A brain decoding method based on the selection of brain regions of interest and continuous wavelet transform

Abstract

The present invention discloses a brain signal decoding method based on the selection of regions of interest in the brain and continuous wavelet transform; this method establishes a signal conduction model by using a publicly available head anatomical template, obtains cortical electroencephalogram (EEG) using the standardized low-resolution tomography method, and then, based on the geometric parameters of source dipole points and the physiological basis of motor imagery EEG signals, more carefully divides the source regions after source imaging. Further, in order to better balance the characteristics of signals in the time domain and frequency domain, continuous wavelet transform is used for feature extraction, and finally, a convolutional neural network is used to complete the automatic classification and selection of features. The method of the present invention can achieve a high accuracy rate in the four-class motor imagery EEG task and has good physiological interpretability. The method for dividing regions of interest in the brain proposed in the present invention helps to improve the efficiency of EEG analysis by researchers, and at the same time, the proposed decoding method helps to establish a more efficient human-computer interaction system.

Description

Technical Field

[0001] The present invention belongs to the technical fields of computer application technology, biomedical engineering, artificial intelligence, and brain science technology. Specifically, it relates to a brain signal decoding method based on the selection of regions of interest in the brain and continuous wavelet transform. Background Art

[0002] Brain-computer interface technology enables communication between the human brain and external devices by decoding human brain activities into control instructions to control external devices. Motor imagery decoding is a hot research topic in the field of brain-computer interfaces. Due to its non-invasiveness, ease of acquisition, and high temporal resolution, EEG has become one of the most important electroencephalogram signals in the research of motor imagery decoding BCI systems. For example, a BCI system can achieve cursor control and the movement of a robotic arm by identifying the electroencephalogram signals of the brain imagining specific actions, and can even complete text input.

[0003] However, the spatial resolution of electroencephalogram signals collected by a limited number of electrodes covering the scalp is very low, which causes the decoding algorithm to be unable to effectively utilize the spatial information of motor imagery electroencephalogram. The reason for this phenomenon is the volume conduction effect. The volume conduction effect refers to the fact that cortical neuron activities can diverge to different positions on the scalp through brain tissue, which greatly weakens the expression of intracranial neuron activities in the scalp space. The volume conduction effect makes the signal contents measured by multiple sensors similar, further weakening the effectiveness of the decoding algorithm. In the past, methods of increasing the number of scalp electrodes have been used to make up for this defect, but there is always a certain distance limitation between electrodes, which limits the upper limit of the number of electrodes. In addition, some methods have tried to combine multi-modal electromagnetic physiological acquisition technologies, such as collecting near-infrared signals while collecting electroencephalogram, and using the advantage of high spatial resolution of near-infrared signals to make up for the weakness of low spatial resolution of electroencephalogram. However, this approach will greatly increase the application cost and is not convenient for application. Electro-physiological source imaging technology is a method to obtain high spatial resolution. By using MRI to obtain an accurate brain model, high-resolution brain spatial information can be obtained at the same time, and this model can be used for a long time. After obtaining an accurate head model, through the inverse solution algorithm and boundary constraint conditions, the mapping relationship between the activities of cortical neurons and scalp electrode signals can be obtained. At the same time, deep learning technology has also been widely used in the decoding of motor imagery electroencephalogram signals. More and more researchers combine source imaging technology with deep learning technology, especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs). In the practical application of brain-computer interface systems, a very important issue is how to select feature extraction methods. Extracting features in multiple domains during electroencephalogram analysis can achieve better decoding effects.

[0004] In addition, many researchers in the past have analyzed the source imaging signals of high-density scalp electrodes in the alpha band and real intracranial signals, and found that there are significant differences between the two. That is to say, their research provides direct evidence that scalp electroencephalogram can sense subcortical signals. Channel selection is a very important issue in the brain-computer interface system. Selecting electrodes that contain more effective information in the sensor domain can usually improve the accuracy of decoding electroencephalogram signals. Moreover, when selecting channels, there are also different divisions for electrodes covering different brain functional regions. For the signals in the source domain, current research analyzes different brain functional regions based on the Desikan-Killiany parcellation, but there is less signal feature extraction and classification for further subdivision of the source domain regions.

[0005] Therefore, due to the above problems, it is very worthy of research how to design an accurate and physiologically interpretable decoding framework for motor imagery electroencephalogram data, which is also the problem to be solved by the present invention. Summary of the Invention

[0006] To solve the above problems, the present invention provides a brain electrogram decoding method based on the selection of regions of interest in the brain and continuous wavelet transform; this method is based on source imaging technology, designs four different source imaging brain regions through mathematical methods to further refine the spatial representation of electroencephalogram signals, and completes the feature extraction and classification of electroencephalogram through continuous wavelet transform and convolutional neural network; the method for dividing regions of interest proposed in the present invention helps to improve the efficiency of scientific researchers in electroencephalogram analysis, and the proposed decoding method for brain-computer interface control and brain-computer interface pattern recognition is beneficial to helping establish a more efficient human-computer interaction system.

[0007] The present invention is realized through the following technical solutions:

[0008] A brain electrogram decoding method based on the selection of regions of interest in the brain and continuous wavelet transform, which realizes the decoding of motor imagery electroencephalogram signals with physiological interpretability based on the selection of regions of interest in non-invasive neuroimaging and continuous wavelet transform; the specific steps are as follows:

[0009] (1) Data preprocessing stage

[0010] Perform band-pass filtering on the electroencephalogram signals to be analyzed, and convert the electrode reference to the common average reference CAR;

[0011] (2) Source signal generation stage

[0012] Use the finite element model FEM to establish a conduction matrix, and then use the dynamic statistical parametric mapping dSPM method to complete the mapping from scalp electroencephalogram to cortical electroencephalogram;

[0013] (3) Region of interest division stage

[0014] Based on the geometric parameters of the source dipole points and the physiological basis of the motor imagery EEG signals, the source regions after source imaging are divided, and the following four regions of interest are designed: E-ROI, B-ROI, G-ROI, and N-ROI; where: E-ROI is the source region close to the scalp signal acquisition electrodes, B-ROI is the source region in the Desikan-Killiany brain atlas partition located in the primary motor cortex, including the precentral gyrus and the postcentral gyrus, G-ROI is the source region located in the gyrus, and N-ROI is the source region close to the scalp surface;

[0015] (4) Source signal feature extraction stage

[0016] The features of the EEG signals are extracted respectively within the different divided regions of interest by the continuous wavelet transform CWT algorithm.

[0017] (5) Feature classification stage

[0018] The classification architecture of the convolutional neural network CNN is used to classify the EEG signal features of the different regions of interest extracted, and the classification accuracy rates between them are compared to obtain the optimal region of interest in the brain, so as to improve the decoding and classification effect of the EEG data.

[0019] In the present invention, in step (1), the EEG signals to be analyzed come from the BCIC IV IIa and High Gamma public EEG data sets; band-pass filtering uses Butterworth band-pass filtering of 8 - 32 Hz, with average re-reference.

[0020] In the present invention, in step (2), the source signal generation stage includes the following steps:

[0021] ① Create a three-layer head model by segmenting the ICBM152 magnetic resonance image;

[0022] ② Use the finite element method FEM to obtain the lead field matrix, which quantitatively describes the changes during the signal volume propagation process;

[0023] ③ Use the dynamic parameter statistical mapping method dSPM to obtain the source imaging mapping kernel, and directly convert the EEG measured by the scalp sensor into cortical EEG.

[0024] In the present invention, in step (2)②, when calculating the finite element method FEM, the electrical conductivities of the scalp, skull, and brain are set to 0.3300 S / m, 0.0220 S / m, and 0.3300 S / m respectively.

[0025] In the present invention, in step (3), the selection process of the E-ROI region of interest is as follows: First, calculate the normal vector of each source plane in the source space, then calculate the angles between the normal vector and the respective vertices of the source, and select the sources located in the gyrus through angle calculation; on the basis of selecting gyrus sources, further calculate the distances from the electrode spatial positions to the three-dimensional planes of each gyrus source, and select the sources close to the scalp electrodes by setting a distance threshold to obtain the E-ROI region;

[0026] The selection process of the G-ROI region of interest is as follows: Considering that during the source imaging process, the signals of the scalp electrodes are back-projected into the electroencephalogram signals in the source domain, and in the finite element model, the source dipole points are designed as a triangular mesh unit. For a single source dipole point, first obtain the coordinates of the vertices of its triangular mesh A(x a ,y a ,z a ), B(x b ,y b ,z b ), C(x c ,y c ,z c ), and the origin O(0, 0, 0) in the model, and then calculate the normal vector corresponding to the three-dimensional plane of this source Then calculate the angles between the vectors formed by the respective vertices of the source and the origin, namely θ1, θ2, θ3, and select the maximum angle max(θ1, θ2, θ3) as the angular attribute of this source. The calculation method is shown in formulas (1) to (4); after determining the angular attribute of the source, calculate the angular attributes of all sources and set an angular threshold θ threshold as the basis for distinguishing gyrus sources and sulcus sources, and further realize the selection of G-ROI:

[0027]

[0028] max(θ1, θ2, θ3) < θ threshold (4)

[0029] When selecting the N-ROI region of interest, considering the possible insufficient contribution of deep brain sources to scalp electroencephalogram and the influence of volume conduction effects, further restrict the source plane in the z direction, and screen the sources with larger z-direction values by setting a percentage.

[0030] In the present invention, in step (4), the wavelet parameter of the continuous wavelet transform adopted in the feature extraction stage is cmor3-3.

[0031] In the present invention, in step (5), the model framework based on the convolutional neural network CNN in the feature classification stage includes two consecutive convolutional layers, two max pooling layers, and two fully connected layers from front to back. The convolutional layer performs dimensionality reduction and feature extraction on the feature data of the extracted EEG signals through convolutional operations. The convolutional kernel slides on a two-dimensional scale, and a padding mechanism is added to keep the resolution and size of the data unchanged. The max pooling layer is sandwiched between two consecutive convolutional layers. The max pooling layer divides the input two-dimensional data into several rectangular regions and outputs the maximum value for each sub-region. The fully connected layer plays the role of a "classifier" in the entire convolutional neural network. The fully connected layer maps the features learned by the CNN to the sample space for classification.

[0032] In the present invention, the model framework includes self-attention mechanisms in the time domain, frequency domain, and spatial domain.

[0033] In the training process of the present invention, the ratio of the training set to the test set is 4:1, the learning rate is set to 0.001, and the number of training epochs is set to 1000.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] The method of the present invention improves the spatial resolution of the original electrode domain signals through electrophysiological signal source imaging technology, captures the components most relevant to the movement intention from the data according to different designed regions of interest, and constructs an efficient decoding and classification framework through the feature extraction method of wavelet transform and convolutional neural network, and can achieve a high accuracy rate and certain physiological interpretability in the four-class motor imagery EEG task. The method for dividing the regions of interest in the brain proposed in the present invention helps to improve the efficiency of EEG analysis by researchers, and at the same time, the proposed decoding method helps to establish a more efficient human-computer interaction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the framework of the entire decoding system.

[0037] Figure 2 It is a diagram showing the geometric features of the source dipole points (three-dimensional grid) of the present invention and the method for selecting gyrus sources and sulcus sources.

[0038] Figure 3 It is a schematic diagram of the CNN network structure of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] The decoding framework proposed by the present invention mainly includes three parts, namely source signal generation, division of brain regions of interest, and feature extraction and classification, corresponding to Figure 1 (a), (b), and (c) in []. First, simple preprocessing is performed on the data. Only the Butterworth filter is used to filter the source signal at 8 - 32 Hz, and average reference rereferencing is used to improve the signal accuracy.

[0041] In the source signal generation part, in the first step, a three - layer head model is created by segmenting the ICBM152 magnetic resonance image. In the second step, based on the head model generated in the previous step, the lead field matrix is obtained using the finite element (FEM) method, which quantitatively describes the changes during the volume propagation of the signal. In the last step, the dynamic parameter statistical mapping method (dSPM) is used to obtain the source imaging mapping kernel, which is a constant - coefficient matrix. Through conventional matrix multiplication calculation, the electroencephalogram measured by scalp sensors can be directly converted into cortical electroencephalogram. The whole process is as Figure 1 shown in (a) in [].

[0042] In the division of brain regions of interest part, E - ROI is the source region selected near the scalp signal acquisition electrodes, B - ROI is the source region located in the primary motor cortex in the Desikan - Killiany brain atlas partition (including the precentral gyrus and the postcentral gyrus), G - ROI is the source region selected on the gyrus, and N - ROI is the source region selected near the scalp surface. The whole process is as Figure 1 shown in (b) in [].

[0043] In the feature extraction and classification part, the source signal is divided into a training set and a test set in a ratio of 4:1. The position of this process in the whole framework is as Figure 1 shown in the left half of (c) in []. The continuous wavelet transform using the cmor3 - 3 wavelet parameters is used for feature extraction.

[0044] More specifically, the present invention provides an electroencephalogram decoding method based on the selection of brain regions of interest and continuous wavelet transform, which realizes the decoding of motor imagery electroencephalogram signals with physiological interpretability based on the selection of brain regions of interest in non - invasive neuroimaging and continuous wavelet transform; it includes the following steps:

[0045] 1. Use the BCIC IV IIa and High Gamma public electroencephalogram datasets as the electroencephalogram signals to be analyzed;

[0046] 2. In the preprocessing stage, first, band-pass filtering of the EEG is performed at 8 - 32 Hz, and the electrode reference is converted to common average reference (CAR).

[0047] 3. Use the finite element (FEM) model to establish the conduction matrix, and then use the dynamic statistical parametric mapping (dSPM) method to complete the mapping from scalp EEG to cortical EEG.

[0048] 4. Design four different regions of interest (ROIs) through mathematical methods, including E - ROI, B - ROI, G - ROI, and N - ROI. The regional positions of each ROI are as shown in (b) of Figure 1 ;

[0049] 5. In the feature extraction part, use the continuous wavelet transform algorithm to extract the features of the EEG signals from different regions of interest respectively.

[0050] 6. In the data feature classification part, use the PyTorch tool to build a classification architecture based on the convolutional neural network (CNN) model, decode and classify the features extracted from different regions of interest, and compare the classification accuracies between them to obtain the optimal region of interest.

[0051] In step 1, the High Gamma and BCIC IV IIa datasets contain 14 and 9 subjects respectively. In the High Gamma dataset, each subject is required to perform four motor imagery tasks: left hand, right hand, foot, and rest, with a total of 640 trials of data for each subject. In the BCIC IV IIa dataset, each subject is required to perform 4 motor imagery (left hand, right hand, foot, and tongue) tasks, with a total of 576 trials of data for each subject.

[0052] In step 2, a Butterworth filter is used for 8 - 32 Hz band-pass filtering. Using band-pass filtering can make the subsequent frequency band division sufficiently average, and at the same time can effectively remove low-frequency drift and power frequency interference. Using average re-reference is beneficial to improving the positioning accuracy of the next source imaging.

[0053] In step 3, the mapping relationship of the whole process can be expressed as formula (5)

[0054] L·S source =S sensor (5) S source and S sensor represent the source dipole and scalp EEG signal respectively, and L represents the mapping matrix.

[0055] However, since the source dipole points usually have more elements than scalp electrodes, this constitutes an "under-determined" problem of "non-uniqueness". That is, the same scalp electrode measurement values may produce different source space pattern signals. The common approach is to find some constraints through linear estimation methods to estimate the most likely source current distribution Equations (6) - (8) describe this process, where E is not just a simple inverse matrix representation of L, but a multi-parameter expression that mixes the regularization parameter λ and the noise covariance C

[0056]

[0057] E MNE = G MNE L(7)

[0058] L T (LL T + λC) -1 = G MNE (8)

[0059] Different from the Minimum Norm Estimation (MNE), the Dynamic Statistical Parametric Mapping (dSPM) used in our invention normalizes the estimated source activity when using the noise covariance matrix. This method thresholds the amplitude of the expected current and converts it into a dimensionless statistical test variable by dividing by the corresponding noise variance. This "depth weighting" of higher amplitude source activities can significantly improve the localization error, and the corresponding weighting matrix is represented as W dSPM . The corresponding formula descriptions are shown in Equations (9) and (10):

[0060] G dSPM = W dSPM G MNE (9)

[0061]

[0062] where diag represents the elements on the diagonal of the matrix. E in Equation (2) MNE can be further replaced by G dSPM .

[0063] In the ROI selection stage of Step 4, in order to avoid the increase in computational complexity caused by the increase in the number of EEG signal channels after electrophysiological source imaging, in the present invention, based on mathematical methods and physiological theory, four different regions of interest in the brain are designed; there are a total of the following four regions of interest in the brain

[0064] 1) G-ROI (the Gyrus region of interset) selection method: As Figure 1In sub - figure (b), the second sub - figure from left to right, G - ROI is to select the source dipole points located in the gyrus region. The selection method of G - ROI is as Figure 2 shown. During the source imaging process, the signals of scalp electrodes are inverse - solved into EEG signals in the source domain, and in the finite - element model, the source dipole points are designed as triangular mesh elements.

[0065] From Figure 2 , it can be seen that for a single source dipole point, we can obtain the coordinates of each vertex of its triangular mesh A(x a , y a , z a ), B(x b , y b , z b ), C(x c , y c , z c ) and the origin O(0, 0, 0) in the model. Then, the normal vector corresponding to the three - dimensional plane of this source is calculated. We calculate the angles θ1, θ2, θ3 between the vectors formed by each vertex of the source and the origin, and select the maximum angle max(θ1, θ2, θ3) as the angular attribute of the source. The calculation method is shown in formulas (1) to (4). After determining the angular attribute of the source, we calculate the angular attributes of all sources and set an angular threshold θ threshold as the basis for us to distinguish gyrus sources and sulcus sources.

[0066]

[0067] max(θ1, θ2, θ3) < θ threshold (4)

[0068] 2) Selection method of N - ROI (the Z - normalized region of interest): Next, the selection idea of N - ROI is introduced, as shown in the first sub - figure from left to right in (b) of Figure 1 . As we introduced before, the source plane has three vertices with their respective spatial positions. Considering that deep - brain sources may contribute insufficiently to scalp EEG and the influence of volume conduction effects, we further restrict the source plane in the z - direction by setting a percentage to screen sources with larger z - direction values.

[0069] 3) Selection method of E - ROI (the region of interst closed to electrode): As shown in Figure 1For the fourth sub - figure from left to right in (b) of [], the selection process of E - ROI is divided into the following steps: First, calculate the normal vector of each source plane in the source space, then calculate the angles between the normal vector and each vertex of the source. Select the sources located in the gyrus through angle calculation. On the basis of selecting gyrus sources, we further calculate the distances from the electrode spatial positions to the three - dimensional planes of each gyrus source, and select the sources close to the scalp electrodes by setting a distance threshold to obtain the E - ROI region.

[0070] 4) Method for selecting B - ROI (the Brodmann region of interest): As Figure 1 For the third sub - figure from left to right in (b) of [], the B - ROI (the Brodmann region of interest) selects the region located in the motor cortex, including three parts: the primary motor cortex, the premotor cortex, and the supplementary motor area. This region is the area in the cerebral cortex that participates in planning, controlling, and executing voluntary movements, and is anatomically called the "precentral gyrus" and "postcentral gyrus". The primary motor cortex is the main part of the brain that generates neural activities and controls the execution of human movements. The premotor cortex and the supplementary motor area are responsible for the accuracy of human movement control, sensory guidance, and the planning and coordination of movements. Select the two regions of the precentral gyrus and the postcentral gyrus of the Desikan - Killiany brain atlas. These two regions are not only related to human cognition and perception, but also include Brodmann area 4 (primary motor cortex) and area 6 (premotor layer), and these two regions have been proven to play a very important role in motor imagery.

[0071] These four different methods of dividing regions are separately used for feature extraction and classification accuracy evaluation in the subsequent operation steps. Different regions of interest have different physiological interpretabilities according to the above introduction, thus having different degrees of influence on the decoding effect. There are different degrees of overlap between these four regions of interest in the brain: G - ROI can be regarded as the part of N - ROI that only retains the gyrus region, B - ROI can be regarded as the part of N - ROI that only retains the precentral gyrus and postcentral gyrus regions, and E - ROI can be regarded as the part of N - ROI that only retains the region around the actual positions of the electrode points.

[0072] In step 5, continuous wavelet transform is used to extract signal features. The specific method is to perform a convolution operation on a signal and a wavelet basis function, decompose the signal into components of different times and frequency bands, and then conduct further analysis. The key point of wavelet transform lies in selecting an appropriate wavelet basis, which can make a set of wavelet bases in the form of translation and scaling constitute an orthogonal basis, and the wavelet basis function is at least continuous or continuously differentiable. For analyzing the signal f(t), CWT is a linear transform with the wavelet ψ(t) as the kernel function. Assuming that the wavelet transform of the signal f(t) with finite energy is to displace the mother wavelet function by b and then take the inner product with the signal f(t) to be analyzed under different scale factors a:

[0073]

[0074] Let f(t) be a square-integrable function, be the wavelet function. The above formula is called the wavelet transform of f(t). Among them, a is the scale factor (a > 0), b is the time shift factor, which can be positive or negative. And represents 's conjugate function, and the symbol <x, y> represents the inner product, whose meaning is

[0075] <x(t), y(t)> = ∫x(t)y * (t)dt (12)

[0076] Since t, a, and b in the formula are all continuous variables, this kind of wavelet transform is called continuous wavelet transform, denoted as CWT (Continuous Wavelet Transform). In the present invention, the wavelet parameter is selected as cmor3-3.

[0077] In step 6, a decoding framework based on a convolutional neural network model is used. Figure 3The network model based on the CNN architecture implemented by the present invention is listed in detail, which includes two convolutional layers, two max pooling layers, and two fully connected layers. The convolutional layer performs dimensionality reduction and feature extraction on the input data through convolutional operations. In the present invention, the convolutional kernel slides on a two-dimensional scale, and a padding mechanism is added to keep the resolution and size of the data unchanged. The pooling layer is sandwiched between consecutive convolutional layers and is used to compress the data and reduce the number of parameters. The max pooling layer divides the input two-dimensional data into several rectangular regions and outputs the maximum value for each sub-region. The pooling layer continuously reduces the spatial size of the data, so the number of parameters and the amount of computation also decrease, which controls overfitting to a certain extent. Generally speaking, pooling layers are periodically inserted into the convolutional layers of CNNs to reduce the mean shift caused by parameter errors in the convolutional layers. The fully connected layer plays the role of a "classifier" in the entire convolutional neural network. The fully connected layer maps the features learned by the CNN to the sample space to achieve classification.

[0078] Table 1 shows the parameter situation in the CNN network model. For the sake of easy display, some matrix sizes are simplified. During model training, the ratio of the training set to the test set is 4:1, the learning rate is set to 0.001, and the epoch size is 1000. Training is completed on NVIDIA RTX 3090 (24G), and the toolkit used to build the network is PyTorch. The features obtained in the feature extraction stage are fed into the introduced model, and then the probabilities belonging to four categories can be calculated. Finally, the label of the category with the highest probability is returned as the predicted value of the feature, and end-to-end data feature classification is achieved through this method.

[0079] Table 1 Parameters of Each Layer of the CNN Network

[0080]

[0081] Table 2 Classification Accuracy Rates under Different ROI Methods in the BCIC IV 2a Dataset

[0082]

[0083] Table 3 Classification Accuracy Rates under Different ROI Methods in the HGD Dataset

[0084]

[0085]

[0086] Tables 2 and 3 compare the impact of different ROI partitioning methods on classification accuracy on the two datasets, respectively. The baseline method used in these comparisons is the original data without ROI partitioning. It can be seen that the B-ROI method achieves the highest accuracy for the BCIC IV IIa dataset, while the G-ROI method achieves the highest classification accuracy for the HGD dataset. These results demonstrate that different ROI partitioning methods can effectively improve EEG data decoding and classification performance across different datasets.

[0087] In summary, the present invention utilizes the above-mentioned design methods and data to develop a motor imagery EEG decoding framework, designs four different ROI brain region division and selection methods, analyzes the contribution of gyral and sulcal sources to EEG signal decoding, provides physiological interpretability for different ROIs, and obtains the optimal brain region division method after comparing classification accuracy. In short, the present invention has the potential to achieve EEG motor imagery decoding based on different ROI region selection or further ROI subdivision, providing a new approach for the portability and accuracy of brain-computer interface systems.

[0088] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, and it should be clear that the present invention is not limited to the scope of the specific embodiments, it is obvious to those skilled in the art that as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

Claims

1. A brain signal decoding method based on the selection of brain regions of interest and continuous wavelet transform, characterized in that, It realizes the decoding of motor imagery electroencephalogram (EEG) signals with physiological interpretability based on the selection of brain regions of interest in non-invasive neuroimaging and continuous wavelet transform; The specific steps are as follows: (1) Data preprocessing stage Perform band-pass filtering on the EEG signals to be analyzed and convert the electrode reference to the common average reference (CAR); (2) Source signal generation stage Use the finite element model (FEM) to establish a conduction matrix, and then use the dynamic statistical parametric mapping (dSPM) method to complete the mapping from scalp EEG to cortical EEG; (3) Region of interest (ROI) division stage Based on the geometric parameters of the source dipole points and the physiological basis of motor imagery EEG signals, divide the source region after source imaging, and design the following four regions of interest: E-ROI, B-ROI, G-ROI, and N-ROI; where: E-ROI is the source region close to the scalp signal acquisition electrodes, B-ROI is the source region in the Desikan-Killiany brain atlas partition located in the primary motor cortex, including the precentral gyrus and the postcentral gyrus, G-ROI is the source region located in the gyrus, and N-ROI is the source region close to the scalp surface; (4) Source signal feature extraction stage After dividing different regions of interest, extract the features of their respective EEG signals through the continuous wavelet transform (CWT) algorithm; (5) Feature classification stage Adopt the classification architecture of the convolutional neural network (CNN) to decode and classify the EEG signal features of different regions of interest extracted, and compare the classification accuracies between them to obtain the optimal brain region of interest, improving the decoding and classification effect of EEG data; where: In step (3), the selection process of the E-ROI region of interest is as follows: First, calculate the normal vector of each source plane in the source space, then calculate the angles between the normal vector and the vertices of the source, and select the source located in the gyrus through angle calculation; on the basis of selecting the gyrus source, further calculate the distance from the electrode spatial position to the three-dimensional plane of each gyrus source, and select the source close to the scalp electrode by setting a distance threshold to obtain the E-ROI region; The process of selecting the G-ROI (region of interest) is as follows: Considering that during the source imaging process, the signals of scalp electrodes are inverse solved into EEG signals in the source domain, and in the finite element model, the source dipole points are designed as triangular mesh elements. For a single source dipole point, first obtain the coordinates of the vertices of its triangular mesh: A(x a , y a , z a ), B(x b , y b , z b ), C(x c , y c , z c ) and the origin O(0, 0, 0) in the model. Then calculate the normal vector corresponding to the three-dimensional plane of this source Next, calculate the angles between the vectors formed by each vertex of the source and the origin, namely θ1, θ2, θ3, and select the maximum angle max(θ1, θ2, θ3) as the angular attribute of this source. The calculation method is shown in formulas (1) to (4); after determining the angular attribute of the source, calculate the angular attributes of all sources and set an angular threshold θ threshold as the basis for distinguishing gyrus sources and sulcus sources, and further realize the selection of the G-ROI: max(θ1, θ2, θ3) < θ hreshold (4) When selecting the N-ROI region of interest, considering the influence of the insufficient contribution probability of deep brain sources to scalp EEG and the volume conduction effect, further restrict the source plane in the z direction, and screen the sources with larger z-direction values by setting a percentage.

2. The brain decoding method according to claim 1, wherein, In step (1), the EEG signals to be analyzed come from the BCIC IV 2a and High Gamma public EEG datasets; band-pass filtering uses Butterworth band-pass filtering with 8 - 32 Hz and average rereferencing.

3. The brain decoding method according to claim 1, characterized in that, In step (2), the source signal generation stage includes the following steps: ① Create a three-layer head model by segmenting the ICBM152 magnetic resonance image; ② Use the finite element method (FEM) to obtain the lead field matrix, which quantitatively describes the changes during the signal volume propagation process; ③ Use the dynamic statistical parametric mapping method (dSPM) to obtain the source imaging mapping kernel, and directly convert the EEG measured by the scalp sensor into cortical EEG.

4. The brain decoding method according to claim 3, wherein In step ②, when calculating the finite element method (FEM), the electrical conductivities of the scalp, skull, and brain are set to 0.3300 S / m, 0.0220 S / m, and 0.3300 S / m, respectively.

5. The brain decoding method according to claim 1, wherein In step (4), the wavelet parameter of the continuous wavelet transform adopted in the feature extraction stage is cmor3-3.

6. The brain decoding method according to claim 1, wherein In step (5), the model framework based on the convolutional neural network (CNN) adopted in the feature classification stage includes two consecutive convolutional layers, two max pooling layers, and two fully connected layers from front to back. The convolutional layer reduces the dimension and extracts features from the feature data of the extracted EEG signals through convolutional operations. The convolutional kernel slides on a two-dimensional scale, and a padding mechanism is added to keep the resolution and size of the data unchanged. The max pooling layer is sandwiched between two consecutive convolutional layers. The max pooling layer divides the input two-dimensional data into several rectangular regions and outputs the maximum value for each sub-region. The fully connected layer plays the role of a "classifier" in the entire convolutional neural network. The fully connected layer maps the features learned by the CNN to the sample space to achieve classification.

7. The brain decoding method according to claim 6, wherein During the training process of the model, the ratio of the training set to the test set is 4:1, the learning rate is set to 0.001, and the number of training epochs is set to 1000.

Images