Deep learning brain power positioning method and system based on residual iteration

Abstract

The invention discloses a deep learning brain power supply positioning method and system based on residual error iteration. The method comprises the following steps: S1, acquiring a current electromagnetic signal; s2, acquiring a preliminary source positioning result based on the current electromagnetic signal; s3, obtaining a sparse source positioning result based on the initial source positioning result; s4, obtaining a maximum intensity activation source based on a sparse source positioning result; and S5, updating the current electromagnetic signal, repeating the steps S2-S5 until a preset condition is met, and obtaining a final source positioning result. The sLORTA is used for preliminary source positioning, and a rough brain source positioning result is provided. Then, a neural network is used for repairing a sLORTA result to recover a sparse source at an accurate position, and the intensity of an activation source is recovered as much as possible. According to the method, the rapid calculation and preliminary positioning capability of the sLORTA and the high-precision repair capability of the neural network are integrated, so that a more accurate and more stable result is realized in source positioning.

Description

Technical Field

[0001] The present invention belongs to the field of brain-computer interface and deep learning technology, and in particular, relates to a deep learning brain source localization method and system based on residual iteration. Background Art

[0002] sLORETA (Low-Resolution Electromagnetic Tomography) is a technology for localizing the source of EEG and MEG signals, and belongs to the category of solving electromagnetic inverse problems. The inverse problem refers to the inference of the spatial distribution of neural activity in the brain from the measured electromagnetic field data. Specifically, sLORETA is used to infer the location of active neural sources inside the brain. It has become a widely used method for localizing brain power sources, especially in neuroscience research and clinical diagnosis for studying brain functional activities and pathological states. It is used to explore the diagnosis of neurological diseases such as cognitive and emotional processes of the brain and epileptic focus localization. The core challenge of source localization lies in the pathological nature of its inverse problem, that is, reconstructing the distribution of current sources inside the brain from limited surface data, and there are multiple possible solutions. Therefore, regularization and other constraints must be relied on to narrow the possible solution space. sLORETA uses a regularization method to solve the inverse problem using the least squares criterion and assumes that the current density of adjacent source points is highly correlated, thereby smoothly reconstructing the brain's power source activity. Its main advantage is to use a low-resolution source localization scheme to improve the stability of the inverse problem solution.

[0003] sLORETA technology has the following disadvantages:

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

[0005] The sLORETA method stabilizes the solution through regularization and spatial smoothness assumptions, but these treatments will cause the numerical intensity of the source to be underestimated or distorted, and the original intensity cannot be accurately restored.

[0006] EEG signals are easily affected by environmental noise and artifacts (such as eye movements, muscle activity, etc.). sLORETA is easily affected when the signal noise is relatively low, resulting in inaccurate source localization. Therefore, a deep learning method and system for EEG source localization based on residual iteration is urgently needed. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a deep learning EEG source localization method and system based on residual iteration. The method is suitable for processing complex multi-source signals, and can gradually analyze and decompose each source in the signal, avoiding the confusion and errors caused by the simultaneous processing of multi-source signals in traditional methods.

[0008] The present invention provides a deep learning method for locating brain power sources based on residual iteration, comprising:

[0009] S1. Obtain current electromagnetic signal;

[0010] S2. obtaining a preliminary source positioning result based on the current electromagnetic signal;

[0011] S3. Based on the preliminary source positioning result, obtaining a sparse source positioning result;

[0012] S4. Based on the sparse source positioning result, obtaining the maximum intensity activation source;

[0013] S5. Update the current electromagnetic signal, repeat S2-S5, until a preset condition is met, and obtain a final source positioning result.

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

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

[0016] Optionally, based on the preliminary source positioning result, obtaining a sparse source positioning result 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 by a training set, and the training set is electroencephalogram (EEG) data and source data.

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

[0019] The encoder is used to extract potential feature vectors in 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 potential feature vector to obtain the sparse source positioning result;

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

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

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

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

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

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

[0028] The potential feature vector is processed by the GRU network to obtain the processed potential feature vector;

[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 potential feature vector using a neural network;

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

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

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

[0034] The activation function is used for upsampling to restore the processed potential 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 positioning result. After the maximum intensity activation source is recorded, the value of the maximum intensity activation source at the corresponding position in the sparse source positioning result is changed to 0, and a residual sparse source matrix is ​​constructed;

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

[0038] Optionally, the preset condition includes: reaching a maximum number of iterations or an L2 norm of the residual vector is 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 iteration source solution module;

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

[0041] The preliminary source positioning module is used to perform preliminary source positioning according to the electromagnetic signal;

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

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

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

[0045] 1. The present invention uses a combination of the sLORETA method and a deep learning model, uses sLORETA for preliminary source localization, and provides a relatively rough brain source localization result. Then the results of sLORETA are repaired using a neural network to restore the sparse sources at the precise location and restore the intensity of the activated source as much as possible. The advantage of this combined method is that it combines the fast calculation and preliminary positioning capabilities of sLORETA and the high-precision repair capabilities of the neural network, thereby achieving more accurate and stable results in source localization.

[0046] 2. The present invention further optimizes the sparse source estimation obtained by the neural network source localization recovery module using the residual iteration method. A significant advantage of the residual iterative solution in brain source localization is its iterative optimization capability. In each iteration, the algorithm only solves the most likely activation position. This step-by-step approximation method makes the positioning result more accurate. Through multiple iterative optimizations, the algorithm can effectively reduce the estimation error and gradually approach the true current source distribution. This method is particularly suitable for processing complex multi-source signals. It can gradually parse and decompose each source in the signal, avoiding the confusion and errors caused by the simultaneous processing of multi-source signals in traditional methods. In addition, the residual iterative solution has a strong noise resistance. The interference of various noises and artifacts cannot be completely eliminated in a single neural network source localization recovery result. The residual iterative solution can gradually filter out the noise effect in each iteration and extract a more stable and reliable source signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0048] Figure 1 This is a flow chart of a deep learning method for locating brain power sources based on residual iteration according to an embodiment of the present invention;

[0049] Figure 2It is a structural diagram of a neural network source positioning and recovery module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

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

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

[0053] S1. Obtain current electromagnetic signal;

[0054] S2. Obtain preliminary source positioning results based on current electromagnetic signals;

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

[0056] S4, based on the sparse source positioning results, obtaining the maximum intensity activation source;

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

[0058] Furthermore, based on the current electromagnetic signal, obtaining a preliminary source positioning result includes:

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

[0060] Specifically, the advantage of the sLORETA method is that it uses a standardization process to overcome the disadvantage of MNE, which prefers weak and surface sources. First, brain electromagnetic MNE tends to select sources with smaller current density during the solution process, which will cause it to prefer weak sources. In addition, due to the conduction characteristics of EEG and MEG signals on the scalp, MNE is more likely to locate surface sources close to the scalp, while ignoring the activities of deep brain areas. sLORETA adjusts the conduction matrix by introducing a standardization matrix so that the current density of each source point is standardized by its standard error, thereby balancing the differences between different source points. In this way, sLORETA effectively weakens the preference for weak and surface sources and enhances the ability to locate deep brain sources and strong sources. The low computational complexity of the sLORETA algorithm is also one of its advantages. Compared with some source localization algorithms that require complex mathematical operations and high computing resources, the calculation process of sLORETA is relatively simple and efficient. This low computational complexity makes sLORETA particularly suitable for real-time applications and large-scale data processing. However, since L2 regularization is used as a constraint, this regularization improves the robustness of source localization to a certain extent, but also leads to over-smoothing of the results. The smoothing process makes the source localization results appear as a large-scale distribution rather than concentrated in a specific point or small area. This means that the source localization results of sLORETA are usually a relatively vague source intensity distribution map, and it is difficult to accurately determine the location of a single neural source. This smoothing effect is particularly unfavorable in the study of scenarios that require high-precision positioning, such as fine structural functional analysis or clinical surgical planning. In addition, sLORETA cannot correctly restore the numerical intensity of source activity, which is also one of its significant defects. The standardization process used by sLORETA, although it improves the comparability of results under different experimental conditions, ignores the actual numerical intensity of source activity. The standardization process involves standardizing the standard error of the source current density. Although this process can eliminate measurement differences between different experiments or individuals, it also leads to the loss of information on the numerical intensity of 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 major limitation in quantitative analysis and functional assessment, especially in studies that need to compare the intensity of source activity under different conditions, sLORETA cannot provide accurate numerical basis. This problem of being unable to restore the numerical intensity of source activity will also affect the understanding and interpretation of brain functional activities. In scientific research, the intensity information of source activity is of great significance for understanding brain area functions, neural network connections, and brain activity patterns. For example, when studying the activation patterns of brain areas, accurate source activity intensity can help researchers distinguish differences in brain area activity under different conditions and explore their neural mechanisms in depth. However, sLORETA only provides relative intensity, which to some extent limits the depth and accuracy of the research. The algorithm implementation steps of sLORETA are shown in Table 1:

[0061] Table 1

[0062]

[0063] Further, based on the preliminary source localization result, obtaining a sparse source localization result includes:

[0064] The preliminary source localization result is input into the neural network model to obtain a sparse source localization result, wherein the neural network model is trained by a training set, and the training set is the EEG data and source data.

[0065] Specifically, this embodiment uses the simulated EEG data and source data as training sets to train the neural network model of this module. In order to effectively train the neural network model, this embodiment selects to use the Adam optimizer to optimize the convolution filter, weights and biases. The Adam optimizer is widely used in the field of deep learning for its good adaptability and efficiency. The Adam optimizer combines the advantages of momentum optimization and RMSProp, and can adaptively adjust the learning rate, handle sparse gradients and noisy gradients, so that it performs well in various situations. Specifically, the Adam optimizer adjusts the learning rate according to the exponential weighted average of past gradients, so that a larger learning rate can be used to converge quickly in the early stage of training, and the learning rate is gradually reduced in the later stage of training to ensure that the model can find the global optimal or near-global optimal solution. In this embodiment, the initial learning rate of the Adam optimizer is set to an initial value so as to quickly reduce the loss function value in the early stage of training. As the number of iterations increases, the learning rate will gradually decrease, so as to avoid the loss function fluctuation caused by excessive learning rate in the later stage of training. The other parameters of the optimizer refer to the default settings recommended by the author. The loss function used in model training is the mean square error (MSE) loss function. The MSE loss function measures the performance of the model by calculating the squared difference between the predicted value and the true value. Using MSE as a loss function in EEG source localization helps quantify the error between the source location predicted by the model and the actual source location.

[0066] Furthermore, if Figure 2 As shown, the neural network includes: an encoder, an intermediate feature processing unit, a decoder and a jump connection;

[0067] An encoder, used to extract latent feature vectors from preliminary source localization results;

[0068] An intermediate feature processing unit, used to process the potential feature vector;

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

[0070] Skip connection is used to concatenate the features of the encoder part with the features of the corresponding layer of the decoder.

[0071] Further, the encoder includes: a convolutional layer, an activation function, and a pooling layer;

[0072] Convolutional layer, the encoder includes: convolutional layer, activation function and pooling layer;

[0073] Convolutional layer, used to extract features from preliminary source localization results;

[0074] Activation function, used to learn the nonlinear features of the preliminary source localization results;

[0075] Pooling layer, used to reduce feature dimensions and obtain potential feature vectors.

[0076] Specifically, the encoding part of this module is the same as the encoder part in the U-Net model, which is a downsampling path. The encoding part has multiple layers, each of which consists of a convolution layer, an activation function, and a pooling layer. The encoding part uses convolution to extract the features of the input source data, and uses pooling to continuously reduce the scale of the features, that is, mapping the high-dimensional source vector to the low-dimensional feature space to obtain the potential feature vector. The encoder part can effectively process and extract the complex features in the sLORETA preliminary localization results. The sLORETA preliminary localization results usually contain a lot of noise and nonlinear signals, which makes it difficult for traditional feature extraction methods to accurately capture the useful information. Through deep convolutional neural networks, the encoder can extract multi-scale features of the sLORETA preliminary localization results layer by layer, thereby capturing valuable information related to brain power activity. The local receptive field and shared weight mechanism of the convolutional layer enable it to efficiently identify and extract spatial information in EEG signals. In addition, the pooling layer of the encoder plays a role in simplifying and strengthening the feature extraction process. The pooling operation reduces the spatial size of the feature map by downsampling, reduces the computational complexity and memory requirements, and retains the most significant feature information. This not only improves the efficiency of the model, but also enhances the robustness of the features, making them less susceptible to noise and small changes. This feature of the pooling layer is particularly useful for brain source data, because the localization results of sLORETA often contain a lot of noise and physiological artifacts.

[0077] Furthermore, processing the potential feature vector includes:

[0078] The potential feature vector is processed by the GRU network to obtain the processed potential feature vector;

[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 potential feature vector using a neural network;

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

[0082] Specifically, the intermediate feature processing part of this module is mainly based on feature fusion and extraction of feature time series information. This embodiment also uses EEG covariance information as prior knowledge to input into the neural network, and fuses it with the potential features extracted by the decoding part, with the purpose of helping the subsequent decoding part to reconstruct sparse sources more accurately. Among them, the EEG covariance information is flattened by the flatten layer, and then reshaped into the same dimension as the potential feature after passing through multiple linear layers and activation functions. The potential feature learns the time series information through a 3-layer GRU network, then performs contact fusion, and finally sends it to the decoding part. The covariance matrix is ​​fused with the potential feature as prior information, providing prior knowledge of the correlation between EEG signals, which can improve the positioning ability of the model. In the brain power source localization task, EEG data is recorded by multiple electrodes, and the signal at each electrode position not only reflects local brain activity, but also contains certain global information. The covariance matrix describes the linear correlation between these signals, revealing the synergistic pattern and spatial relationship between different electrode signals. This rich contextual information can help deep learning models better understand and utilize the intrinsic structure of data and improve the accuracy of source localization. In addition, the covariance matrix has significant advantages in noise reduction and feature extraction. EEG signals usually contain a lot of noise and interference, such as electromyographic noise, eye movement artifacts, and electromagnetic interference, which can seriously affect the accuracy of source localization. By learning covariance prior information, the deep learning model can effectively identify and filter out the noise components and extract features that reflect real brain activity. GRU is integrated into the U-Net model to process intermediate features and enhance the ability of the RI-DSLNet framework to capture temporal information. First, the U-Net model extracts rich spatial features through convolutional layers and pooling layers, but for data types such as brain source signals that are highly temporally related, simple spatial feature extraction is not enough. 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 make full use of the temporal information in brain source signals. When processing intermediate features, GRU uses its gating mechanism to effectively control information flow and state updates. The reset gate and update gate mechanisms of GRU enable it to selectively remember and forget information, which has significant advantages in processing important and irrelevant information in long time series data. In the brain source localization task, 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 these irrelevant information and retain the key features related to EEG activity, thereby improving the accuracy and reliability of brain source localization.

[0083] Further, the decoder includes: a number of transposed convolutional layers and activation functions;

[0084] The transposed convolution layer is used to upsample and restore the processed latent feature vector to its original size;

[0085] Activation function, used for upsampling to restore the processed latent feature vector to its original size.

[0086] Specifically, the decoding part of this module is an upsampling path, just like the decoder part in the U-Net model. 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 vector back to the high-dimensional sparse source data space. The decoding part consists of multiple transposed convolutional layers and activation functions. This step-by-step upsampling process ensures that the final output EEG source localization map has high resolution, so that the localization results can reflect the neural activities of different brain regions in detail. The jump connection part of this module is to splice the features of the encoder part with the features of the corresponding layer of the decoder. The first important role of the jump connection is to retain and transmit high-resolution detail information. In the U-Net model, the encoder extracts features through a series of convolution and pooling operations, while gradually reducing the spatial resolution of the feature map. Although this compression process helps to extract high-level abstract features, it also loses some detail information. The jump connection makes up for this defect by directly passing the intermediate feature map of the encoder to the corresponding layer of the decoder. In addition, the jump connection can enhance the effect of feature fusion and improve the learning ability and performance of the model. By directly fusing shallow features with deep features, the jump connection not only alleviates the problem of gradient vanishing, but also enriches the feature representation, enabling the model to better learn brain power information at different levels. Finally, the jump connection also has advantages in assisting feature interpretation and visualization, making it easier to track and analyze the fusion process of features at different levels.

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

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

[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 positioning result to obtain the residual sparse source matrix.

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

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

[0092] Specifically, in the sparse source estimation after recovery, the location of the activation source with the maximum intensity value is estimated accurately, and the location accuracy of the activation source with the secondary intensity value is reduced, that is, the location will be more difficult and complicated when multiple sources are activated. Since the location of the activation source with the maximum intensity value is accurate in the sparse source estimation, the present invention will use a cyclic method to select only the most likely activated source each time. For all time points (t=1→T), find the position of the most likely activation source and record the intensity value of the position in the 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 conduction matrix L, i.e. the j-th loop is as follows:

[0093] Y j =Y-LX new

[0094] Among them, Y j is the residual EEG vector obtained in the jth cycle, L is the conduction field matrix, X new is the vector recording the activation source. After obtaining the residual EEG vector Yj, Y j Re-input into the RI-DSLNet framework of the present invention, select the jth most likely activated source, and update it to X new vector until the condition ends, i.e. ||Y j ||2 is less than the threshold or reaches the maximum number of cycles. After the cycle ends, X new is the final source estimation result.

[0095] This embodiment uses the residual iteration method to further optimize the sparse source estimate obtained by the neural network source localization and recovery module. Through successive iterative optimization, each generation only finds the most likely activation position, saves it in a new vector, and generates a new residual EEG for the next iteration. In each iteration, the residual iterative solution also generates a new residual EEG signal. This step is achieved by subtracting the contribution of the current estimated activation position from the original EEG signal. Specifically, the algorithm calculates the impact of the current source distribution of the current activation position on the EEG signal, and subtracts it from the original signal to obtain a residual signal. This residual EEG signal represents the unexplained part and contains information about other current sources that have not yet been located. By processing the residual EEG signal, the algorithm can continue to find new activation positions in the next iteration, gradually improving the positioning results to make 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 the most likely activation position. This step-by-step approximation method makes the positioning result more accurate. Through multiple iterative optimizations, the algorithm can effectively reduce the estimation error and gradually approach the true current source distribution. This method is particularly suitable for processing complex multi-source signals because it can gradually parse and decompose each source in the signal, avoiding the confusion and errors caused by the simultaneous processing of multiple source signals in traditional methods. In addition, residual iterative solution has strong noise resistance. The interference of various noises and artifacts cannot be completely eliminated in a single neural network source localization recovery result. The residual iterative solution can gradually filter out the noise influence in each iteration and extract a more stable and reliable source signal. It is worth noting that the entire RI-DSLNet module only needs to train the neural network source localization recovery module. After the training is completed, simulated EEG data and real EEG data can be used for testing, such as Figure 1 The recorded EEG signal is input into the trained RI-DSLNet framework, which will output the estimated source time series X = [x1, ..., x t , ..., x T ]∈R m×T . In addition, sparse EEG source localization assumes that only a small number of brain regions have significant activity at a certain point in time, while most brain regions remain static or in a state of low activity. The residual loop solution finds only one most likely activation location in each iteration, and the progressive approximation method is very suitable for processing sparse signals. Finding a source each time and updating the residual signal helps to accurately identify and locate sparse EEG activity sources. This step-by-step iterative strategy can not only effectively reduce the complexity of the problem, but also more accurately focus on the most potential activation location at each step, thereby improving the overall positioning accuracy.

[0097] Table 2

[0098]

[0099] This embodiment also discloses a deep learning brain 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 iteration source solution module;

[0100] A signal acquisition module, used for acquiring electromagnetic signals;

[0101] A preliminary source positioning module, used for performing preliminary source positioning according to electromagnetic signals;

[0102] Neural network source location recovery module, used to accurately process preliminary source location and obtain sparse sources;

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

[0104] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A deep learning method for localizing brain power sources based on residual iteration, characterized in that: include: S1. Obtain current electromagnetic signal; S2. obtaining a preliminary source positioning result based on the current electromagnetic signal; S3. Based on the preliminary source positioning result, obtaining a sparse source positioning result; S4. Based on the sparse source positioning result, obtaining the maximum intensity activation source; S5. Update the current electromagnetic signal, repeat S2-S5, until a preset condition is met, and obtain a final source positioning result.

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

3. According to the residual iteration based deep learning method for brain power source localization in claim 1, it is characterized in that: Based on the preliminary source positioning result, obtaining a sparse source positioning result 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 by a training set, and the training set is electroencephalogram (EEG) data and source data.

4. According to the residual iteration based deep learning method for brain power source localization according to claim 3, it is characterized in that: The neural network comprises: an encoder, an intermediate feature processing unit, a decoder and a skip connection; The encoder is used to extract potential feature vectors in 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 potential feature vector to obtain the sparse source positioning result; The jump connection is used to concatenate the features of the encoder part with the features of the corresponding layer of the decoder.

5. According to the residual iteration based deep learning method for brain power source localization according to claim 4, it is characterized in that: The encoder includes: a convolution layer, an activation function and a pooling layer; The convolution layer is used to extract features from the preliminary source localization result; The activation function is used to learn the nonlinear characteristics of the preliminary source localization result; The pooling layer is used to reduce the feature dimension and obtain the potential feature vector.

6. According to the residual iteration based deep learning method for brain power source localization in claim 4, it is characterized in that: Processing the potential feature vector includes: The potential feature vector is processed by the GRU network to obtain the processed potential feature vector; Obtain the covariance matrix of the current electromagnetic signal; The current covariance matrix is ​​used to obtain features with the same dimension as the potential feature vector using a neural network; The processed latent feature vector is fused with the current covariance matrix of the same dimension as the latent feature vector to obtain the processed feature vector.

7. The method for localizing brain power sources based on deep learning and residual iteration according to claim 4 is characterized in that: The decoder includes: a plurality of transposed convolutional layers and activation functions; The transposed convolution layer is used for upsampling to restore the processed potential feature vector to its original size; The activation function is used for upsampling to restore the processed potential feature vector to its original size.

8. The deep learning method for locating brain power sources based on residual iteration according to claim 1 is characterized in that: Updating the current electromagnetic signal includes: The maximum intensity activation source is recorded in the final positioning result. After the maximum intensity activation source is recorded, the value of the maximum intensity activation source at the corresponding position in the sparse source positioning result is changed to 0, and a residual sparse source matrix is ​​constructed; The current electromagnetic signal is updated based on the residual sparse source matrix.

9. The deep learning method for locating brain power sources based on residual iteration according to claim 8 is characterized in that: The preset conditions include: reaching a maximum number of iterations or an L2 norm of a residual vector being less than a preset threshold.

10. A deep learning brain power source localization system based on residual iteration, characterized in that: include: Signal acquisition module, preliminary source localization module, neural network source localization recovery module and residual iterative solution source module; The signal acquisition module is used to acquire electromagnetic signals; The preliminary source positioning module is used to perform preliminary source positioning according to the electromagnetic signal; The neural network source positioning recovery module is used to accurately process the preliminary source positioning to obtain sparse sources; The residual iterative solution source module is used to iteratively process the sparse source to obtain a source matrix.

Images