A stereotactic electroencephalogram brain network generation method based on multi-source information

Abstract

The application relates to a stereotactic electroencephalogram brain network generation method based on multi-source information, and relates to the field of brain science research. 1) stereotactic electroencephalogram signals, magnetic resonance imaging (MRI) and CT imaging data after implantation of electrodes are acquired; 2) electrode positioning is performed to obtain the position coordinates of each electrode point in the brain and the brain area to which the electrode point belongs; 3) data processing is performed to obtain preprocessed electroencephalogram signals; 4) wavelet transform is used to analyze the time domain and the frequency domain of the stereotactic electroencephalogram signals, corresponding wavelet coefficients are extracted, and a local brain network is constructed; 5) a multi-source self-encoder model network is trained by using a backward propagation parameter updating mode; 6) multi-layer perception machine model is used to obtain multi-source global latent features by combining the global multi-source electrode spatial coordinates and the local spatial coordinates in a high-dimensional latent space; and global electroencephalogram signals are obtained by decoding through a decoder and relevant functional networks are generated. Global latent features are effectively obtained, and the application has great potential in the application of cognitive function research.

Description

Technical Field

[0001] This invention relates to the field of brain science research, and in particular to a method for generating stereotactic electroencephalogram brain networks based on multi-source information. Background Technology

[0002] The brain is an extremely complex biological system, and brain network research has become a hot topic in the field of neuroscience. Brain networks enable the study of the connection mechanisms between brain regions at the system level, revealing the brain's internal organizational patterns and disease mechanisms. To accurately locate the epileptogenic focus in patients with refractory epilepsy, electrodes are typically implanted deep into the brain—stereotactic electroencephalography (SEEG)—to obtain spatial information from the deep brain. This greatly facilitates the study of neural information extraction, decoding, and regulation of cognitive functions such as perception, memory, language, and emotion, revealing the brain's structure, function, and network mechanisms in cognitive processes.

[0003] While brain networks help us better understand changes in brain connectivity during specific cognitive functions, stereotactic electroencephalography (EEG) presents challenges due to the varying number and placement of implanted electrodes. This makes it difficult to perform group analyses on multiple subjects to uncover general patterns, and also presents significant difficulties in applying artificial intelligence methods for cross-subject analysis or model training. These challenges hinder subsequent research. Therefore, this paper proposes a method for generating stereotactic EEG brain networks based on multi-source information. Summary of the Invention

[0004] The purpose of this invention is to provide a method for generating a stereotactic electroencephalogram (EEG) brain network based on multi-source information. This method addresses the challenge of directly training the features of stereotactic EEG signals in cases where the number and location of implanted electrodes vary. The goal is to generate a global brain network containing multi-source information from local data containing important potential information, thus preparing the ground for revealing the functional connectivity network of cognitive processes. This method has great potential for applications in cognitive function research.

[0005] A method for generating stereotactic electroencephalogram (EEG) brain networks based on multi-source information includes the following steps:

[0006] 1) Data acquisition: Stereotactic EEG signals, MRI images, and CT images after electrode implantation are acquired from existing databases;

[0007] 2) Electrode localization: The location coordinates of each electrode point in the brain and the brain region to which it belongs are obtained by processing MRI and CT images after electrode implantation.

[0008] 3) Data processing: The stereotactic EEG signal is preprocessed to obtain the preprocessed EEG signal;

[0009] 4) Local brain network construction: The time and frequency domains of the stereotactic EEG signals are analyzed using wavelet transform to extract the corresponding wavelet coefficients, and the local brain network is constructed using these wavelet coefficients.

[0010] 5) Multi-source encoder model training: The brain network is input into the multi-source encoder model, and the network of the multi-source encoder model is trained by using the backpropagation parameter update method until the training of the multi-source encoder model is completed.

[0011] 6) Generation of multi-source global brain networks: Using the high-dimensional latent space obtained from the trained multi-source encoder model, combined with the global multi-source electrode spatial coordinates and local spatial coordinates, a multilayer perceptron model is used to obtain multi-source global latent features. The global EEG signal is then decoded by a decoder to generate related functional networks.

[0012] In step 2), the specific steps for electrode localization can be as follows: using FreeSurfer software to process the patient's MRI images, reconstructing and visualizing the cerebral cortex, then using BioImage software to perform electrode localization on the CT scan after electrode implantation, using the SPM12 toolkit to register the MRI and CT scans together onto the MNI template, and finally using the iELVis toolkit to visualize the results.

[0013] In step 3), the specific steps of the data processing can be as follows: performing channel selection on the EEG signal, removing bad channels and channels accompanied by epileptic seizures, downsampling the EEG signal after channel selection to 500Hz, and obtaining the preprocessed stereotactic EEG signal.

[0014] In step 4), the construction process of the local brain network includes:

[0015] Step S1: Filter the preprocessed stereotactic EEG signal using discrete wavelet transform and extract wavelet coefficients;

[0016] Step S2: Calculate the spectral coherence value between multi-channel signals using the extracted wavelet coefficients and frequency bands. The spectral coherence value ranges from [0,1]. If the signals from two electrode channels are identical, their coherence value is 1; otherwise, their spectral coherence value approaches 0. Let the functional connectivity strength between two electrode channels be the spectral coherence value. Construct a graph set G = {G1, G2, ..., G(N)} from the stereotactic EEG data. For the graph Gi = (Vi, Ei, Xi), Vi represents the set of nodes, Ei represents the set of edges, and Xi represents the node attributes. Each node in the graph represents an electrode channel, and the attributes of each node represent the EEG signal of a local brain region. The edges in the graph represent the multi-channel topology based on the functional connectivity network.

[0017] Step S3: Use spectral decomposition to perform feature decomposition on each graph Gi, select the electrode with the largest feature value from the set of electrode points in each brain region as the representative electrode of that brain region, and reconstruct the local graph based on the representative electrode.

[0018] Step S4: Using the spatial coordinates of all representative electrodes obtained by the electrode positioning module, calculate the mean center of the spatial coordinates of the same brain region electrodes of all patients to obtain the global center coordinates of all brain region representative electrodes.

[0019] In step 5), the specific training process of the multi-source encoder model training can be as follows:

[0020] Step S1: Construct the network layers of the multi-source encoder model, which consists of two modules: encoder and decoder. Set the number of network layers and network parameter values, including weights and biases.

[0021] Step S2: Input the local map, temporal features, and spatial coordinates into the encoder to train the model. The output of the model is the reconstructed temporal features and the local functional connectivity brain network. Using the backpropagation algorithm, minimize the mean square error between the input and output, update the weights and the bias to obtain the updated weights and biases, and finally obtain the corresponding high-dimensional latent space to complete the training of the multi-source encoder model.

[0022] Further, in step S1, the initial value of the weight is a random number between [0,1], and the initial vector of the deviation is 0; the encoder is composed of a long short-term memory network layer for extracting temporal features and a graph convolutional layer for extracting brain network features, and the decoder is composed of a multilayer perceptron and a long short-term memory network layer.

[0023] In step 6), the specific steps for generating the multi-source global brain network include:

[0024] (1) Local-to-global mapping: Using the obtained global center coordinates, the spatial coordinates of local graph nodes and the local latent features obtained by the encoder, the mapping relationship between local nodes and the global is learned, and the latent features of global nodes in the latent space are obtained.

[0025] (2) Global feature generation: The global node latent features obtained by the above local-global mapping are input into the decoder module that originates from the encoder model to obtain the global temporal features and the global brain network.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention utilizes an autoencoder model based on multi-source information, employing a local brain network constructed from stereotactic electroencephalogram (EEG) signals using discrete wavelet transform as training input. The trained autoencoder model maps the temporal features of each channel to a high-dimensional latent space and reconstructs the original signal using a decoder, effectively obtaining global latent features. By statistically mapping the local brain networks and temporal features of multiple patients, a global brain network for each patient can be generated. This provides significant convenience for subsequent cross-subject analysis and training of artificial intelligence models, revealing the brain structure, function, and network mechanisms of cognitive processes. Attached Figure Description

[0028] Figure 1 This is a framework diagram of a stereotactic electroencephalogram brain network generation method based on multi-source information according to an embodiment of the present invention;

[0029] Figure 2 This is a visualization result of the electrode positioning module described in an embodiment of the present invention;

[0030] Figure 3 This is a framework diagram of the local brain network construction module described in an embodiment of the present invention;

[0031] Figure 4 The embodiments of this invention are mostly derived from the encoder model schematic diagram;

[0032] Figure 5 This is the specific framework of the multi-source global brain network generation module in an embodiment of the present invention;

[0033] Figure 6 This is a schematic diagram of the input and output of the brain network generation method according to an embodiment of the present invention. Detailed Implementation

[0034] The following embodiments, in conjunction with the accompanying drawings, will further illustrate the present invention. The descriptions of exemplary embodiments are merely illustrative and are in no way intended to limit the invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0035] See Figure 1 A method for generating stereotactic electroencephalogram (EEG) brain networks based on multi-source information, comprising:

[0036] Data acquisition module: used to acquire stereotactic EEG signals, MRI images, and CT images after electrode implantation from existing databases.

[0037] The existing database used in this study comes from a hospital. The stereotactic EEG signals are usually manually marked by experts or doctors based on their experience, marking the epileptic seizure segments. The large amount of epileptic seizure data marking work will inevitably cause fatigue for medical staff, resulting in some segments with epileptic seizures not being marked. The data needs to be carefully checked to distinguish between epileptic seizure and non-seizure segments.

[0038] Electrode localization module: It is used to process MRI and CT images after electrode implantation to obtain the location coordinates of each electrode point in the brain and the brain region to which it belongs.

[0039] The specific process of the electrode localization module is as follows: The patient's MRI images are processed using FreeSurfer software to reconstruct and visualize the cerebral cortex. Then, the CT scans after electrode implantation are used with BioImage software for electrode localization. The SPM12 toolkit is used to register both the MRI and CT scans onto the MNI template. Finally, the results are visualized using the iELVis toolkit. (See [link to relevant documentation]). Figure 2 .

[0040] Data processing module: used to preprocess the stereotactic EEG signal to obtain preprocessed EEG signal.

[0041] The specific processing procedure of the data processing module is as follows: channel selection is performed on the EEG signal to remove bad channels and channels accompanied by epileptic seizures, and the EEG signal after channel selection is downsampled to obtain a preprocessed stereotactic EEG signal.

[0042] The acquisition of stereotactic EEG signals typically depends on the specific condition of the epilepsy patient. The acquisition time and frequency vary from patient to patient. Therefore, the signals cannot be used directly during network training and require standardized preprocessing.

[0043] Step 1: Extract stereotactic EEG signals from all channels of 7 patients in a certain hospital.

[0044] Step 2: The window size is 1s (500Hz*1s), with 500 data points. The step size is also 1s. The original stereotactic EEG data is then subjected to a sliding window operation, and the excess data is discarded.

[0045] Local brain network construction module: used to analyze the time and frequency domains of the stereotactic EEG signal using wavelet transform, extract the corresponding wavelet coefficients, and use the wavelet coefficients to construct a local brain network.

[0046] like Figure 3 As shown, the specific construction process of the local brain network construction module includes:

[0047] Step 1: Filter the preprocessed stereotactic EEG signal using discrete wavelet transform and extract wavelet coefficients.

[0048] Discrete wavelet transform is used for five downsampling operations to simultaneously capture information in both the time and frequency domains. Wavelet transform can extract arbitrary signal frequency bands by changing the size and position of the time window. The equation is as follows:

[0049]

[0050] Among them, 2 j and k*2 j Let represent the time scale parameter and the translation scale parameter, respectively. ψ(t) represents the generating function, which here is the db4 wavelet function. The discrete wavelet transform decomposes the original signal x[n] through high-pass and low-pass filters at the sampling frequency fs, resulting in approximate coefficients A1 with a frequency range of [0, fs / 2] and detail coefficients D1 with a frequency range of [fs / 2, fs]. After five downsampling operations, we obtain A1, D1, A2, D2, A3, D3, A4, D4, A5, D5.

[0051] Step 2: Calculate the spectral coherence value between multi-channel signals using the extracted wavelet coefficients, divided by frequency band. The spectral coherence value ranges from [0,1]. If the signals from two electrode channels are identical, their coherence value is 1; otherwise, their spectral coherence value approaches 0. Let the functional connectivity strength between two electrode channels be the spectral coherence value. Construct a set of graphs, G = {G1, G2, ..., G(N)}, from the stereotactic EEG data.

[0052] For the graph Gi = (Vi, Ei, Xi), Vi represents the set of nodes, Ei represents the set of edges, and Xi represents node attributes. Each node in the graph represents an electrode channel, and the attributes of each node represent the EEG signal of a local brain region. The edges in the graph represent a multi-channel topology based on a functional connectivity network. The formula for calculating the spectral coherence value is as follows:

[0053]

[0054] Among them, P xy P represents the mutual information between signals x and y. xx and P yy These represent the wavelet coefficients of signals x and y, respectively.

[0055] Step 3: Use spectral decomposition to perform feature decomposition on each graph Gi, select the electrode with the largest feature value from the set of electrode points in each brain region as the representative electrode of that brain region, and reconstruct the local graph based on the representative electrode.

[0056] Step 4: Using the spatial coordinates of all representative electrodes obtained by the electrode positioning module, calculate the mean center of the spatial coordinates of the same brain region electrodes of all patients to obtain the global center coordinates of all representative brain region electrodes.

[0057] The stereotactic electroencephalogram (EEG) signals from all channels of seven patients in a certain hospital involved 36 different brain regions. The center coordinates of the representative electrodes for all brain regions are shown below:

[0058]

[0059]

[0060] Among them, unkonwn refers to electrode channels that the electrode positioning module cannot accurately locate. These channels are generally deleted to avoid affecting subsequent generation and research.

[0061] Multi-source encoder model training module: used to input the brain network into the multi-source encoder model, and train the network of the multi-source encoder model using a backpropagation parameter update method until the training of the multi-source encoder model is completed.

[0062] The specific training process of the multi-source encoder model training module is as follows:

[0063] Step 5: Build the network layers of the multi-source encoder model, which consists of two modules: encoder and decoder. Set the number of network layers and the network parameter values, including weights and biases.

[0064] like Figure 4 As shown, the encoder consists of a long short-term memory network layer for extracting temporal features and four graph convolutional layers for extracting brain network features, with feature dimensions set to 10, 32, 64, 64, and 64, respectively. The decoder consists of two multilayer perceptron layers and a long short-term memory network layer, with feature dimensions set to 32, 10, and 1, respectively.

[0065] The initial value of the weight is a random number between [0,1], and the initial vector of the deviation is 0;

[0066] Step 6: Input the local map, temporal features, and spatial coordinates into the encoder to train the model. The model output is the reconstructed temporal features and the local functional connectivity brain network. Using the backpropagation algorithm, minimize the mean square error between the input and output, update the weights and biases to obtain the updated weights and biases, and finally obtain the corresponding high-dimensional latent space, completing the training of the multi-source encoder model.

[0067] Multi-source global brain network generation module: It is used to obtain multi-source global latent features by combining the high-dimensional latent space obtained by the trained multi-source encoder model with the global multi-source electrode spatial coordinates and the local spatial coordinates using a multilayer perceptron model. Then, it is decoded by a decoder to obtain the global EEG signal and generate related functional networks.

[0068] See Figure 5 The multi-source global brain network generation module specifically includes:

[0069] Local-Global Hidden Image Unit: Using the obtained global center coordinates, the spatial coordinates of local graph nodes, and the local latent features obtained by the encoder, the mapping relationship between local nodes and the global graph is learned, and the latent features of global nodes in the latent space are obtained.

[0070] The mapping relationship between the local nodes and the global nodes in the latent space is as follows:

[0071]

[0072] Where k1, k2, and k3 are the three neighboring nodes of global node i in the local nodes. It is the latent representation of local node t in the latent space, x q and Let f represent the spatial coordinates of global node q and local node t, respectively. θ It is a linear layer model, w t These are learnable parameters, initially set to random numbers between [0,1]. The latent representation I(x) of the global node in the latent space can be obtained using this formula. q ).

[0073] Global Feature Generation Unit: Inputs the global node latent features obtained by the above local-global mapping unit into the decoder module of the multi-source encoder model to obtain global temporal features and global brain network.

[0074] In this embodiment, the global brain network generated by each stereotactic electroencephalogram is a 36*36 weighted matrix, which helps to explore the information interaction and topological network of multiple functional brain regions, and helps to deeply understand the network structure and topological characteristics of brain functional states.

[0075] Following the electroencephalogram (EEG) acquisition and analysis process described in this invention, the functional connectivity network properties of the patient's brain in a resting state are further investigated. Figure 6The original brain network, reconstructed brain network, and generated global brain network of a patient are presented. Using SimiNet, the network similarity between the reconstructed and original brain networks was calculated to be 0.9533, and the network similarity between the generated global brain network and the original brain network was 0.9492. This demonstrates that this global brain network construction method can generate functional connections between functional brain regions of a patient without electrode access points, which will greatly facilitate research into brain function.

[0076] Specific examples are used in this invention to illustrate the principles and implementation methods of the invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the invention.

Claims

1. A method for generating stereotactic electroencephalogram (EEG) brain networks based on multi-source information, characterized in that... Includes the following steps: 1) Data acquisition: Stereotactic electroencephalogram (EEG) signals, magnetic resonance imaging (MRI) images, and CT images after electrode implantation are acquired from existing databases; 2) Electrode localization: The location coordinates of each electrode point in the brain and the brain region to which it belongs are obtained by processing MRI and CT images after electrode implantation. 3) Data processing: The stereotactic EEG signal is preprocessed to obtain the preprocessed EEG signal; 4) Local brain network construction: The time and frequency domains of the stereotactic EEG signals are analyzed using wavelet transform to extract the corresponding wavelet coefficients, and the local brain network is constructed using these wavelet coefficients. The construction process of the local brain network includes: Step S1: Filter the preprocessed stereotactic EEG signal using discrete wavelet transform and extract wavelet coefficients; Step S2: Calculate the spectral coherence value between multi-channel signals by dividing the frequency bands using the extracted wavelet coefficients; the range of the spectral coherence value is [0,1]. If the signals of two electrode channels are consistent, their coherence value is 1; otherwise, their spectral coherence value approaches 0; let the functional connectivity strength between two electrode channels be the spectral coherence value, and construct the stereotactic EEG data as a graph set G={G1,G2,…,G(N)}; for the graph Gi=(Vi,Ei,Xi), Vi represents the set of nodes, Ei represents the set of edges, and Xi represents the node attributes; each node in the graph represents an electrode channel, the attributes of each node represent the EEG signal of a local brain region, and the edges in the graph represent the multi-channel topology based on the functional connectivity network; Step S3: Use spectral decomposition to perform feature decomposition on each graph Gi, select the electrode with the largest feature value from the set of electrode points in each brain region as the representative electrode of that brain region, and reconstruct the local graph based on the representative electrode. Step S4: Using the spatial coordinates of all representative electrodes obtained by the electrode localization module, calculate the mean center of the spatial coordinates of the same brain region electrodes of all patients to obtain the global center coordinates of all brain region representative electrodes. 5) Multi-source encoder model training: The brain network is input into the multi-source encoder model, and the network of the multi-source encoder model is trained by using the backpropagation parameter update method until the training of the multi-source encoder model is completed. 6) Generation of multi-source global brain networks: Using the high-dimensional latent space obtained from the trained multi-source encoder model, combined with the global multi-source electrode spatial coordinates and local spatial coordinates, a multilayer perceptron model is used to obtain multi-source global latent features. The global EEG signal is then decoded by a decoder to generate related functional networks.

2. The method for generating stereotactic electroencephalogram brain networks based on multi-source information as described in claim 1, characterized in that... In step 2), the specific steps for electrode localization can be as follows: using FreeSurfer software to process the patient's MRI images, reconstruct and visualize the cerebral cortex, then using BioImage software to perform electrode localization on the CT scan after electrode implantation, using the SPM12 toolkit to register the MRI and CT scans together onto the MNI template, and finally using the iELVis toolkit to visualize the results.

3. The method for generating stereotactic electroencephalogram brain networks based on multi-source information as described in claim 1, characterized in that... In step 3), the specific steps of the data processing are as follows: channel selection is performed on the EEG signal to remove bad channels and channels accompanied by epileptic seizures, and the EEG signal after channel selection is downsampled to 500Hz to obtain the preprocessed stereotactic EEG signal.

4. The method for generating stereotactic electroencephalogram brain networks based on multi-source information as described in claim 1, characterized in that... In step 5), the specific training process of the multi-source encoder model training is as follows: Step S1: Construct the network layers of the multi-source encoder model, which consists of two modules: encoder and decoder. Set the number of network layers and network parameter values, including weights and biases. Step S2: Input the local map, temporal features, and spatial coordinates into the encoder to train the model. The output of the model is the reconstructed temporal features and the local functional connectivity brain network. Using the backpropagation algorithm, minimize the mean square error between the input and output, update the weights and the bias to obtain the updated weights and biases, and finally obtain the corresponding high-dimensional latent space to complete the training of the multi-source encoder model.

5. The method for generating stereotactic electroencephalogram brain networks based on multi-source information as described in claim 4, characterized in that... In step S1, the initial value of the weight is a random number between [0, 1], and the initial vector of the deviation is 0.

6. The method for generating stereotactic electroencephalogram brain networks based on multi-source information as described in claim 4, characterized in that... In step S1, the encoder consists of a long short-term memory network layer for extracting temporal features and a graph convolutional layer for extracting brain network features, and the decoder consists of a multilayer perceptron and a long short-term memory network layer.

7. The method for generating stereotactic electroencephalogram brain networks based on multi-source information as described in claim 1, characterized in that... In step 6), the specific steps for generating the multi-source global brain network include: (1) Local-to-global mapping: Using the obtained global center coordinates, the spatial coordinates of local graph nodes and the local latent features obtained by the encoder, the mapping relationship between local nodes and the global is learned, and the latent features of global nodes in the latent space are obtained. (2) Global feature generation: The global node latent features obtained by the above local-global mapping are input into the decoder module of the multi-source encoder model to obtain the global temporal features and global brain network.

Images