A deep learning electroencephal source localization method and system based on residual iteration

Abstract

The application discloses a kind of deep learning brain electromagnetic source positioning method and system based on residual iteration, comprising: S1, current electromagnetic signal is obtained;S2, based on current electromagnetic signal, obtain preliminary source positioning result;S3, based on preliminary source positioning result, obtain sparse source positioning result;S4, based on sparse source positioning result, obtain maximum intensity activation source;S5, update current electromagnetic signal, repeat S2-S5 until reach preset condition, obtain final source positioning result. Carry out preliminary source positioning using sLORETA, provide a more rough brain source positioning result. Then the result of sLORETA is repaired using neural network, to restore the sparse source of accurate position, and as far as possible restore the intensity of activation source. The application comprehensively sLORETA's fast calculation and preliminary positioning ability, and the high-precision repair capability of neural network, so as to realize more accurate, more stable result in source positioning.

Description

Technical Field

[0001] This invention belongs to the fields of brain-computer interface and deep learning technology, and in particular relates to a deep learning brain power source localization method and system based on residual iteration. Background Technology

[0002] sLORETA (Low-Resolution Electromagnetic Tomography) is a technique for locating sources of EEG and MEG signals, falling under the category of solving the electromagnetic inverse problem. The inverse problem refers to deducing the spatial distribution of neural activity in the brain from measured electromagnetic field data; specifically, sLORETA is used to infer the location of active neural sources within the brain. It has become a widely used method for locating brain power sources, particularly in neuroscience research and clinical diagnosis, for studying brain function and pathological states. It has been used to explore cognitive and emotional processes, localize epileptic foci, and diagnose neurological disorders. The core challenge of source localization lies in the pathological nature of its inverse problem—reconstructing the distribution of internal brain current sources from limited surface data results in multiple possible solutions. Therefore, regularization and other constraints are necessary to narrow down the possible solution space. sLORETA solves the inverse problem using a regularization method and the least squares criterion, assuming a high correlation between the current densities of adjacent source points, thus smoothly reconstructing the brain's electrical activity. Its main advantage is the use of a low-resolution source localization scheme to improve the stability of the inverse problem solution.

[0003] The sLORETA technology has the following drawbacks:

[0004] sLORETA assumes that the current densities of adjacent sources are highly correlated, smoothly reconstructing the brain's electrical activity. Therefore, the resulting source localization results have low spatial resolution and cannot accurately distinguish spatially close sources of brain activity.

[0005] The sLORETA method stabilizes the solution through regularization and spatial smoothing assumptions, but these treatments can lead to an underestimation or distortion of the numerical strength of the source, making it impossible to accurately recover the original strength.

[0006] EEG signals are easily affected by environmental noise and artifacts (such as eye movements and muscle activity), and sLORETA is also easily affected when the signal-noise ratio is low, leading to inaccurate source localization. Therefore, there is an urgent need for a deep learning-based method and system for EEG source localization based on residual iteration. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a deep learning-based brain power source localization method and system based on residual iteration. This method is suitable for processing complex multi-source signals, and can progressively analyze and decompose each source in the signal, avoiding the confusion and errors caused by processing multiple source signals simultaneously in traditional methods.

[0008] This invention provides a deep learning-based brain power source localization method based on residual iteration, comprising:

[0009] S1. Obtain the current electromagnetic signal;

[0010] S2. Based on the current electromagnetic signal, obtain preliminary source localization results;

[0011] S3. Based on the preliminary source localization results, obtain the sparse source localization results;

[0012] S4. Based on the sparse source localization results, obtain the activation source with the maximum intensity;

[0013] S5. Update the current electromagnetic signal, repeat S2-S5 until the preset conditions are met, and obtain the final source location result.

[0014] Optionally, obtaining the preliminary source localization result based on the current electromagnetic signal includes:

[0015] The current electromagnetic signal is processed using the sLORETA method to obtain the preliminary source localization result.

[0016] Optionally, based on the preliminary source localization results, obtaining sparse source localization results includes:

[0017] The preliminary source localization result is input into a neural network model to obtain the sparse source localization result, wherein the neural network model is trained using a training set, which consists of EEG data and source data.

[0018] Optionally, the neural network comprises an encoder, an intermediate feature processing unit, a decoder, and skip connections;

[0019] The encoder is used to extract the latent feature vector from the preliminary source localization result;

[0020] The intermediate feature processing unit is used to process the potential feature vector;

[0021] The decoder is used to reconstruct the processed latent feature vector to obtain the sparse source localization result;

[0022] The skip connection is used to concatenate the features of the encoder portion with the features of the corresponding layer of the decoder.

[0023] Optionally, the encoder includes: a convolutional layer, an activation function, and a pooling layer;

[0024] The convolutional layer is used to extract features from the preliminary source localization results;

[0025] The activation function is used to learn the nonlinear characteristics of the preliminary source localization results;

[0026] The pooling layer is used to reduce the feature dimension and obtain the potential feature vector.

[0027] Optionally, processing the latent feature vector includes:

[0028] The latent feature vectors are processed by the GRU network to obtain the processed latent feature vectors;

[0029] Obtain the covariance matrix of the current electromagnetic signal;

[0030] The current covariance matrix is ​​used to obtain features with the same dimension as the latent feature vector using a neural network;

[0031] The processed latent feature vectors are fused with the current covariance matrix of the same dimension as the latent feature vectors to obtain the processed feature vectors.

[0032] Optionally, the decoder includes: several transposed convolutional layers and activation functions;

[0033] The transposed convolutional layer is used for upsampling to restore the processed latent feature vector to its original size;

[0034] The activation function is used to upsample and restore the processed latent feature vector to its original size.

[0035] Optionally, updating the current electromagnetic signal includes:

[0036] The maximum intensity activation source is recorded in the final localization result. After the maximum intensity activation source is recorded, the value of the maximum intensity activation source at the corresponding position will be changed to 0 in the sparse source localization result, and a residual sparse source matrix will be constructed.

[0037] The current electromagnetic signal is updated based on the residual sparse source matrix.

[0038] Optionally, the preset conditions include: reaching the maximum number of iterations or the L2 norm of the residual vector being less than a preset threshold.

[0039] The present invention also provides a deep learning brain source localization system based on residual iteration, comprising: a signal acquisition module, a preliminary source localization module, a neural network source localization recovery module, and a residual iterative source solution module;

[0040] The signal acquisition module is used to acquire electromagnetic signals;

[0041] The preliminary source localization module is used to perform preliminary source localization based on the electromagnetic signal;

[0042] The neural network source localization and recovery module is used to accurately process the preliminary source localization to obtain sparse sources;

[0043] The residual iterative source solving module is used to iteratively process the sparse source to obtain the source matrix.

[0044] Compared with the prior art, the present invention has the following advantages and technical effects:

[0045] 1. This invention employs a combination of the sLORETA method and a deep learning model. sLORETA is used for initial source localization, providing a relatively coarse result. Then, a neural network is used to refine the sLORETA result, recovering sparse sources at precise locations and restoring the intensity of activation sources as much as possible. The advantage of this combined approach lies in integrating the rapid computation and initial localization capabilities of sLORETA with the high-precision refining capabilities of neural networks, thereby achieving more accurate and stable results in source localization.

[0046] 2. This invention utilizes a residual iteration method to further optimize the sparse source estimation obtained from the neural network source localization and recovery module. A significant advantage of residual iteration in brain power source localization is its iterative optimization capability. In each iteration, the algorithm only solves for the most probable activation location; this step-by-step approximation method makes the localization results more accurate. Through multiple iterations, the algorithm can effectively reduce estimation errors and gradually approximate the true current source distribution. This method is particularly suitable for processing complex multi-source signals, capable of progressively analyzing and decomposing each source in the signal, avoiding the confusion and errors caused by processing multiple source signals simultaneously in traditional methods. Furthermore, residual iteration has strong noise resistance. A single neural network source localization and recovery result cannot completely eliminate the interference of various noises and artifacts; residual iteration can progressively filter out noise influences in each iteration, extracting more stable and reliable source signals. Attached Figure Description

[0047] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0048] Figure 1 This is a flowchart of a deep learning brain power source localization method based on residual iteration according to an embodiment of the present invention;

[0049] Figure 2This is a structural diagram of the neural network source localization and recovery module according to an embodiment of the present invention. Detailed Implementation

[0050] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0052] This embodiment proposes a deep learning-based brain power source localization method based on residual iteration, such as... Figure 1 As shown, it specifically includes:

[0053] S1. Obtain the current electromagnetic signal;

[0054] S2. Based on the current electromagnetic signal, obtain preliminary source localization results;

[0055] S3. Based on the preliminary source localization results, obtain the sparse source localization results;

[0056] S4. Based on the sparse source localization results, obtain the activation source with the maximum intensity;

[0057] S5. Update the current electromagnetic signal, repeat S2-S5 until the preset conditions are met, and obtain the final source localization result.

[0058] Furthermore, based on the current electromagnetic signal, the preliminary source localization results include:

[0059] The sLORETA method is used to process the current electromagnetic signal to obtain preliminary source localization results.

[0060] Specifically, the sLORETA method's advantage lies in using a normalization process to overcome the bias of MNE (Mechanical Electrification Array) towards weak and surface sources. First, MNE tends to select sources with lower current densities during the solution process, leading to a bias towards weak sources. Furthermore, due to the conduction characteristics of EEG and MEG signals on the scalp, MNE is more likely to locate surface sources near the scalp, neglecting activity in deeper brain regions. sLORETA introduces a normalization matrix to adjust the conduction matrix, ensuring that the current density at each source point is normalized by its standard error, thus balancing the differences between different source points. In this way, sLORETA effectively reduces the bias towards weak and surface sources, enhancing its ability to locate deep and strong brain sources. The low computational complexity of the sLORETA algorithm is also one of its advantages. Compared to some source localization algorithms that require complex mathematical operations and high computational resources, sLORETA's computation process is relatively simple and efficient. This low computational complexity makes sLORETA particularly suitable for real-time applications and large-scale data processing. However, while L2 regularization improves the robustness of source localization to some extent, it also leads to overly smoothed results. This smoothing results in a broad distribution of source localization rather than a concentration at specific points or small regions. This means that sLORETA's source localization results are typically a rather vague distribution of source intensity, making it difficult to accurately pinpoint the location of individual neurons. This smoothing effect is particularly detrimental in scenarios requiring high-precision localization, such as fine structural-functional analysis or clinical surgical planning. Furthermore, sLORETA fails to accurately recover the numerical intensity of source activity, which is another significant drawback. While the standardization process used by sLORETA improves the comparability of results under different experimental conditions, it ignores the actual numerical intensity of source activity. The standardization process involves standardizing the source current density using standard errors. While this eliminates measurement differences between different experiments or individuals, it also leads to the loss of numerical intensity information about source activity. The final result presents a relative value rather than an absolute value, making it difficult for sLORETA to provide the true numerical intensity of source activity. This is a significant limitation in quantitative analysis and functional assessment, especially in studies requiring comparisons of source activity intensity under different conditions, where sLORETA cannot provide accurate numerical data. This inability to recover numerical source activity intensity also affects the understanding and interpretation of brain functional activity. In scientific research, information on the intensity of source activity is crucial for understanding brain region function, neural network connectivity, and brain activity patterns. For example, when studying activation patterns in brain regions, accurate source activity intensity helps researchers distinguish differences in brain region activity under different conditions and explore their neural mechanisms in depth. sLORETA only provides relative intensity, which to some extent limits the depth and precision of the research. The algorithm implementation steps of sLORETA are shown in Table 1:

[0061] Table 1

[0062]

[0063] Furthermore, based on the preliminary source localization results, the sparse source localization results are obtained as follows:

[0064] The preliminary source localization results are input into the neural network model to obtain sparse source localization results. The neural network model is trained using a training set, which consists of EEG data and source data.

[0065] Specifically, this embodiment uses simulated EEG data and source data as the training set to train the neural network model of this module. To effectively train the neural network model, this embodiment chooses to use the Adam optimizer to optimize the convolutional filter, weights, and biases. The Adam optimizer is widely used in the field of deep learning due to its good adaptability and efficiency. The Adam optimizer combines the advantages of momentum optimization and RMSProp, and can adaptively adjust the learning rate to handle sparse and noisy gradients, making it perform well in various situations. Specifically, the Adam optimizer adjusts the learning rate based on the exponentially weighted average of past gradients. This allows for rapid convergence with a larger learning rate in the early stages of training, while gradually decreasing the learning rate in the later stages to ensure that the model can find a globally optimal or near-global optimal solution. In this embodiment, the initial learning rate of the Adam optimizer is set to an initial value to quickly reduce the loss function value in the early stages of training. As the number of iterations increases, the learning rate will gradually decrease, which can avoid fluctuations in the loss function due to an excessively large learning rate in the later stages of training. Other parameters of the optimizer refer to the default settings suggested by the authors. The loss function used for model training is the mean squared error (MSE) loss function. The MSE loss function measures model performance by calculating the squared difference between predicted and true values. Using MSE as a loss function in EEG source localization helps quantify the error between the model's predicted source location and the actual source location.

[0066] Furthermore, such as Figure 2 As shown, the neural network consists of an encoder, intermediate feature processing units, a decoder, and skip connections.

[0067] The encoder is used to extract latent feature vectors from the preliminary source localization results;

[0068] Intermediate feature processing unit, used to process latent feature vectors;

[0069] The decoder is used to reconstruct the processed latent feature vectors to obtain the sparse source localization results;

[0070] Skip connections are used to concatenate features from the encoder portion with features from the corresponding layer of the decoder.

[0071] Furthermore, the encoder includes: convolutional layers, activation functions, and pooling layers;

[0072] Convolutional layers, encoders include: convolutional layers, activation functions, and pooling layers;

[0073] Convolutional layers are used to extract features from the preliminary source localization results;

[0074] An activation function is used to learn the nonlinear characteristics of the initial source localization results;

[0075] Pooling layers are used to reduce feature dimensionality and obtain potential feature vectors.

[0076] Specifically, the encoding part of this module, like the encoder part in the U-Net model, follows a downsampling path. The encoding part has multiple layers, each consisting of convolutional layers, activation functions, and pooling layers. The encoding part uses convolution to extract features from the input source data and utilizes pooling to continuously reduce the scale of the features, mapping high-dimensional source vectors to low-dimensional feature spaces to obtain latent feature vectors. The encoder part can effectively process and extract complex features from the preliminary sLORETA localization results. Preliminary sLORETA localization results typically contain a large amount of noise and nonlinear signals, making it difficult for traditional feature extraction methods to accurately capture useful information. Through deep convolutional neural networks, the encoder can extract multi-scale features from the preliminary sLORETA localization results layer by layer, thereby capturing valuable information related to brain electrical activity. The local receptive fields and shared weight mechanism of the convolutional layers enable them to efficiently identify and extract spatial information from EEG signals. Furthermore, the pooling layers of the encoder play a simplifying and enhancing role in the feature extraction process. Pooling operations reduce the spatial size of the feature map through downsampling, lowering computational complexity and memory requirements while retaining the most salient feature information. This not only improves the model's efficiency but also enhances the robustness of features, making them less susceptible to noise and minor variations. This property of pooling layers is particularly useful for brain-derived data, as sLORETA localization results often contain significant noise and physiological artifacts.

[0077] Further processing of the latent feature vectors includes:

[0078] The latent feature vectors are processed by the GRU network to obtain the processed latent feature vectors;

[0079] Obtain the covariance matrix of the current electromagnetic signal;

[0080] The current covariance matrix is ​​used to obtain features with the same dimension as the latent feature vector using a neural network;

[0081] The processed latent feature vectors are fused with the current covariance matrix of the same dimension as the latent feature vectors to obtain the processed feature vectors.

[0082] Specifically, the intermediate feature processing section of this module mainly focuses on feature fusion and extraction of temporal information. This embodiment also utilizes EEG covariance information as prior knowledge input into the neural network, fusing it with the latent features extracted by the decoding section. This aims to help the subsequent decoding section more accurately reconstruct the sparse source. Specifically, the EEG covariance information is flattened through a flattening layer, then reshaped to the same dimension as the latent features after passing through multiple linear layers and activation functions. The latent features learn temporal information through a 3-layer GRU network, then undergo contact fusion, and finally are fed into the decoding section. The covariance matrix, as prior information, provides prior knowledge of the correlation between EEG signals, improving the model's localization ability. In the brain source localization task, EEG data is recorded by multiple electrodes. The signal at each electrode location not only reflects local brain activity but also contains certain global information. The covariance matrix describes the linear correlation between these signals, revealing the cooperative patterns and spatial relationships between signals from different electrodes. This rich contextual information helps deep learning models better understand and utilize the inherent structure of the data, improving the accuracy of source localization. Furthermore, the covariance matrix has significant advantages in noise reduction and feature extraction. EEG signals typically contain a large amount of noise and interference, such as electromyographic noise, eye movement artifacts, and electromagnetic interference, which can severely affect the accuracy of source localization. By learning prior information about covariance, deep learning models can effectively identify and filter out noise components, extracting features that reflect real brain activity. Integrating GRU into the U-Net model to process intermediate features enhances the RI-DSLNet framework's ability to capture temporal information. First, the U-Net model extracts rich spatial features through convolutional and pooling layers, but for highly temporally correlated data types like brain source signals, simple spatial feature extraction is insufficient. Brain source signals are time-series data of brain activity, and signals at different time points have important dependencies. Therefore, introducing GRU into the intermediate feature processing of U-Net can fully utilize the temporal information in brain source signals. When processing intermediate features, GRU uses its gating mechanism to effectively control information flow and state updates. GRU's reset and update gate mechanisms enable it to selectively remember and forget information, which has significant advantages in processing important and irrelevant information in long-term series data. In brain source localization tasks, the sLORETA localization results at different time points may contain noise and interference. Through the gating mechanism of GRU, the model can automatically filter out this irrelevant information and retain key features related to brain electrical activity, thereby improving the accuracy and reliability of brain source localization.

[0083] Furthermore, the decoder includes: several transposed convolutional layers and activation functions;

[0084] Transposed convolutional layers are used for upsampling to restore the processed latent feature vectors to their original size;

[0085] The activation function is used for upsampling to restore the processed latent feature vector to its original size.

[0086] Specifically, the decoding part of this module, like the decoder part in the U-Net model, follows an upsampling path. The main purpose of this decoding part is to reconstruct the processed latent features back to the sparse source, that is, to map the latent feature vectors back to the high-dimensional sparse source data space. The decoding part consists of multiple transposed convolutional layers and activation functions. This progressive upsampling process ensures that the final output brain power localization map has high resolution, allowing the localization results to reflect the neural activity of different brain regions in detail. The skip connection part of this module concatenates the features from the encoder part with the features from the corresponding layers of the decoder. The first important function of skip connections is to preserve and pass high-resolution detail information. In the U-Net model, the encoder extracts features through a series of convolution and pooling operations, while progressively reducing the spatial resolution of the feature maps. While this compression process helps extract high-level abstract features, it also loses some detail information. Skip connections compensate for this deficiency by directly passing the intermediate feature maps of the encoder to the corresponding layers of the decoder. Furthermore, skip connections can enhance the effect of feature fusion, improving the model's learning ability and performance. Skip connections, by directly fusing shallow and deep features, not only alleviate the vanishing gradient problem but also enrich feature representations, enabling the model to better learn brain electrical information at different levels. Finally, skip connections also have advantages in assisting feature interpretation and visualization, making it easier to track and analyze the fusion process of features at different levels.

[0087] Furthermore, updating the current electromagnetic signal includes:

[0088] 1. In S4, find the source with the highest average activation intensity at t time points from the sparse source localization results (i.e., the sparse source matrix), record it, and add it to the final source localization results.

[0089] 2. After recording the value of the maximum intensity activation source vector, the value of the maximum intensity activation source vector at that position will be changed to 0 in the current sparse source localization result, thus obtaining the residual sparse source matrix.

[0090] 3. Update the current EEG / MEG signal by using the guided field matrix to generate a new residual EEG / MEG signal from the residual sparse source matrix. (That is, subtract the contribution of the currently estimated activation position from the current electromagnetic signal to obtain the new electromagnetic signal).

[0091] Furthermore, the preset conditions include: reaching the maximum number of iterations and the norm of the residual vector satisfying the regularization matrix.

[0092] Specifically, in the restored sparse source estimation, the location of the activation source with the highest intensity value is accurate, while the accuracy of the location of the activation source with the second highest intensity value decreases, meaning that location becomes more difficult and complex when multiple sources are activated. Since the location of the activation source with the highest intensity value is accurate in sparse source estimation, this invention uses a cyclical method to select only the most likely activation source each time. For all time points (t = 1 → T), the location of the most likely activation source is found, and the intensity value at that location is recorded in a new vector X. new In, that is, X new (idx, 1: T) = X(idx, 1: T). Then, the residual EEG vector Y is calculated using the guiding field matrix L, i.e., the j-th iteration, as shown below:

[0093] Y j =Y-LX new

[0094] Among them, Y j Let X be the residual EEG vector obtained in the j-th iteration, L be the guiding field matrix, and X be the vector vector. new To record the vector of the activation source. After obtaining the residual EEG vector Yj, Y is... j Re-enter the data into the RI-DSLNet framework of this invention, select the j-th most likely activated source, and update X. new In the vector, until the condition ends, i.e., ||Y j ||2 is less than the threshold or the maximum number of iterations has been reached. After the loop ends, X new This is the final source estimation result.

[0095] This embodiment utilizes a residual iteration method to further optimize the sparse source estimation obtained by the neural network source localization and recovery module. Through successive iterations, each generation only finds the most probable activation locations, saves them to a new vector, and generates a new residual EEG for the next iteration. In each iteration, the residual iteration solution also generates a new residual EEG signal. This step is achieved by subtracting the contribution of the currently estimated activation location from the original EEG signal. Specifically, the algorithm calculates the influence of the current source distribution at the current activation location on the EEG signal and subtracts it from the original signal to obtain the residual signal. This residual EEG signal represents the uninterpreted portion, containing information about other current sources that have not yet been localized. By processing the residual EEG signal, the algorithm can continue to search for new activation locations in the next iteration, gradually improving the localization results and making them closer to the actual brain source distribution.

[0096] In this embodiment, a significant advantage of residual iterative solution in brain source localization is its iterative optimization capability. In each iteration, the algorithm only solves for the most probable activation location; this stepwise approximation method makes the localization results more accurate. Through multiple iterations, the algorithm can effectively reduce estimation errors and gradually approximate the true current source distribution. This method is particularly suitable for processing complex multi-source signals because it can progressively analyze and decompose the individual sources in the signal, avoiding the confusion and errors caused by processing multiple sources simultaneously in traditional methods. Furthermore, residual iterative solution has strong noise resistance. A single neural network source localization recovery result cannot completely eliminate the interference of various noises and artifacts; residual iterative solution can progressively filter out noise effects in each iteration, extracting more stable and reliable source signals. It is worth noting that the entire RI-DSLNet module only needs to train the neural network source localization recovery module. After training, it can be tested using simulated EEG data and real EEG data, such as... Figure 1 As shown in Table 2, the recorded EEG signals are input into the trained RI-DSLNet framework, which will output the estimated source time series X = [x1, ..., x2] according to Table 2. t , ..., x T ]∈R m×T Furthermore, sparse brain source localization assumes that only a small number of brain regions show significant activity at a given time point, while most brain regions remain quiescent or inactive. The iterative residual solution, by finding only one most probable activation location in each iteration, is well-suited for handling sparse signals. Finding one source and updating the residual signal each time helps to accurately identify and localize sparse EEG activity sources. This iterative strategy not only effectively reduces the complexity of the problem but also allows for more precise focus on the most promising activation location at each step, thereby improving overall localization accuracy.

[0097] Table 2

[0098]

[0099] This embodiment also discloses a deep learning brain power source localization system based on residual iteration, including: a signal acquisition module, a preliminary source localization module, a neural network source localization recovery module, and a residual iterative source solution module;

[0100] Signal acquisition module, used to acquire electromagnetic signals;

[0101] The preliminary source localization module is used to perform preliminary source localization based on electromagnetic signals.

[0102] The neural network source localization and recovery module is used to accurately process the preliminary source localization and obtain sparse sources;

[0103] The residual iterative source solving module is used to iteratively process sparse sources to obtain the source matrix.

[0104] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based brain power source localization method based on residual iteration, characterized in that, include: S1. Obtain the current electromagnetic signal; S2. Based on the current electromagnetic signal, obtain preliminary source localization results; S3. Based on the preliminary source localization results, obtain the sparse source localization results; Based on the preliminary source localization results, obtaining sparse source localization results includes: The preliminary source localization result is input into a neural network model to obtain the sparse source localization result, wherein the neural network model is trained using a training set, which consists of EEG data and source data. The neural network comprises an encoder, an intermediate feature processing unit, a decoder, and skip connections. The encoder is used to extract the latent feature vector from the preliminary source localization result; The intermediate feature processing unit is used to process the potential feature vector; The decoder is used to reconstruct the processed latent feature vector to obtain the sparse source localization result; The skip connection is used to concatenate the features of the encoder part with the features of the corresponding layer of the decoder. The encoder includes: a convolutional layer, an activation function, and a pooling layer; The convolutional layer is used to extract features from the preliminary source localization results; The activation function is used to learn the nonlinear characteristics of the preliminary source localization results; The pooling layer is used to reduce the feature dimension and obtain the latent feature vector; Processing the latent feature vector includes: The latent feature vectors are processed by the GRU network to obtain the processed latent feature vectors; Obtain the covariance matrix of the current electromagnetic signal; The current covariance matrix is ​​used to obtain features with the same dimension as the latent feature vector using a neural network; The processed latent feature vectors are fused with the current covariance matrix of the same dimension as the latent feature vectors to obtain the processed feature vectors. The decoder includes: several transposed convolutional layers and activation functions; The transposed convolutional layer is used for upsampling to restore the processed latent feature vector to its original size; The activation function is used to upsample and restore the processed latent feature vector to its original size; S4. Based on the sparse source localization results, obtain the activation source with the maximum intensity; S5. Update the current electromagnetic signal, repeat S2-S5 until the preset condition is met, and obtain the final source location result; Updating the current electromagnetic signal includes: The maximum intensity activation source is recorded in the final localization result. After the maximum intensity activation source is recorded, the value of the maximum intensity activation source at the corresponding position will be changed to 0 in the sparse source localization result, and a residual sparse source matrix will be constructed. The current electromagnetic signal is updated based on the residual sparse source matrix.

2. The deep learning-based brain power source localization method based on residual iteration according to claim 1, characterized in that, Based on the current electromagnetic signal, obtaining the preliminary source localization result includes: The current electromagnetic signal is processed using the sLORETA method to obtain the preliminary source localization result.

3. The deep learning-based brain power source localization method based on residual iteration according to claim 1, characterized in that, The preset conditions include: reaching the maximum number of iterations or the L2 norm of the residual vector being less than a preset threshold.

4. A deep learning-based brain power source localization system based on residual iteration, used to implement the method as described in any one of claims 1-3, characterized in that, It includes: a signal acquisition module, a preliminary source localization module, a neural network source localization recovery module, and a residual iterative solution source module; The signal acquisition module is used to acquire electromagnetic signals; The preliminary source localization module is used to perform preliminary source localization based on the electromagnetic signal; The neural network source localization and recovery module is used to accurately process the preliminary source localization to obtain sparse sources; The residual iterative source solving module is used to iteratively process the sparse source to obtain the source matrix.

Images