Generative artificial intelligence-based brain data augmentation method and device

Abstract

The present invention relates to the field of brain functional data augmentation, and in particular, to a generative artificial intelligence-based brain data augmentation method and device. The method and device comprise: correcting missing channels in low-resolution electroencephalogram data to obtain coarse-grained high-resolution electroencephalogram data; taking the low-resolution electroencephalogram data as input to output corresponding temporal pattern encodings and spatial pattern encodings of the electroencephalogram data; and taking the low-resolution electroencephalogram data and the temporal pattern encodings and the spatial pattern encodings of the electroencephalogram data as input to output fine-grained high-resolution electroencephalogram data, which is then superimposed with the coarse-grained high-resolution electroencephalogram data to obtain reconstructed high-resolution electroencephalogram data. Compared with existing methods, the method of the present invention is capable of preserving and restoring key information and local details in the data, thereby generating more realistic and higher-quality high-spatial-resolution brain data.

Description

A brain data enhancement method and device based on generative artificial intelligence Technical Field

[0001] The present invention relates to the field of brain function data enhancement, and in particular to a method and device for brain data enhancement based on generative artificial intelligence. Background Art

[0002] Brain data refers to information related to brain structure and function obtained through neuroscience and brain imaging technologies. The collection and analysis of brain data is crucial for understanding brain workings, cognitive processes, and neurological diseases. However, data enhancement in four scenarios remains challenging: enhancing the spatial resolution of EEG and iEEG, cross-modal enhancement from EEG to iEEG, cross-modal enhancement from EEG to fMRI, and cross-modal enhancement from fNIRS to fMRI. These scenarios involve four types of brain data: EEG, iEEG, fMRI, and fNIRS.

[0003] EEG is a noninvasive technique for recording electrical activity in the brain. Electrodes are placed at specific locations on the scalp to record differences in electrical potential on the surface of the brain, generating corresponding multi-channel EEG signals. iEEG involves performing an invasive surgical procedure and directly implanting electrodes within the patient's brain to record electrical activity. fNIRS and fMRI are also two commonly used functional brain imaging techniques for studying brain function. fNIRS uses the absorption properties of near-infrared light on hemoglobin and oxyhemoglobin to estimate changes in cerebral blood oxygenation, indirectly reflecting brain activity. fNIRS has high temporal resolution (milliseconds) and low spatial resolution (centimeter levels), making it suitable for studying functional connectivity and blood oxygenation changes in localized brain regions. fMRI, on the other hand, detects changes in magnetic properties to locate and visualize changes in cerebral blood oxygenation, thereby revealing activity across entire brain regions. fMRI has high spatial resolution (millimeters) and low temporal resolution (seconds), making it suitable for studying functional activity across entire brain regions. For psychiatric conditions such as focal epilepsy and depression, which do not exhibit obvious structural abnormalities, structural imaging cannot pinpoint the location of the lesion, requiring high spatial resolution fMRI to infer the location of the lesion network.

[0004] Spatial resolution enhancement of EEG and iEEG involves converting low-spatial-resolution EEG (LR-EEG) and low-spatial-resolution iEEG (LR-iEEG) signals into corresponding high-spatial-resolution iEEG (HR-iEEG) and high-spatial-resolution iEEG (HR-iEEG) signals. Generally speaking, EEG and iEEG acquisition devices have a small number of electrodes, resulting in relatively low spatial resolution, which cannot provide comprehensive information on brain activity and cannot meet the requirements of existing technologies. However, high-throughput EEG acquisition equipment is expensive, has cumbersome operation steps, and is not portable, so enhancing the spatial resolution of EEG data is an effective method. Existing EEG spatial resolution enhancement technologies, such as Tang et al. (2022), proposed a graph convolutional network-based EEG data super-resolution framework that associates brain structure and functional connectivity to achieve enhanced high-spatial-resolution EEG data.

[0005] The cross-modal enhancement task from EEG to iEEG refers to the use of EEG signals to reconstruct iEEG signals. Both EEG signals and iEEG signals reflect brain electrical activity. Although EEG signals and iEEG signals differ in electrode positions, they have potential connections. Although EEG signals have low acquisition costs and a wide range of applications, they cannot obtain accurate brain activity information; although iEEG signals can provide more accurate and detailed brain activity information, data acquisition is limited by surgical costs and risks. Therefore, EEG signals can be used to achieve non-invasive acquisition of iEEG signals. Existing EEG to iEEG cross-modal enhancement technologies, such as Hu et al. (2022), use generative adversarial networks (GANs) as an EEG data enhancement framework, which performs cross-modal generation of stereotactic EEG (SEEG) data from EEG data.

[0006] Cross-modal enhancement from EEG to fMRI involves using EEG signals to reconstruct fMRI signals or enhance the temporal resolution of fMRI signals. Both EEG and fMRI signals reflect brain activity, and while the two activities take different forms, they share underlying characteristic connections. EEG signals offer advantages such as high temporal resolution, low acquisition cost, and limited usage. However, fMRI data acquisition is expensive, has low temporal resolution, and is subject to limitations such as requiring the subject to remain still and the absence of metal implants. Therefore, EEG signals are used to reconstruct and enhance fMRI signals. Current fMRI enhancement techniques, such as the machine learning-based Amygdala EFP model proposed by Keynan et al. (2016), utilize EEG signals to predict the amygdala BOLD signal. However, these techniques only reconstruct a subset of regions, and no technology has yet been developed to enhance the temporal resolution of fMRI data.

[0007] Cross-modal enhancement from fNIRS to fMRI involves using fNIRS signals to reconstruct fMRI signals or enhance the temporal resolution of fMRI signals. Both fNIRS and fMRI signals detect brain blood oxygenation levels and share similar hemodynamic origins. Compared to fMRI, fNIRS offers greater portability, cost advantages, temporal resolution (inferior to EEG), and reduced motion artifacts. It is also unrestricted for clinical populations and can be monitored while moving. Using fNIRS to reconstruct or enhance fMRI signals offers higher accuracy than using EEG, but at a higher cost. Current fMRI reconstruction techniques, such as those by Yan et al. (2020), utilize DCGANs to reconstruct localized BOLD signal loss, demonstrating that the reconstructed BOLD signal is temporally correlated with the original signal and retains the individual specificity of the BOLD signal. However, this only partially reconstructs the fMRI signal and cannot provide a method for acquiring fMRI data in patients who are unable to undergo fMRI.

[0008] Summary of the Invention

[0009] The embodiments of the present invention provide a method and device for brain data enhancement based on generative artificial intelligence, so as to at least solve the technical problem that the existing spatial super-resolution technology only relies on the information of adjacent channels without considering the spatial correlation of the data.

[0010] According to one embodiment of the present invention, a method for enhancing brain data based on generative artificial intelligence is provided, comprising the following steps:

[0011] S100: Correct the missing channels of the low-resolution EEG to obtain coarse-grained high-resolution EEG data;

[0012] S200: taking low-resolution EEG data as input and outputting corresponding EEG temporal pattern coding and EEG spatial pattern coding;

[0013] S300: Taking low-resolution EEG data and the EEG temporal pattern coding and EEG spatial pattern coding as input, outputting fine-grained high-resolution EEG data, and superimposing them with the coarse-grained high-resolution EEG data to obtain reconstructed high-resolution EEG data.

[0014] Furthermore, the step S100 specifically includes:

[0015] S101: Locate the missing channels in the low-resolution EEG data according to the preset EEG regional electrode distribution, and output a low-resolution EEG data missing channel matrix, denoted as S = [s ij ] c×c (i, j = 1, 2, ..., c), c represents the number of channels;

[0016] S102: Calculating the distance between each missing channel and its adjacent channels and the correlation between signals based on the low-resolution EEG data missing channel matrix and the three-dimensional EEG scalp electrode model, thereby determining the weight of each adjacent channel;

[0017] S103: Perform spatial inference operations on the missing EEG channel signals according to the weights of adjacent known electrodes to correct the missing channels of low-resolution EEG and obtain coarse-grained high-resolution EEG data.

[0018] Furthermore, the step S200 specifically includes:

[0019] S201: Based on the time-series-aware Transformer of multi-channel EEG data, the global time series representation of EEG data is performed. The features of each time point are compared and weighted with the features of other time points using the parallel self-attention mechanism to obtain the hidden layer representation of the global time series of EEG data and obtain the corresponding EEG time series pattern encoding;

[0020] S202: Learning the spatial relationship between brain regions based on a spatial pattern perception network of multi-channel EEG data, and characterizing the spatial relationship between channels in the EEG data to obtain a spatial pattern encoding of a specific subject in the EEG data signal.

[0021] Furthermore, step S300 specifically includes:

[0022] S301: Mapping the EEG temporal pattern code and the EEG spatial pattern code into the same feature space by using EEG feature alignment technology;

[0023] S302: Dynamically aggregate the aligned EEG temporal pattern codes and EEG spatial pattern codes using an adaptive weighting algorithm based on multi-dimensional features of EEG data;

[0024] S303: Using the EEG data hidden layer representation encoder to map the low-resolution EEG data into a latent space EEG data variable, then performing a noise operation on the latent space EEG data variable through a forward diffusion process, and converting it into a Gaussian distribution variable;

[0025] S304: Using a U-Net model guided by EEG spatiotemporal features for denoising, the EEG spatiotemporal fusion code is fed into the cross-attention layer of the U-Net model, guiding the model to gradually generate latent space variables with the characteristics of the subject's brain activity;

[0026] S305: The latent space variables generated by the model are passed through the EEG data hidden layer inference decoder to obtain fine-grained low-resolution EEG data, which are then superimposed with the obtained coarse-grained high-resolution EEG data through residual connection to obtain reconstructed high-resolution EEG data.

[0027] Furthermore, the EEG data is selected from one of EEG and iEEG.

[0028] Furthermore, the method further comprises the steps of:

[0029] S400: Construct a high-spatial-resolution and low-cost EEG data enhancement technology framework based on EEG spatiotemporal representation, and apply the framework to several EEG data reconstruction scenarios.

[0030] Furthermore, the method further comprises the steps of:

[0031] S500: EEG head-table mapping technology and a matching algorithm based on EEG spatial feature similarity are used to align the brain regions corresponding to the channels in EEG and iEEG data; the topological structure pattern and functional activity pattern of EEG are represented by a perception network based on EEG whole-brain topology structure and a perception network based on EEG whole-brain functional activity; the EEG topological structure features and EEG functional activity features are compressed and represented through a dual-encoder structure, and used as conditions to guide the EEG-based deep brain electrical activity signal reconstruction module to obtain reconstructed iEEG data.

[0032] Furthermore, the method further comprises the steps of:

[0033] S600: Extract EEG feature codes at different sampling rates from EEG data and input them as conditions to the BOLD signal inference module controlled by EEG conditions to obtain reconstructed fMRI data with the same sampling rate as the EEG feature codes; synthesize EEG data and high spatial resolution fMRI data pairs using the EEG feature codes as conditions; optimize the BOLD signal inference module controlled by EEG conditions during training by distinguishing between the reconstructed EEG-fMRI data pairs and the real EEG-fMRI data pairs, and strengthen the correspondence between the reconstructed fMRI data and the EEG data.

[0034] Furthermore, the method further comprises the steps of:

[0035] S700: resampling the fNIRS data according to a preset sampling rate, and then performing mapping from the fNIRS to the fMRI whole-brain BOLD signal modality mapping module to obtain an fMRI signal of the set resolution.

[0036] According to another embodiment of the present invention, a brain data enhancement device based on generative artificial intelligence is provided, comprising:

[0037] The low-resolution EEG missing channel spatial interpolation module is used to correct the missing channels of low-resolution EEG to obtain coarse-grained high-resolution EEG data;

[0038] The original EEG temporal-spatial pattern representation module is used to take low-resolution EEG data as input and output the corresponding EEG temporal pattern code and EEG spatial pattern code;

[0039] The EEG spatial super-resolution restoration module based on spatiotemporal condition guidance is used to take low-resolution EEG data and the EEG temporal pattern coding and EEG spatial pattern coding as input, output fine-grained high-resolution EEG data, and superimpose it with the coarse-grained high-resolution EEG data to obtain reconstructed high-resolution EEG data.

[0040] A storage medium storing a program file capable of implementing any of the above-mentioned brain data enhancement methods based on generative artificial intelligence.

[0041] A processor is used to run a program, wherein when the program is run, any one of the above-mentioned brain data enhancement methods based on generative artificial intelligence is executed.

[0042] The brain data enhancement method and device based on generative artificial intelligence in the embodiments of the present invention solve the problem that the existing spatial super-resolution technology only relies on the information of adjacent channels without considering the spatial correlation of the data by fully exploiting the spatiotemporal correlation of EEG data. It can also deeply fuse the temporal characteristics and spatial characteristics of EEG data, and use EEG fusion spatiotemporal coding as a conditional guidance module to reconstruct high-spatial-resolution EEG data. Compared with existing methods, the method of the present invention can retain and restore key information and local details of the data, generating more realistic and higher-quality high-spatial-resolution brain data. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] FIG1 is a flow chart of a method for enhancing brain data based on generative artificial intelligence according to the present invention;

[0044] FIG2 is a schematic diagram of a method for spatial super-resolution reconstruction of EEG data according to the present invention;

[0045] FIG3 is a schematic diagram of the cross-modal enhancement method from EEG to iEEG according to the present invention;

[0046] FIG4 is a modal conversion module integrating structural and functional features of the present invention;

[0047] FIG5 is a schematic diagram of the cross-modality enhancement method from EEG to fMRI according to the present invention;

[0048] FIG6 is a schematic diagram of EEG signal extraction with controllable sampling rate according to the present invention;

[0049] FIG7 is a detailed diagram of the implementation of cross-modality enhancement from EEG to fMRI according to the present invention;

[0050] FIG8 is a schematic diagram of the fNIRS to fMRI enhancement method of the present invention;

[0051] FIG9 is a schematic diagram of a controllable sampling rate fNIRS whole-brain signal conditional characterization synthesis module of the present invention;

[0052] FIG10 is a detailed diagram of the modal translation module integrating temporal feature learning according to the present invention. DETAILED DESCRIPTION

[0053] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0054] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0055] Currently, high spatial resolution non-invasive electroencephalogram (EEG) and invasive EEG (iEEG) acquisition equipment are expensive, difficult to operate, and not portable, while low spatial resolution signals can lead to large biases in decision-making tasks that require fine-grained spatial information.

[0056] Based on this, the present invention provides a brain data enhancement method based on generative artificial intelligence, referring to Figures 1 and 2, comprising the following steps:

[0057] S100: Correct the missing channels of the low-resolution EEG to obtain coarse-grained high-resolution EEG data;

[0058] S200: taking low-resolution EEG data as input and outputting corresponding EEG temporal pattern coding and EEG spatial pattern coding;

[0059] S300: Taking low-resolution EEG data and the EEG temporal pattern coding and EEG spatial pattern coding as input, outputting fine-grained high-resolution EEG data, and superimposing them with the coarse-grained high-resolution EEG data to obtain reconstructed high-resolution EEG data.

[0060] The present invention uses a high spatial resolution EEG data low-cost enhancement technology based on EEG spatiotemporal representation to generate high spatial resolution EEG data using low spatial resolution EEG data, thereby solving the problem of high cost of high-throughput EEG data acquisition.

[0061] Furthermore, the step S100 specifically includes:

[0062] S101: Locate the missing channels in the low-resolution EEG data according to the preset EEG regional electrode distribution, and output a low-resolution EEG data missing channel matrix, denoted as S = [s ij ] c×c (i, j = 1, 2, ..., c), c represents the number of channels;

[0063] S102: Calculating the distance between each missing channel and its adjacent channels and the correlation between signals based on the low-resolution EEG data missing channel matrix and the three-dimensional EEG scalp electrode model, thereby determining the weight of each adjacent channel;

[0064] S103: Perform spatial inference operations on the missing EEG channel signals according to the weights of adjacent known electrodes to correct the missing channels of low-resolution EEG and obtain coarse-grained high-resolution EEG data.

[0065] Furthermore, the step S200 specifically includes:

[0066] S201: Based on the time-series-aware Transformer of multi-channel EEG data, the global time series representation of EEG data is performed. The features of each time point are compared and weighted with the features of other time points using the parallel self-attention mechanism to obtain the hidden layer representation of the global time series of EEG data and the corresponding EEG time series pattern encoding, which implicitly captures the temporal correlation between EEG signals at different time points and the dynamic change characteristics of EEG signals over the entire time series.

[0067] S202: A spatial pattern perception network based on multi-channel EEG data is used to learn the spatial relationship between brain regions, and the spatial relationship between channels in the EEG data is characterized to obtain the spatial pattern encoding of a specific subject in the EEG data signal, thereby reflecting the connection strength between different brain regions.

[0068] Furthermore, step S300 specifically includes:

[0069] S301: Mapping the EEG temporal pattern code and the EEG spatial pattern code into the same feature space by using EEG feature alignment technology;

[0070] S302: Using the adaptive weighting algorithm based on the multi-dimensional features of EEG data to dynamically aggregate the aligned EEG temporal pattern coding and EEG spatial pattern coding to improve the effect of EEG spatiotemporal feature fusion. During the fusion process, the adaptive weighting algorithm based on the multi-dimensional features of EEG aggregates the EEG temporal pattern coding EEG by learning adaptive weights. t and EEG spatial pattern coding s Add the weights to get the EEG spatiotemporal fusion code E c , the calculation process can be expressed as: E c =W t *E t +W s *E s

[0071] Among them, W t represents the adaptive weight of EEG temporal pattern encoding, W s represents the adaptive weights of EEG spatial pattern encoding;

[0072] S303: Using the EEG data hidden layer representation encoder to map the low-resolution EEG data into a latent space EEG data variable, then performing a noise operation on the latent space EEG data variable through a forward diffusion process, and converting it into a Gaussian distribution variable;

[0073] S304: Using a U-Net model guided by EEG spatiotemporal features for denoising, the EEG spatiotemporal fusion code is fed into the cross-attention layer of the U-Net model, guiding the model to gradually generate latent space variables with the characteristics of the subject's brain activity;

[0074] S305: The latent space variables generated by the model are passed through the EEG data hidden layer inference decoder to obtain fine-grained low-resolution EEG data, which are then superimposed with the obtained coarse-grained high-resolution EEG data through residual connection to obtain reconstructed high-resolution EEG data, completing the spatial super-resolution enhancement of the EEG data.

[0075] Furthermore, the EEG data is selected from one of EEG and iEEG.

[0076] Furthermore, the method further comprises the steps of:

[0077] S400: Construct a high-spatial-resolution and low-cost EEG data enhancement technology framework based on EEG spatiotemporal representation, and apply the framework to several EEG data reconstruction scenarios.

[0078] Specifically, in order to facilitate the explanation of the training process of this technical framework, the enhancement of EEG spatial resolution is described as an example.

[0079] The training of the framework includes the steps:

[0080] Step A100: The present invention synchronously acquires paired HR-EEG and LR-EEG data using a high-throughput / low-throughput non-invasive EEG synchronous acquisition device, divides the data into a training set and a validation set in proportion, and performs basic parameter setting for model training;

[0081] Step A200: The LR-EEG data in the training set are preprocessed accordingly and then fed into the low-resolution EEG missing channel correction module. The missing channels of the LR-EEG are spatially located and corrected according to the pre-set EEG brain region electrode distribution to obtain reconstructed coarse-grained HR-EEG data.

[0082] Step A300: The preprocessed LR-EEG data is fed into the original EEG temporal-spatial pattern representation module. The temporal dynamic change pattern of the EEG and the spatial distribution pattern of the brain regions are learned through the multi-channel EEG-based temporal perception Transformer and the multi-channel EEG-based spatial pattern perception network, respectively, to obtain the EEG temporal pattern encoding and the EEG spatial temporal pattern encoding;

[0083] Step A400: The EEG temporal pattern encoding, EEG spatial temporal pattern encoding and pre-processed LR-EEG data are fed into the EEG spatial super-resolution restoration module guided by spatiotemporal conditions, and the HR-EEG data are reconstructed by the diffusion learning strategy guided by the EEG spatiotemporal conditions; the spatiotemporal perception loss between the HR-EEG data reconstructed by this module and the corresponding HR-EEG data in the training set is calculated. Reconstructing EEG distribution loss with high spatial resolution To form the model loss:

[0084] right Perform backpropagation to guide the original EEG temporal-spatial pattern representation module and the EEG spatial super-resolution restoration module guided by spatiotemporal conditions to update parameters;

[0085] Step A500: Repeat steps A200 to A400 to iteratively train the model on the training set. During each iteration, the model is evaluated and verified using the validation set. After the iteration, the optimal spatiotemporal-guided EEG spatial super-resolution restoration module is selected based on the verification results.

[0086] After model training is completed, the high-spatial-resolution, low-cost EEG data enhancement technology proposed in this paper based on EEG spatiotemporal representation will be applied to two scenarios: high-spatial-resolution non-invasive EEG data reconstruction and high-spatial-resolution invasive EEG data reconstruction. The specific implementation steps are as follows:

[0087] Step B100: For subjects whose HR-EEG data cannot be collected using a high-throughput non-invasive EEG acquisition device, a low-throughput non-invasive EEG acquisition device is used to acquire LR-EEG data; for subjects whose HR-iEEG data cannot be collected using a high-throughput invasive EEG acquisition device, a low-throughput invasive EEG acquisition device is used to acquire the subject's LR-iEEG data, and corresponding preprocessing operations are performed at the same time;

[0088] Step B200: The acquired LR-EEG data and LR-iEEG data are sent to the missing channel correction module of low spatial resolution EEG, and the missing EEG channel signals are spatially inferred by adjusting the preset brain area electrode distribution to obtain the corresponding coarse-grained HR-EEG and coarse-grained HR-iEEG data.

[0089] Step B300: The acquired LR-EEG data and LR-iEEG data are sent to the original EEG temporal-spatial pattern representation module, and the EEG temporal pattern coding and EEG spatial pattern coding of the corresponding EEG data are obtained through the multi-channel EEG-based temporal perception Transformer and the multi-channel EEG-based spatial pattern perception network.

[0090] Step B400: The extracted EEG temporal pattern codes and EEG spatial pattern codes are sent to the trained EEG spatial super-resolution restoration module based on spatiotemporal conditions. The obtained EEG temporal pattern codes and EEG spatial pattern codes are first projected into the same feature space through EEG feature alignment technology, and then the aligned temporal features and spatial features are deeply aggregated using an adaptive weighting algorithm based on EEG multi-dimensional features to obtain EEG spatiotemporal fusion codes.

[0091] Step B500: Fix the parameters of the spatiotemporal-condition-guided EEG spatial super-resolution restoration module, use the EEG hidden layer representation encoder to map the LR-EEG and LR-iEEG data into the latent space, and then generate Gaussian distributed latent variables by adding noise through the forward diffusion process; use the obtained EEG spatiotemporal fusion encoding as a condition to guide the U-Net model based on the EEG spatiotemporal feature guidance to perform the reverse denoising process, and gradually generate latent space variables that have the characteristics of the subject's brain activity and are similar to the feature distribution of the high spatial resolution EEG data;

[0092] Step B600: Similar to step B500, the latent space variables generated by the model pass through the hidden layer inference decoder to obtain fine-grained HR-EEG and fine-grained HR-iEEG data, which are then superimposed with the obtained coarse-grained HR-EEG and coarse-grained HR-iEEG data through residual connection to obtain model-reconstructed HR-EEG and HR-iEEG data.

[0093] Furthermore, existing iEEG acquisition technology is invasive and requires the implantation of electrodes in the subject's brain tissue, which can cause discomfort to the patient and may even cause permanent damage. However, iEEG data can provide extremely accurate and detailed information on brain electrical activity, which is an effective tool for analyzing connectivity and activity between brain regions. Based on this, as shown in Figure 3, the method also includes the following steps:

[0094] S500: EEG head-table mapping technology and a matching algorithm based on EEG spatial feature similarity are used to align the brain regions corresponding to the channels in EEG and iEEG data; the topological structure pattern and functional activity pattern of EEG are represented by a perception network based on EEG whole-brain topology structure and a perception network based on EEG whole-brain functional activity; the EEG topological structure features and EEG functional activity features are compressed and represented through a dual-encoder structure, and used as conditions to guide the EEG-based deep brain electrical activity signal reconstruction module to obtain reconstructed iEEG data.

[0095] The present invention proposes a non-invasive cross-modal mapping technology of iEEG signals based on EEG structural and functional characteristics, which generates iEEG data through the patient's EEG data, thereby achieving non-invasive acquisition of the patient's iEEG data, thereby solving the problems of high cost of iEEG acquisition equipment and high risk of invasive surgery.

[0096] Further, referring to FIG4 , the step S500 specifically includes the following steps:

[0097] Step C100: using head table mapping technology to visualize the electrode arrangement of the EEG data and the pre-set iEEG intracranial electrode distribution, thereby determining the brain area corresponding to each electrode position;

[0098] Step C200: Mapping the spatial positions of corresponding electrodes in the EEG and iEEG data to the same reference coordinate system using a matching algorithm based on EEG spatial feature similarity;

[0099] Step C300: aligning the electrode positions representing the same brain region, establishing a mapping relationship between EEG and iEEG in terms of spatial brain region distribution, and obtaining spatial channel-aligned EEG data;

[0100] Step C400: This module uses a perceptual network based on the EEG whole-brain topology to characterize the structural pattern of the topological map composed of each EEG channel, capturing the spatial correlation within the EEG data channel, thereby effectively extracting the topological structural features of the EEG data;

[0101] Step C500: Use the Pearson coefficient and mutual information to measure the functional connectivity between the brain regions corresponding to the EEG data channels and obtain the corresponding functional connectivity matrix;

[0102] Step C600: This module characterizes the functional connectivity patterns between brain regions corresponding to different channels in the functional connectivity matrix through a perceptual network based on EEG whole-brain functional activity, thereby obtaining corresponding EEG functional activity characteristics;

[0103] Step C700: Using the EEG-fused topological conditional encoder, the obtained EEG topological structure features are mapped into the latent space to obtain the latent variables of the implicit brain region topological structure connectivity features, and the latent variables are used as conditions to input into the EEG-based deep brain electrical activity signal reconstruction module;

[0104] Step C800: Mapping the obtained EEG functional activity features into the latent space using the EEG-fused functional activity condition encoder to obtain latent variables that imply the functional activity pattern characteristics of the brain region, and inputting them as conditions into the EEG-based intracranial deep brain electrical activity signal reconstruction module;

[0105] Step C900: Input the spatial channel-registered EEG data into the EEG-based deep brain electrical activity signal reconstruction module. This submodule uses a score learning strategy to reconstruct the deep brain electrical activity signal. First, different levels of noise are used to perturb the original data distribution and the scores corresponding to different levels of noise are obtained through denoising score matching. Then, a conditional score network based on EEG structure-function characteristics is used to predict the scores corresponding to all noise levels.

[0106] Step C1000: Using the latent variables output by the fused EEG topological structure conditional encoder and the fused EEG functional activity conditional encoder as conditions to guide the EEG structure-function feature-guided conditional score network to gradually reduce the noise level and obtain reconstructed iEEG data.

[0107] Furthermore, the training of the non-invasive cross-modal mapping technology framework of iEEG signals based on EEG structural and functional characteristics includes the following steps:

[0108] Step D100: The present invention uses a non-invasive EEG-invasive EEG synchronous acquisition device to synchronously acquire paired EEG and iEEG data from several subjects, and divides the data into a training set and a validation set in proportion to set basic parameters for model training;

[0109] Step D200: After the EEG data undergoes corresponding preprocessing operations, it is sent to the brain region channel registration module based on EEG head table mapping, and the EEG and iEEG brain region electrodes are registered according to the predefined iEEG intracranial electrode distribution to obtain spatial channel registered EEG data;

[0110] Step D300: sending the pre-processed EEG data to the EEG-based whole-brain structure and function integrated pattern representation module to obtain the corresponding EEG topological structure features and EEG functional activity features;

[0111] Step D400: EEG topological structure features, EEG functional activity features, and spatial channel registration EEG data are fed into a modality transfer module that converts EEG signals into high spatial resolution iEEG signals. This module reconstructs intracranial deep EEG signals, i.e., iEEG data, using a fractional learning strategy guided by EEG structural and functional feature conditions.

[0112] Step D500: Reconstruction loss between iEEG data reconstructed by the modal transfer module from EEG signal to high spatial resolution iEEG signal and input iEEG data, reconstruction loss of intracranial deep brain electrical signal and distribution loss based on KL divergence To form the model loss; and Perform backpropagation to guide the parameter update of the EEG-based whole-brain structure and function integrated pattern representation module and the modality transfer module from EEG signals to high spatial resolution iEEG signals;

[0113] Step D600: Repeat steps D200 to D500 to iteratively train the model on the training set. During each iteration, the model is validated using the validation set. After the iteration, the optimal modality transfer module for EEG signals to high-spatial-resolution iEEG signals is selected based on the validation results.

[0114] After the model training is completed, the non-invasive cross-modal mapping technology of iEEG signals based on EEG structural and functional characteristics proposed in this invention is applied to the reconstruction of intracranial deep EEG signals with high spatial resolution. The specific implementation process is as follows:

[0115] Step E100: For subjects whose iEEG data cannot be obtained using invasive EEG acquisition equipment, non-invasive EEG acquisition equipment is used to collect the EEG data of the subject and perform corresponding preprocessing operations;

[0116] The collected EEG data is input into the EEG structure-function feature extraction module to obtain the corresponding structural features and functional features;

[0117] Step E200: sending the pre-processed EEG data to a brain region channel registration module based on EEG head table mapping, and obtaining spatial channel registration EEG data by adjusting the pre-set iEEG intracranial electrode distribution;

[0118] Step E300: sending the pre-processed EEG data to the EEG-based whole-brain structure and function integrated pattern representation module to obtain the corresponding EEG topological structure features and EEG functional activity features;

[0119] Step E400: The characterized EEG topological structure features and EEG functional activity features are sent to the optimal EEG signal to high spatial resolution iEEG signal modality transfer module to reconstruct a high spatial resolution intracranial deep brain signal.

[0120] Furthermore, existing functional magnetic resonance imaging (fMRI) data acquisition technology has limitations such as low temporal resolution, high equipment cost, and the need for the subject to remain still during use. Based on this, the method further includes the steps of:

[0121] S600: Extract EEG feature codes at different sampling rates from EEG data and input them as conditions to the BOLD signal inference module controlled by EEG conditions to obtain reconstructed fMRI data with the same sampling rate as the EEG feature codes; synthesize EEG data and high spatial resolution fMRI data pairs using the EEG feature codes as conditions; optimize the BOLD signal inference module controlled by EEG conditions during training by distinguishing between the reconstructed EEG-fMRI data pairs and the real EEG-fMRI data pairs, and strengthen the correspondence between the reconstructed fMRI data and the EEG data.

[0122] Using EEG-based cross-modal enhancement of fMRI signal temporal and spatial resolution, fMRI data can be generated from EEG data. This technology enhances the temporal and spatial resolution of fMRI data through a controllable sampling rate EEG feature representation module and an EEG-conditioned BOLD signal inference module. This allows for the acquisition of difficult-to-obtain fMRI data using lower-cost EEG, while also addressing the issues of fMRI's low temporal resolution and the inability of certain populations to undergo fMRI examinations.

[0123] Further, referring to FIG. 5-7 , step S600 specifically includes:

[0124] Step F100: Segment the input EEG data by adjusting the window position, according to the sampling rate fs of the input EEG signal. EEG The ratio of n to the set sampling rate fs fs Extract EEG signals for intervals;

[0125] Step F200: performing microstate extraction on the extracted EEG signal to obtain a microstate sequence of the EEG signal;

[0126] Step F300: Input the EEG signal microstate sequence obtained in step F200 into the EEG timing initialization characterization module to obtain the EEG feature code with a sampling rate of fs, which serves as the EEG condition input of the BOLD signal inference module of EEG condition control;

[0127] Step F400: taking the latent vector with Gaussian distribution and the EEG temporal feature encoding condition as input, adding them together to obtain the BOLD signal latent spatial encoding feature of the fused EEG condition;

[0128] Step F500: extracting the temporal and spatial features of the BOLD signal latent spatial coding using a BOLD latent spatial coding spatiotemporal feature upsampling network composed of LSTM and convolutional layers and performing dimension expansion on the BOLD signal latent spatial coding, and merging the dimensionally expanded BOLD signal latent spatial coding features with the EEG temporal feature coding to perform feature fusion;

[0129] Step F600: Repeat step F500 to obtain a BOLD signal latent space code of the same dimension as the cortical BOLD feature code, merge it with the EEG temporal feature code, and pass it through the EEG-BOLD signal output layer to obtain a reconstructed cortical BOLD signal feature code and a reconstructed EEG temporal feature code;

[0130] Step F700: Combining the EEG temporal feature code and the cortical BOLD feature code is input into the EEG-BOLD combined feature discrimination module.

[0131] Step F800: The EEG-BOLD combined feature is input into an EEG-BOLD spatiotemporal feature extraction network composed of LSTM and residual convolution blocks to obtain the temporal and spatial features of the EEG-BOLD combined feature.

[0132] Step F900: The temporal and spatial features of the EEG-BOLD combined feature are fused, and the result of the identification of whether the input EEG-BOLD combined feature is real data or data generated by the BOLD signal inference module controlled by EEG conditions is obtained through the EEG-BOLD signal identification output network.

[0133] Furthermore, the training of the EEG-based fMRI signal spatiotemporal resolution cross-modal enhancement technology framework includes the following steps:

[0134] Step G100: Use the EEG-fMRI data acquisition system to synchronously acquire paired EEG and BOLD signals from the subject. The acquired signal pairs are divided into training and validation sets proportionally. Basic parameters are set, including the number of training sessions.

[0135] Step G200: Input the training set EEG data into a controllable sampling rate EEG feature representation module, and set the module's sampling rate to the same sampling rate as the training set fMRI data. Obtain the EEG feature encoding output by this module. Simultaneously, preprocess the training set fMRI data by spatially registering the BOLD signal acquired from the fMRI data with a standard cortical template to obtain a high-spatial-resolution cortical BOLD signal feature encoding.

[0136] Step G300: The EEG temporal feature code obtained in step G200 is combined with the Gaussian hidden vector as a condition and input into the EEG condition-controlled BOLD signal inference module to obtain the reconstructed cortical BOLD signal feature code and the reconstructed EEG temporal feature code.

[0137] Step G400: Input the reconstructed cortical BOLD signal feature code and the reconstructed EEG timing feature code into the EEG-BOLD combined feature identification module to obtain an identification result of whether the input EEG-BOLD combined feature is real data or data generated by the BOLD signal inference module controlled by EEG conditions.

[0138] Step G500: The cross entropy loss between the true and false labels output by the EEG-BOLD combined feature identification module and the actual situation (i.e., whether the input EEG-BOLD combined feature comes from the training set or the BOLD signal inference module controlled by the EEG condition) is calculated. and the EEG temporal feature reconstruction loss between the reconstructed EEG temporal feature encoding generated by the BOLD signal inference module controlled by the EEG condition and the corresponding EEG temporal feature encoding in the training set BOLD temporal reconstruction loss of reconstructed cortical BOLD signal feature encoding and input cortical BOLD signal feature encoding The total loss of the BOLD signal inference module for EEG conditional control is:

[0139] Update the parameters of the BOLD signal inference module controlled by EEG conditions.

[0140] Step G600: Based on the cross entropy loss between the true and false labels output by the EEG-BOLD combined feature identification module and the actual situation (i.e., whether the input EEG-BOLD combined feature comes from the training set or the BOLD signal inference module controlled by the EEG condition) The parameters of the BOLD signal inference module controlled by EEG conditions are fixed, and the parameters of the EEG-BOLD combined feature identification module are updated.

[0141] Step G700: The EEG-based fMRI signal spatiotemporal resolution cross-modal enhancement framework is trained using a training set of simultaneously acquired EEG and fMRI signals for the number of training cycles in step G100. During each iteration, the framework is validated using a validation set of simultaneously acquired EEG and fMRI signals. After the iteration, the optimal BOLD signal inference module controlled by the EEG condition is selected based on the validation results.

[0142] After model training is completed, the specific application process includes:

[0143] Step H100: Use an EEG device to collect the EEG data of the subject and input it into the controllable sampling rate EEG feature characterization module to obtain the EEG time series feature code. Set the sampling rate fs lower than the sampling rate of the subject's EEG data as required to obtain the EEG time series feature code with the sampling rate fs;

[0144] Step H200: The EEG temporal feature code with a sampling rate of fs is combined with the Gaussian hidden vector as a condition and inputted into the BOLD signal inference module controlled by the optimal EEG condition selected in step G700 to obtain the reconstructed cortical BOLD signal feature code with a sampling rate of fs and the reconstructed EEG temporal feature code;

[0145] Step H300: Fix the parameters of the BOLD signal inference module controlled by the EEG condition, and reconstruct the EEG signal timing feature signal and the EEG timing feature signal obtained in step H100. To update the Gaussian hidden vector in step H200;

[0146] Step H400: The Gaussian hidden vector after several iterations and the EEG temporal feature code with a sampling rate of fs are input into the EEG conditional BOLD signal inference module to obtain a high spatial resolution BOLD signal feature code with a sampling rate of fs. Finally, this code is converted into a high spatial resolution cortical BOLD signal with a sampling rate of fs using the standard cortical template in step G200.

[0147] Furthermore, acquisition of fMRI signals has limitations such as high cost and the need for the subject to remain still. Based on this, the method further includes the steps of:

[0148] S700: resampling the fNIRS data according to a preset sampling rate, and then performing mapping from the fNIRS to the fMRI whole-brain BOLD signal modality mapping module to obtain an fMRI signal of the set resolution.

[0149] Both functional near-infrared spectroscopy (fNIRS) and fMRI detect time-series data of brain blood oxygen activity. fNIRS has many advantages, such as portability, the ability to move during detection, low cost, and fewer user restrictions. Using high-temporal and spatial resolution BOLD signal mapping from fNIRS to fMRI, it is possible to generate corresponding fMRI data from fNIRS data, thereby resolving issues such as the difficulty of fMRI acquisition and low temporal resolution.

[0150] Further, referring to FIG. 8 to FIG. 10 , step S700 specifically includes:

[0151] Step K100: During the training process of the high-temporal-resolution fNIRS to fMRI BOLD signal mapping technology framework (steps G100-G700), the simultaneously acquired paired fNIRS and fMRI data are input into the controllable sampling rate fNIRS whole-brain signal conditional representation synthesis module, and the sampling rate fs is set to the sampling rate of the fMRI data in the data pair. During the high-temporal-resolution cortical BOLD signal reconstruction process (steps H100-H400), the subject's fNIRS data is input into the controllable sampling rate fNIRS whole-brain signal conditional representation synthesis module, and the sampling rate fs is set to meet the requirements but less than the sampling rate of the fNIRS data.

[0152] Step K200: The controllable sampling rate fNIRS whole-brain signal conditional representation synthesis module synthesizes the regional spatial locations of fNIRS-detected cortical activity into an fNIRS whole-brain BOLD signal representation under the condition of time alignment, and its data dimension is the same as the cortical BOLD signal extracted from fMRI;

[0153] Step K300: extracting the time points corresponding to the whole-brain cortical activity features according to the set sampling rate fs, and outputting the resampled fNIRS whole-brain BOLD signal representation;

[0154] Step K400: During the training process of the fNIRS to fMRI high-temporal-resolution BOLD signal mapping technology framework (steps G100-G700), the fNIRS whole-brain BOLD signal representation and the cortical BOLD signal extracted from fMRI are input into the fNIRS to fMRI whole-brain BOLD signal modality mapping module. During the high-temporal-resolution cortical BOLD signal reconstruction process (steps H100-H400), only the fNIRS whole-brain BOLD signal representation is input;

[0155] Step K500: During the training process of the fNIRS to fMRI high-temporal-resolution BOLD signal mapping technology framework (steps G100-G700), the fMRI cortical BOLD signal is input into the fMRI cortical BOLD signal conditional representation module to obtain the fMRI cortical BOLD signal conditional representation. Start and end markers are added to the beginning and end of the fMRI cortical BOLD signal conditional representation sequence;

[0156] Step K600: During the training process of the fNIRS to fMRI high-temporal-resolution BOLD signal mapping technology framework (steps G100-G700), the fMRI cortical BOLD signal representation and the fNIRS whole-brain BOLD signal representation obtained in step K500 are input into the BOLD signal representation modality mapping module to obtain the cortical BOLD signal hidden feature encoding. During the high-temporal-resolution cortical BOLD signal reconstruction process (steps H100-H400), the fNIRS whole-brain BOLD signal and the onset marker are input into the BOLD signal representation modality mapping module to iteratively obtain the cortical BOLD signal hidden feature encoding.

[0157] Step K700: Input the hidden feature encoding of the cortical BOLD signal obtained in step K600 into the fMRI cortical BOLD signal inference and restoration module to obtain the reconstructed fMRI cortical BOLD signal as the output of the fNIRS to fMRI whole-brain BOLD signal modality mapping module.

[0158] Furthermore, the training of the high spatiotemporal resolution BOLD signal mapping technology framework from fNIRS to fMRI includes the following steps:

[0159] Step M100: Using the fNIRS-fMRI data synchronization acquisition system, synchronously acquire paired fNIRS and BOLD signals from the subject, divide the acquired signal pairs into a training set and a validation set in proportion, and perform basic parameter settings (such as setting the number of training times);

[0160] Step M200: Preprocess the fMRI data of the training set and spatially register the BOLD signals collected from the fMRI data with the standard cortical template to obtain a high-spatial-resolution fMRI cortical BOLD signal. Execute steps K100-K700 to obtain the reconstructed fMRI cortical BOLD signal.

[0161] Step M300: Based on the reconstruction loss L between the reconstructed fMRI cortical BOLD signal in step M200 and the preprocessed high spatial resolution fMRI cortical BOLD signal rec and KL divergence loss L KL The overall loss function of the fNIRS to fMRI whole-brain BOLD signal modality mapping module is: L = Lrec +L KL

[0162] Based on this, update the parameters of the fNIRS to fMRI whole-brain BOLD signal modality mapping module;

[0163] Step M400: Train the model using the training set for the same number of training cycles as in Step M100. Validate the model using the fNIRS-fMRI validation set during each iteration. After the iterations are complete, select the optimal fNIRS-to-fMRI whole-brain BOLD signal mapping module based on the validation results.

[0164] After training is completed, the specific application process includes:

[0165] Step N100: Use the fNIRS device to collect the subject's fNIRS data, and input it into the controllable sampling rate fNIRS whole-brain signal conditional representation synthesis module according to steps K100-K300 to obtain the fNIRS whole-brain BOLD signal representation with a sampling rate of fs;

[0166] Step N200: Execute steps K400-K700, input the fNIRS whole-brain BOLD signal representation with a sampling rate of fs into the optimal fNIRS to fMRI whole-brain BOLD signal modality mapping module selected in step M400, and obtain the fMRI cortical BOLD signal with a sampling rate of fs.

[0167] According to another embodiment of the present invention, a brain data enhancement device based on generative artificial intelligence is provided, comprising:

[0168] The low-resolution EEG missing channel spatial interpolation module is used to correct the missing channels of low-resolution EEG to obtain coarse-grained high-resolution EEG data;

[0169] The original EEG temporal-spatial pattern representation module is used to take low-resolution EEG data as input and output the corresponding EEG temporal pattern code and EEG spatial pattern code;

[0170] The EEG spatial super-resolution restoration module based on spatiotemporal condition guidance is used to take low-resolution EEG data and the EEG temporal pattern coding and EEG spatial pattern coding as input, output fine-grained high-resolution EEG data, and superimpose it with the coarse-grained high-resolution EEG data to obtain reconstructed high-resolution EEG data.

[0171] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the system embodiments described above are only exemplary. For example, the division of units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0172] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected to achieve the purpose of the present embodiment according to actual needs.

[0173] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0174] If the integrated unit is implemented in the form of 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 the present invention, 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. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc. Various media that can store program codes.

[0175] In summary, the present invention utilizes a generative AI brain data enhancement method to solve a series of technical problems in the acquisition and processing of brain time series data. By innovatively adopting spatial super-resolution reconstruction technology and cross-modal enhancement technology, the present invention can convert low spatial resolution EEG data into higher resolution data, thereby achieving non-invasive acquisition of iEEG and fMRI data. At the same time, the present invention also achieves temporal resolution enhancement of fMRI data and generates corresponding fMRI data using fNIRS data, thereby significantly improving the quality and accuracy of the data while reducing costs and alleviating equipment dependence. In general, this invention provides a new perspective to understand and solve problems in the acquisition and processing of brain time series data by introducing generative AI technology, bringing new possibilities to neuroscience and clinical diagnosis.

[0176] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A brain data enhancement method based on generative artificial intelligence, characterized in that: The following steps are involved: S100: Correct the missing channels of the low-resolution EEG to obtain coarse-grained high-resolution EEG data; S200: taking low-resolution EEG data as input and outputting corresponding EEG temporal pattern coding and EEG spatial pattern coding; S300: Taking low-resolution EEG data and the EEG temporal pattern coding and EEG spatial pattern coding as input, outputting fine-grained high-resolution EEG data, and superimposing them with the coarse-grained high-resolution EEG data to obtain reconstructed high-resolution EEG data.

2. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: The step S100 specifically includes: S101: Locate the missing channels in the low-resolution EEG data according to the preset EEG regional electrode distribution, and output a low-resolution EEG data missing channel matrix, denoted as S = [s ij ] c×c (i, j = 1, 2, ..., c), c represents the number of channels; S102: Calculating the distance between each missing channel and its adjacent channels and the correlation between signals based on the low-resolution EEG data missing channel matrix and the three-dimensional EEG scalp electrode model, thereby determining the weight of each adjacent channel; S103: Perform spatial inference operations on the missing EEG channel signals according to the weights of adjacent known electrodes to correct the missing channels of low-resolution EEG and obtain coarse-grained high-resolution EEG data.

3. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: The step S200 specifically includes: S201: Based on the time-series-aware Transformer of multi-channel EEG data, the global time series representation of EEG data is performed. The features of each time point are compared and weighted with the features of other time points using the parallel self-attention mechanism to obtain the hidden layer representation of the global time series of EEG data and obtain the corresponding EEG time series pattern encoding; S202: Learning the spatial relationship between brain regions based on a spatial pattern perception network of multi-channel EEG data, and characterizing the spatial relationship between channels in the EEG data to obtain a spatial pattern encoding of a specific subject in the EEG data signal.

4. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: Step S300 specifically includes: S301: Mapping the EEG temporal pattern code and the EEG spatial pattern code into the same feature space by using EEG feature alignment technology; S302: Dynamically aggregate the aligned EEG temporal pattern codes and EEG spatial pattern codes using an adaptive weighting algorithm based on multi-dimensional features of EEG data; S303: Using the EEG data hidden layer representation encoder to map the low-resolution EEG data into a latent space EEG data variable, then performing a noise operation on the latent space EEG data variable through a forward diffusion process, and converting it into a Gaussian distribution variable; S304: Using a U-Net model guided by EEG spatiotemporal features for denoising, the EEG spatiotemporal fusion code is fed into the cross-attention layer of the U-Net model, guiding the model to gradually generate latent space variables with the characteristics of the subject's brain activity; S305: The latent space variables generated by the model are passed through the EEG data hidden layer inference decoder to obtain fine-grained low-resolution EEG data, which are then superimposed with the obtained coarse-grained high-resolution EEG data through residual connection to obtain reconstructed high-resolution EEG data.

5. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: The EEG data is selected from one of EEG and iEEG.

6. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: Also includes the steps: S400: Construct a high-spatial-resolution and low-cost EEG data enhancement technology framework based on EEG spatiotemporal representation, and apply the framework to several EEG data reconstruction scenarios.

7. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: Also includes the steps: S500: EEG head-table mapping technology and a matching algorithm based on EEG spatial feature similarity are used to align the brain regions corresponding to the channels in EEG and iEEG data; the topological structure pattern and functional activity pattern of EEG are represented by a perception network based on EEG whole-brain topology structure and a perception network based on EEG whole-brain functional activity; the EEG topological structure features and EEG functional activity features are compressed and represented through a dual-encoder structure, and used as conditions to guide the EEG-based deep brain electrical activity signal reconstruction module to obtain reconstructed iEEG data.

8. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: Also includes the steps: S600: Extract EEG feature codes at different sampling rates from EEG data and input them as conditions to the BOLD signal inference module controlled by EEG conditions to obtain reconstructed fMRI data with the same sampling rate as the EEG feature codes; synthesize EEG data and high spatial resolution fMRI data pairs using the EEG feature codes as conditions; optimize the BOLD signal inference module controlled by EEG conditions during training by distinguishing between the reconstructed EEG-fMRI data pairs and the real EEG-fMRI data pairs, and strengthen the correspondence between the reconstructed fMRI data and the EEG data.

9. The method for brain data enhancement based on generative artificial intelligence according to claim 1, characterized in that: Also includes the steps: S700: resampling the fNIRS data according to a preset sampling rate, and then performing mapping from the fNIRS to the fMRI whole-brain BOLD signal modality mapping module to obtain an fMRI signal of the set resolution.

10. A brain data enhancement device based on generative artificial intelligence, characterized in that: include: The low-resolution EEG missing channel spatial interpolation module is used to correct the missing channels of low-resolution EEG to obtain coarse-grained high-resolution EEG data; The original EEG temporal-spatial pattern representation module is used to take low-resolution EEG data as input and output the corresponding EEG temporal pattern code and EEG spatial pattern code; The EEG spatial super-resolution restoration module based on spatiotemporal condition guidance is used to take low-resolution EEG data and the EEG temporal pattern coding and EEG spatial pattern coding as input, output fine-grained high-resolution EEG data, and superimpose it with the coarse-grained high-resolution EEG data to obtain reconstructed high-resolution EEG data.

Images