Deep learning brain power supply imaging method based on improved MWCNN

Abstract

The invention discloses a deep learning brain power supply imaging method based on an improved MWCNN, and the method comprises the steps: synthesizing a time signal and space activation information of a brain source signal to obtain brain source signal data, constructing a standard boundary element head model, carrying out the forward calculation of the synthesized brain source signal data and the boundary element head model, adding noise, and obtaining a brain source signal model; synthesizing to obtain electroencephalogram signal data; the electroencephalogram signal data are mapped to a regularized brain source space through sLORTA, and dispersed pseudo source signals with physical information constraints are obtained; inputting the pseudo source signal into the improved MWCNN for training, calculating an error by using a multi-target loss function in back propagation, and carrying out multiple iterations until the loss value is minimum to obtain a network with optimal performance; and inputting electroencephalogram signal data to be detected into the network with the optimal performance, so that an electroencephalogram imaging result with high space-time reconstruction performance can be obtained. According to the method, the space-time performance and reconstruction robustness of brain power source imaging can be effectively improved, and the reasoning speed can be increased.

Description

Technical Field

[0001] This invention relates to the technical field of deep learning brain power imaging, and in particular to a deep learning brain power imaging method based on an improved MWCNN. Background Technology

[0002] EEG Source Imaging (ESI) is an important technique for retrieving the location and intensity of neural activity sources within the brain by recording electroencephalogram (EEG) signals from the scalp. This facilitates more precise tracking of neural signals and has wide applications in neuroscience research and the diagnosis of brain diseases. Currently, ESI methods are mainly divided into three categories: 1. Regularized methods: These use Tikhonov regularization to introduce additional constraints that allow the imaging to converge to a stable solution. A drawback of this method is that manually designed regularization terms cannot effectively constrain the solution space. 2. Bayesian methods: These implicitly integrate physical constraints into the prior distribution assumptions of the sources, deriving the posterior distribution of the sources within a Bayesian inference framework. While Bayesian methods offer excellent imaging performance, they also rely on hyperparameter control design, and the iterative solution process is generally time-consuming. 3. Deep learning methods: These use synthetic data and leverage the powerful nonlinear modeling capabilities of neural networks to directly model and inversely solve the process, offering the advantage of rapid inference. However, deep learning methods depend on the strategy of synthesizing data, requiring the synthesized data to fully simulate the neural activity of real brain source signals; otherwise, they exhibit instability across different datasets. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a deep learning brain power imaging method based on an improved MWCNN, which can effectively improve the spatiotemporal performance and reconstruction robustness of brain power imaging, as well as improve inference speed.

[0004] To achieve the above objectives, the technical solution provided by this invention is as follows: a deep learning-based brain energy imaging method based on an improved MWCNN. This improved MWCNN adds a feature extraction module to the original multi-scale wavelet convolutional network MWCNN, improves the original Skip-connection module, and designs a multi-objective loss function for brain energy imaging. The feature extraction module is a graph convolutional network (GCN) module, which aggregates neighboring node information in the brain's anatomical space through a graph adjacency matrix, enhancing the network's ability to capture the brain's topological structure information. The improvement to the Skip-connection module involves using a depthwise separable convolutional module to fuse features between the encoding and decoding modules, replacing the original addition operation, and obtaining a more flexible multi-scale feature fusion mechanism in a data-driven manner. A multi-objective loss function is designed for brain energy imaging, which simultaneously constrains the network's reconstruction errors for whole-brain source signals, whole-brain electroencephalogram (EEG) signals, and brain source signals from activated regions, thereby guiding the network to focus on high-fidelity reconstruction of the temporal dynamic waveforms of activated brain regions.

[0005] The specific implementation of the deep learning brain power imaging method includes the following steps:

[0006] 1) Synthesize training data, including synthesizing the temporal signal and spatial activation information of brain source signals to obtain brain source signal data, construct a standard boundary element head model, perform forward calculation on the synthesized brain source signal data and boundary element head model and add noise to synthesize EEG signal data.

[0007] 2) The EEG signal data is mapped to a regularized brain source space using the standardized low-resolution EEG tomography method sLORETA to obtain diffuse pseudo-source signals with physical information constraints.

[0008] 3) Input the pseudo-source signal into the improved MWCNN for training. Use the designed multi-objective loss function to calculate the error during backpropagation. After multiple iterations, the loss value is minimized, and the network with the best performance is obtained.

[0009] 4) By inputting the EEG signal data to be tested into the network with the best performance, brain power imaging results with high spatiotemporal reconstruction performance can be obtained.

[0010] Further, in step 1), the real EEG signal is decomposed into a time basis vector using the singular value decomposition method (SVD). A time signal is generated by random sampling in the time domain space formed by the time basis vector. The spatial activation information is obtained by selecting a large number of random cortical dipoles as seed points. Based on the seed points as the center, a reasonable size is randomly sampled to determine the activation region. Finally, the activation position and activation size of the brain source signal are obtained. Brain source signal data is synthesized by combining the time signal and the spatial activation information.

[0011] A standard boundary element head model is constructed using the boundary element method to obtain the lead matrix required for brain power imaging. The synthesized brain-derived signal data is then subjected to forward calculation to obtain synthesized EEG signal data. The forward calculation process is as follows:

[0012] ;

[0013] In the formula, This represents synthesized electroencephalogram (EEG) signal data. This represents synthesized brain-derived signal data. This represents Gaussian noise signal data;

[0014] To characterize the rich and varied properties of brain-derived signals, brain-derived signal data... Based on this, pink noise is added to simulate pseudo-source activity, resulting in a noisy brain source signal. Then, the noisy synthetic EEG signal data is obtained through forward calculation. :

[0015] ;

[0016] The levels of pink noise and Gaussian noise can be controlled by explicitly specifying the signal-to-noise ratio (SNR):

[0017] ;

[0018] In the formula, Indicates noise power. The signal power is represented by SNR, and the controlled noise level is represented by SNR. Noise is simulated by using a large number of different SNRs to enrich the training samples and enhance the robustness of the network.

[0019] Furthermore, in step 2), sLORETA is used to map the synthesized EEG signal data to a regularized brain source space. By eliminating spatial bias, sLORETA makes the estimated variance of each brain source dipole equal and standardizes the solution of the minimum norm estimation method, which can achieve zero localization error for a single source and obtain a diffuse pseudo-source signal with physical information constraints.

[0020] Furthermore, in step 3), the improved MWCNN includes a GCN module and an encoder-decoder backbone network. The pseudo-source signal is input into the improved MWCNN for processing, as detailed below:

[0021] The GCN module, operating on a time-step basis, models the physical topology of the brain's network connections using a graph adjacency matrix. It spatially aggregates information from adjacent nodes, strengthens connection-based activation features, and suppresses noise interference to encode neighborhood channel feature relationships, overcoming the limitation of the receptive field in two-dimensional convolution. For time step t, the output of the m-th layer of the GCN... as follows:

[0022] ;

[0023] In the formula, ReLU represents the activation function. This represents a degree matrix with self-connections. This represents an adjacency matrix with self-connections. Indicates the first Layer weight parameters, Indicates at time step No. The output of the layer is shown in Figure 0. The 0th layer represents the input of the GCN module, i.e., the pseudo-source signal. To avoid over-smoothing, a three-layer GCN is used for feature extraction to obtain the output features. ;

[0024] The encoder-decoder backbone network includes an encoding module, a decoding module, a process module, and an improved Skip-connection module; in the encoding module, features First, the features are processed by the Conv module, which consists of a 3×3 convolutional layer and a ReLU activation function. Then, the data alternately passes through the DownConv and DWT modules. The DownConv module, composed of two 3×3 convolutional layers with different dilation factors and ReLU activation functions, is responsible for convolutional processing of the feature information and compressing channel information at different scales to form a more compact information representation. The DWT module performs wavelet transform decomposition on the features at different scales to obtain high- and low-frequency features, completing the downsampling process. This process not only preserves the low-frequency contour of the signal but also captures high-frequency transient components, ensuring that key details of neural activity are not lost. This adapts to the nonlinear and non-stationary characteristics of brain-derived signal data, and since brain-derived signal data, as a physiological signal, has features in each frequency band directly related to different... The neural oscillation mode utilizes the DWT module to fully leverage information for encoding. In the process module, six 3×3 convolutional layers complete the final processing of the encoding module, outputting features to the decoding module. In the decoding module, features alternately enter the UpConv and IWT modules. The UpConv module consists of two 3×3 convolutional layers with dilation factors corresponding to DownConv and a ReLU activation function. The IWT module performs inverse wavelet transform along the channel dimension to complete the upsampling process. An improved Skip-connection module fuses encoded and decoded features at different scales between the encoding and decoding modules. This improved Skip-connection module uses depthwise separable convolutional modules for feature fusion, and the fusion process is as follows:

[0025] ;

[0026] In the formula, Indicates the first The encoded features obtained from the layer Indicates the first Decoding features obtained from the layer, This indicates that the encoded and decoded features are concatenated along the channel dimension. Table depth-separable convolution functions Indicates the first Layer depth can separate the parameters of the convolution kernel, and feature fusion yields the first... Layer decoding features The shallowest decoded features in the upsampling process enter the Conv module, and the network output is obtained through convolution, which is the reconstructed brain source signal data;

[0027] The following multi-objective loss function is designed for brain power imaging:

[0028] ;

[0029] In the formula, Represents the mean squared error function. This represents brain-derived signal data reconstructed using MSE calculations. With synthetic brain-derived signal data The loss, to The lead matrix is ​​obtained again through the standard boundary element head model in step 1). EEG signal data reconstructed by mapping to the observation signal space Simultaneously, the synthesized brain-derived signal data Performing the same mapping yields noise-free EEG signal data. Calculated via MSE and loss The study focuses on the accuracy of reconstructed brain-derived signal waveforms in activated regions, and calculates the reconstructed brain-derived signal data in activated regions using MSE (Mean Sequence of Effect). Synthesized brain-derived signal data in the activated regions loss The three losses are fused together using designed weights to form the overall loss function for training. , , express , , The corresponding weights This represents the final multi-objective loss function.

[0030] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0031] 1. This invention uses the SVD decomposition time base method to construct brain source signals, and adds pink noise perturbation to the source signals to make the simulated brain source signals more realistic and reliable, providing a more complex simulation background for neural network training.

[0032] 2. This invention uses the standardized low-resolution EEG method sLORETA combined with deep learning methods for reconstruction. By using the standardized low-resolution EEG method sLORETA to embed prior information about brain anatomy into diffuse pseudo-source signals, it has anti-interference capabilities for different populations, data collection, and experimental configurations. It allows the combined deep learning method to be trained without modifying structural parameters, effectively improving the network's adaptability. At the same time, training on the basis of pseudo-source signals also improves the convergence speed of the neural network. The pseudo-source signals serve as the basis for candidate solutions, enabling the neural network to specifically decode the activation relationships between brain regions. Moreover, the nonlinear modeling capability of the neural network can also overcome the limitations of the standardized low-resolution EEG method sLORETA, identifying activation regions outside of the pseudo-source signals and obtaining better spatiotemporal reconstruction performance.

[0033] 3. This invention uses an improved MWCNN, employing a GCN module to model the physical topology of brain network connections using a graph adjacency matrix on a time-step basis. It spatially aggregates information from adjacent nodes, strengthens connection-based activation features, and suppresses noise interference to encode neighborhood channel feature relationships, overcoming the limitation of the receptive field in two-dimensional convolution. An improved Skip-connection module facilitates the fusion of encoded and decoded features at different scales between the encoding and decoding modules. This improved Skip-connection module uses depthwise separable convolution to learn the fused features, achieving a personalized fusion process at different scales. The use of a multi-objective loss function further improves the spatiotemporal reconstruction performance of brain power imaging. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the logical flow of the method of the present invention.

[0035] Figure 2 This diagram illustrates the improved structure of MWCNN. In the diagram, GCN represents the graph convolution module, Conv represents the convolution module, DownConv represents the downsampling convolution module, DWT represents the wavelet transform module, Process represents the feature processing module, IWT represents the inverse wavelet transform module, UpConv represents the upsampling convolution module, and Skip-connection represents the skip connection module (which uses depthwise separable convolution for feature fusion learning).

[0036] Figure 3 This is a schematic diagram of the improved Skip-connection module; in the diagram, Conv represents the convolution operation. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0038] like Figures 1 to 3As shown, this embodiment discloses a deep learning-based brain energy imaging method based on an improved MWCNN. This improved MWCNN adds a feature extraction module to the original multi-scale wavelet convolutional network MWCNN, improves the original Skip-connection module, and designs a multi-objective loss function for brain energy imaging. The feature extraction module is a graph convolutional network (GCN) module, which aggregates the information of adjacent nodes in the brain's anatomical space through a graph adjacency matrix, improving the network's ability to capture the brain's topological structure information. The improvement to the Skip-connection module is that a depthwise separable convolutional module is used to fuse the features of the encoding and decoding modules, replacing the original addition operation, and obtaining a more flexible multi-scale feature fusion mechanism in a data-driven manner. A multi-objective loss function is designed for brain energy imaging. This function simultaneously constrains the reconstruction errors of the network for whole-brain source signals, whole-brain electroencephalogram (EEG) signals, and brain source signals of activated regions, thereby guiding the network to focus on high-fidelity reconstruction of the temporal dynamic waveforms of activated brain regions.

[0039] The specific implementation of this deep learning-based brain energy imaging method includes the following steps:

[0040] 1) Synthesize training data, including synthesizing the temporal signal and spatial activation information of brain source signals to obtain brain source signal data, construct a standard boundary element head model, perform forward calculations on the synthesized brain source signal data and add noise, and synthesize EEG signal data, as follows:

[0041] The real EEG signal is decomposed into a time basis vector using the singular value decomposition method (SVD). The time signal is generated by random sampling in the time domain space formed by the time basis vector. The spatial activation information is obtained by selecting a large number of random cortical dipoles as seed points. Based on the seed points, the activation region is determined by random sampling of reasonable size. Finally, the activation position and activation size of the brain source signal are obtained. The brain source signal data is synthesized by combining the time signal and the spatial activation information.

[0042] A standard boundary element head model is constructed using the boundary element method to obtain the lead matrix required for brain power imaging. The synthesized brain-derived signal data is then subjected to forward calculation to obtain synthesized EEG signal data. The forward calculation process is as follows:

[0043] ;

[0044] In the formula, This represents synthesized electroencephalogram (EEG) signal data. This represents synthesized brain-derived signal data. This represents Gaussian noise signal data;

[0045] To characterize the rich and varied properties of brain-derived signals, brain-derived signal data... Based on this, pink noise is added to simulate pseudo-source activity, resulting in a noisy brain source signal. Then, the noisy synthetic EEG signal data is obtained through forward calculation. :

[0046] ;

[0047] The levels of pink noise and Gaussian noise can be controlled by explicitly specifying the signal-to-noise ratio (SNR):

[0048] ;

[0049] In the formula, Indicates noise power. The signal power is represented by SNR, and the controlled noise level is represented by SNR. Noise is simulated by using a large number of different SNRs to enrich the training samples and enhance the robustness of the network.

[0050] 2) The EEG signal data is mapped to a regularized brain source space using the standardized low-resolution EEG computed tomography method sLORETA to obtain diffuse pseudo-source signals with physical information constraints, as follows:

[0051] By using sLORETA to map the synthesized EEG signal data to a regularized brain source space, sLORETA eliminates spatial bias, making the estimated variance of each brain source dipole equal. By standardizing the solution of the minimum norm estimation method, it can achieve zero localization error for a single source and obtain diffuse pseudo-source signals with physical information constraints.

[0052] 3) The pseudo-source signal is input into the improved MWCNN for training. During backpropagation, the designed multi-objective loss function is used to calculate the error. After multiple iterations until the loss value is minimized, the network with optimal performance is obtained. The improved MWCNN includes a GCN module and an encoder-decoder backbone network. The pseudo-source signal is input into the improved MWCNN for processing, as detailed below:

[0053] The GCN module, operating on a time-step basis, models the physical topology of the brain's network connections using a graph adjacency matrix. It spatially aggregates information from adjacent nodes, strengthens connection-based activation features, and suppresses noise interference to encode neighborhood channel feature relationships, overcoming the limitation of the receptive field in two-dimensional convolution. For time step t, the output of the m-th layer of the GCN... as follows:

[0054] ;

[0055] In the formula, ReLU represents the activation function. This represents a degree matrix with self-connections. This represents an adjacency matrix with self-connections. Indicates the first Layer weight parameters, Indicates at time step No. The output of the layer is shown in Figure 0. The 0th layer represents the input of the GCN module, i.e., the pseudo-source signal. To avoid over-smoothing, a three-layer GCN is used for feature extraction to obtain the output features. ;

[0056] The encoder-decoder backbone network includes an encoding module, a decoding module, a process module, and an improved Skip-connection module; in the encoding module, features First, the features are processed by the Conv module, which consists of a 3×3 convolutional layer and a ReLU activation function. Then, the data alternately passes through the DownConv and DWT modules. The DownConv module, composed of two 3×3 convolutional layers with different dilation factors and ReLU activation functions, is responsible for convolutional processing of the feature information and compressing channel information at different scales to form a more compact information representation. The DWT module performs wavelet transform decomposition on the features at different scales to obtain high- and low-frequency features, completing the downsampling process. This process not only preserves the low-frequency contour of the signal but also captures high-frequency transient components, ensuring that key details of neural activity are not lost. This adapts to the nonlinear and non-stationary characteristics of brain-derived signal data, and since brain-derived signal data, as a physiological signal, has features in each frequency band directly related to different... The neural oscillation mode utilizes the DWT module to fully leverage information for encoding. In the process module, six 3×3 convolutional layers complete the final processing of the encoding module, outputting features to the decoding module. In the decoding module, features alternately enter the UpConv and IWT modules. The UpConv module consists of two 3×3 convolutional layers with dilation factors corresponding to DownConv and a ReLU activation function. The IWT module performs inverse wavelet transform along the channel dimension to complete the upsampling process. An improved Skip-connection module fuses encoded and decoded features at different scales between the encoding and decoding modules. This improved Skip-connection module uses depthwise separable convolutional modules for feature fusion, and the fusion process is as follows:

[0057] ;

[0058] In the formula, Indicates the first The encoded features obtained from the layer Indicates the first Decoding features obtained from the layer, This indicates that the encoded and decoded features are concatenated along the channel dimension. Table depth-separable convolution functions Indicates the first Layer depth can separate the parameters of the convolution kernel, and feature fusion yields the first... Layer decoding features The shallowest decoded features in the upsampling process enter the Conv module, and the network output is obtained through convolution, which is the reconstructed brain source signal data;

[0059] The following multi-objective loss function is designed for brain power imaging:

[0060] ;

[0061] In the formula, Represents the mean squared error function. This represents brain-derived signal data reconstructed using MSE calculations. With synthetic brain-derived signal data The loss, to The lead matrix is ​​obtained again through the standard boundary element head model in step 1). EEG signal data reconstructed by mapping to the observation signal space Simultaneously, the synthesized brain-derived signal data Performing the same mapping yields noise-free EEG signal data. Calculated via MSE and loss The study focuses on the accuracy of reconstructed brain-derived signal waveforms in activated regions, and calculates the reconstructed brain-derived signal data in activated regions using MSE (Mean Sequence of Effect). Synthesized brain-derived signal data in the activated regions loss The three losses are fused together using designed weights to form the overall loss function for training. , , express , , The corresponding weights This represents the final multi-objective loss function.

[0062] 4) By inputting the EEG signal data to be tested into the network with the best performance, brain power imaging results with high spatiotemporal reconstruction performance can be obtained.

[0063] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A deep learning-based brain energy imaging method based on improved MWCNN, characterized in that, This improved MWCNN adds a feature extraction module to the original multi-scale wavelet convolutional network MWCNN, improves the original Skip-connection module, and designs a multi-objective loss function for brain power imaging. The feature extraction module is a graph convolutional network (GCN) module, which aggregates neighboring node information in the brain's anatomical space through a graph adjacency matrix, enhancing the network's ability to capture the brain's topological structure. The improvement to the Skip-connection module involves using a depthwise separable convolutional module to fuse features between the encoding and decoding modules, replacing the original addition operation, and obtaining a more flexible multi-scale feature fusion mechanism through a data-driven approach. A multi-objective loss function is designed for brain power imaging, which simultaneously constrains the network's reconstruction errors for whole-brain source signals, whole-brain electroencephalogram (EEG) signals, and brain source signals from activated regions, thereby guiding the network to focus on high-fidelity reconstruction of the temporal dynamic waveforms of activated brain regions. The specific implementation of the deep learning brain power imaging method includes the following steps: 1) Synthesize training data, including synthesizing the temporal signal and spatial activation information of brain source signals to obtain brain source signal data, construct a standard boundary element head model, perform forward calculation on the synthesized brain source signal data and boundary element head model and add noise to synthesize EEG signal data. 2) The EEG signal data is mapped to a regularized brain source space using the standardized low-resolution EEG tomography method sLORETA to obtain diffuse pseudo-source signals with physical information constraints. 3) Input the pseudo-source signal into the improved MWCNN for training. Use the designed multi-objective loss function to calculate the error during backpropagation. After multiple iterations, the loss value is minimized, and the network with the best performance is obtained. 4) By inputting the EEG signal data to be tested into the network with the best performance, brain power imaging results with high spatiotemporal reconstruction performance can be obtained.

2. The deep learning-based brain energy imaging method based on improved MWCNN according to claim 1, characterized in that, In step 1), the real EEG signal is decomposed into a time basis vector using the singular value decomposition method (SVD). A time signal is generated by random sampling in the time domain space formed by the time basis vector. The spatial activation information is obtained by selecting a large number of random cortical dipoles as seed points. Based on the seed points, the activation region is determined by randomly sampling at a reasonable size. Finally, the activation position and activation size of the brain source signal are obtained. Brain source signal data is synthesized by combining the time signal and the spatial activation information. A standard boundary element head model is constructed using the boundary element method to obtain the lead matrix required for brain power imaging. The synthesized brain-derived signal data is then subjected to forward calculation to obtain synthesized EEG signal data. The forward calculation process is as follows: ； In the formula, This represents synthesized electroencephalogram (EEG) signal data. This represents synthesized brain-derived signal data. This represents Gaussian noise signal data; To characterize the rich and varied properties of brain-derived signals, brain-derived signal data... Based on this, pink noise is added to simulate pseudo-source activity, resulting in a noisy brain source signal. Then, the noisy synthetic EEG signal data is obtained through forward calculation. : ； The levels of pink noise and Gaussian noise can be controlled by explicitly specifying the signal-to-noise ratio (SNR): ； In the formula, Indicates noise power. The signal power is represented by SNR, and the controlled noise level is represented by SNR. Noise is simulated by using a large number of different SNRs to enrich the training samples and enhance the robustness of the network.

3. The deep learning-based brain energy imaging method based on improved MWCNN according to claim 1, characterized in that, In step 2), sLORETA is used to map the synthesized EEG signal data to a regularized brain source space. By eliminating spatial bias, sLORETA makes the estimated variance of each brain source dipole equal and standardizes the solution of the minimum norm estimation method, which can achieve zero localization error for a single source and obtain a diffuse pseudo-source signal with physical information constraints.

4. The deep learning-based brain energy imaging method based on improved MWCNN according to claim 1, characterized in that, In step 3), the improved MWCNN includes a GCN module and an encoder-decoder backbone network. The pseudo-source signal is input into the improved MWCNN for processing, as detailed below: The GCN module, operating on a time-step basis, models the physical topology of the brain's network connections using a graph adjacency matrix. It spatially aggregates information from adjacent nodes, strengthens connection-based activation features, and suppresses noise interference to encode neighborhood channel feature relationships, overcoming the limitation of the receptive field in two-dimensional convolution. For time step t, the output of the m-th layer of the GCN... as follows: ； In the formula, ReLU represents the activation function. This represents a degree matrix with self-connections. This represents an adjacency matrix with self-connections. Indicates the first Layer weight parameters, Indicates at time step No. The output of the layer is shown in Figure 0. The 0th layer represents the input of the GCN module, i.e., the pseudo-source signal. To avoid over-smoothing, a three-layer GCN is used for feature extraction to obtain the output features. ; The encoder-decoder backbone network includes an encoding module, a decoding module, a process module, and an improved Skip-connection module; in the encoding module, features The features are first processed by the Conv module, which consists of a 3×3 convolutional layer and a ReLU activation function. Then, the features alternately pass through the DownConv and DWT modules. The DownConv module consists of two 3×3 convolutional layers with different dilation factors and a ReLU activation function. It is responsible for convolutional processing of the feature information and compressing channel information at different scales to form a more compact information representation. The DWT module performs wavelet transform decomposition on the features at different scales to obtain high and low frequency features and complete the downsampling process. This process not only preserves the low-frequency contour of the signal but also captures high-frequency transient components, ensuring that key details of neural activity are not lost. It adapts to the nonlinear and non-stationary characteristics of brain source signal data. Moreover, as a physiological signal, the features of each frequency band of brain source signal data are directly related to different neural oscillation modes. The DWT module helps the network to make full use of information for encoding. In the process module, the final processing of the encoding module is completed through six 3×3 convolutional layers, and then the features are output to the decoding module. In the decoding module, features are alternately fed into the UpConv and IWT modules multiple times. The UpConv module consists of two 3×3 convolutional layers with dilation factors corresponding to DownConv and a ReLU activation function. The IWT module performs inverse wavelet transform along the channel dimension to complete the upsampling process. Between the encoding and decoding modules, an improved Skip-connection module is used to fuse encoded and decoded features at different scales. This improved Skip-connection module employs a depthwise separable convolutional module for feature fusion, and the fusion process is as follows: ； In the formula, Indicates the first The encoded features obtained from the layer Indicates the first Decoding features obtained from the layer, This indicates that the encoded and decoded features are concatenated along the channel dimension. Table depth-separable convolution functions Indicates the first Layer depth can separate the parameters of the convolution kernel, and feature fusion yields the first... Layer decoding features The shallowest decoded features in the upsampling process enter the Conv module, and the network output is obtained through convolution, which is the reconstructed brain source signal data; The following multi-objective loss function is designed for brain power imaging: ； In the formula, Represents the mean squared error function. This represents brain-derived signal data reconstructed using MSE calculations. With synthetic brain-derived signal data The loss, to The lead matrix is ​​obtained again through the standard boundary element head model in step 1). EEG signal data reconstructed by mapping to the observation signal space Simultaneously, the synthesized brain-derived signal data Performing the same mapping yields noise-free EEG signal data. Calculated via MSE and loss The study focuses on the accuracy of reconstructed brain-derived signal waveforms in activated regions, and calculates the reconstructed brain-derived signal data in activated regions using MSE (Mean Sequence of Effect). Synthesized brain-derived signal data in the activated regions loss The three losses are fused together using designed weights to form the overall loss function for training. , , express , , The corresponding weights This represents the final multi-objective loss function.

Images