Threshold-adaptive deep expansion brain source imaging method and system

Abstract

The invention discloses a deep expansion brain source imaging method and system of an adaptive threshold, and relates to the technical field of brain source imaging, and the method comprises the steps: obtaining a training electroencephalogram signal observation matrix and a real source signal matrix; inputting the training electroencephalogram signal observation matrix into the initial iteration shrinkage threshold deep expansion network, and outputting a training reconstruction source signal matrix; determining a loss function value according to the real source signal matrix and the training reconstruction source signal matrix, and determining a target iterative shrinkage threshold depth expansion network by minimizing the loss function value; and an actually-measured electroencephalogram signal observation matrix is collected, the actually-measured electroencephalogram signal observation matrix is solved through the target iteration shrinkage threshold depth expansion network, and a target reconstruction source signal matrix is output. And a traditional ISTA algorithm is expanded into a trainable deep neural network, so that the brain source imaging reliability can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of brain source imaging, and in particular to a method and system for deep expansion brain source imaging with an adaptive threshold. Background Art

[0002] Brain source imaging technology is a technology that visualizes EEG activity in the brain. It is of great significance for cognitive neuroscience research, disease diagnosis and treatment, and brain-computer interface system development. It includes invasive and non-invasive methods. Invasive brain source imaging is a method of recording EEG activity by implanting devices such as electrodes or optical fibers. Since electrodes or optical fibers need to be implanted on the skull or under the skin, this traumatic operation may cause patients to have adverse reactions such as infection and bleeding. In addition, implants can usually only detect neuronal activity close to them, and cannot fully understand the activity of the entire brain area. At the same time, the measurement results are easily affected by the location and number of implants. Therefore, non-invasive brain source imaging technology based on electroencephalography (EEG) and magnetoencephalography (MEG) that detects human brain function with millisecond time resolution is widely used.

[0003] Source imaging with deep learning neural networks SIFNet network is a non-invasive brain source imaging technology based on EEG or MEG signals for feature extraction and source reconstruction, such as Figure 1 As shown in the figure, the SIFNet network is mainly based on the residual structure. After the convolution block, five residual blocks are cascaded, followed by a pooling layer and a fully connected layer. The output size is the same as the number of source regions. Each residual block contains two Conv layers and batch normalization is performed to eliminate the covariance shift during training. The original input of each residual block is added back to the output and fed to the next module through the rectified linear unit (ReLU) activation function, which converts the brain source imaging problem into a classification problem. Although the SIFNet network has improved the accuracy of source reconstruction to a certain extent through deep learning, it relies on the characteristics of convolution operations, which easily leads to the loss of signal details in some cases. It relies on simulation data and a large amount of labeled data, and its regularization parameters are usually fixed, which may limit the generalization ability of the network. The training and feature learning process of the network is usually a black box operation, which easily leads to low trust in the source reconstruction results. At the same time, the calculation is more complex, which reduces the reliability of brain source imaging as a whole. Summary of the invention

[0004] The present invention provides a method and system for deep unfolding brain source imaging with an adaptive threshold, which improves the technical problem of low reliability of brain source imaging in the existing source imaging method of deep learning neural network.

[0005] The first aspect of the present invention provides a method for deep unfolding brain source imaging with an adaptive threshold, comprising:

[0006] Obtain training EEG signal observation matrix and real source signal matrix;

[0007] The training EEG signal observation matrix is ​​used to input an initial iterative shrinkage threshold depth expansion network, and a training reconstructed source signal matrix is ​​output;

[0008] Determine a loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and determine a target iterative shrinkage threshold depth expansion network by minimizing the loss function value;

[0009] The measured EEG signal observation matrix is ​​collected, and the measured EEG signal observation matrix is ​​solved through a target iterative shrinkage threshold depth expansion network to output a target reconstructed source signal matrix.

[0010] Optionally, obtaining a training EEG signal observation matrix and a real source signal matrix includes:

[0011] Acquire multiple real source signals;

[0012] Performing simulation based on each of the real source signals through the Brainstorm toolbox, and outputting a real source signal matrix and a training lead field matrix;

[0013] Based on the training lead field matrix, the real source signal matrix is ​​linearly transformed and noise is added to generate a training EEG signal observation matrix.

[0014] Optionally, the step of using the training EEG signal observation matrix as input into an initial iterative shrinkage threshold depth expansion network and outputting a training reconstructed source signal matrix comprises:

[0015] Construct a signal loss function about the EEG signal observation matrix;

[0016] Iteratively perform gradient descent operation and soft threshold operation on the signal loss function based on the training EEG signal observation matrix, and output a training reconstructed source signal matrix;

[0017] The signal loss function includes:

[0018] ;

[0019] In the formula, is the signal loss function, is the square of the Frobenius norm, is the lead field matrix, is the source signal matrix, is the EEG signal observation matrix, For the source signal, is the number of source signals, is the regularization parameter, is the source signal matrix source signal, is the number of iterations;

[0020] The gradient descent operation includes:

[0021] ;

[0022] In the formula, is the gradient, is the Lipschitz constant, Lead field matrix The transpose of

[0023] The soft threshold operation includes:

[0024] ;

[0025] in, , ;

[0026] In the formula, is the variable of the soft threshold operator, is the threshold parameter, is the soft threshold operator, is a symbolic function.

[0027] Optionally, determining a loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and determining a target iterative shrinkage threshold depth expansion network by minimizing the loss function value, comprises:

[0028] The real source signal matrix and the training reconstruction source signal matrix are used to calculate the loss function value;

[0029] When the loss function value does not meet the early stopping condition, the model parameters are updated by using the Adam optimizer with L2 regularization to minimize the loss function value, and the hyperparameters are dynamically adjusted through cross-validation to determine the intermediate iterative shrinkage threshold depth to expand the network;

[0030] The training EEG signal observation matrix is ​​used to train the intermediate iterative contraction threshold depth expansion network until the loss function value meets the early stopping condition, and the target iterative contraction threshold depth expansion network is determined.

[0031] A second aspect of the present invention provides an adaptive threshold depth unfolding brain source imaging system, comprising:

[0032] A training data acquisition module is used to obtain the training EEG signal observation matrix and the real source signal matrix;

[0033] A network training module, used to use the training EEG signal observation matrix to input an initial iterative shrinkage threshold depth expansion network, and output a training reconstructed source signal matrix;

[0034] A network iteration module is used to determine a loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and to determine a target iterative shrinkage threshold depth to expand the network by minimizing the loss function value;

[0035] The source signal reconstruction module is used to collect the measured EEG signal observation matrix, solve the measured EEG signal observation matrix through the target iterative shrinkage threshold depth expansion network, and output the target reconstructed source signal matrix.

[0036] Optionally, the training data acquisition module is specifically used to:

[0037] Acquire multiple real source signals;

[0038] Performing simulation based on each of the real source signals through the Brainstorm toolbox, and outputting a real source signal matrix and a training lead field matrix;

[0039] Based on the training lead field matrix, the real source signal matrix is ​​linearly transformed and noise is added to generate a training EEG signal observation matrix.

[0040] Optionally, the network training module is specifically used for:

[0041] Construct a signal loss function about the EEG signal observation matrix;

[0042] Iteratively perform gradient descent operation and soft threshold operation on the signal loss function based on the training EEG signal observation matrix, and output a training reconstructed source signal matrix;

[0043] The signal loss function includes:

[0044] ;

[0045] In the formula, is the signal loss function, is the square of the Frobenius norm, is the lead field matrix, is the source signal matrix, is the EEG signal observation matrix, For the source signal, is the number of source signals, is the regularization parameter, is the source signal matrix source signal, is the number of iterations;

[0046] The gradient descent operation includes:

[0047] ;

[0048] In the formula, is the gradient, is the Lipschitz constant, Lead field matrix The transpose of

[0049] The soft threshold operation includes:

[0050] ;

[0051] in, , ;

[0052] In the formula, is the variable of the soft threshold operator, is the threshold parameter, is the soft threshold operator, is a symbolic function.

[0053] A third aspect of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the adaptive threshold deep unfolding brain source imaging method as described in any one of the above items.

[0054] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the adaptive threshold deep unfolding brain source imaging method as described in any one of the above items.

[0055] A fifth aspect of the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the adaptive threshold deep unfolding brain source imaging method as described in any one of the above.

[0056] It can be seen from the above technical solutions that the present invention has the following advantages:

[0057] The above scheme of the present invention provides a method for deep expansion of brain source imaging with an adaptive threshold, including: obtaining a training EEG signal observation matrix and a real source signal matrix; using the training EEG signal observation matrix to input an initial iterative shrinking threshold deep expansion network, and outputting a training reconstructed source signal matrix; determining a loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and determining a target iterative shrinking threshold deep expansion network by minimizing the loss function value; collecting a measured EEG signal observation matrix, solving the measured EEG signal observation matrix through a target iterative shrinking threshold deep expansion network, and outputting a target reconstructed source signal matrix. The traditional ISTA algorithm is expanded into a trainable deep neural network, and the regularization parameters of signal reconstruction can be adaptively adjusted through network iterative training, which is beneficial to improving the accuracy of brain source imaging. The optimization process of each iteration can be explicitly displayed to make it interpretable. Multiple iterations can be efficiently processed in parallel during each optimization update to facilitate improving computational efficiency. These advantages help improve the reliability of brain source imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0059] Figure 1 This is the SIFNet network structure diagram;

[0060] Figure 2 A flowchart of the steps of a method for deep unfolding brain source imaging with an adaptive threshold provided by an embodiment of the present invention;

[0061] Figure 3 A simplified schematic diagram of a deep expansion network provided by an embodiment of the present invention;

[0062] Figure 4 A structural block diagram of a deep unfolding brain source imaging system with an adaptive threshold provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The embodiments of the present invention provide a method and system for deep unfolding brain source imaging with an adaptive threshold, which are used to improve the technical problem of low reliability of brain source imaging in the existing source imaging method of deep learning neural network.

[0064] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] See also Figure 2 , Figure 2 A flowchart of the steps of a method for deep unfolding brain source imaging with an adaptive threshold provided by an embodiment of the present invention.

[0066] The present invention provides an adaptive threshold depth unfolding brain source imaging method, comprising:

[0067] Step 101: Obtain a training EEG signal observation matrix and a real source signal matrix.

[0068] Step 101 includes the following sub-steps:

[0069] Acquire multiple real source signals;

[0070] Through the Brainstorm toolbox, simulation is performed based on each real source signal, and the real source signal matrix and training lead field matrix are output;

[0071] Based on the training lead field matrix, the real source signal matrix is ​​linearly transformed and noise is added to generate the training EEG signal observation matrix.

[0072] It should be noted that the training EEG signal observation matrix and the real source signal matrix are used to construct the training set, so as to facilitate the subsequent network training; in one implementation method, it can be directly obtained from known data; in another implementation method, multiple real source signals can be determined first, and the real source signal matrix and the training lead field matrix can be determined based on each real source signal through the Brainstorm toolbox: the brain model is loaded through the Brainstorm toolbox, such as using the ICBM 152 default MRI template, and the ICBM 152 default MRI template is downsampled to 6004 voxels, and then the brain model is determined by modeling. There are multiple potential dipoles on the brain model. One or more dipoles are randomly selected as seed dipoles in the cortical grid of the brain model, and adjacent units are iteratively absorbed into the patch based on various sub-dipoles for expansion until the required activation area is met, thereby obtaining the source space, setting the neural current direction of each dipole in the source space to be perpendicular to the cortical surface, and assigning a real source signal to each dipole, and forming a real source signal matrix based on each real source signal; based on the 64-channel Neuroscan The sensor configuration in the Quik-cap system is determined by calculating the lead field matrix as the training lead field matrix through the OpenMEEG software package embedded in Brainstorm, in which two reference electrodes are removed and the three-layer head surface is obtained by the boundary element method (BEM). The specific process can refer to the existing technology; according to the obtained training lead field matrix, according to the observation model Perform linear transformation on the real source signal matrix and add noise, where: is the EEG signal observation matrix, is the lead field matrix, is the source signal matrix, is the noise to generate the training EEG signal observation matrix.

[0073] Step 102: Use the training EEG signal observation matrix to input the initial iterative shrinkage threshold depth expansion network, and output the training reconstructed source signal matrix.

[0074] Step 102 includes the following sub-steps:

[0075] Construct a signal loss function about the EEG signal observation matrix;

[0076] Based on the training EEG signal observation matrix, the signal loss function is iteratively subjected to gradient descent and soft threshold operations to output the training reconstructed source signal matrix.

[0077] It should be noted that the deep unfolding network is a method that combines traditional optimization algorithms and deep learning. The simplified schematic diagram is as follows Figure 3As shown, it forms an end-to-end trainable network structure by expanding each step of the iterative algorithm into a layer of a deep network. This method aims to inherit the interpretability of traditional optimization algorithms while utilizing the powerful modeling capabilities of deep learning. In this embodiment, the iterative process of ISTA (iterative shrinkage threshold algorithm) is expanded into a layer structure of a deep neural network to construct an initial iterative shrinkage threshold deep expansion network:

[0078] 1) Input layer: input EEG signal observation matrix ;

[0079] 2) Hidden layer: In order to consider both data fidelity and the sparsity of the source signal, the signal loss function is defined:

[0080] ;

[0081] In the formula, is the signal loss function, is the square of the Frobenius norm, is a regularization parameter to control the sparsity, For the source signal, is the number of source signals, is the source signal matrix source signal, is the number of iterations;

[0082] By iteratively shrinking the threshold to solve the above optimization problem, the iterative process of ISTA is expanded into a layer structure of a deep neural network. Each layer corresponds to an iterative update. Each iterative update performs two operations, linear transformation and nonlinear transformation. Gradient descent operation is performed in linear transformation, and soft threshold operator is applied in nonlinear transformation to perform soft threshold operation to achieve sparsity constraint.

[0083] The gradient descent operation consists of:

[0084] ;

[0085] In the formula, is the number of iterations, is the gradient, is the source signal matrix, is the Lipschitz constant, Lead field matrix The transpose of is the EEG signal observation matrix;

[0086] The soft threshold operation includes:

[0087] ;

[0088] in, , ;

[0089] In the formula, is the regularization parameter, is the variable of the soft threshold operator, is the threshold parameter, is the soft threshold operator, is the sign function. The soft threshold operator is Less than threshold The components are set to 0, and the remaining components are reduced by the magnitude Implementing sparsity constraints;

[0090] 3) Output layer: After multiple layers of iterative updates, if the signal iteration termination condition is met, the reconstructed source signal matrix is ​​output according to the optimal solution of the current signal loss function. .

[0091] Step 103: determine the loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and determine the target iterative shrinkage threshold depth to expand the network by minimizing the loss function value.

[0092] Step 103 includes the following sub-steps:

[0093] The loss function value is calculated using the real source signal matrix and the training reconstructed source signal matrix;

[0094] When the loss function value does not meet the early stopping condition, the goal is to minimize the loss function value, use the Adam optimizer with L2 regularization to update the model parameters, and dynamically adjust the hyperparameters through cross-validation to determine the intermediate iteration shrinkage threshold depth to expand the network;

[0095] The training EEG signal observation matrix is ​​used to train the intermediate iterative shrinkage threshold depth expansion network until the loss function value meets the early stopping condition, and the target iterative shrinkage threshold depth expansion network is determined.

[0096] It should be noted that the training data is used to train the initial iterative shrinkage threshold depth expansion network expansion model, and the mean square error loss function is used to measure the difference between the training reconstructed source signal matrix and the real source signal matrix:

[0097] ;

[0098] In the formula, is the loss function value, For the training samples, is the number of training samples, For the The true source signal matrix of training samples, For the The training reconstructed source signal matrix of training samples;

[0099] An early stopping strategy is adopted during the training process. When the loss function value does not meet the early stopping condition, the Adam optimizer with adaptive matrix estimation is used to update the model parameters with the goal of minimizing the loss function value. L2 regularization is introduced in the process of updating the model parameters to prevent overfitting. At the same time, hyperparameters such as the learning rate are dynamically adjusted through cross-validation. After the optimized intermediate iterative shrinkage threshold deep expansion network is obtained, the training data is re-used for training until the loss function value meets the early stopping condition. The target iterative shrinkage threshold deep expansion network is determined so that it can effectively reconstruct the source signal.

[0100] During the network training process, the iteratively updated model parameters include data such as weights and threshold parameters. The threshold parameters of each layer are learned and adjusted based on the back propagation of the loss function value, so that the network can dynamically adjust the sparsity regularization parameters according to the characteristics of different data (such as noise level, signal strength, etc.). By introducing adaptive threshold regularization, it can better adapt to different sparsities, noise levels and non-uniformity of brain source signals. For strong signal components, the adaptive threshold can lower the threshold to reduce the suppression of real signals. For weak signals or noise components, the adaptive threshold can increase the threshold to better suppress noise, so that the network has stronger noise adaptability and data robustness, and can provide higher source reconstruction accuracy under different data quality and noise environments. It is particularly suitable for EEG / MEG data with large noise in clinical practice.

[0101] Step 104: collect the measured EEG signal observation matrix, solve the measured EEG signal observation matrix through the target iterative shrinkage threshold depth expansion network, and output the target reconstructed source signal matrix.

[0102] It should be noted that the collected measured EEG signal observation matrix is ​​input into the trained target iterative shrinkage threshold deep expansion network, and after forward propagation, the target reconstructed source signal matrix is ​​output. Through the nonlinear transformation of the deep network, the step effect in the traditional method is reduced and more important detail information is retained.

[0103] In an embodiment of the present invention, the traditional ISTA algorithm is expanded into a trainable deep neural network, which has good prospects for improving the reliability of brain source imaging and is particularly suitable for brain source imaging tasks that require high precision and real-time processing: 1. The optimization process of each iteration can be explicitly displayed, which improves the transparency of the source signal reconstruction process, allowing users to understand how each layer adjusts the signal and why the network makes a specific decision; 2. Multiple iterations can be efficiently processed in parallel during each optimization update, avoiding the computational bottleneck of step-by-step iteration in traditional ISTA, significantly improving computational efficiency, and having significant advantages in processing large-scale data; 3. This method combines the advantages of model-based methods and data-driven methods, and can optimize the reconstruction process by adaptively adjusting the regularization parameters of signal reconstruction through network iterative training while maintaining the constraints of the physical model, significantly improving the accuracy of brain source imaging while avoiding the traumatic operation risks of invasive brain source imaging.

[0104] See also Figure 4 , Figure 4 A structural block diagram of a deep unfolding brain source imaging system with an adaptive threshold provided in an embodiment of the present invention.

[0105] The present invention provides an adaptive threshold depth unfolding brain source imaging system, comprising:

[0106] The training data acquisition module 401 is used to acquire the training EEG signal observation matrix and the real source signal matrix;

[0107] A network training module 402 is used to use a training EEG signal observation matrix to input an initial iterative shrinkage threshold depth expansion network and output a training reconstructed source signal matrix;

[0108] A network iteration module 403 is used to determine a loss function value according to a real source signal matrix and a training reconstructed source signal matrix, and to determine a target iterative shrinkage threshold depth to expand the network by minimizing the loss function value;

[0109] The source signal reconstruction module 404 is used to collect the measured EEG signal observation matrix, solve the measured EEG signal observation matrix through the target iterative shrinkage threshold depth expansion network, and output the target reconstructed source signal matrix.

[0110] Optionally, the training data acquisition module 401 is specifically used for:

[0111] Acquire multiple real source signals;

[0112] Through the Brainstorm toolbox, simulation is performed based on each real source signal, and the real source signal matrix and training lead field matrix are output;

[0113] Based on the training lead field matrix, the real source signal matrix is ​​linearly transformed and noise is added to generate the training EEG signal observation matrix.

[0114] Optionally, the network training module 402 is specifically used for:

[0115] Construct a signal loss function about the EEG signal observation matrix;

[0116] Based on the training EEG signal observation matrix, the signal loss function is iteratively subjected to gradient descent and soft threshold operations, and the training reconstructed source signal matrix is ​​output;

[0117] The signal loss function includes:

[0118] ;

[0119] In the formula, is the signal loss function, is the square of the Frobenius norm, is the lead field matrix, is the source signal matrix, is the EEG signal observation matrix, For the source signal, is the number of source signals, is the regularization parameter, is the source signal matrix source signal, is the number of iterations;

[0120] The gradient descent operation consists of:

[0121] ;

[0122] In the formula, is the gradient, is the Lipschitz constant, Lead field matrix The transpose of

[0123] The soft threshold operation includes:

[0124] ;

[0125] in, , ;

[0126] In the formula, is the variable of the soft threshold operator, is the threshold parameter, is the soft threshold operator, is a symbolic function.

[0127] Optionally, the network iteration module 403 is specifically used for:

[0128] The loss function value is calculated using the real source signal matrix and the training reconstructed source signal matrix;

[0129] When the loss function value does not meet the early stopping condition, the goal is to minimize the loss function value, use the Adam optimizer with L2 regularization to update the model parameters, and dynamically adjust the hyperparameters through cross-validation to determine the intermediate iteration shrinkage threshold depth to expand the network;

[0130] The training EEG signal observation matrix is ​​used to train the intermediate iterative shrinkage threshold depth expansion network until the loss function value meets the early stopping condition, and the target iterative shrinkage threshold depth expansion network is determined.

[0131] An embodiment of the present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the adaptive threshold depth unfolding brain source imaging method as in any of the above embodiments.

[0132] An embodiment of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the adaptive threshold depth unfolding brain source imaging method as in any of the above embodiments are implemented.

[0133] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the adaptive threshold depth unfolding brain source imaging method as in any of the above embodiments.

[0134] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0135] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0136] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0137] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0138] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0139] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for deep unwrapping of brain source imaging with an adaptive threshold, characterized in that: include: Obtain training EEG signal observation matrix and real source signal matrix; The training EEG signal observation matrix is ​​used to input an initial iterative shrinkage threshold depth expansion network, and a training reconstructed source signal matrix is ​​output; Determine a loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and determine a target iterative shrinkage threshold depth expansion network by minimizing the loss function value; The measured EEG signal observation matrix is ​​collected, and the measured EEG signal observation matrix is ​​solved through a target iterative shrinkage threshold depth expansion network to output a target reconstructed source signal matrix.

2. The method for deep unwrapping of brain source imaging with adaptive threshold according to claim 1, characterized in that: The step of obtaining a training EEG signal observation matrix and a real source signal matrix includes: Acquire multiple real source signals; Performing simulation based on each of the real source signals through the Brainstorm toolbox, and outputting a real source signal matrix and a training lead field matrix; Based on the training lead field matrix, the real source signal matrix is ​​linearly transformed and noise is added to generate a training EEG signal observation matrix.

3. The method for deep unwrapping of brain source imaging with adaptive threshold according to claim 1, characterized in that: The method of using the training EEG signal observation matrix as input into an initial iterative shrinkage threshold depth expansion network and outputting a training reconstructed source signal matrix comprises: Construct a signal loss function about the EEG signal observation matrix; Iteratively perform gradient descent operation and soft threshold operation on the signal loss function based on the training EEG signal observation matrix, and output a training reconstructed source signal matrix; The signal loss function includes: ； In the formula, is the signal loss function, is the square of the Frobenius norm, is the lead field matrix, is the source signal matrix, is the EEG signal observation matrix, For the source signal, is the number of source signals, is the regularization parameter, is the source signal matrix source signal, is the number of iterations; The gradient descent operation includes: ； In the formula, is the gradient, is the Lipschitz constant, Lead field matrix The transpose of The soft threshold operation includes: ； in, , ; In the formula, is the variable of the soft threshold operator, is the threshold parameter, is the soft threshold operator, is a symbolic function.

4. The method for deep unwrapping of brain source imaging with adaptive threshold according to claim 1, characterized in that: The method of determining a loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and determining a target iterative shrinkage threshold depth expansion network by minimizing the loss function value, comprises: The real source signal matrix and the training reconstruction source signal matrix are used to calculate the loss function value; When the loss function value does not meet the early stopping condition, the model parameters are updated by using the Adam optimizer with L2 regularization to minimize the loss function value, and the hyperparameters are dynamically adjusted through cross-validation to determine the intermediate iteration shrinkage threshold depth to expand the network; The training EEG signal observation matrix is ​​used to train the intermediate iterative contraction threshold depth expansion network until the loss function value meets the early stopping condition, and the target iterative contraction threshold depth expansion network is determined.

5. An adaptive threshold deep unfolding brain source imaging system, characterized in that: include: A training data acquisition module is used to obtain the training EEG signal observation matrix and the real source signal matrix; A network training module, used to use the training EEG signal observation matrix to input an initial iterative shrinkage threshold depth expansion network, and output a training reconstructed source signal matrix; A network iteration module is used to determine a loss function value according to the real source signal matrix and the training reconstructed source signal matrix, and to determine a target iterative shrinkage threshold depth to expand the network by minimizing the loss function value; The source signal reconstruction module is used to collect the measured EEG signal observation matrix, solve the measured EEG signal observation matrix through the target iterative shrinkage threshold depth expansion network, and output the target reconstructed source signal matrix.

6. The adaptive threshold depth unwrapping brain source imaging system according to claim 5, characterized in that: The training data acquisition module is specifically used for: Acquire multiple real source signals; Performing simulation based on each of the real source signals through the Brainstorm toolbox, and outputting a real source signal matrix and a training lead field matrix; Based on the training lead field matrix, the real source signal matrix is ​​linearly transformed and noise is added to generate a training EEG signal observation matrix.

7. The adaptive threshold depth unwrapping brain source imaging system according to claim 5, characterized in that: The network training module is specifically used for: Construct a signal loss function about the EEG signal observation matrix; Iteratively perform gradient descent operation and soft threshold operation on the signal loss function based on the training EEG signal observation matrix, and output a training reconstructed source signal matrix; The signal loss function includes: ； In the formula, is the signal loss function, is the square of the Frobenius norm, is the lead field matrix, is the source signal matrix, is the EEG signal observation matrix, For the source signal, is the number of source signals, is the regularization parameter, is the source signal matrix source signal, is the number of iterations; The gradient descent operation includes: ； In the formula, is the gradient, is the Lipschitz constant, Lead field matrix The transpose of The soft threshold operation includes: ； in, , ; In the formula, is the variable of the soft threshold operator, is the threshold parameter, is the soft threshold operator, is a symbolic function.

8. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the adaptive threshold depth unfolding brain source imaging method as described in any one of claims 1 to 4.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the adaptive threshold depth unfolding brain source imaging method according to any one of claims 1 to 4 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the adaptive threshold depth unfolding brain source imaging method according to any one of claims 1 to 4 are implemented.

Images