A time-frequency multi-sparse prior champagne method

Abstract

The application discloses a time-frequency multi-sparse priori champagne method, and belongs to the technical field of neural oscillation activity reconstruction. The method comprises the following steps: collecting brain neural activity original data by using a sensor and performing pretreatment; dividing a cerebral cortex source space into a plurality of parcel regions; processing the parcel region data by using a Bayesian sparse priori algorithm; estimating an optimal source covariance matrix and an optimal noise covariance matrix by using a maximum likelihood principle; based on a spatial priori, introducing a result after Bayesian sparse priori processing under a TFC framework, respectively obtaining optimal sensor covariance matrices and optimal source variance matrices of control sections, task sections and comprehensive sections, calculating and comparing source power changes, and realizing estimation on oscillation activity. The application shows performance superior to other benchmark algorithms, especially under low signal-to-noise ratio and multi-source conditions, and shows its potential and practicability in neural oscillation activity reconstruction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of neural oscillation activity reconstruction, and particularly relates to a time-frequency multi-sparse prior champagne method. BACKGROUND

[0002] With the development of neuroscience and brain imaging technology, non-invasive brain function imaging technologies such as magnetoencephalography (MEG), optical pumped magnetoencephalography (OPM-MEG) and electroencephalography (EEG) are increasingly widely used in the fields of cognitive neuroscience and neurological disease diagnosis. These technologies help researchers and clinicians analyze and reconstruct neural oscillation activity in the brain by measuring electromagnetic activity in the brain. However, in practical applications, existing brain source reconstruction algorithms face various challenges, especially in the case of multiple sources, strongly correlated sources and low signal-to-noise ratio, it is very difficult to accurately reconstruct brain activity.

[0003] In clinical applications, such as neurological diseases such as epilepsy, Parkinson's disease and Alzheimer's disease, abnormal changes in neural oscillation activity can be used as biomarkers for disease diagnosis and monitoring. Therefore, it is crucial to accurately reconstruct and analyze neural oscillation activity in the time-frequency domain. However, due to the presence of background activity in the brain, neural oscillation signals are usually weak and have very low signal-to-noise ratio. At the same time, the simultaneous activation of multiple strongly correlated sources can cause energy leakage, false sources or positioning errors in traditional reconstruction algorithms, thereby affecting the accuracy of diagnosis.

[0004] Existing brain source reconstruction algorithms, such as beamformers and sparse Bayesian learning (SBL) algorithms, have limitations in dealing with these problems. Traditional beamforming algorithms rely on pre-set forward models and are sensitive to noise, especially when faced with multiple strongly correlated sources, the reconstructed energy is easily spread, resulting in low reconstruction accuracy. Although sparse Bayesian learning can solve the sparsity problem to some extent, it still faces the risk of overfitting and false sources in low signal-to-noise ratio or multiple source conditions. Therefore, how to achieve high-precision neural oscillation activity reconstruction in low signal-to-noise ratio and complex source environment has become a key technical problem. SUMMARY

[0005] To solve the above technical problems, the present application provides a time-frequency multi-sparse prior champagne method, which mainly includes pre-segmenting the cerebral cortex source space into multiple parcel regions; introducing spatial prior to improve the accuracy of source reconstruction; introducing Bayesian sparse prior assumption, estimating the source covariance matrix and noise covariance matrix using hierarchical Bayesian model through maximum likelihood principle; using TFC framework to estimate oscillation activity; using Bayesian sparse prior algorithm to obtain sensor coherence matrix and source variance of control segment, task segment and comprehensive segment respectively; superimposing the coherence matrices of different trials through Bayesian network learning to obtain the reconstruction results of oscillation activity at different times and frequencies.

[0006] To achieve the above object, the present application adopts the following technical solutions:

[0007] A time-frequency multi-sparse prior champagne method, comprising:

[0008] Step S1: collecting brain nerve activity raw data using a sensor and performing preprocessing;

[0009] Step S2: dividing a cerebral cortex source space into a plurality of parcel regions;

[0010] Step S3: processing the plurality of parcel region data using a Bayesian sparse prior algorithm, estimating an optimal source covariance matrix and an optimal noise covariance matrix through a maximum likelihood principle;

[0011] Step S4: based on a spatial prior, introducing a result of the Bayesian sparse prior processing under a TFC framework, respectively obtaining optimal sensor covariance matrices and optimal source covariance matrices of control sections, task sections and comprehensive sections, calculating and comparing changes in source power, and realizing estimation of oscillation activity.

[0012] In another aspect, the present application provides a time-frequency multi-sparse prior champagne device, comprising:

[0013] A preprocessing module for collecting brain nerve activity raw data using a sensor and performing preprocessing;

[0014] A segmentation module for dividing a cerebral cortex source space into a plurality of parcel regions;

[0015] An estimation module for processing the plurality of parcel region data using a Bayesian sparse prior algorithm, estimating an optimal source covariance matrix and an optimal noise covariance matrix through a maximum likelihood principle;

[0016] A calculation module for based on a spatial prior, introducing a result of the Bayesian sparse prior processing under a TFC framework, respectively obtaining optimal sensor covariance matrices and optimal source covariance matrices of control sections, task sections and comprehensive sections, calculating and comparing changes in source power, and realizing estimation of oscillation activity.

[0017] In a third aspect, the present application provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned time-frequency multi-sparse prior champagne method.

[0018] In a fourth aspect, the present application provides a computer readable storage medium, which stores executable instructions, and the instructions, when executed by a processor, enable the processor to implement the time-frequency multi-sparse prior champagne method described above.

[0019] The present application has the following beneficial effects:

[0020] The time-frequency multi-sparse prior champagne algorithm of the present application has significant advantages in the reconstruction of neural oscillation activities, especially in the case of multiple sources, strongly correlated sources, and low signal-to-noise ratio environments. By introducing source space pre-segmentation (parcel segmentation) and sparse prior assumptions, combined with adaptive learning of the source and noise covariance matrix in the Bayesian framework, the mutual interference between multiple strongly correlated sources can be effectively reduced, and the accuracy of source reconstruction can be significantly improved. This method can still ensure the concentration of reconstructed energy in the case of multiple sources, avoiding the problem of energy leakage. Through diagonal heteroscedastic noise learning and multi-modal prior information fusion (such as the joint use of MEG and fMRI data), high-quality source reconstruction can still be ensured under extremely low signal-to-noise ratio conditions, and the method has strong robustness and noise resistance. In the TFC framework, combined with the time-frequency beamformer (TFBF) for source activity reconstruction in the time-frequency domain, the oscillation activities of different frequency bands in the brain can be accurately captured, especially the estimation of instantaneous power changes is more accurate, and the centralization error caused by traditional methods is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The flowchart of the time-frequency multi-sparse prior champagne method of the present application;

[0022] Figure 2 The flowchart of the oscillation source DDP algorithm;

[0023] Figure 3 The performance comparison chart of the time-frequency multi-sparse prior champagne method and the benchmark algorithm under different signal-to-noise ratios and source numbers. DETAILED DESCRIPTION

[0024] The present application will be further described below in conjunction with the drawings and examples.

[0025] In the following description, the terms "first", "second", etc. are only for the purpose of description and cannot be understood as indicating or implying relative importance. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0026] To solve the problem of realizing high-precision neural oscillation activity reconstruction in low signal-to-noise ratio and complex source environment, some researchers have proposed a dynamic imaging of coherent sources (DICS) technology. Some researchers have also calculated the average covariance matrix of the task time period and the control time period, and introduced a pseudo-statistic F-ratio to compare the differences in rhythm changes, thereby reducing the possibility of pseudo-source generation. Some researchers have combined the SBL-BF method and the TFBF method to propose a TFC framework, which learns the coherence matrix through multi-layer Bayesian structure and improves the robustness of the algorithm. However, the above methods may cause energy leakage and pseudo-source in low signal-to-noise ratio or strong correlation source, and the imaging effect is not good in low signal-to-noise ratio and multiple correlation source activity. In combination with the work done by previous researchers, a time-frequency multi-sparse prior champagne method is disclosed, which can efficiently and accurately reconstruct the neural oscillation activity of the brain in the time-frequency domain, and is particularly suitable for complex environments with multiple sources, strong correlation sources and low signal-to-noise ratio. The process is as shown in Figure 1 By fusing multi-modal prior information such as MEG, OPM-MEG and fMRI data, the algorithm significantly improves the spatial resolution and temporal accuracy of neural oscillation imaging, and can provide more reliable tools for neuroscience research and clinical applications. Specifically, the following steps are included:

[0027] Step S1, using a sensor to collect brain neural activity data, pre-processing the collected raw data, filtering, artifact removal and time locking, etc. to reduce noise and enhance signal quality.

[0028] Step S2, the cerebral cortex source space is divided into multiple parcel regions. Based on the prior model or data-driven method, the parcel regions are generated using the point cloud growth method and the adjacency relationship of the cortex vertices. Through parcel segmentation, the model parameters to be estimated are reduced, and spatial prior is introduced to improve the accuracy of source reconstruction.

[0029] In particular, the pre-segmentation method for non-locked oscillation data, oscillation source DDP algorithm, is used to divide the cerebral cortex source space into multiple parcel regions, as shown in Figure 2 If multiple non-locked oscillation trials are collected, multiple bandpass filters are designed according to the frequency band of interest, and each trial is filtered. Considering that the design of the bandpass filter may cause phase distortion, a 200-order FIR filter is used for filtering, and zero-padding extension and bidirectional filtering are used to alleviate the edge effect and time delay caused by the FIR filter. The MSP method is used to calculate the activation probability of the filtered data, and the average score of each dipole is calculated. Then, the point cloud growth method and the adjacency relationship of the cortex vertices are used to generate multiple parcel regions.

[0030] For the multiple parcel regions, the The prior distribution of the source is defined as an independent Gaussian distribution, denotes the total number of time instances, wherein the source prior covariance , denotes the number of parcel regions, denotes the th parcel sub-covariance matrix, denotes the th parcel region corresponding weight; wherein can be divided into a series of sub-regional index sets by prior experience, data-driven clustering, and obtained through a spatial smoothing filter and a vertex adjacency matrix.

[0031] Step S3, the optimization problem is described using a Bayesian sparse prior algorithm:

[0032] ,

[0033] wherein, is the sensor covariance, is the noise covariance matrix, denotes the conditional probability distribution of and under the condition of ;

[0034] The objective function is expressed as:

[0035] ,

[0036] In the formula, the sensor covariance estimate , denotes the matrix trace operation, denotes the lead field matrix, and the sensor prior covariance The superscript T denotes the transpose of the matrix, ensures the sparsity of the model structure, reflects the similarity measure between the sensor covariance estimate value and the model learning value.

[0037] For ease of calculation, the auxiliary matrix is defined, denotes the matrix cholesky decomposition, denotes the matrix rearrangement encoding, denotes the non-singular covariance sub-matrix generated by the th parcel sub-covariance matrix corresponding to the index, denotes the dimension of the th parcel sub-covariance matrix;

[0038] The parcel regions obtained by different modalities or data-driven methods are denoted by , denotes the dipole index set composed of the first parcel region;

[0039] The sub-matrix corresponding to the dipole index set is extracted from , where , the elements of the sub-matrix are vectorized into a vector in column-major order:

[0040] ,

[0041] The superscript T represents the transpose of the matrix;

[0042] According to the indexes of the dipole index set , the row index and the column index required for filling the matrix are generated:

[0043] ,

[0044] wherein denotes a full-1 vector, and each element of is filled into the matrix , The prior integration matrix is obtained by merging the matrices generated by different parcels ;

[0045] The source prior covariance and the noise covariance matrix are alternately updated for parameter optimization, including: first updating the source prior covariance , and fixing :

[0046] ,

[0047] ,

[0048] In the formula, denotes the estimated value of the source at time , and the matrix denotes , is the lead field matrix, and the variable is the superscript representing the updated value for the first time, and the matrix denotes the dipole index set​ Corresponding lead field matrix All column elements;

[0049] Then update ,fixed :

[0050] ,

[0051] In the formula, Representation matrix The diagonal elements, Represents the lead field matrix The row element, Indicates the first One diagonal element;

[0052] Through iteration and Two equations estimate brain activity and noise time series until the target loss function converges, and then output the optimal source prior covariance matrix and the optimal noise covariance matrix;

[0053] Step S4: Use spatial priors to improve the accuracy of source reconstruction. At the same time, introduce the results after Bayesian sparse prior processing under the TFC framework to improve the robustness of source reconstruction. Obtain the sensor coherence matrix and source variance of the control segment, task segment and synthesis segment respectively, and finally realize the estimation of oscillation activity, calculate and compare the changes in source power.

[0054] Specifically, the TFC framework is used to output the raw data through a bandpass filter in the target frequency band and to divide the time period of interest into control segment data of the same length. c (t) and task segment data b t (t), and construct the comprehensive data b o (t)=[b c (t),b t [(t)] to improve the robustness of the algorithm and solve the division-to-zero error of traditional time-frequency beamers.

[0055] The output power of the control segment and the mission segment is calculated, and pseudo-statistics are calculated to mitigate the impact of beamformer centering. The rhythmic changes in different time periods are compared, and the results are as follows: Figure 3 As shown:

[0056] ,

[0057] In the formula, This represents the estimate of the rhythmic power variation of the nth source. , , The sensor covariance expectation of multiple trials is described respectively, wherein the subscript o corresponds to the comprehensive section, the subscript t corresponds to the task section, the subscript c corresponds to the control section, f refers to a certain specific frequency, the optimal source prior covariance matrix and the optimal noise covariance matrix are iterated to obtain the sensor covariance matrix when the objective function converges, The multi-trial variance expectation corresponding to the source is represented, and the coherence matrix of different trials is superimposed to improve the estimation accuracy through Bayesian network learning.

[0058] In another aspect, the present application provides a time-frequency multi-sparse prior champagne device, which comprises various modules capable of performing the steps of the aforementioned method, comprising:

[0059] A preprocessing module is used to collect raw data of brain neural activity using sensors and perform preprocessing;

[0060] A segmentation module is used to divide the cerebral cortex source space into multiple parcel regions;

[0061] An estimation module is used to process the data of the multiple parcel regions using a Bayesian sparse prior algorithm, and estimate the optimal source covariance matrix and the optimal noise covariance matrix through the maximum likelihood principle;

[0062] A calculation module is used to introduce the results of Bayesian sparse prior processing under the TFC framework based on spatial prior, to obtain the optimal sensor covariance matrix and the optimal source covariance matrix of the control section, the task section and the comprehensive section respectively, to calculate and compare the changes of source power, and to realize the estimation of oscillatory activity.

[0063] In a third aspect, the present application provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned time-frequency multi-sparse prior champagne method.

[0064] In a fourth aspect, the present application provides a computer-readable storage medium having stored executable instructions, which when executed by a processor, can enable the processor to implement the aforementioned time-frequency multi-sparse prior champagne method.

[0065] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only for specific embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A time-frequency multi-sparse prior champagne method, characterized in that, The method comprises: Step S1: collecting brain nerve activity raw data using a sensor and performing preprocessing; Step S2: dividing the cerebral cortex source space into a plurality of parcel regions based on an oscillation source DDP algorithm, comprising: collecting a plurality of non-locked phase oscillation trials, respectively designing a plurality of band-pass filters according to a frequency band of interest, filtering each trial, calculating the activation probability of the filtered trial using an MSP method, calculating the average score of each dipole, and generating a plurality of parcel regions using a point cloud growth method and the adjacency relationship of the cortex vertices; Step S3: processing the parcel region data using a Bayesian sparse prior algorithm, estimating the optimal source covariance matrix and the optimal noise covariance matrix through the maximum likelihood principle; Step S4: based on the spatial prior, introducing the results of the Bayesian sparse prior processing under the TFC framework, respectively obtaining the optimal sensor covariance matrix and the optimal source covariance matrix of the control section, the task section and the comprehensive section, calculating and comparing the changes in source power, and realizing the estimation of oscillation activity.

2. The time-frequency multi-sparse prior champagne method of claim 1, wherein, The preprocessing in step S1 includes filtering, artifact removal and time locking.

3. The time-frequency multi-sparse prior champagne method of claim 1, wherein, The band-pass filter uses a 200-order FIR filter for filtering, and zero-padding extension and bidirectional filtering are used to alleviate the edge effect and time delay caused by the FIR filter.

4. The time-frequency multi-sparse prior champagne method of claim 1, wherein, For the plurality of parcels regions, different time instants Next, the prior distribution of the source is defined as an independent Gaussian distribution, , denotes the total number of time instants, where the source prior covariance , denotes the number of parcel regions, denotes the th parcel sub-covariance matrix, denotes the th parcel region corresponding weight.

5. The time-frequency multi-sparse prior champagne method of claim 4, wherein, In step S3, the Bayesian sparse prior algorithm is used to describe the optimization problem as: ， wherein, is the sensor covariance, is the noise covariance matrix, denotes the conditional probability distribution of and given the conditionals Objective function is represented as: ， In the formula, the sensor covariance estimation , denotes the matrix trace operation, denotes the lead field matrix, the sensor prior covariance the superscript T represents the transpose of the matrix, guarantees the sparsity of the model structure, reflects the similarity measure of the sensor covariance estimation value and the model learning value; defining an auxiliary matrix , cholesky decomposition of a matrix, reordering encoding of a matrix, denotes a non-singular covariance sub-matrix generated by the th parcel sub-covariance matrix corresponding index, denotes the dimension of the th parcel sub-covariance matrix; The parcel regions obtained by different modalities or data-driven methods are denoted by , denotes the dipole index set consisting of the th parcel region. extracting from the set of dipole indices the corresponding sub-matrix where vectorizing the elements of the sub-matrix in column-major fashion into a vector ， Generating row indices and column indices needed for the padding matrix according to the dipole index set :​​ ； wherein, represents an all-ones vector, and each element of the vector is filled into a matrix , the matrix generated by different parcels is merged to obtain a prior integration matrix ; alternating updates of source prior covariances , noise covariance matrix performing parameter optimization, including: first updating source prior covariances , fixing : ， ， wherein denotes an estimate of the time of origin, matrix denotes , is the lead field matrix, variable the superscript denotes the th update, matrix denotes the dipole index set corresponding to all column elements of the lead field matrix . Then update , fix : ， wherein denotes the diagonal elements of the matrix , denotes the lead field matrix , the element of the th row of the th diagonal element; By iterating and Two equations estimate the brain activity and noise time series until the target loss function converges, and then output the optimal source prior covariance matrix and the optimal noise covariance matrix.

6. The method of claim 1, wherein, In the step S4, the original data is outputted through a band-pass filter of the target frequency band using a TFC framework and the time period of interest is segmented to obtain control segment data of the same length and task segment data , and comprehensive data is constructed ; Calculate the output power of the control section and the task section, and the pseudo-statistics, and compare the rhythm changes in different time sections: ， wherein represents the rhythm power variation estimate of the nth source, , , respectively describe the sensor covariance expectation of multiple trials, wherein subscript o corresponds to the integration segment, subscript t corresponds to the task segment, subscript c corresponds to the control segment, f refers to a certain specific frequency, the optimal source prior covariance matrix and the optimal noise covariance matrix are iterated to obtain the sensor covariance matrix when the objective function converges, represents the multi-trial variance expectation of the corresponding source, and the coherence matrix of different trials is superimposed through Bayesian network learning to improve the estimation accuracy.

7. A time-frequency multi-sparse prior champagne device, characterized in that, It comprises: A preprocessing module for collecting brain nerve activity raw data using a sensor and performing preprocessing; A segmentation module for dividing the cerebral cortex source space into a plurality of parcel regions based on an oscillation source DDP algorithm, comprising: collecting a plurality of non-locked phase oscillation trials, respectively designing a plurality of band-pass filters according to a frequency band of interest, filtering each trial, calculating the activation probability of the filtered trial using an MSP method, calculating the average score of each dipole, and generating a plurality of parcel regions using a point cloud growth method and the adjacency relationship of the cortex vertices; An estimation module for processing the parcel region data using a Bayesian sparse prior algorithm, estimating the optimal source covariance matrix and the optimal noise covariance matrix through the maximum likelihood principle; A calculation module for introducing the results of the Bayesian sparse prior processing under the TFC framework based on the spatial prior, respectively obtaining the optimal sensor covariance matrix and the optimal source covariance matrix of the control section, the task section and the comprehensive section, calculating and comparing the changes in source power, and realizing the estimation of oscillation activity.

8. An electronic device, comprising: It comprises: One or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the time-frequency multi-sparse prior champagne method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Executable instructions are stored thereon, which instructions, when executed by a processor, enable the processor to implement the time-frequency multi-sparse prior champagne method of any one of claims 1-6.

Images