Brain electromagnetic brain source imaging system based on Bayesian model and deep learning
By combining Bayesian model and deep learning technology, using the ResNet model to learn the relationship between noise accuracy matrix and voxel time series, the problem of inaccurate brain activity reconstruction in traditional methods is solved, and higher accuracy and robustness are achieved brain-electromagnetic brain source imaging.
Patent Information
- Application Number
- CN202411073060.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-15
AI Technical Summary
Due to the uncertainty of the inversion problem, traditional brain electromagnetic brain source imaging methods need to introduce prior constraints to obtain a unique solution, but these constraints cannot truly reflect the neurophysiological significance of brain signals, resulting in heavy activity in the brain on spatial and time scales. Inaccurate enough structure.
Using the brain electromagnetic brain source imaging system based on Bayesian model and deep learning, data is obtained through the signal processing module, probability generation model is constructed, and the relationship between noise accuracy matrix and voxel time series is learned by using the deep learning module to generate brain source activity time Sequences are trained in combination with ResNet model to improve reconstruction accuracy and robustness.
It significantly improves the accuracy and robustness of brain-source activities, especially in high noise and interference, and shows stronger stability, avoids cumbersome solutions and hyperparameter settings, and enhances the repeatability of imaging technology. and accessibility.
Smart Images

Figure CN118967859A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of brain electromagnetic brain source imaging, and more specifically, relates to a brain electromagnetic brain source imaging system based on Bayesian model and deep learning. Background Art
[0002] Electromagnetic brain source imaging is a technology that reconstructs brain activity in space and time. Electromagnetic brain source imaging is of great significance to cognitive neuroscience and clinical research: on the one hand, it facilitates human exploration of brain activity mechanisms, improves the understanding of brain information processing, and provides possibilities for the development of artificial intelligence equipment in various fields; on the other hand, in clinical research, electromagnetic brain source imaging equipment helps improve the clinical diagnosis of brain physiological diseases, such as epilepsy, stroke or abnormal brain function, and provides new possibilities for rehabilitation treatment.
[0003] Due to the uncertainty of the inversion problem in EEM brain source imaging, traditional methods usually need to introduce prior constraints to obtain a unique solution. However, these constraints are often only for mathematical simplification and cannot truly reflect the neurophysiological significance of brain signals. Therefore, how to more accurately reconstruct brain activities in space and time scales by collecting information is a difficult problem that urgently needs to be solved in EEM brain source imaging research. Summary of the invention
[0004] In response to the above defects or improvement needs of the prior art, the present invention provides a brain electromagnetic brain source imaging system based on a Bayesian model and deep learning, which can achieve accurate reconstruction of the spatiotemporal characteristics of brain source activity.
[0005] To achieve the above object, according to one aspect of the present invention, a brain electromagnetic source imaging system based on Bayesian model and deep learning is provided, comprising: A signal processing module, used to obtain brain electromagnetic sampling data; A model building module is used to build a brain electromagnetic sampling data generation model based on the brain electromagnetic sampling data, and convert the brain electromagnetic sampling data generation model into a brain electromagnetic sampling data probability generation model that satisfies the Gaussian distribution, wherein the brain electromagnetic sampling data probability generation model includes a noise precision matrix parameter and a voxel time series precision parameter; A training sample set construction module is used to construct a training sample set, wherein the training sample set includes a plurality of brain electromagnetic sampling data samples, each of which is marked with a noise precision matrix label and a voxel time series precision label; A deep learning module, used for learning the relationship between brain electromagnetic sampling data samples and noise precision matrix labels and voxel time series precision labels from the training sample set, and outputting predicted values of the noise precision matrix and voxel time series precision based on the input brain electromagnetic sampling data; The brain source activity reconstruction module is used to generate a brain source activity time series based on the noise accuracy matrix and the predicted value of the voxel time series accuracy output by the deep learning module and the probability generation model of the brain electromagnetic sampling data.
[0006] Furthermore, the deep learning module adopts a ResNet model.
[0007] Furthermore, the deep learning module includes a first deep learning module and a second deep learning module, wherein the first deep learning module is used to learn the relationship between the brain electromagnetic sampling data samples and the voxel time series accuracy labels from the training sample set, and output the predicted value of the voxel time series accuracy according to the input brain electromagnetic sampling data, and the second deep learning module is used to learn the relationship between the brain electromagnetic sampling data samples and the noise accuracy matrix labels from the training sample set, and output the predicted value of the noise accuracy matrix according to the input brain electromagnetic sampling data; The training function of the first deep learning module is: ; in, represents the loss function of the first deep learning module, is the batch size, The first deep learning module The loss value of the batch, ≤ , For the The loss weight of the batch, For the voxel-wise time series accuracy of batch predictions, For the The batch's voxel-wise time series precision labels, Represents matrix transpose; The training function of the second deep learning module is: ; represents the loss function of the second deep learning module, is the batch size, The second deep learning module The loss value of the batch, ≤ , For the The noise precision matrix of the batch predictions, For the Noisy precision matrix labels for the batch.
[0008] Furthermore, the voxels of the whole brain are divided into multiple brain regions, and the brain electromagnetic sampling data samples are divided according to the brain regions; The deep learning module includes multiple brain region deep learning modules and a splicing module, and each brain region deep learning module is trained using the brain electromagnetic sampling data samples corresponding to each brain region and the noise precision matrix labels and voxel time series precision labels corresponding to the samples; Each brain region deep learning module is used to output the noise precision matrix label and the predicted value of the voxel time series precision matrix of the corresponding brain region according to the input brain electromagnetic sampling data samples of the corresponding brain region; The splicing module is used to splice the noise precision matrix and the predicted value of voxel time series precision of each brain region into the noise precision matrix and the predicted value of voxel time series precision of the whole brain.
[0009] Furthermore, the calculation steps of the noise precision matrix label and the voxel time series precision matrix label corresponding to each brain electromagnetic sampling data sample include: simulating brain activity through a dipole model, which includes the noise precision matrix label and the voxel time series precision label.
[0010] Furthermore, before inputting the brain electromagnetic sampling data into the deep learning module, the brain electromagnetic sampling data is converted into a grayscale value image.
[0011] Furthermore, the brain electromagnetic sampling data generation model is: is the brain source activity time series, is the brain electromagnetic sampling data, , , Indicates Brain activity at each sampling point, Indicates Brain electromagnetic sampling data of sampling points, 1≤ ≤ , Represents the total number of sampling points, where represents the guiding field matrix, Indicates noise; ; ; ; in, represents the probability distribution, It means that it satisfies Gaussian distribution. express The mean of represents the noise precision matrix, represents the voxel time series accuracy, , Represents matrix transpose.
[0012] Furthermore, the brain electromagnetic sampling data is preprocessed before being input into the deep learning module.
[0013] In general, the above technical scheme conceived by the present invention has beneficial effects compared with the prior art: by adopting a method of learning voxel time series precision matrix and noise precision matrix based on deep learning, and reconstructing and locating brain source activity areas based on the Bayesian model, the accuracy of brain source activity reconstruction can be greatly improved, and it can show stronger robustness in the face of high noise and interference. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a schematic diagram of the module composition of the brain electromagnetic brain source imaging system based on the Bayesian model and deep learning in an embodiment of the present invention; Figure 2 It is a schematic diagram of the principle of a brain electromagnetic brain source imaging system based on a Bayesian model and deep learning according to an embodiment of the present invention; Figure 3 It is an imaging schematic diagram of a brain electromagnetic brain source imaging system based on a Bayesian model and deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION
[0015] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0016] Unless otherwise specified, "plurality" means two or more.
[0017] The terms "including" and "having" and any variations thereof in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product or equipment comprising a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products or equipment.
[0018] like Figure 1 As shown, an electroencephalographic source imaging system based on a Bayesian model and deep learning according to an embodiment of the present invention includes: A signal processing module, used to obtain brain electromagnetic sampling data; A model building module is used to build a brain electromagnetic sampling data generation model based on the brain electromagnetic sampling data, and convert the brain electromagnetic sampling data generation model into a brain electromagnetic sampling data probability generation model that satisfies the Gaussian distribution, wherein the brain electromagnetic sampling data probability generation model includes a noise precision matrix parameter and a voxel time series precision parameter; A training sample set construction module is used to construct a training sample set, wherein the training sample set includes a plurality of brain electromagnetic sampling data samples, each of which is marked with a noise precision matrix label and a voxel time series precision label; A deep learning module is used to learn the relationship between brain electromagnetic sampling data samples and noise precision matrix labels and voxel time series precision labels from the training sample set, and output the predicted values of the noise precision matrix and voxel time series precision based on the input brain electromagnetic sampling data; The brain source activity reconstruction module is used to generate a brain source activity time series based on the noise accuracy matrix and the predicted value of the voxel time series accuracy output by the deep learning module and the probability generation model of the brain electromagnetic sampling data.
[0019] The implementation principle of each module is explained in detail below.
[0020] (1) Signal processing module Input brain electromagnetic sampling data , , S is the number of original signal sampling points, Indicates time Sampling data, 1≤ ≤ S , ,here is the number of channels at the acquisition point, Indicates the sampling data of sampling channel 1. Indicates the sampling data of sampling channel 2. Similarly, Indicates the sampling channel The sampling data of .
[0021] Furthermore, the brain electromagnetic sampling data is preprocessed before being input into the deep learning module.
[0022] Furthermore, the input data can be initially screened to ensure the validity and integrity of the sampled data, such as data format conversion and parameter confirmation, including sampled data information.
[0023] Furthermore, signal preprocessing can be performed on the brain electromagnetic sampling data.
[0024] Preprocessing includes baseline correction, interference removal and filtering.
[0025] The baseline correction process can be expressed as: ; Interference removal processing needs to be performed based on the characteristics of the sampled data.
[0026] Preferably, based on independent component analysis (ICA) or spatial projection, such as ICA-based interference removal (including electrooculography, etc.), it can be expressed as: ; in is the decomposition weight matrix, As an independent component, N represents the number of independent components, that is, the total number of brain source voxels, If the features contain significant interference or noise, the sampled data is denoised It can be expressed as: ; Preferably, the signal after ICA-based interference removal is further processed using a projection-based interference removal method. The projection-based interference removal method requires finding a projection matrix, which can be based on features such as brain structure spatial information. Assume that the projection matrix is , and its interference removal formula is: ; Filtering: The purpose of filtering the collected data is to eliminate frequency bands that are not related to brain activity and to achieve signal processing and improvement. The main method used is time-frequency filtering based on Fourier or wavelet transform to retain data in the frequency band of interest. The preprocessed data can ensure the smooth construction of subsequent data sets.
[0027] (2) Model building module Using Gaussian priors, we get Voxel value , whose precision (variance) is specific to each ,Right now ; In the formula, It means that it satisfies Gaussian distribution. Indicates the corresponding Variance representation of the voxel-wise prior distribution.
[0028] Defining voxel time series ,Sensor time series Using set expressions and , represents the entire voxel time series and the entire sensor time series . Indicates Brain activity at each sampling point, Indicates Brain source activity at sampling points, 1≤ ≤ , Represents the total number of sampling points. and The relationship is expressed as ; in represents the guiding field matrix, Indicates noise.
[0029] Define column vector , represents the voxel time series accuracy, , Corresponding to The variance of the prior distribution of the voxels is expressed as 1≤ ≤ , Represents the total number of brain-derived voxels and the posterior distribution of the voxel time series It obeys Gaussian distribution, which is ; in, represents the probability distribution, , It means that it satisfies Gaussian distribution. express The mean of is the noise precision matrix, where and The expression is ; ; At the same time, in order to further deduce the accuracy of the pixel time series , let's define : ; in is the identity matrix.
[0030] In the existing Bayesian-based brain source activity solution method, it is necessary to construct a solution The cost function is expressed as ; The value of is obtained by minimizing this cost function, and the voxel time series is solved by combining equation (8) and equation (9) .
[0031] In the embodiment of the present invention, the existing Bayesian-based brain source activity solution method is not used, but deep learning is combined with Bayesian to solve the brain source activity, bypassing the tedious solution and complex hyperparameter setting, and having reproducibility. Combining equation (5) and equation (6) to generate the sensor time series , you can define the sensor time series The specific definition can be set by the user and used as a training set. In order to localize brain source activity, the noise precision matrix Voxel-wise time series accuracy As the label of the training set. Repeating the above process can get a large-scale training set. It is worth noting that a smaller validation set will be generated independently instead of cross-validation to test the generalization performance of the network.
[0032] The deep learning model will learn the sensor time series With the noise precision matrix , voxel time series accuracy The implicit relationship between the sensor time series accuracy Input into the trained deep learning model to estimate the expected noise accuracy matrix and voxel-wise time series accuracy , avoiding Complicated solution. When performing matrix restoration, the brain activity reconstruction module needs to face several characteristics of the inverse problem of brain activity: small values, extremely sparse, and large dimensions.
[0033] The preferred implementation methods of the training sample set construction module and the deep learning module are described in detail below.
[0034] (3) Training sample set construction module The training sample set includes multiple brain electromagnetic sampling data samples. Each sample is represented as ,in As a brain electromagnetic sampling data sample. Each brain electromagnetic sampling data sample corresponds to a sensor time series In addition, each sample is also labeled with the noise precision matrix and voxel-wise time series accuracy labels .
[0035] The noise precision matrix label and the voxel time series precision label can simulate brain activity through a dipole model, which includes the noise precision matrix label and the voxel time series precision label, and also includes the obtained brain electromagnetic sampling data samples. Other methods in the prior art can also be used to obtain the noise precision matrix label and the voxel time series precision label.
[0036] (4) Deep Learning Module Furthermore, the deep learning module adopts the ResNet model.
[0037] Traditional machine learning models have problems such as low accuracy, difficulty in solving, and complex prior settings when facing the inverse problem of brain activity. Brain activity is complex, and the accuracy of sensor time series is With the noise precision matrix and voxel-wise time series accuracy The deep connections between the two are difficult to mine, so deep neural networks are chosen to solve the inverse problem of brain source activity. However, due to the problem of gradient vanishing, deep network training becomes difficult. When the gradient is back-propagated to the previous layer, repeated multiplication may cause the gradient to become very small, causing the network performance to saturate or even decline rapidly as the number of layers increases.
[0038] To overcome these problems, ResNet was selected as the training model. ResNet introduces skip connections while maintaining a deep neural network, which can effectively solve the challenges faced by traditional machine learning.
[0039] During the data loading phase, since the precision matrix of the MEG data is not significant, the sensor time series needs to be Perform an effective normalization, where the magnetoencephalogram data is normalized to ; in It is a small factor to prevent equal This results in numerical calculation errors.
[0040] The sensor time series accuracy Convert it to a grayscale image, and use convolutional neural network to extract features by referring to computer vision technology. and voxel-wise time series accuracy As labels, a dataset is constructed. This method aims to mine the accuracy of sensor time series With the noise precision matrix And voxel-wise time series accuracy This is a new approach to solving the inverse problem of brain source activity.
[0041] During the model training phase, the noise precision matrix was learned during the derivation of brain source imaging. and voxel-wise time series accuracy The matrix actually has nonzero elements only on the diagonal. Therefore, and The output of can be represented by reducing the dimensions of only the elements on the diagonal into a one-dimensional vector.
[0042] Noise Accuracy Matrix and voxel-wise time series accuracy The two are trained independently. The deep learning module includes a first deep learning module and a second deep learning module, the first deep learning module is used to learn the relationship between the brain electromagnetic sampling data samples and the voxel time series accuracy labels from the training sample set, and output the predicted value of the voxel time series accuracy based on the input brain electromagnetic sampling data, and the second deep learning module is used to learn the relationship between the brain electromagnetic sampling data samples and the noise accuracy matrix labels from the training sample set, and output the predicted value of the noise accuracy matrix based on the input brain electromagnetic sampling data.
[0043] First, this problem can be abstracted as a multi-label classification problem. Since multiple brain source activities are distributed in different brain regions throughout the brain, each voxel is regarded as a binary classification task, that is, to determine whether the voxel generates source activity. If the voxel generates source activity, a larger value is output at the position of this voxel relative to other positions.
[0044] Therefore, the loss function used is Binary Cross-Entropy Loss, referred to as BCE Loss below, where BCE Loss is ; in is the batch size, The first deep learning module Batch loss Loss of value , ≤ , For the The loss weight of the batch, Can be manually readjusted, For the voxel-wise time series accuracy of batch predictions, For the Batch voxel time series accuracy labels. It is worth noting that BCE Loss makes each voxel independent, that is, when the first voxel judges whether it generates source activity, it will not affect the model's judgment on whether the second voxel generates source activity.
[0045] Estimated noise precision matrix As a regression task, the loss function constructed is mean-square error (MSE), where MSE is ; in represents the loss function of the second deep learning module, is the batch size, The second deep learning module The loss value of the batch, ≤ , For the The noise precision matrix of the batch predictions, For the Noisy precision matrix labels for the batch.
[0046] Furthermore, the voxels of the whole brain are divided into multiple brain regions, and the deep learning module includes multiple brain region deep learning modules and a splicing module, and each brain region deep learning module is trained separately according to the brain region.
[0047] Considering the extreme sparsity and high-dimensionality of the inverse problem of brain source activity, the idea of the Beamformer method is borrowed. In the complex computational context of high-resolution whole-brain source localization, Beamformer provides an effective idea: the voxels of the whole brain are divided into multiple brain regions according to the brain regions, and the same training is performed in each brain region. In the evaluation stage, the results obtained from the training of each brain region are spliced to reduce the complexity of model training.
[0048] First, the training samples of the training sample set also need to be divided according to brain regions, and each brain electromagnetic sampling data sample, noise precision matrix label and voxel time series precision label need to be divided according to brain regions.
[0049] Then, the deep learning module of each brain region is trained using the brain electromagnetic sampling data samples corresponding to each brain region and the noise precision matrix labels and voxel time series precision labels corresponding to the brain region samples. After the training of each brain region deep learning module is completed, the noise precision matrix labels and predicted values of the voxel time series precision matrix of the corresponding brain region can be output according to the input brain electromagnetic sampling data samples of the corresponding brain region. The splicing module is used to splice the predicted values of the noise precision matrix and voxel time series precision of each brain region into the predicted values of the noise precision matrix and voxel time series precision of the whole brain.
[0050] The voxels of the whole brain are divided into Q brain regions, ; Each model accepts sensor time series accuracy for each brain region As input, 1≤ ≤ Q , and output the noise precision matrix corresponding to each brain region and voxel-wise time series accuracy , expressed as . It is the noise precision matrix corresponding to the model output of each brain region and voxel-wise time series accuracy The noise precision matrix of the whole brain can be obtained by splicing formula (16): and voxel-wise time series accuracy .
[0051] Furthermore, each brain region deep learning module may include a first deep learning module and a second deep learning module to calculate the noise precision matrix and voxel-wise time series accuracy The two are trained separately and independently.
[0052] (5) The brain source activity reconstruction module is used to obtain the brain source activity time series inside the brain from the noise accuracy matrix and the predicted value of the voxel time series accuracy output by the deep learning module.
[0053] The covariance matrix of the brain signal sequence Input into the trained deep learning model to obtain the expected corresponding noise accuracy matrix and voxel-wise time series accuracy . Noise Accuracy Matrix and voxel-wise time series accuracy Combining equation (8) with equation (9), This is the time series of brain activity that is needed.
[0054] The imaging module is used to image brain source activity output and imaging according to the brain source activity time series obtained by the brain source activity reconstruction module.
[0055] Obtaining time series of brain electrical activity After that, the results need to be displayed. Furthermore, the display results are based on the existing model parameters, such as the source location after brain space segmentation and brain structure information. This patent uses MRI templates to render source activities and general brain models. The running results are as follows Figure 3 shown.
[0056] By adopting the method of learning the voxel time series precision matrix and noise precision matrix based on deep learning in the EEG source imaging system of Bayesian estimation and deep learning, and reconstructing and locating the brain source activity area based on the Bayesian model, since the deep learning method is based on the algorithm of characterizing and learning the data, the complex brain dynamics can be embedded in the training data and will be implicitly reflected as the weights and nonlinear connections in the trained neural network without explicitly defining the regularization term. Secondly, after the time-consuming training process, when the neural network is applied to new EEG data measurement instances, the evaluation is effective without searching for hyperparameters for each new EEG data measurement instance. This will also increase the reproducibility of EEG source imaging results and avoid the ambiguity of selecting hyperparameters. Once the deep learning model is trained, it can be directly applied to real test data with little user intervention, because the prior information is "memorized" in the weights of the trained network. This increases the reproducibility and accessibility of EEG source imaging technology, because the expertise in parameter selection or model adjustment represents a bottleneck for the widespread application of EEG source imaging technology, making it easier to use source imaging evaluation as a standard tool in various clinical settings. There is also a Bayesian estimation method that can estimate the noise precision matrix without a "baseline" or "control" measurement, so it shows stronger robustness in the face of high noise and interference. Even in the case of low EEG density, complex brain source activity can still be reconstructed significantly. The accuracy and accessibility of brain source activity reconstruction are greatly improved. Therefore, the present invention organically combines deep learning with the Bayesian model, which is particularly suitable for brain source activity imaging.
[0057] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A brain electromagnetic source imaging system based on Bayesian model and deep learning, characterized in that: include: A signal processing module, used to obtain brain electromagnetic sampling data; A model building module is used to build a brain electromagnetic sampling data generation model based on the brain electromagnetic sampling data, and convert the brain electromagnetic sampling data generation model into a brain electromagnetic sampling data probability generation model that satisfies the Gaussian distribution, wherein the brain electromagnetic sampling data probability generation model includes a noise precision matrix parameter and a voxel time series precision parameter; A training sample set construction module is used to construct a training sample set, wherein the training sample set includes a plurality of brain electromagnetic sampling data samples, each of which is marked with a noise precision matrix label and a voxel time series precision label; A deep learning module, used for learning the relationship between brain electromagnetic sampling data samples and noise precision matrix labels and voxel time series precision labels from the training sample set, and outputting predicted values of the noise precision matrix and voxel time series precision based on the input brain electromagnetic sampling data; The brain source activity reconstruction module is used to generate a brain source activity time series based on the noise accuracy matrix and the predicted value of the voxel time series accuracy output by the deep learning module and the probability generation model of the brain electromagnetic sampling data.
2. The brain electromagnetic source imaging system based on Bayesian model and deep learning as claimed in claim 1, characterized in that: The deep learning module adopts the ResNet model.
3. The brain electromagnetic source imaging system based on Bayesian model and deep learning as claimed in claim 1, characterized in that: The deep learning module includes a first deep learning module and a second deep learning module, wherein the first deep learning module is used to learn the relationship between the brain electromagnetic sampling data samples and the voxel time series accuracy labels from the training sample set, and output the predicted value of the voxel time series accuracy according to the input brain electromagnetic sampling data, and the second deep learning module is used to learn the relationship between the brain electromagnetic sampling data samples and the noise accuracy matrix labels from the training sample set, and output the predicted value of the noise accuracy matrix according to the input brain electromagnetic sampling data; The training function of the first deep learning module is: ; in, represents the loss function of the first deep learning module, is the batch size, The first deep learning module The loss value of the batch, ≤ , For the The loss weight of the batch, For the voxel-wise time series accuracy of batch predictions, For the The batch's voxel-wise time series precision labels, Represents matrix transpose; The training function of the second deep learning module is: ; represents the loss function of the second deep learning module, is the batch size, The second deep learning module The loss value of the batch, ≤ , For the The noise precision matrix of the batch predictions, For the Noisy precision matrix labels for the batch.
4. The brain electromagnetic source imaging system based on Bayesian model and deep learning as claimed in claim 1, characterized in that: Divide the voxels of the whole brain into multiple brain regions, and divide the brain electromagnetic sampling data samples according to brain regions; The deep learning module includes multiple brain region deep learning modules and a splicing module, and each brain region deep learning module is trained using the brain electromagnetic sampling data samples corresponding to each brain region and the noise precision matrix labels and voxel time series precision labels corresponding to the samples; Each brain region deep learning module is used to output the noise precision matrix label and the predicted value of the voxel time series precision matrix of the corresponding brain region according to the input brain electromagnetic sampling data samples of the corresponding brain region; The splicing module is used to splice the noise precision matrix and the predicted value of voxel time series precision of each brain region into the noise precision matrix and the predicted value of voxel time series precision of the whole brain.
5. A brain electromagnetic source imaging system based on Bayesian model and deep learning as claimed in claim 1, characterized in that: The calculation steps of the noise precision matrix label and the voxel time series precision matrix label corresponding to each brain electromagnetic sampling data sample include: simulating brain activity through a dipole model, which includes the noise precision matrix label and the voxel time series precision label.
6. The brain electromagnetic source imaging system based on Bayesian model and deep learning as claimed in claim 1, characterized in that: Before inputting the brain electromagnetic sampling data into the deep learning module, the brain electromagnetic sampling data is converted into a grayscale value image.
7. The brain electromagnetic source imaging system based on Bayesian model and deep learning as claimed in claim 1, characterized in that: The brain electromagnetic sampling data generation model is: is the brain source activity time series, is the brain electromagnetic sampling data, , , Indicates Brain activity at each sampling point, Indicates Brain electromagnetic sampling data of sampling points, 1≤ ≤ , Represents the total number of sampling points, where represents the guiding field matrix, Indicates noise; ; ; ; in, represents the probability distribution, It means that it satisfies Gaussian distribution. express The mean of represents the noise precision matrix, represents the voxel time series accuracy, , Represents matrix transpose.
8. The brain electromagnetic source imaging system based on Bayesian model and deep learning as claimed in claim 1, characterized in that: Before inputting the brain electromagnetic sampling data into the deep learning module, the brain electromagnetic sampling data is preprocessed.
Citation Information
Cited By
Model training method and device for brain magnetic system and source tracing and positioning method
CN119719681A
Model training method and device for magnetoencephalography system and source localization method
CN119719681B
Electroencephalogram signal source space function connection estimation system and method based on deep learning
CN120763621A
Deep learning based electroencephalogram source spatial functional connectivity estimation system and method
CN120763621B
