information input and output; The communication interface 704 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 705 transmits information between various components of the device (e.g., processor 701, memory 702, input / output interface 703, and communication interface 704); The processor 701, memory 702, input / output interface 703, and communication interface 704 are connected to each other within the device via bus 705.

[0095] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0096] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0097] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0098] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0099] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented in the embodiments of this program product are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. The executable computer program code or "code" used to perform the various embodiments can be written in high-level programming languages ​​such as C, C++, Python, Smalltalk, Java, JavaScript, Visual Basic, Structured Query Language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0100] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0101] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0102] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0103] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0104] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0105] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0106] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0107] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0108] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0109] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0110] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for detecting visual impairment, characterized in that, The method includes the following steps: The subject's electroencephalogram (EEG) signals were acquired and preprocessed. Source spatial localization of preprocessed EEG signals was performed to reconstruct neural activity sequences in the source space of the cerebral cortex; Based on standard brain atlases, the source space is functionally partitioned, the cortical vertices are divided into multiple brain regions, and the time series of each brain region in a specific functional network are extracted. Based on the time series of each brain region in the specific functional network, the brain functional connectivity matrix is ​​calculated to construct the spatiotemporal feature tensor; The spatiotemporal feature tensor is input into a trained multi-scale controllable graph convolutional network model, and the detection or prediction results of visual impairment are output. The multi-scale controllable graph convolutional network model includes at least a graph feature extraction module. The graph feature extraction module uses Chebyshev-like polynomial connections and combines them with dual gating units to aggregate information of neighbor nodes of different orders in the graph structure.

2. The method according to claim 1, characterized in that, The functional partitioning of the source space based on standard brain atlas includes: Calculate the Euclidean distance from each neural source vertex in the source space to the centroid of each region of interest in the standard brain atlas; A dynamic k-nearest neighbor algorithm based on inverse distance weighting is used to adaptively assign each neural source vertex to a region of interest.

3. The method according to claim 1 or 2, characterized in that, The specific functional network includes a visual network and a default mode network; the time series is derived from the EEG signals. The signal extracted from the frequency band.

4. The method according to claim 3, characterized in that, The computation of the brain functional connectivity matrix includes: The time series is divided into multiple time segments using a sliding time window; Within each time segment, the Pearson correlation coefficients between each pair of brain regions in the visual network and the default mode network are calculated to form a dynamic functional connectivity matrix.

5. The method according to claim 1, characterized in that, In the graph feature extraction module, the output of the Chebyshev-like polynomial connection... Represented as: in, For activation function, The maximum order of graph convolution. For the improved Chebyshev join function, As input features, for The adjacency matrix corresponding to the order, for The weights corresponding to the order; And, when At that time, through the expression Establish an invariant mapping.

6. The method according to claim 1, characterized in that, The dual-gate control unit includes a first gate control unit and a second gate control unit; The first gating unit includes an update unit, a candidate hiding unit, and an output unit; the update unit adjusts the amount of information retained from previous neighboring orders and the degree to which new data from the current order is incorporated into the hidden state, and the calculation formula is as follows: in, This represents the output representation of the update unit at order t. The weights are represented as follows, and the outputs corresponding to orders t-1 and t are respectively expressed as and ; As input features, for The adjacency matrix corresponding to the order; For activation functions; Candidate hidden states serve as intermediate representations. Information is refreshed by combining the current input with the previous hidden state, and the calculation formula is as follows: in, Indicates the candidate hidden state. Indicates element-wise multiplication; The output unit controls the final hidden state output and coordinates the weighted interaction between the candidate hidden state and its previous state. The calculation formula is as follows: in, The output of the t-th order is represented by the output unit; The second gating unit is a forget gate, used to control the importance of each order of neighbor information in the final summation, and its output... Represented as: in, For activation function, As for the weight of the Forgotten Gate, This represents the result of each order of graph convolution.

7. The method according to claim 1, characterized in that, The multi-scale controllable graph convolutional network model also includes a multi-channel feature extractor; The multi-channel feature extractor is based on a residual network structure and incorporates a dilated convolution module to form a residual dilated module, which is used to extract multi-scale deep spatiotemporal features from the output of the graph feature extraction module.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images