A method and system for tracing internal brain neural activity based on scalp electroencephalogram signals

Abstract

The application discloses a scalp EEG signal-based brain internal nerve activity tracing method and system, and the method comprises the following steps: a virtual brain model is built by using a neurocluster model with biological interpretation, a whole brain map and brain physiological connection, and the number of the neurocluster model is recorded; brain internal nerve activity generated by the virtual brain model is converted into scalp EEG signals according to an EEG measurement model; a brain internal nerve activity tracing neural network is constructed; internal brain activity data with spatial characteristics and time characteristics are generated by using the virtual brain model, and corresponding EEG signals are obtained through the EEG measurement model; the obtained internal brain activity data and the corresponding EEG signals are used to construct a data set for training the brain internal nerve activity tracing neural network. The brain internal nerve activity tracing neural network is constructed by cascading a spatial module and a time module, so that the stability and accuracy of brain internal nerve activity tracing are improved.

Description

Technical fields:

[0001] This invention relates to a method and system for tracing the source of internal neural activity in the brain based on scalp electroencephalogram (EEG) signals, particularly for tracing the source of internal cortical brain activity using EEG signals obtained through non-invasive EEG acquisition devices. Background technology:

[0002] Developing non-invasive brain imaging techniques to provide a high spatial resolution, high temporal resolution, and whole-brain coverage method for imaging complex and distributed brain activity is crucial for understanding brain function and dysfunction, as well as for the clinical management of various brain diseases. Among numerous measurement techniques, electroencephalography (EEG) signals have become the most widely studied technique due to their low cost, good temporal resolution, and ease of use. However, due to volume conduction effects and low signal-to-noise ratio, EEG is limited in terms of spatial resolution for measuring neural activity, necessitating accurate estimation of brain activity in both time and space through tracing the source of intracranial neural activity.

[0003] Traditional methods for tracing the source of intracranial neural activity often use equivalent current dipole models to simulate brain activity, relying on prior knowledge of brain activity to determine regularization terms in the source model. Existing deep learning methods, due to the lack of good correspondence data between intracranial neural activity and scalp EEG data, are limited to tasks such as classification and feature extraction. Therefore, it is necessary to develop a framework for tracing the source of intracranial neural activity based on virtual brain simulation data and deep learning, to extract accurate and robust estimates of intracranial neural activity at the temporal and spatial levels from EEG. Summary of the Invention:

[0004] The purpose of this invention is to provide a method and system for tracing the source of internal neural activity in the brain based on scalp electroencephalogram (EEG) signals. A virtual brain model is constructed using a neural cluster model to provide data for neural network optimization, and a neural network for tracing brain activity is established to achieve the tracing of the source of internal neural activity in the brain.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention provides a method for tracing the source of internal neural activity in the brain based on scalp electroencephalogram (EEG) signals, the method comprising the following:

[0007] Part 1: Building a Virtual Brain Model

[0008] A virtual brain model was built using physiologically interpretable neural cluster models, whole-brain maps, and brain physiological connections, and the number of neural cluster models was recorded.

[0009] The internal neural activity of the brain generated based on the virtual brain model was converted into scalp electroencephalogram (EEG) signals using an EEG measurement model.

[0010] Part Two: Constructing a neural network for tracing the source of neural activity within the brain:

[0011] The input of the neural network for tracing the source of brain activity is the scalp EEG signal, and the output is the brain activity. The neural network for tracing the source of brain activity is obtained by cascading a spatial module and a temporal module. The spatial module includes a transformation matrix and five layers of deconvolution in series. The input scalp EEG signal is processed by the five layers of deconvolution on one hand, and multiplied by the transformation matrix on the other hand. The output of the spatial module is obtained by adding the result of the five layers of deconvolution multiplication with the result of the transformation matrix.

[0012] The time module comprises three layers of recurrent neural network (RNN) connected in series.

[0013] The number of rows in the transformation matrix corresponds to the number of leads in the scalp EEG signal, and the number of columns corresponds to the number of neural cluster models in the virtual brain model.

[0014] Part Three: Using a virtual brain model to generate internal brain activity data with spatial and temporal characteristics, and acquiring corresponding EEG signals through an EEG measurement model; constructing a dataset from the obtained internal brain activity data and corresponding EEG signals to train a neural network for tracing the source of internal brain activity, and obtaining the trained neural network for tracing the source of internal brain activity.

[0015] During training, the neural network for tracing the source of neural activity within the brain uses temporal accuracy loss (TL) and spatial accuracy loss (SL) as loss functions.

[0016] The crossover ratio of source regions is used to measure the spatial accuracy of tracing neural activity within the brain, and the difference between the crossover ratio of source regions and 1 is calculated as the spatial accuracy loss.

[0017] Temporal similarity is used to measure the temporal accuracy of tracing the origins of neural activity within the brain, and the difference between temporal similarity and 1 is calculated as the loss of temporal accuracy.

[0018]

[0019]

[0020] t s The sampling interval is t, where t is time, T is the length of the time window, and d is the number of brain regions whose neural activity is to be traced. H represents the brain neural activity estimated by the source neural network for the θ-th brain region at time t; H(994-θ,t) represents the actual brain neural activity of the 994-θ-th brain region at time t; and H(θ,Tt) represents the actual brain neural activity of the θ-th brain region at time Tt.

[0021] The EEG measurement model is I = Q(L(θ))H(θ) + Q(L(θ)). (i) ))H i (θ (i) )+Q(L(θ (b) ))H b (θ (b) )+n m ,in, L(θ) represents the electroencephalogram (EEG) signals collected from the electrodes of the EEG acquisition device during time T, starting from time t, where m is the number of leads and L(θ) ∈ R. 3 It is the location of the target brain activity source and θ∈(θ1,...,θ) d ), where θ is a brain region, Q(L(θ)) is the spatial conductance matrix of the target brain activity, and H(θ)∈R d It is the amplitude of the target brain activity, which is also the amplitude of the target equivalent current dipole, L(θ). (i) )∈R 3 It is the location of the source of interference with brain activity and Q(L(θ (i) H is the spatial conductance matrix that interferes with brain activity. i (θ (i) )∈R k There are k interfering brain activities, where i represents the interfering brain and b represents the background brain, L(θ) (b) )∈R 3 It is the location of background brain activity and Q(L(θ (b) H is the spatial conductance matrix of background brain activity. b (θ (b) )∈R p There are p background brain activities, n m The measurement noise of the electrode is expressed in the form of Gaussian white noise.

[0022] Secondly, the present invention provides a brain internal neural activity tracing system based on scalp electroencephalogram (EEG) signals, the system comprising a virtual brain model, an EEG measurement module, and a brain internal neural activity tracing neural network;

[0023] The virtual brain model is constructed based on a neural cluster model, a whole-brain map, and brain physiological connections, and is used to generate internal neural activity in the brain.

[0024] The EEG measurement module is used to convert the internal neural activity of the brain generated based on the virtual brain model into scalp EEG signals.

[0025] The neural network for tracing the source of brain activity takes scalp electroencephalogram (EEG) signals as input and outputs brain activity as output. This neural network is obtained by cascading spatial and temporal modules.

[0026] A virtual brain model is used to generate internal brain activity data with spatial and temporal characteristics, and the corresponding electroencephalogram (EEG) signals are obtained through an EEG measurement module. A dataset of internal brain activity data and corresponding EEG signals is constructed to train a neural network for tracing the source of internal brain activity. The trained neural network is then used for tracing the source of internal brain activity.

[0027] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described thereon.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] (1) The present invention employs a neural network for tracing the source of neural activity in the brain, which is constructed by cascading spatial modules and temporal modules. The spatial modules are constructed using spatial features in EEG signals, and the temporal modules are constructed using temporal features in EEG signals, thereby improving the stability and accuracy of tracing the source of neural activity in the brain.

[0030] (2) This invention uses a virtual brain model based on a neural cluster model to generate neural activity inside the brain and calculates the scalp EEG signals in sequence to obtain training and optimization data with physiological interpretability for the neural activity tracing neural network inside the brain, thereby improving the robustness and generalization ability of the neural activity tracing neural network inside the brain.

[0031] (3) The present invention uses temporal accuracy loss and spatial accuracy loss as loss functions of the neural network for tracing the source of neural activity in the brain, thereby improving the accuracy of tracing the source of neural activity in the brain from both spatial and temporal aspects. Attached image description:

[0032] Figure 1 This is a schematic diagram illustrating the training data for the neural network that generates the source neural activity of the brain using a neural cluster model and an electroencephalogram (EEG) measurement model, as described in this invention.

[0033] Figure 2 This is a schematic diagram of the architecture of the brain neural activity tracing method based on scalp electroencephalogram (EEG) signals of the present invention. Detailed implementation method:

[0034] The present invention will be further explained below with reference to examples and accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0035] This invention relates to a method for tracing the source of intrabrain neural activity based on scalp electroencephalogram (EEG) signals, the method comprising the following:

[0036] Part 1: Building a Virtual Brain Model and an EEG Measurement Model

[0037] Utilizing physiologically interpretable neural cluster models (NMMs), whole-brain maps, and brain physiological connectivity (KP) θ and a d (θ) Construct a virtual brain model and record the number of neural cluster models;

[0038] The EEG measurement model is I = Q(L(θ))H(θ) + Q(L(θ)). (i) ))H i (θ (i) )+Q(L(θ (b) ))H b (θ (b) )+n m ,in, L(θ)∈R is the electroencephalogram (EEG) signal collected from the electrodes of the EEG acquisition device during time T starting from time t. 3 It is the location of the target brain activity source and θ∈(θ1,...,θ) d Q(L(θ)) is the spatial conductance matrix of the target brain activity, H(θ)∈R d It is the amplitude of the target brain activity, which is also the amplitude of the target equivalent current dipole, L(θ). (i) )∈R 3 It is the location of the source of interference with brain activity and Q(L(θ (i) H is the spatial conductance matrix that interferes with brain activity. i (θ (i) )∈R k There are k interfering brain activities, L(θ) (b) )∈R 3 It is the location of background brain activity and Q(L(θ (b) H is the spatial conductance matrix of background brain activity. b (θ (b) )∈R p There are p background brain activities, n m The measurement noise of the electrode is expressed in the form of Gaussian white noise;

[0039] Part Two: Constructing a neural network for tracing the source of neural activity within the brain (represented by STSIN):

[0040] The input of the neural network for tracing the internal brain activity is the scalp EEG signal, and the output is the internal brain activity. The neural network for tracing the internal brain activity is obtained by cascading a spatial module and a temporal module. The spatial module adopts a deconvolutional neural network as its basic structure, including a transformation matrix and five layers of deconvolution in series. The input scalp EEG signal is processed by the five layers of deconvolution on one hand, and multiplied by the transformation matrix on the other hand. The output of the spatial module is obtained by adding the result of the five layers of deconvolution multiplication with the result of the transformation matrix.

[0041] The time module uses a recurrent neural network as its basic structure, which includes three layers of recurrent neural network (RNN) connected in series.

[0042] The number of rows in the transformation matrix corresponds to the number of leads in the scalp EEG signal, and the number of columns corresponds to the number of neural cluster models in the virtual brain model.

[0043] The spatial module of the neural network for tracing the source of internal brain activity is designed to address the spatial multi-lead characteristics of scalp EEG signals, extracting potential spatial features; the temporal module is designed to address the low-frequency temporal characteristics of scalp EEG signals, extracting potential temporal features. Through these spatial and temporal modules, the neural network for tracing the source of internal brain activity overcomes the instability and strong interference of scalp EEG signals, improving the localization and temporal accuracy of tracing the source of internal brain activity.

[0044] Part Three: Using a virtual brain model to generate internal brain activity data with spatial and temporal characteristics, and acquiring corresponding EEG signals through an EEG measurement model; constructing a dataset from the obtained internal brain activity data and corresponding EEG signals to train a neural network for tracing the source of internal brain activity, and obtaining the trained neural network for tracing the source of internal brain activity.

[0045] Methods for calculating and optimizing the temporal and spatial accuracy losses of neural networks used to trace neural activity within the brain:

[0046] The source region crossover ratio is used to measure the spatial accuracy of tracing the source of neural activity in the brain. The difference between the source region crossover ratio and 1 is calculated as the spatial accuracy loss. The source region crossover ratio is: the area of ​​overlap between the brain region with active neural activity after tracing and the actual active brain region divided by the area of ​​the combined region.

[0047] Temporal similarity is used to measure the temporal accuracy of tracing the source of neural activity in the brain. The difference between temporal similarity and 1 is calculated as the loss of temporal accuracy. Temporal similarity is the phase synchronization index between the sequence of neural activity in the brain estimated by the neural network for tracing the source of neural activity and the sequence of neural activity generated by the virtual brain model.

[0048] By using internal brain activity data with spatial and temporal characteristics and EEG signals to train a neural network for tracing the source of internal brain activity, the accuracy of localization and temporal sequencing of tracing the source of internal brain activity is improved by enhancing the physiological interpretability of the neural network.

[0049] Example 1

[0050] This embodiment of the method for tracing the source of internal neural activity in the brain based on scalp electroencephalogram (EEG) signals includes the following steps:

[0051] The first step involves constructing a Spatial-Temporal Source Imaging Network (STSIN) using cascaded spatial and temporal modules to estimate source activity (SA) of the brain based on scalp electroencephalogram (EEG) signals. The spatial module, consisting of five layers of deconvolution and residual connections in the transformation matrix, leverages the multi-lead spatial characteristics of EEG to upscale 76 leads to 994 dimensions. The number of rows in the transformation matrix corresponds to the number of leads, and the number of columns corresponds to the number of neural clusters in the brain model. This spatial module extracts latent spatial features from the EEG. The temporal module uses a recurrent neural network (RNN) as its basic structure, employing a three-layer GRU directly cascaded configuration to address the low-frequency temporal characteristics of EEG. This low-order network structure satisfies the requirements of low-frequency temporal characteristics, allowing the temporal module to extract latent temporal features from the EEG.

[0052] The neural network that traces the source of neural activity within the brain uses n within a time window of length T. l The EEG signal I(t) composed of several connected EEG signals is the input, and its mathematical description is as follows:

[0053] V(t)=STSIN(I(t)) (1)

[0054] STSIN(I(t))=GRU1(GRU2(GRU3(TConvs(I(t)))+W f I(t)))) (2)

[0055]

[0056] Among them, t s For the sampling interval, W f It is the transformation matrix, where T is the time window length and n is the number of nodes. l This refers to the number of EEG leads. The EEG signal map I(t) contains both the positional relationships between the various EEG acquisition leads and temporal sequence features. TConvs is a five-layer deconvolutional network, and GRU is a recurrent network layer. STSIN generates vector V from I(t). s (t)=[v1(t)v2(t)...v d [(t)] is a temporal sequence with spatial characteristics, where v1(t) is the neural activity inside the first brain region, and d is the number of brain regions whose neural activity is to be traced, which is a number from 1 to 994.

[0057] The second step involves constructing a virtual brain model using neural mass models. This virtual brain model consists of 994 neural mass models (NNMs), each representing the neural activity patterns of a group of neurons in the cerebral cortex of a specific brain region. Based on network connectivity measurements across these 994 brain regions obtained through brain science research, including connectivity matrices and connection delays, the 944 NMMs are connected together to form the virtual brain model.

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] Among them, H i (θ,t) is the i-th NMM state of brain region θ at time t, where i∈[1,2,3,...,8], K θ It is a connectivity matrix of the corresponding brain region θ from the whole brain map, a d (θ) represents the connection delay of the corresponding brain region θ derived from the whole-brain map and brain physiological connectivity; Sigmoid(.) is the sigmoid activation function; A is the average action potential; B is the synaptic gain; a is the synaptic membrane potential time constant; b is the dendritic time delay; C1-C4 are the number of synaptic connections in the four connectivity scenarios of excitation-cone, inhibition-cone, cone-excitation, and cone-inhibition in NMM. K θ a d (θ) and C1-C4 are the parameters used by the virtual brain model to generate neural activity. The values ​​used are obtained from the input experiment, and θ takes the integer value from 1 to 994.

[0067] The third step is to use the EEG measurement model shown in formula (4) to convert the internal neural activity generated by the virtual brain model into scalp electroencephalogram (EEG) signals I(t):

[0068] I=Q(L(θ))H(θ)+Q(L(θ (i) ))H i (θ (i) )+Q(L(θ (b) ))Hb (θ (b) )+n m (4)

[0069] in, L(θ)∈R is the electroencephalogram (EEG) signal collected from the electrodes of the EEG acquisition device during time T starting from time t. 3 It is the location of the target brain activity source and θ∈(θ1,...,θ) d Q(L(θ)) is the spatial conductance matrix of the target brain activity, H(θ)∈R d It is the amplitude of the target brain activity, which is also the amplitude of the target equivalent current dipole, L(θ). (i) )∈R 3 It is the location of the source of interference with brain activity and Q(L(θ (i) H is the spatial conductance matrix that interferes with brain activity. i (θ (i) )∈R k There are k interfering brain activities, L(θ) (b) )∈R 3 It is the location of background brain activity and Q(L(θ (b) H is the spatial conductance matrix of background brain activity. b (θ (b) )∈R p There are p background brain activities, n m The measurement noise of the electrode is expressed in the form of Gaussian white noise, d+k+p=994.

[0070] The fourth step involves training STSIN using brain activity generated by a virtual brain model and scalp EEG signals generated by an EEG measurement model. The source region crossover ratio (CRR) is calculated as: the area of ​​overlap between the actively active brain region after tracing the source and the actual active brain region, divided by the area of ​​the combined region. In this embodiment, the source region CRR is calculated according to... Equivalent calculation

[0071] Temporal similarity is used to measure the temporal accuracy of tracing the origins of neural activity within the brain. The difference between the temporal similarity and 1 is calculated as the loss of temporal accuracy. Temporal similarity is the phase synchronization index between the sequence of neural activity estimated by the neural network for tracing the origins of neural activity within the brain and the sequence of neural activity generated by the virtual brain model. In this embodiment, temporal similarity is calculated according to... calculate.

[0072] The training of the brain's internal neural activity source network uses a newly defined spatial accuracy loss SL (the difference between the source region intersection-union ratio and 1) and temporal accuracy loss TL (the difference between the temporal similarity and 1) as the loss function L. The calculation model is as follows:

[0073]

[0074]

[0075] L = SL + TL;

[0076] in, H(994-θ,t) represents the estimated internal neural activity of the θ-th brain region at time t, and H(θ,Tt) represents the actual internal neural activity of the 994-θ-th brain region at time t.

[0077] The training process of STSIN employs error backpropagation and stochastic gradient descent. When the loss stabilizes, the training of the neural activity source network within the brain is complete.

[0078] The trained brain neural activity tracing network can trace brain neural activity when a measured EEG signal is obtained.

[0079] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for tracing the source of intracranial neural activity based on scalp electroencephalogram (EEG) signals, characterized in that, The method includes the following: Part 1: Building a Virtual Brain Model A virtual brain model was built using physiologically interpretable neural cluster models, whole-brain maps, and brain physiological connections, and the number of neural cluster models was recorded. The internal neural activity of the brain generated based on the virtual brain model was converted into scalp electroencephalogram (EEG) signals using an EEG measurement model. Part Two: Constructing a neural network for tracing the source of neural activity within the brain: The input of the neural network for tracing the source of brain activity is the electroencephalogram (EEG) signal from the scalp, and the output is the neural activity of the brain. The neural network for tracing the source of brain activity is obtained by cascading a spatial module and a temporal module. The spatial module includes a transformation matrix and five layers of deconvolution in series. The input scalp EEG signal is processed by the five layers of deconvolution on one hand, and multiplied by the transformation matrix on the other hand. The output of the spatial module is obtained by adding the result of the five layers of deconvolution to the result of the multiplication of the transformation matrix. The time module comprises three layers of recurrent neural network (RNN) connected in series. The number of rows in the transformation matrix corresponds to the number of leads in the scalp EEG signal, and the number of columns corresponds to the number of neural cluster models in the virtual brain model. Part Three: Using a virtual brain model to generate internal brain activity data with spatial and temporal characteristics, and acquiring corresponding EEG signals through an EEG measurement model; constructing a dataset from the obtained internal brain activity data and corresponding EEG signals to train a neural network for tracing the source of internal brain activity, and obtaining the trained neural network for tracing the source of internal brain activity.

2. The method for tracing the source of intracranial neural activity based on scalp electroencephalogram (EEG) signals according to claim 1, characterized in that, During training, the neural network tracing the source of neural activity within the brain loses temporal accuracy. With spatial accuracy loss As a loss function, The crossover ratio of source regions is used to measure the spatial accuracy of tracing neural activity within the brain, and the difference between the crossover ratio of source regions and 1 is calculated as the spatial accuracy loss. Temporal similarity is used to measure the temporal accuracy of tracing the origins of neural activity within the brain, and the difference between temporal similarity and 1 is calculated as the loss of temporal accuracy.

3. The method for tracing the source of intracranial neural activity based on scalp electroencephalogram (EEG) signals according to claim 2, characterized in that, ； ； The sampling interval is t, where t is time. It is the length of the time window. To trace the number of brain regions involved in neural activity, The neural activity of the θ-th brain region at time t is estimated by a source neural network. For the 994th The actual internal neural activity of each brain region at time t; This represents the actual internal neural activity of the θ-th brain region at time Tt.

4. The method for tracing the source of intracranial neural activity based on scalp electroencephalogram (EEG) signals according to claim 1, characterized in that, The EEG measurement model is , in, From From this moment on The EEG signals collected on the electrodes of the EEG acquisition device within a given time period, where m is the number of leads. It is the location of the target brain activity source and has , For brain regions, It is the spatial conductance matrix of the target brain activity. yes The brain activity of a target is also the amplitude of the target's equivalent current dipole. It is the location of the source of interference with brain activity and , It is a spatial conductance matrix that interferes with brain activity. yes Let i represent the interfering brain activity and b represent the background brain activity. It is the location of background brain activity and , It is the spatial conductance matrix of background brain activity. yes Background brain activity, The measurement noise of the electrode is expressed in the form of Gaussian white noise.

5. A system for tracing the source of intracranial neural activity based on scalp electroencephalogram (EEG) signals, characterized in that, The system includes a virtual brain model, an EEG measurement module, and a neural network for tracing the source of neural activity within the brain. The virtual brain model is constructed based on a neural cluster model, a whole-brain map, and brain physiological connections, and is used to generate internal neural activity in the brain. The EEG measurement module is used to convert the internal neural activity of the brain generated based on the virtual brain model into scalp EEG signals. The input of the neural network for tracing the source of brain activity is the electroencephalogram (EEG) signal from the scalp, and the output is the neural activity of the brain. The neural network for tracing the source of brain activity is obtained by cascading a spatial module and a temporal module. The spatial module includes a transformation matrix and five layers of deconvolution in series. The input scalp EEG signal is processed by the five layers of deconvolution on one hand, and multiplied by the transformation matrix on the other hand. The output of the spatial module is obtained by adding the result of the five layers of deconvolution to the result of the multiplication of the transformation matrix. The time module comprises three layers of recurrent neural network (RNN) connected in series. The number of rows in the transformation matrix corresponds to the number of leads in the scalp EEG signal, and the number of columns corresponds to the number of neural cluster models in the virtual brain model. A virtual brain model is used to generate internal brain activity data with spatial and temporal characteristics, and the corresponding electroencephalogram (EEG) signals are obtained through an EEG measurement module. A dataset of internal brain activity data and corresponding EEG signals is constructed to train a neural network for tracing the source of internal brain activity. The trained neural network is then used for tracing the source of internal brain activity.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1-4.

Images