Motion-related electroencephalogram signal source positioning method and system based on MNE-GAN
By constructing a generative adversarial network based on MNE-GAN and combining it with the minimum norm inverse solution, the problem of low spatial resolution of EEG signals is solved, more accurate and stable EEG source positioning is achieved, and dynamic neural processes can be tracked in real time.
Patent Information
- Application Number
- CN202510744870.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-05
- Publication Date
- 2025-09-19
- Estimated Expiration
- 2045-06-05
Smart Images

Figure CN120661157A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of brain source localization technology, and in particular to a method and system for localizing motion-related EEG signal sources based on MNE-GAN. Background Art
[0002] Electroencephalogram (EEG) is a technology that uses scalp electrodes to capture the weak electrical activity of groups of neurons in the cerebral cortex. Since German scientist Hans Berger first recorded this signal on the human scalp, its millisecond-level temporal resolution has made it a key tool for tracking dynamic changes in brain function (including cognitive process analysis, emotional state recognition, and motor intention decoding) and clinical diagnosis (such as epileptic lesion localization and sleep stage analysis). However, due to the attenuation effect of skull conductivity and the volume conductor effect, the spatial resolution of scalp EEG signals is significantly limited, making it difficult to achieve precise spatial localization of intracranial neural activity.
[0003] To address the low spatial resolution of EEG signals, EEG source localization technology has gradually developed. This technology inverts the three-dimensional spatial distribution of neural electrical activity in the brain from the scalp potential distribution by establishing a forward head conduction model and solving an inverse problem. However, solving this problem is essentially a mathematical "inverse problem," prone to multiple solutions and highly dependent on the accuracy of the head conduction model. Existing source localization methods have significant limitations. Traditional algorithms, such as minimum norm estimation (MNE) and standardized low-resolution electromagnetic tomography (sLORETA), improve the solution space by introducing physical prior constraints, but oversimplify the head geometry model, which can easily introduce systematic localization errors. Furthermore, traditional algorithms lack robustness and are sensitive to measurement noise such as electromyographic artifacts and electrode impedance variations, which can easily distort source localization results. Furthermore, existing methods often assume static neural activity and fail to effectively capture the temporal and spatial evolution of dynamic neural processes.
[0004] Deep learning demonstrates significant advantages in EEG localization, overcoming the limitations of traditional methods that rely on simplified models. It automatically learns the complex relationship between EEG signals and neural activity through a data-driven approach. Its dynamic modeling capabilities enable real-time tracking of instantaneous neural activity. Combined with noise-reduction design and multimodal fusion techniques, it effectively improves localization reliability. Its end-to-end architecture also significantly improves computational efficiency, enabling real-time imaging and promoting precision medicine applications. These technical characteristics make deep learning a key research direction for optimizing EEG localization performance. Summary of the Invention
[0005] In response to the shortcomings of the existing technology, the present invention provides a method and system for localizing the source of motion-related EEG signals based on MNE-GAN, which introduces the minimum norm constraint into deep neural networks for EEG source localization, providing a new possibility for in-depth analysis of EEG signals.
[0006] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:
[0007] In the first aspect, a method for locating the source of motion-related EEG signals based on MNE-GAN includes the following steps:
[0008] (1) Generate virtual EEG power data in motion based on Brodmann partitioning, calculate the EEG forward propagation process based on the real head model to obtain the lead matrix, map the virtual EEG power data to the scalp electrode position, and add electromyography, electrocardiography, and electrooculography noise after the mapping process to generate EEG signals that simulate the real acquisition environment;
[0009] (2) Preprocessing the EEG signals, including re-referencing, bandpass filtering, downsampling, artifact removal, and slice segmentation, and extracting effective EEG features;
[0010] (3) Constructing a generative adversarial network (MNE-GAN) model based on minimum norm estimation, in which the generator uses preprocessed virtual EEG data as a training set, learns the spatial-temporal characteristics of EEG signals through multi-layer convolution operations, outputs high-precision source space average EEG signals in blocks according to Brodmann partitioning, and establishes a mapping model from EEG signals to brain source distribution; the discriminator uses the source space average EEG signals generated by the generator as false samples and the EEG signals simulated in step (1) as true samples to perform adversarial training, adopts binary cross entropy loss as the loss function, calculates the generator loss through the discrimination results, and introduces the minimum norm inverse solution into the generator loss;
[0011] (4) A phased training strategy is used to train the MNE-GAN model: first, the discriminator is trained using a noise signal, then the generator and discriminator are trained alternately until the model converges, and finally the minimum norm inverse solution is introduced as a supervision term for joint optimization;
[0012] (5) The real EEG data is preprocessed according to the method in step (2) and then input into the trained MNE-GAN model. The model outputs the signal of the corresponding partition of the cerebral cortex, and then locates it according to the output signal intensity.
[0013] The second aspect is a motion-related EEG signal source localization system based on MNE-GAN, including:
[0014] The virtual brain power data generation module is used to generate virtual brain power data in motion based on Brodmann partitioning, calculate the EEG forward propagation process based on the real head model to obtain the lead matrix, map the virtual brain power data to the scalp electrode positions, and add electromyography, electrocardiography, and electrooculography noise after the mapping process to generate EEG signals that simulate the real acquisition environment;
[0015] The EEG signal preprocessing module is used to preprocess EEG signals, including re-referencing, bandpass filtering, downsampling, artifact removal, and slice segmentation, and to extract effective EEG features;
[0016] An MNE-GAN model construction module is used to construct a generative adversarial network (MNE-GAN) model based on minimum norm estimation. In the MNE-GAN model, the generator uses preprocessed virtual EEG data as a training set, learns the spatial-temporal characteristics of EEG signals through multi-layer convolution operations, outputs high-precision source space average EEG signals based on Brodmann partitioning, and establishes a mapping model from EEG signals to brain source distribution; the discriminator uses the source space average EEG signals generated by the generator as false samples and the EEG signals simulated by the virtual EEG source data generation module as true samples to perform adversarial training, adopts binary cross entropy loss as the loss function, calculates the generator loss based on the discrimination results, and introduces the minimum norm inverse solution into the generator loss;
[0017] The model training module is used to train the MNE-GAN model using a phased training strategy: first, the discriminator is trained using a noise signal, then the generator and discriminator are trained alternately until the model converges, and finally the minimum norm inverse solution is introduced as a supervision term for joint optimization;
[0018] The source localization module is used to preprocess the real EEG data according to the method of the EEG signal preprocessing module and then input it into the trained MNE-GAN model. The model outputs the signal of the corresponding partition of the cerebral cortex, and then locates the source based on the output signal strength.
[0019] In a third aspect, the present invention also provides a computer device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the method for locating the source of motion-related EEG signals based on MNE-GAN as described in the first aspect are implemented.
[0020] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for locating the source of motion-related EEG signals based on MNE-GAN as described in the first aspect.
[0021] Beneficial effects:
[0022] (1) The present invention uses the generative adversarial network (GAN) method to construct an EEG signal inverse solution model. Compared with the traditional EEG source localization method, this method breaks through the limitation of the traditional method relying on simplified models, which is conducive to enhancing the model positioning accuracy and stability, and provides a new idea for the EEG source localization method.
[0023] (2) Through the adversarial training mechanism, the present invention can identify and suppress the influence of noise, thereby improving the robustness of the inverse solution. Combined with the minimum norm constraint, it can effectively reduce the interference of noise on the inverse solution while ensuring the accuracy of the solution.
[0024] (3) The present invention emphasizes the importance of generating simulated EEG signals. It designs a method for generating simulated EEG signals by combining the brain activity mechanism with Brodmann partitioning, which can more reasonably simulate real brain activity.
[0025] (4) The present invention uses the minimum norm inverse solution as a constraint, which limits the scope of the inverse solution, avoids overfitting and enhances the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 It is the main flow chart of the method of the present invention;
[0027] Figure 2 Schematic diagram of movement-related brain areas;
[0028] Figure 3 Schematic diagram for constructing the generator model;
[0029] Figure 4 Schematic diagram of building a discriminator model. DETAILED DESCRIPTION
[0030] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings.
[0031] The embodiment of the present invention provides a method for locating the source of motion-related EEG signals based on MNE-GAN. Figure 1 , the method comprises the following steps:
[0032] (1) Generate virtual EEG power data: Use the MNE-Python toolkit to generate virtual EEG power data based on Brodmann partitioning, calculate the EEG forward propagation process based on the real head model to obtain the lead matrix, map the virtual EEG power data to the scalp electrode position, and add electromyography, electrocardiography, and electrooculography noise after the mapping process to generate an EEG signal R(x) that simulates the real acquisition environment, and save the EEG power data and simulated EEG data;
[0033] Specifically, step (1) includes:
[0034] (1.1) Schematic diagram of movement-related brain areas Figure 2 As shown in the figure, the highlighted parts are 26 movement-related brain areas selected according to Brodmann partitioning, including the primary somatomotor cortex (G_precentral), the extra / premotor area (S_precentral), the supplementary motor area (G_and_S_paracentral), the cortical-cortical motor integration zone (G_and_S_subcentral) and other limbic motor areas; the primary somatomotor cortex and the extra / premotor area in the upper limb movement-related areas are set as the main activation areas, the postcentral gyrus (G_postcentral), the supplementary motor area (G_and_S_paracentral), the central sulcus (S_central), and the cortical-cortical motor integration zone (G_and_S_subcentral) are set as the collaborative areas, and the remaining movement-related areas are set as the background areas;
[0035] (2.2) The dipoles of the selected 26 movement-related brain regions were initialized using a blank signal with a duration of 5 seconds and a sampling frequency of 256 Hz. In the selected main activation area, a sine wave with a random frequency of 8-13 Hz and a random amplitude of 5-10 μV was used to simulate the μ rhythm in the period 0-1 s, and a sine wave with a random frequency of 13-30 Hz and a random amplitude of 10-20 μV was used to simulate the β rhythm. The amplitudes of the μ rhythm and β rhythm were gradually reduced to 30% of the original amplitude in the period 1-3 s. The μ rhythm was kept unchanged in the period 3-5 s, and the amplitude of the β rhythm was gradually increased back to the source amplitude. In the synergistic area, 50% of the amplitude of the main activation area was used to simulate the EEG activity during movement. In the background area, 0.1-2 μV random noise was used to simulate background noise and other low-frequency activities. The dipoles in different areas were assigned corresponding signals to obtain the overall simulated signal.
[0036] (2.3) Based on the forward BEM head model, the dielectric conductivity of the three-layer model is set (0.3S / m for scalp, 0.006S / m for skull, and 0.3S / m for brain), and the forward solution of the EEG signals in the movement-related area is calculated. The EEG signals in the source space are mapped to the electrodes. In the process, electrocardiogram, electromyography, electrooculogram noise, and random noise are added to simulate the real EEG signals.
[0037] (2) Preprocessing the EEG signal generated in step (1), including re-referencing, bandpass filtering, downsampling, artifact removal, and slice segmentation, and extracting effective EEG features;
[0038] Specifically, step (2) includes:
[0039] (2.1) Perform 0.3Hz-45Hz bandpass filtering on the data to remove outliers, and use linear interpolation to repair the abnormal channel data using the channel data near the abnormal channel;
[0040] (2.2) Use the independent component analysis (ICA) algorithm to remove the myoelectric, electrocardiographic and oculoscopic interference from the EEG signal;
[0041] (2.3) Segment the processed signal and perform baseline correction and common average reference.
[0042] (3) Construct a convolutional neural network (CNN) based on minimum norm estimation (MNE) as a generator, referring to Figure 3 , using the virtual EEG data preprocessed in step (2) as the training set, learning the spatial-temporal characteristics of the EEG signal through multi-layer convolution operations, outputting high-precision source average EEG signals according to Brodmann partitioning and block, and establishing a mapping model from EEG signals to brain source distribution;
[0043] Specifically, step (3) includes:
[0044] (3.1) The generator uses a convolutional neural network (CNN) combined with a fully connected layer to extract EEG signal features, and converts the preprocessed EEG signal X∈R with c electrode channels and t time length into c×t As input, it is passed into the convolution layer for feature extraction and dimensionality reduction through the pooling layer. The input feature of the kth layer is h k-1 , the convolution calculation of this layer can be expressed as: h k =f(Conv(h k-1 ,ω k ,b k )), f(·) represents the ReLu activation function, ω is the weight parameter, b is the bias, and a pooling operation is performed after each two convolution operations to reduce the overall computational complexity. A total of five convolution operations are used to finally obtain the EEG encoding data s;
[0045] (3.2) Use the fully connected layer to decode the data s encoded by the convolution layer, and use two layers of fully connected layers to calculate. The process can be expressed as h m =f(ω m *h m-1 +b m ), f(·) represents the ReLu activation function, ω is the weight parameter, b is the bias, and a linear activation function is used to finally obtain the EEG features;
[0046] (3.3) The obtained EEG features are mapped into blocks, and the motion-related areas are selected according to the Brodmann partition. Each block is mapped separately using a single-layer convolution. Assuming that each block has r signal sources, the mapping obtains the single source average EEG signal A(x)∈R1*t Corresponding to the average signal of r signal sources in the source space, all blocks are spliced together to output the overall source average EEG signal G(x).
[0047] In the context of the present invention, source space average EEG signal and source average EEG signal are used interchangeably, referring to the EEG signal obtained by partition averaging. The source average EEG signal in step (1) refers to the signal obtained by averaging the dipole signals within each partition. The signal in step (3) is the source average EEG signal directly generated by the generator; the source EEG signal refers to the initial signal containing multiple dipole signals in each partition that has not been averaged.
[0048] (4) Build the discriminator model, refer to Figure 4 , the source average EEG signal generated in step (3) is marked as a false sample (label 0), and the source average EEG signal generated in step (1) is obtained by dipole averaging according to the partition, which is marked as a true sample (label 1). The two are input into the discriminator for adversarial training; the discriminator parameters are updated by binary cross entropy loss, and the generator loss is calculated by the discrimination result. The traditional minimum norm solution (MNE) is introduced as a supervisory signal in the generator loss to enhance the physical rationality of the source space distribution;
[0049] Specifically, step (4) includes:
[0050] (4.1) Assume that the input signal to the discriminator is S(x). The signal passes through three convolutional layers and two fully connected layers, and then passes through the Sigmoid activation function. The output is the probability D(S(x)) that the input sample is a true sample.
[0051] (4.2) The loss function of the discriminator is calculated by binary cross entropy. The loss of the real sample is -log(D(R(x))), and the probability of successful real sample discrimination D(R(x)) should be as close to 1 as possible. The loss of the generated sample is -log(1-D(G(x))), and the probability of the generated sample being judged as a real sample D(G(x)) should be as close to 0 as possible. The final loss of the discriminator is the average of the sum of the generated sample loss and the real sample loss:
[0052]
[0053] (4.3) The probability D(S(x)) that the input sample output by the discriminator is a true sample can be used to calculate the adversarial loss:
[0054]
[0055] In order to avoid large fluctuations in the early stages of training, the initial loss function of the generator is only the adversarial loss. When the generated results tend to be stable, the minimum norm constraint loss function (MNE) is added:
[0056]
[0057] Where G(x i ) is the generator output, For data x i The corresponding minimum norm inverse solution is calculated using the python toolkit to obtain the inverse solution MNE(x). The source signal of the motion-related area is selected as the final result according to the Brodmann partition, and the minimum norm inverse solution reference is obtained by averaging the regional signals. After adding the minimum norm constraint loss, it can provide a clear reference guide for the generated EEG data, further enhancing the reliability of the generated EEG data. The final loss function is obtained by weighting the adversarial loss and the minimum norm constraint loss:
[0058] Loss Generator =Loss GAN +γ mne *Loss mne
[0059] γ mne is the weight coefficient of the minimum norm loss, which is used to balance the impact of various losses.
[0060] (5) A phased training strategy is used to train the MNE-GAN model: first, the discriminator and generator are trained alternately until the model reaches initial convergence, and then physical constraints and supervision terms are introduced for joint optimization to ensure model stability and positioning accuracy;
[0061] Specifically, step (5) includes:
[0062] (5.1) Pre-train the GAN network, fix the generator parameters, use noise as fake samples, and use the virtual source average EEG signal as the real sample to iteratively update the discriminator parameters;
[0063] (5.2) Use the pre-trained discriminator and the designed generator for adversarial training, alternating the parameters until convergence;
[0064] (5.3) After the model initially converges, Loss is introduced mne , γ mne Conduct joint training until Loss Generator Descent convergence.
[0065] (6) Perform preprocessing on real EEG data, including re-referencing, bandpass filtering, downsampling, artifact removal, and slice segmentation, and extract effective EEG features;
[0066] (7) The real EEG data processed in step (6) is input into the trained MNE-GAN model for source localization. The model outputs the collected EEG signals as signals of the corresponding partitions of the cerebral cortex, i.e., source signals, and then locates the source signals based on the signal strength. Furthermore, the source signals obtained by this model and the source signals obtained by the MNE method are simultaneously used in the action classification task to compare the classification accuracy of the two methods. The generation effect is evaluated using the classification accuracy.
[0067] Based on the same technical concept as the method embodiment, another embodiment of the present invention provides a motion-related EEG signal source localization system based on MNE-GAN, including:
[0068] The virtual brain power data generation module is used to generate virtual brain power data in motion based on Brodmann partitioning, calculate the EEG forward propagation process based on the real head model to obtain the lead matrix, map the virtual brain power data to the scalp electrode positions, and add electromyography, electrocardiography, and electrooculography noise after the mapping process to generate EEG signals that simulate the real acquisition environment;
[0069] The EEG signal preprocessing module is used to preprocess EEG signals, including re-referencing, bandpass filtering, downsampling, artifact removal, and slice segmentation, and to extract effective EEG features;
[0070] An MNE-GAN model construction module is used to construct a generative adversarial network (MNE-GAN) model based on minimum norm estimation. In the MNE-GAN model, the generator uses preprocessed virtual EEG data as a training set, learns the spatial-temporal characteristics of EEG signals through multi-layer convolution operations, outputs high-precision source space average EEG signals based on Brodmann partitioning, and establishes a mapping model from EEG signals to brain source distribution; the discriminator uses the source space average EEG signals generated by the generator as false samples and the EEG signals simulated by the virtual EEG source data generation module as true samples to perform adversarial training, adopts binary cross entropy loss as the loss function, calculates the generator loss based on the discrimination results, and introduces the minimum norm inverse solution into the generator loss;
[0071] The model training module is used to train the MNE-GAN model using a phased training strategy: first, the discriminator is trained using a noise signal, then the generator and discriminator are trained alternately until the model converges, and finally the minimum norm inverse solution is introduced as a supervision term for joint optimization;
[0072] The source localization module is used to preprocess the real EEG data according to the method of the EEG signal preprocessing module and then input it into the trained MNE-GAN model. The model outputs the signal of the corresponding partition of the cerebral cortex, and then locates the source based on the output signal strength.
[0073] It should be understood that the MNE-GAN-based motion-related EEG signal source localization system in this embodiment can implement all the technical solutions in the above-mentioned method embodiments, and the functions of its various functional modules can be specifically implemented according to the methods in the above-mentioned method embodiments. The specific implementation process can refer to the relevant description in the above-mentioned embodiments, which will not be repeated here.
[0074] Another embodiment of the present invention provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the method for localizing the source of motion-related EEG signals based on MNE-GAN are implemented as described above.
[0075] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method for locating a source of motion-related EEG signals based on MNE-GAN are implemented as described above.
[0076] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, apparatus (systems), computer devices, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0077] The present invention is described with reference to flowcharts of methods according to embodiments of the present invention. It should be understood that each process in the flowcharts and combinations of processes in the flowcharts can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts. Figure 1 A device that specifies functions in a process or multiple processes.
[0078] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A function specified in a process or multiple processes.
[0079] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 The steps of a specified function in a process or multiple processes.
Claims
1. A method for locating the source of motion-related EEG signals based on MNE-GAN, characterized in that: The following steps are involved: (1) Generate virtual EEG power data in motion based on Brodmann partitioning, calculate the EEG forward propagation process based on the real head model to obtain the lead matrix, map the virtual EEG power data to the scalp electrode position, and add electromyography, electrocardiography, and electrooculography noise after the mapping process to generate EEG signals that simulate the real acquisition environment; (2) Preprocessing the EEG signals, including re-referencing, bandpass filtering, downsampling, artifact removal, and slice segmentation, and extracting effective EEG features; (3) Constructing a generative adversarial network (MNE-GAN) model based on minimum norm estimation, in which the generator uses preprocessed virtual EEG data as a training set, learns the spatial-temporal characteristics of EEG signals through multi-layer convolution operations, outputs high-precision source space average EEG signals in blocks according to Brodmann partitioning, and establishes a mapping model from EEG signals to brain source distribution; the discriminator uses the source space average EEG signals generated by the generator as false samples and the EEG signals simulated in step (1) as true samples to perform adversarial training, adopts binary cross entropy loss as the loss function, calculates the generator loss through the discrimination results, and introduces the minimum norm inverse solution into the generator loss; (4) A phased training strategy is used to train the MNE-GAN model: first, the discriminator is trained using a noise signal, then the generator and discriminator are trained alternately until the model converges, and finally the minimum norm inverse solution is introduced as a supervision term for joint optimization; (5) The real EEG data is preprocessed according to the method in step (2) and then input into the trained MNE-GAN model. The model outputs the signal of the corresponding partition of the cerebral cortex, and then locates it according to the output signal intensity.
2. The method according to claim 1, characterized in that Step (1) includes: According to Brodmann partitioning, motion-related areas were selected, brain area labels were set, and motion-related areas were divided into main activation areas, collaborative areas, and background areas; Initialize the source dipole. In the selected activation area, use sine waves of different frequencies and amplitudes in different partitions to simulate brain activity in different time periods according to the label. In the background area, use random noise to simulate noise and low-frequency activity. Use the simulated signal to give the source dipole amplitude. According to the forward model, the forward solution of the EEG signal in the movement-related area is calculated, the EEG signal in the source space is mapped to the electrode, and electrocardiogram, electromyography, electrooculogram noise and random noise are added to simulate the real collected EEG signal.
3. The method according to claim 1, characterized in that Step (2) includes: The data was band-pass filtered from 0.3 Hz to 45 Hz to remove outliers, and the abnormal channel data was repaired by interpolation; Independent component analysis (ICA) algorithm is used to remove electromyographic, electrocardiographic and oculoscopic interference from EEG signals. Segment the processed signal and perform baseline correction with a common mean reference.
4. The method according to claim 3, characterized in that The generator uses a convolutional neural network combined with a fully connected layer to extract EEG signal features, including: The EEG signal X∈R with c electrode channels and t time length after preprocessing c×t As input, it is passed into the convolution layer for feature extraction and dimensionality reduction through the pooling layer. The input feature of the kth layer is h k-1 , the convolution calculation of this layer is expressed as: h k =f(Conv(h k-1 ,ω k ,b k )), f(·) represents the ReLu activation function, ω is the weight parameter, b is the bias, and the final EEG encoding data s is obtained; The fully connected layer is used to decode the encoded data s encoded by the convolution layer. This process is represented as h m =f(ω m *h m-1 +b m ), and finally get the EEG features; Using block mapping, we select motion-related areas according to Brodmann partitioning and map each block separately. When each block has r signal sources in the source space, we map the single source average EEG signal A(x)∈R 1*t Corresponding to the average signal of r signal sources in the source space, all blocks are spliced together to output the overall source average EEG signal G(x).
5. The method according to claim 1, wherein In the discriminator model, the signal input to the discriminator is denoted as S(x). After the signal passes through the convolutional layer and the fully connected layer, it is output through the Sigmoid activation function to obtain the probability D(S(x)) that the input sample is a true sample. The loss function of the discriminator is calculated by binary cross entropy. The loss of the real sample is -log(D(R(x))), and the loss of the generated sample is -log(1-D(G(x))). The final loss of the discriminator is the average of the two terms: Where N is the number of samples.
6. The method according to claim 5, characterized in that The loss of the generator is as follows: The adversarial loss is calculated by the probability D(S(x)) that the input sample output by the discriminator is a real sample: The initial loss function of the generator is only the adversarial loss. When the generated results tend to be stable, the minimum norm constraint loss function MNE is added: Where G(x i ) is the generator output, For data x i The corresponding minimum norm inverse solution is calculated to obtain the inverse solution MNE(x). The source signal of the motion-related area is selected as the final result according to the Brodmann partition, and the minimum norm inverse solution reference is obtained by averaging the regional signals. The final generator loss is obtained by weighting the adversarial loss and the minimum norm constraint loss: Loss Generator =Loss GAN +γ mne *Loss mne where γ mne is the weight coefficient of the minimum norm loss, which is used to balance the impact of various losses.
7. The method according to claim 6, characterized in that Step (5) includes: Pre-train the model, fix the generator parameters, use noise as fake samples, and use the virtual source average EEG signal as the real sample to iteratively update the discriminator parameters; Use the pre-trained discriminator and the designed generator for adversarial training, alternating the parameters until convergence; Loss is introduced after the model initially converges mne , γ mne Conduct joint training until Loss Generator Descent convergence.
8. A motion-related EEG signal source localization system based on MNE-GAN, characterized by: include: The virtual brain power data generation module is used to generate virtual brain power data in motion based on Brodmann partitioning, calculate the EEG forward propagation process based on the real head model to obtain the lead matrix, map the virtual brain power data to the scalp electrode positions, and add electromyography, electrocardiography, and electrooculography noise after the mapping process to generate EEG signals that simulate the real acquisition environment; The EEG signal preprocessing module is used to preprocess EEG signals, including re-referencing, bandpass filtering, downsampling, artifact removal, and slice segmentation, and to extract effective EEG features; An MNE-GAN model construction module is used to construct a generative adversarial network (MNE-GAN) model based on minimum norm estimation. In the MNE-GAN model, the generator uses preprocessed virtual EEG data as a training set, learns the spatial-temporal characteristics of EEG signals through multi-layer convolution operations, outputs high-precision source space average EEG signals based on Brodmann partitioning, and establishes a mapping model from EEG signals to brain source distribution; the discriminator uses the source space average EEG signals generated by the generator as false samples and the EEG signals simulated by the virtual EEG source data generation module as true samples to perform adversarial training, adopts binary cross entropy loss as the loss function, calculates the generator loss based on the discrimination results, and introduces the minimum norm inverse solution into the generator loss; The model training module is used to train the MNE-GAN model using a phased training strategy: first, the discriminator is trained using a noise signal, then the generator and discriminator are trained alternately until the model converges, and finally the minimum norm inverse solution is introduced as a supervision term for joint optimization; The source localization module is used to preprocess the real EEG data according to the method of the EEG signal preprocessing module and then input it into the trained MNE-GAN model. The model outputs the signal of the corresponding partition of the cerebral cortex, and then locates the source based on the output signal strength.
9. A computer device comprising: one or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the method for localizing the source of motion-related EEG signals based on MNE-GAN are implemented as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the method for locating a motion-related EEG signal source based on MNE-GAN are implemented as described in any one of claims 1 to 7.
Citation Information
Patent Citations
Electroencephalogram signal decoding method based on deep convolutional generative adversarial neural network
CN112001306A
Motor imagery electroencephalogram data generator based on generative adversarial network data enhancement
CN118708058A
Source affine reconstruction for medical imaging
US20090182220A1
Identification of a psychotherapy-predictive electroencephalographic signature in ptsd
WO2024145587A1
