Magnetoencephalogram signal noise suppression method and device

Abstract

The invention belongs to the technical field of biological signal processing, and discloses a magnetoencephalogram signal noise suppression method and device, and the method comprises the steps: obtaining a target magnetoencephalogram signal; noise in the target magnetoencephalogram signal is determined according to a pre-generated neural network, the neural network is generated through training of historical magnetoencephalogram signals, and the training cut-off condition of the neural network is that the difference between the noise determined through training and the average value of the noise predetermined in the historical magnetoencephalogram signals is smaller than a preset value; the input of a residual convolution module in the neural network is the difference between the target magnetoencephalogram signal and the output of the neural network; the output of the neural network is the noise determined by the training; and suppressing the noise of the target magnetoencephalogram according to the average value of the noise. According to the method, data with higher quality can be provided for brain magnetic research.

Description

Technical Field

[0001] The present invention belongs to the technical field of bio-signal processing, and particularly relates to a method and device for suppressing brain magnetogram signal noise. Background Art

[0002] In the prior art, magnetoencephalography (MEG) denoising algorithms generally rely on the principle of subspace projection. The subspace projection algorithm divides neural signals and interference signals by projecting MEG signals onto the internal brain subspace and the external interference subspace. Although there are various subspace projection algorithms based on the signal space domain, time domain, or frequency domain and their performance has been widely verified, the above algorithms still have certain limitations. Specifically, the above algorithms require prior information such as empty room noise data or sensor positions, and the denoising effect depends on the accuracy of the prior information.

[0003] Specifically, the subspace projection algorithm can be further divided into the Signal Space Projection (SSP) algorithm and the Signal Space Separation (SSS) algorithm. Methods such as signal space projection obtain noise projection vectors by decomposing the empty room noise data collected without a subject. This method assumes that the empty room noise is highly similar to the noise in the actual experimental signal. However, the interference outside the shielded room may change at any time, which is very common in long-term MEG experiments. In addition, the entry and exit of the subject into and out of the shielded room will also change the distribution of the background magnetic field. Therefore, the empty room noise data does not always accurately reflect the noise characteristics in the actual signal. The signal space separation method requires prior knowledge of the accurate position of the sensors relative to the subject's head. However, in MEG, especially in Optically Pumped Magnetometer-based Magnetoencephalography (OPM-MEG) experiments, it is very difficult for the subject to remain stationary for a long time. Therefore, the positions of the sensors will also change during the experiment, thus reducing the accuracy of subspace separation. Summary of the Invention

[0004] An object of the present invention is to provide a method for suppressing brain magnetogram signal noise, aiming to solve the technical pain point that the existing methods for suppressing brain magnetogram signal noise rely on prior information and cannot perform denoising only based on the data characteristics of the brain magnetogram signal.

[0005] Another object of the present invention is to provide a device for suppressing the noise of magnetoencephalogram signals. Still another object of the present invention is to provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method for suppressing the noise of magnetoencephalogram signals are implemented. Still another object of the present invention is to provide a readable medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method for suppressing the noise of magnetoencephalogram signals are implemented.

[0006] To solve the technical problems in the background art of the present invention, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for suppressing the noise of magnetoencephalogram signals, including:

[0008] Obtain a target magnetoencephalogram signal;

[0009] Determine the noise in the target magnetoencephalogram signal according to a pre-generated neural network, wherein the neural network is generated by training with historical magnetoencephalogram signals, and the training cut-off condition of the neural network is that the difference between the noise determined by training and the average value of the pre-determined noise in the historical magnetoencephalogram signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target magnetoencephalogram signal and the output of the neural network; the output of the neural network is the noise determined by training;

[0010] Suppress the noise of the target magnetoencephalogram according to the average value of the noise.

[0011] In some embodiments of the present invention, the fully connected layer structure of the residual attention convolution module of the neural network is one-dimensional convolution; the residual convolution module and the residual attention convolution module are used to learn the noise characteristics in the historical magnetoencephalogram signals.

[0012] In some embodiments of the present invention, the historical magnetoencephalogram signals include training data and test data; the average value of the pre-determined noise belongs to the test data; the noise in the training data and the noise in the test data are not correlated with each other.

[0013] In some embodiments of the present invention, the noise in the training data and the noise in the test data follow a zero-mean distribution, and the real magnetoencephalogram signals in the training data are the same as the real magnetoencephalogram signals in the test data.

[0014] In a second aspect, the present invention provides a device for suppressing the noise of magnetoencephalogram signals, and the device includes:

[0015] A target signal acquisition module, configured to obtain a target magnetoencephalogram signal;

[0016] A noise extraction module, configured to determine the noise in the target magnetoencephalogram (MEG) signal according to a pre-generated neural network, where the neural network is generated by training with historical MEG signals, and the training cut-off condition of the neural network is that the difference between the noise determined by training and the average value of the pre-determined noise in the historical MEG signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target MEG signal and the output of the neural network; the output of the neural network is the noise determined by training.

[0017] A noise suppression module, configured to suppress the noise of the target MEG according to the average value of the noise.

[0018] In some embodiments of the present invention, the fully connected layer structure of the residual attention convolution module of the neural network is one-dimensional convolution; the residual convolution module and the residual attention convolution module are used to learn the noise characteristics in the historical MEG signals.

[0019] In some embodiments of the present invention, the historical MEG signals include training data and test data; the average value of the pre-determined noise belongs to the test data; the noise in the training data is uncorrelated with the noise in the test data.

[0020] In some embodiments of the present invention, the noise in the training data and the noise in the test data follow a zero-mean distribution, and the real MEG signals in the training data are the same as the real MEG signals in the test data.

[0021] In a third aspect, the present invention provides a computer program product, including a computer program / instructions, which when executed by a processor, implement the steps of a method for suppressing the noise of MEG signals.

[0022] In a fourth aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor implements the steps of a method for suppressing the noise of MEG signals when executing the program.

[0023] In a fifth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and the computer program implements the steps of a method for suppressing the noise of MEG signals when executed by a processor.

[0024] From the above description, it can be seen that an embodiment of the present invention provides a method and device for suppressing noise in a magnetoencephalogram (MEG) signal. The corresponding method includes: first, obtaining a target MEG signal; then, determining the noise in the target MEG signal based on a pre-generated neural network, wherein the neural network is generated by training historical MEG signals, and the training cutoff condition of the neural network is that the difference between the noise determined by training and the average value of the noise pre-determined in the historical MEG signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target MEG signal and the output of the neural network; the output of the neural network is the noise determined by training; finally, suppressing the noise of the target MEG signal based on the average value of the noise.

[0025] The present invention first proposes a model training strategy that uses the readily available noise average signal as the target signal. It then utilizes residual connections and an attention mechanism module to improve the traditional convolutional network architecture, enabling the model to extract and remove noise features from the original signal. The present invention also validates the algorithm using simulation and real-world experimental data. The proposed method outperforms traditional denoising algorithms in signal-to-noise ratio, root mean square error, source localization, and waveform performance, potentially providing higher-quality data for subsequent MEG research. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 A schematic flow chart of a method for suppressing noise in a magnetoencephalogram signal according to an embodiment of the present invention;

[0028] Figure 2 Schematic diagram of the structure of the MEGNSnet model in an embodiment of the present invention;

[0029] Figure 3 Schematic diagram of the traditional structure of the MEGNSnet model in an embodiment of the present invention (consisting of two layers of convolutional modules);

[0030] Figure 4 Schematic diagram of the structure of the convolution module in an embodiment of the present invention;

[0031] Figure 5 Schematic diagram of the residual connection structure in an embodiment of the present invention;

[0032] Figure 6 Schematic diagram of the combination of the residual structure and the ECA attention mechanism module in an embodiment of the present invention;

[0033] Figure 7 Schematic diagram of the principle of the simulation experiment system in the specific embodiment of the present invention;

[0034] Figure 8 Schematic diagram of the relative position between the sensor and the head in the specific embodiment of the present invention;

[0035] Figure 9 Schematic diagram of the amplitude of the simulation source in the specific embodiment of the present invention;

[0036] Figure 10 Schematic diagram of the waveform of the simulation signal in the specific embodiment of the present invention;

[0037] Figure 11 Schematic diagram of the noise power spectral density in the simulation experiment data in the specific embodiment of the present invention;

[0038] Figure 12 Schematic diagram of the comparison of waveforms of different algorithms in the simulation experiment of the present invention Figure 1 (simulated neural signal);

[0039] Figure 13 Schematic diagram of the comparison of waveforms of different algorithms in the simulation experiment of the present invention Figure 2 (SSP);

[0040] Figure 14 Schematic diagram of the comparison of waveforms of different algorithms in the simulation experiment of the present invention Figure 3 (HFC);

[0041] Figure 15 Schematic diagram of the comparison of waveforms of different algorithms in the simulation experiment of the present invention Figure 4 (N2CUnet);

[0042] Figure 16 Schematic diagram of the comparison of waveforms of different algorithms in the simulation experiment of the present invention Figure 5 (N2CUnet+residual+ECA);

[0043] Figure 17 Schematic diagram of the comparison of waveforms of different algorithms in the simulation experiment of the present invention Figure 6 (N2Nevoke MEGNSnet (the algorithm provided by the present application));

[0044] Figure 18 Schematic diagram of the comparison of waveforms of the somatosensory evoked experiment in the specific embodiment of the present invention Figure 1 (only band-pass);

[0045] Figure 19 Schematic diagram of comparison of somatosensory evoked experiment waveforms in the specific embodiment of the present invention Figure 2 (SSP);

[0046] Figure 20 Schematic diagram of comparison of somatosensory evoked experiment waveforms in the specific embodiment of the present invention Figure 3 (HFC);

[0047] Figure 21 Schematic diagram of comparison of somatosensory evoked experiment waveforms in the specific embodiment of the present invention Figure 4 (MEGNSnet (algorithm provided in this application));

[0048] Figure 22 Schematic diagram of comparison of somatosensory evoked experiment source localization results in the specific embodiment of the present invention Figure 1 (Only band-pass);

[0049] Figure 23 Schematic diagram of comparison of somatosensory evoked experiment source localization results in the specific embodiment of the present invention Figure 2 (SSP);

[0050] Figure 24 Schematic diagram of comparison of somatosensory evoked experiment source localization results in the specific embodiment of the present invention Figure 3 (HFC);

[0051] Figure 25 Schematic diagram of comparison of somatosensory evoked experiment source localization results in the specific embodiment of the present invention Figure 4 (MEGNSnet (algorithm provided in this application));

[0052] Figure 26 Schematic diagram of comparison of auditory evoked experiment waveforms in the specific embodiment of the present invention Figure 1 (Only band-pass);

[0053] Figure 27 Schematic diagram of comparison of auditory evoked experiment waveforms in the specific embodiment of the present invention Figure 2 … (HFC);

[0054] Figure 28 Schematic diagram of comparison of auditory evoked experiment waveforms in the specific embodiment of the present invention Figure 3 … (MEGNSnet (algorithm provided in this application));

[0055] Figure 29 Schematic diagram of comparison of auditory evoked experiment source localization results in the specific embodiment of the present invention Figure 1 (Only band-pass);

[0056] Figure 30Schematic diagram for comparing source localization results of auditory evoked experiments in the specific embodiments of the present invention Figure 2 (HFC);

[0057] Figure 31 Schematic diagram for comparing source localization results of auditory evoked experiments in the specific embodiments of the present invention Figure 3 (MEGNSnet (the algorithm provided in this application));

[0058] Figure 32 Block diagram of a device for suppressing magnetoencephalogram signal noise in the embodiments of the present invention;

[0059] Figure 33 Schematic structural diagram of an electronic device in the embodiments of the present invention. Specific embodiments

[0060] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0061] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of an all-hardware embodiment, an all-software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned accompanying drawings are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices. Without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0063] Magnetoencephalography (MEG) is a brain neuroimaging technique with high temporal and spatial resolution. With the development of optically pumped magnetometers (OPMs) in recent years, wearable OPM-MEG systems are not only more flexible than traditional MEG devices but also have a better signal sensitivity level. Therefore, more complex neuroscience experiments can be conducted, and it has broad prospects in brain neuroscience and clinical applications. However, one of the most important challenges in MEG signal analysis is how to remove the background noise interference in the signal, which usually comes from moving vehicles, mechanical vibrations, or nearby metal objects. Since MEG signals are very weak, even when measured in a magnetically shielded room, the background noise interference in the signal may still be eight orders of magnitude larger than the neural signal, which can seriously affect the subsequent source localization and the results of brain function analysis. Therefore, noise suppression algorithms are very important in MEG signal processing.

[0064] The denoising effect of traditional algorithms based on subspace projection usually depends on the accuracy of prior information such as empty room noise data or sensor positions. While the data-driven neural network method does not require complex prior information, on the one hand, neural networks usually need clean (neural) signals as the target signals to be restored during training, and in real MEG experiments, it is difficult to obtain noise-free neural signals. On the other hand, traditional network structures are difficult to accurately extract the features of neural signals from noisy background noise.

[0065] Specifically, using neural networks to learn signal features is a feasible idea, and such methods have been widely applied to denoising tasks such as one-dimensional speech signals and two-dimensional images. However, such methods are less applied to the suppression of MEG background noise because it is difficult to obtain accurate neural signals.

[0066] One is that traditional neural networks need to form training pairs with the original signal and its corresponding noise-free neural (target) signal during training, enabling the network to learn the mapping from the noisy original signal to the clean signal, which is called the Noise2Clean (N2C) training strategy. However, different from image or speech signals, it is almost impossible to obtain noise-free neural signals such as MEG signals, so it is difficult to apply the N2C strategy to train network models.

[0067] The second is that the brain neural signals induced by different experimental stimuli are also different, and it is impossible to include all signal patterns during training. Therefore, when the tested signal data is different from the training data, the denoising effect of traditional networks is not good;

[0068] Thirdly, the amplitude of the background noise in the raw MEG signal is relatively large, and it is difficult for the network to effectively learn the neural signal features therein. Therefore, the existing neural network-based methods mainly use the idea of classification to remove the physiological artifacts (eye movements or heartbeats) with fixed morphological features in the signal. Based on this and to solve at least some of the technical problems in the existing technology, an embodiment of the present invention provides a specific implementation manner of a method for suppressing MEG signal noise, see Figure 1 , and the method specifically includes the following contents:

[0069] Step 100: Obtain the target MEG signal;

[0070] Step 200: Determine the noise in the target MEG signal according to a pre-generated neural network, where the neural network is generated by training with historical MEG signals, and the training cut-off condition of the neural network is that the difference between the noise determined by training and the average value of the pre-determined noise in the historical MEG signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target MEG signal and the output of the neural network; the output of the neural network is the noise determined by training;

[0071] Step 300: Suppress the noise of the target MEG according to the average value of the noise.

[0072] As can be seen from the above description, an embodiment of the present invention provides a method for suppressing MEG signal noise, which is used to remove the interference of background noise in the MEG signal. First, the target signal of the model is replaced by the averaged MEG signal in N2C. Then, through residual connection and attention mechanism, the network learns the noise features instead of the weak neural signal features, and subtracts the estimated noise from the original signal to obtain the denoised signal. In addition, the present invention compares with the traditional denoising algorithm on simulation and two groups of real experimental data respectively, and verifies its effectiveness.

[0073] It can be understood that the target MEG signal in step 100 is the MEG signal whose noise is to be suppressed, which contains both the real MEG signal (clean MEG signal) and the background noise.

[0074] For step 200, the training strategy of Noise2Clean (N2C) requires a clean signal as the target signal to enable the network to learn the mapping relationship between the original noisy MEG signal and the noise-free clean MEG signal. However, it is theoretically difficult to obtain noise-free brain magnetic nerve signals, which hinders the application of deep learning-based methods in MEG background noise suppression.

[0075] Regarding the problem of lacking clean MEG signals, step 200 proposes a training method for the original input signal and the clean MEG signal without noise, satisfying the following assumptions. Using another MEG signal containing noise as the clean MEG signal in the training strategy proposed in step 200 has basically the same denoising effect as the N2C training strategy using the clean signal as the target signal.

[0076] Assumption 1: The clean (neural) signals in the input signal and the target signal are the same.

[0077] Assumption 2: The noise in the input signal and the target signal follows a zero-mean distribution and is uncorrelated with each other.

[0078] Obviously, the evoked neural signals in each evoked cycle segment of the MEG evoked state data are similar and uncorrelated with the noise, satisfying Assumption 1. And baseline correction and the large amount of random noise contained in the signal make the MEG signal also satisfy Assumption 2. Suppose the input signal y consists of the neural signal x and the background noise n, where the sensor noise much smaller than the background noise is ignored:

[0079] y = x + n (1)

[0080] Then when the input signal is y1 = x + n1, the denoising task of N2C is usually to solve the training optimization problem of the neural network using the mean square error loss function, denoted as L n2c :

[0081]

[0082] where θ is the network model parameter and f θ (y) is the output predicted by the network. Since the accurate x cannot be obtained in the experiment, step 200 uses another signal y2 = x + n2 with the same neural signal but uncorrelated noise as the target signal in the loss function to replace x in L n2c in, and the loss function is denoted as L n2n :

[0083]

[0084] Expanding formula (3) gives:

[0085]

[0086]

[0087] Obviously, the first three terms of formula (5) are equal to L n2c . Since n2 satisfies Assumption 1 and Assumption 2, that is so the fourth and fifth terms in formula (5) are zero, and the last term is the noise variance.

[0088] L n2n = L n2c + Var(n) (6)

[0089] Since the variance is inversely proportional to the training sample size, as the sample size of the dataset increases, L n2n and L n2c are equivalent, and the training effect corresponding to the training method proposed in step 200 is basically the same as that of N2C.

[0090] For step 300, specifically, subtract the noise extracted in step 200 from the original signal to obtain the final denoised signal.

[0091] It can be understood that the premise for the training strategy proposed in step 200 to be equivalent to N2C is to have a large number of training samples. However, the sample size of magnetoencephalogram data is usually small. Therefore, here the output signal of the neural network is further replaced by the average noise output sample from another segment of noise samples. The average noise output sample is the average of multiple segments of noisy signals added together, that is, the data obtained after averaging the superimposed magnetoencephalogram signals. This improved training strategy is called N2N here. evoke .

[0092] First, a simple proof is carried out using the average value of two signals. y3 is the average value of y1 and y2, and can be written as the sum of the neural signal x and the average value of two noises (n1, n2):

[0093]

[0094] The corresponding loss function optimization process is defined as:

[0095]

[0096] Similar to formulas (4) and (5), expanding formula (8) gives:

[0097]

[0098] Since the noises are uncorrelated with each other, L3 is expressed as:

[0099]

[0100] It can be seen that L3 can converge to L n2n using less data than L n2c . Therefore, applying it to the evoked signal y n , which is obtained by averaging the superimposition of n original magnetoencephalogram signals.

[0101]

[0102] Correspondingly, the loss function L of N2Nevoke evoke is similar to formulas (8)-(10):

[0103]

[0104] Therefore, using the data obtained by superimposing and averaging the magnetoencephalogram signals corresponding to the input signals as the output signals in the dataset makes N2Nevoke converge faster than the original training method proposed in step 200, requires fewer data samples, and is more conducive to applying the method to the actual magnetoencephalogram signal denoising scenario.

[0105] As Figure 2 shown, the network structure of the existing denoising neural network is improved from the Unet network, which is also the basic model of most current denoising methods. However, there are still certain limitations when applied to magnetoencephalogram signal denoising. As Figure 3 shown, the network module in the basic Unet model consists of two convolutional modules ( Figure 4 ) and the model directly predicts the denoised signal (clean magnetoencephalogram signal), that is, f in formula (2) θ (y) is equal to However, since the background noise n in the magnetoencephalogram signal is much larger than the neural signal x, and the morphologies of neural signals induced by different experiments are also different, it is difficult for traditional convolutional modules to accurately learn the features of neural signals in the input signal. Therefore, targeted improvements have been made to the convolutional modules.

[0106] As Figure 5 shown, first, the traditional convolutional module is improved using residual connections (directly passing the input y of the residual convolutional module to the final output end, subtracting it from the predicted output f(y) of the network module at the output end to obtain the final output). The new module is called the residual convolutional module. The output of the residual convolutional module is obtained by subtracting the noisy input y from the network prediction f θ (y). The network trained using this structure no longer directly predicts neural signals, but learns and predicts the noise in the original signal, thus solving the limitations of traditional convolutional modules. Correspondingly, L evoke 's optimization process is expressed as:

[0107]

[0108] In some embodiments of the present invention, the fully connected layer structure of the residual attention convolutional module of the neural network is a one-dimensional convolution; the residual convolutional module and the residual attention convolutional module are used to learn the noise features in the historical magnetoencephalogram signals.

[0109] Considering the complex form of background noise, in order to enable the model to learn noise features more efficiently, an attention module is added on the basis of the residual convolution module, called the residual attention convolution module (see Figure 6 ). The residual attention convolution module replaces the fully connected layer in the traditional attention module with a one-dimensional convolution, avoiding the destruction of the direct correspondence between channels and their weights caused by the dimensionality reduction operation of the fully connected layer, enabling the model to learn noise features more accurately.

[0110] In some embodiments of the present invention, the historical magnetoencephalogram signals include training data and test data; the average value of the pre-determined noise belongs to the test data; the noise in the training data is uncorrelated with the noise in the test data.

[0111] In some embodiments of the present invention, the noise in the training data and the noise in the test data follow a zero-mean distribution, and the true magnetoencephalogram signals in the training data are the same as the true magnetoencephalogram signals in the test data.

[0112] To further illustrate the solution, the present invention also provides a simulation experiment of a method for suppressing magnetoencephalogram signal noise, specifically including the following content.

[0113] There is a lot of environmental noise interference in magnetoencephalogram signals (especially OPM), making it difficult to directly analyze the neural signals therein. Currently, a large number of studies choose subspace projection-based denoising algorithms to preprocess the signals, but such algorithms require certain prior information for denoising, and the accuracy of the prior information has a great impact on the final denoising effect. For example, the SSP algorithm relies on the noise data of an empty room collected in advance to construct a noise subspace, and algorithms such as the SSS series algorithms and DSSP, HFC, etc. require accurate position information of the sensor relative to the head. In actual experiments, the background noise is changing at any time, so the noise in an empty room does not always accurately contain all types of noise in the signal. In addition, in OPM experiments, the relative position of the sensor and the head is likely to change with the movement of the subject. Therefore, when the prior information is inaccurate, it is necessary to consider using a fully data-driven deep learning method to remove the noise in the magnetoencephalogram signal. However, traditional neural networks also have limitations. The noise-free clean neural signals required in the training strategy of N2C are difficult to obtain in magnetoencephalogram experiments, and a simple network structure is difficult to extract the features of the neural signals to be restored from a large amount of background noise. Therefore, the present invention proposes an improved neural network-based method, MEGNSnet. The present invention designs a training strategy called N2N evoke , different from N2C, N2N evokeThe target signal uses the result of superimposing and averaging easily obtainable original signals. Additionally, the network structure is improved using residual connections and attention mechanism modules, enabling the model to focus on extracting the features of the noise during training and subtracting it from the original signal to obtain the final denoised signal.

[0114] Specifically, the present invention tested the performance of the algorithm provided in this application on simulation experiment data, as well as real somatosensory and auditory evoked data.

[0115] The composition of the signal data for the simulation experiment is as Figure 7 , Figure 8 , Figure 9 and Figure 10 shown. The magnetoencephalogram signal in the simulation experiment consists of simulated neural signals and actually measured empty room noise. Among them, the simulated neural signals are generated by a signal source placed inside a real head model, which is set as a current dipole with a peak amplitude of 50 nAm and a waveform of a 10 Hz sine signal. With a 1-second induction period, the dipole is activated 500 ms before the period ( Figure 9 ). And through co-registration, the OPM sensors outside the brain obtain the corresponding waveforms ( : , where the curves of different colors in this figure represent the data in different sensor channels, Figure 10 and the meaning of the curves of different colors in Figures 12 to 21 is the same). The empty room noise data is collected using the same sensor array layout inside a shielded room with a background field amplitude less than 13 nT. The simulated signal and the empty room noise signal are superimposed to form the total signal. Then, the signal is divided into m epoch data according to the induction period (please note that here it means periodic signal, different from the training epochs of the neural network). Finally, the signal of each sensor in the epoch is used as the data form constituting the dataset. When there are N sensors, a segment of signal generates a total of N×m data, and the size of each data is 1×t, where t is the signal sampling rate. The input or target data in the dataset of the present invention is generated in the above manner.

[0116] Dataset composition: This application collected three segments of noise data at different times and divided the signal data into three datasets accordingly. The power spectral density of the three segments of noise is as Figure 10 shown. Dataset 1 and Dataset 2 are used as the training set, and Dataset 3 is used as the test set, as shown in Table 1. The dataset consists of training pairs composed of the input data to be denoised and the corresponding target data. Finally, the training set has a total of 23,840 pairs of data, and the test set has a total of 7,748 pairs of data.

[0117] Since the noise-free neural signal in the simulation experiment is known, in order to verify N2N evokeThe effectiveness of the strategy was also verified by training and testing the same network structure using the N2C strategy. For different training strategies, the dataset needs to be adjusted accordingly, as shown in Table 1. N2C and N2N evoke have the same input data, which is obtained from the original signal using the method in "Dataset Composition" and is called Noisy_epoch_data here. N2N evoke The target data in is obtained by superimposing and averaging Noisy_epoch_data and is called Noisy_evoke_data. While N2C directly uses the clean simulated neural signals and is called Sim_clean_data.

[0118] Table 1 Simulation experiment data and dataset composition

[0119]

[0120] Training and testing methods:

[0121] 1) Experimental settings: The data sampling rate in the article is 1000Hz, and the real empty room data is band-pass filtered from 1 - 40Hz. All models are trained and tested using the same GPU, with a training batch size of 16, and trained for about 120 epochs until the validation set converges.

[0122] 2) Ablation experiments: In addition to training the network model using N2C to test the effectiveness of N2N evoke , ablation experiments were also conducted to verify the effects of the residual and ECA modules.

[0123] 3) Benchmark algorithms and evaluation metrics: The SSP and HFC algorithms were used as benchmark algorithms to compare the denoising effects with MEGNSnet trained using N2C and N2N respectively evoke . The algorithms were all applied to the test set data. For the SSP, the prior data of the empty room noise selected was a different empty room noise 1 from the empty room noise 3 in the test set data. Two objective evaluation metrics, SNR and RMSE, were selected. To ensure the consistency of the results of each algorithm, they were uniformly calculated on continuous data. Since the output of MEGNSnet is the epoch data of a single sensor with a size of 1×t, all outputs were concatenated to obtain continuous data. The calculation formulas for SNR and RMSE are as follows:

[0124]

[0125] where B clean (n) is the simulated neural signal, and P clean (n) is the power of B clean (n). B recon (n) is the reconstructed signal after denoising, and Precon-clean (n) is the noise power in the reconstructed signal. N is the number of sensors, and t is the number of signal sampling points. In addition, the waveform results of the superposition average of each algorithm are also compared.

[0126] The signal-to-noise ratio and root mean square error results of each algorithm in the simulation experiment are shown in Table 2. These results are obtained on Dataset 3 composed of the empty room noise 3. Figure 11 It can be seen that although the three sections of noise are collected in the same environment, different acquisition times make the empty room noise 3 have certain differences in power spectral density compared with the empty room noise 1 and 2 used in the training set data. This makes the effect of the SSP algorithm on the original signal superimposed with the empty room noise 3 poor when using the empty room noise 1 as the noise prior. This is also a common situation in long-term experiments, that is, the actual noise does not completely match the empty room noise prior measured before and after the experiment, indicating that the denoising effect of SSP is greatly affected by the accuracy of the noise prior information. In addition, since there is no active magnetic shielding coil used in the shielding room and the test environment background noise is complex, the effect of HFC on this non-uniform background noise is limited. In addition, compared with the original Unet network structure, the MEGNSnet we proposed has significantly improved the model's denoising ability after improving and adding residual and ECA modules. At the same time, it is noted that under the same network structure, the effect of the MEGNSnet model trained using the N2N evoke strategy is slightly worse than that of the model trained using N2C, but the difference is not significant. This proves the effectiveness of the N2N evoke strategy proposed in this application. At the same time, since there is no clean signal as the target signal for training in the real experiment, theoretically, the MEGDnet using the N2N evoke training strategy is more applicable in the actual experiment.

[0127] Table 2 Comparison of evaluation indicators in the simulation experiment

[0128]

[0129] The superposition average waveform results of each algorithm in the simulation experiment are as Figure 11 shown. It can be seen that compared with SSP and HFC, the neural network-based method ( Figures 12 to 17 ) removes most of the noise. In addition, relying on the residual and ECA modules, the waveform distortion of the result obtained by MEGNSnet is smaller than that of the original Unet network structure, indicating that the improvement enables the network to more accurately identify and remove noise. At the same time, the denoising effect of the MEGNSnet trained using N2N evoke ( Figures 15 to 17 ) is similar to the waveform reconstruction effect of the model trained using N2C ( Figure 17 ), further proving the effectiveness of N2N evoke .

[0130] Experimental results: including somatosensory evoked experiments and auditory evoked experiments.

[0131] The somatosensory evoked experiment data was obtained by actual measurement. The experimental type was median nerve stimulation. The measurement location was the shielded room used to collect the noise of the empty room in the simulation experiment. Therefore, the model trained by the simulation experiment was used as the pre-trained model, and then transfer learning was carried out using 20% of the measured data (2300 pairs). The data processing was the same as that of the simulation experiment.

[0132] Since there is no clean real nerve signal as the target signal in the measured data, the results of SSP, HFC, and MEGNSnet trained using N2N evoke were compared in terms of the superimposed average waveform, shielding factor, and source localization distribution. Among them, the shielding factor SF is expressed as the ratio of the vector 2-norm between the signals before and after denoising.

[0133]

[0134] The source localization method used the minimum norm estimation (MNE) based on the principle of distributed source imaging to view the source distribution of the data processed by different algorithms.

[0135] The shielding factor results of each algorithm in the somatosensory evoked experiment are shown in Table 3. The result of MEGNSnet is better than that of SSP and HFC. It shows that MEGNSnet suppresses the amplitude of the noise lower. Figure 16 The results of waveform reconstruction of different algorithms are shown. Affected by factors such as the number of sensors, position, subjects, and noise, it is difficult to obtain accurate nerve signals in actual magnetoencephalography experiments like in simulation experiments. Therefore, only subjective comparison can be carried out. The main evoked component of the median nerve stimulation experiment is M35 (at the black dotted line in the figure) that occurs around 0.035 s. It can be seen that MEGNSnet does not cause obvious distortion of the M35 evoked waveform due to more noise suppression. In addition, the waveforms with lower amplitude and smoother in the remaining time periods also indicate that MEGNSnet suppresses more random background noise. Especially for the blue sensor signal at 0.125 s, the waveform results obtained after processing by different algorithms are different, indicating that this should not be an obvious evoked nerve signal, and it is probably noise or signal distortion caused by algorithm processing, while MEGNSnet suppresses this signal well.

[0136] Table 3 Comparison of shielding factors in real experiments

[0137]

[0138] Next, refer to Figures 18 to 21, which is the comparison result of the source activation region and intensity at time M35. The neural source activities of M35 are mainly distributed in the postcentral gyrus of the central sulcus where the somatosensory evoked region is located (within the blue circle). Compared with the results of other algorithms, the result of MEGNSnet has a more concentrated activation region and fewer other activated regions, indicating less influence by noise and more sufficient noise suppression.

[0139] Auditory evoked experiment: The algorithm was verified using the publicly available auditory dataset in the prior art. The experimental type was pure tone auditory stimulation, and the specific experimental acquisition method was as follows. Since the acquisition environment was different from that of the somatosensory experiment, the pre-trained model was not used, and 40% of the data (19,448 pairs) in the dataset was used to train the network model. In addition, since no empty room noise data was provided, only the HFC method was compared. The remaining data processing methods and comparison metrics were the same as those in the somatosensory evoked experiment.

[0140] The shielding factor results of the algorithm on the publicly available auditory experimental data are shown in Table 3, and the result of MEGNSnet is also better than that of HFC. Figures 22 to 25 The results of waveform reconstruction by different algorithms are shown. The main evoked component of pure tone auditory stimulation is M100 (at the black dotted line in the figure) occurring around 0.1 s. Similar to the conclusion of the somatosensory evoked experiment, MEGNSnet not only retains the M100 component without obvious distortion, but also better suppresses the noise in the time period of 0.2 s - 0.4 s.

[0141] Figures 26 to 28 which is the comparison result of the source activation region and intensity at time M100. The neural source activities of M100 are mainly distributed in the temporal lobe region where the auditory evoked region is located (within the blue circle). Similarly, the result of MEGNSnet has a more concentrated activation region and fewer other activated regions, indicating better denoising effect of MEGNSnet.

[0142] In addition, besides the denoising effect, the MEGNSnet based on neural network proposed in the present invention has an advantage in time cost compared with traditional algorithms. Taking SSP as an example, MEGNSDnet can omit the step of collecting empty room noise in each subsequent experiment after the model is trained, improving the experimental efficiency while the denoising effect is not affected by the quality of the empty room noise data.

[0143] As can be seen from the above description, the specific embodiment of the present invention provides a method for suppressing the noise of magnetoencephalogram signals. Aiming at the problem that the denoising effect of the traditional subspace projection algorithm is restricted by the accuracy of prior information, a data-driven neural network denoising method is proposed. Specifically, the present invention first improves the training strategy, making it use the average signal as the target signal instead of the noise-free magnetoencephalogram signal that is difficult to obtain. Then, the network structure is improved through residual connections and attention mechanism modules, enabling the network model to extract and eliminate noise from the original signal. Finally, simulation and two sets of real experimental data are used to verify that the denoising effect of the proposed method is better than that of traditional denoising algorithms, effectively reducing the background noise in the signal. Therefore, after denoising the signal using the method proposed by the present invention, the accuracy of subsequent brain research can be improved.

[0144] Based on the same inventive concept, the embodiment of the present invention also provides a device for suppressing the noise of magnetoencephalogram signals, which can be used to implement the method described in the above embodiment, as shown in the following embodiment. Since the principle of the device for suppressing the noise of magnetoencephalogram signals to solve the problem is similar to that of the method for suppressing the noise of magnetoencephalogram signals, the implementation of the device for suppressing the noise of magnetoencephalogram signals can refer to the implementation of the method for suppressing the noise of magnetoencephalogram signals, and the repeated parts will not be described again. Hereinafter, the term "unit" or "module" may be a combination of software and / or hardware that can achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0145] The embodiment of the present invention provides a specific implementation manner of a device for suppressing the noise of magnetoencephalogram signals that can implement the method for suppressing the noise of magnetoencephalogram signals. Refer to Figures 29 to 31 , a device for suppressing the noise of magnetoencephalogram signals specifically includes the following contents:

[0146] A target signal acquisition module 10, configured to acquire a target magnetoencephalogram signal;

[0147] A noise extraction module 20, configured to determine the noise in the target magnetoencephalogram signal according to a pre-generated neural network, where the neural network is generated by training historical magnetoencephalogram signals, and the training cut-off condition of the neural network is that the difference between the noise determined by training and the average value of the pre-determined noise in the historical magnetoencephalogram signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target magnetoencephalogram signal and the output of the neural network; the output of the neural network is the noise determined by training;

[0148] A noise suppression module 30, configured to suppress the noise of the target magnetoencephalogram according to the average value of the noise.

[0149] In some embodiments of the present invention, the fully connected layer structure of the residual attention convolution module of the neural network is a one-dimensional convolution; the residual convolution module and the residual attention convolution module are used to learn the noise features in the historical magnetoencephalogram signals.

[0150] In some embodiments of the present invention, the historical magnetoencephalogram signals include training data and test data; the average value of the pre-determined noise belongs to the test data; the noise in the training data is uncorrelated with the noise in the test data.

[0151] In some embodiments of the present invention, the noise in the training data and the noise in the test data follow a zero-mean distribution, and the real magnetoencephalogram signals in the training data are the same as the real magnetoencephalogram signals in the test data.

[0152] Embodiments of the present invention also provide a specific implementation manner of an electronic device that can implement all steps in the method for suppressing the noise of magnetoencephalogram signals in the above embodiments. Refer to Figure 32 Figure 33 , the electronic device specifically includes the following contents:

[0153] A processor (1201), a memory (1202), a communication interface (1203), and a bus (1204);

[0154] Among them, the processor (1201), the memory (1202), and the communication interface (1203) complete mutual communication through the bus (1204); the communication interface (1203) is used to implement information transmission between related devices such as server-side devices and user-side devices;

[0155] The processor (1201) is used to call a computer program in the memory (1202). When the processor executes the computer program, it implements all steps in the method for suppressing the noise of magnetoencephalogram signals in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0156] Step 100: Obtain a target magnetoencephalogram signal;

[0157] Step 200: Determine the noise in the target magnetoencephalogram signal according to a pre-generated neural network, where the neural network is generated by training with historical magnetoencephalogram signals, and the training cut-off condition of the neural network is that the difference between the determined noise in training and the average value of the pre-determined noise in the historical magnetoencephalogram signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target magnetoencephalogram signal and the output of the neural network; the output of the neural network is the determined noise in training;

[0158] Step 300: Suppress the noise of the target magnetoencephalogram (MEG) based on the average value of the noise.

[0159] An embodiment of the present invention also provides a computer-readable storage medium capable of implementing all steps in the method for suppressing the noise of the MEG signal in the above embodiment. A computer program is stored on the computer-readable storage medium. When the computer program is executed by a processor, all steps of the method for suppressing the noise of the MEG signal in the above embodiment are implemented. For example, when the processor executes the computer program, the following steps are implemented:

[0160] Step 100: Obtain a target MEG signal;

[0161] Step 200: Determine the noise in the target MEG signal according to a pre-generated neural network. The neural network is generated by training with historical MEG signals, and the training termination condition of the neural network is that the difference between the noise determined by training and the average value of the pre-determined noise in the historical MEG signals is less than a preset value. The input of the residual convolution module in the neural network is the difference between the target MEG signal and the output of the neural network. The output of the neural network is the noise determined by training;

[0162] Step 300: Suppress the noise of the target MEG based on the average value of the noise.

[0163] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments. In particular, for the hardware + program type embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0164] The specific embodiments of this specification are described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be executed in a different order from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the figures do not necessarily require the specific order or continuous order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0165] Although the present invention provides method operation steps such as in the embodiments or flowcharts, more or fewer operation steps may be included based on routine or non-creative labor. The order of steps listed in the embodiments is only one way among the execution orders of numerous steps and does not represent the only execution order. When the actual device or client product is executed, it may be executed in the method order shown in the embodiments or the drawings or executed in parallel (such as in an environment of parallel processors or multi-threaded processing).

[0166] For the convenience of description, when describing the above device, it is divided into various modules according to functions for separate description. Of course, when implementing the embodiments of this specification, the functions of each module can be implemented in the same or multiple software and / or hardware, or the modules implementing the same function can be realized by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in electrical, mechanical or other forms.

[0167] Those skilled in the art also know that in addition to implementing the controller in the form of pure computer-readable program code, the method steps can be logically programmed to enable the controller to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. to achieve the same function. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be regarded as the structures within the hardware component. Or even, the devices for implementing various functions can be regarded as either software modules for implementing the method or structures within the hardware component.

[0168] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0169] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM). The memory is an example of computer-readable media.

[0170] The various embodiments in this specification are described in a progressive manner. For the parts that are the same or similar among the various embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for the relevant parts, reference can be made to the partial description of the method embodiments. In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the embodiments of this specification. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0171] The above is only the embodiment of the embodiments of this specification and is not used to limit the embodiments of this specification. For those skilled in the art, various changes and modifications can be made to the embodiments of this specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of this specification shall be included within the scope of the claims of the embodiments of this specification.

Claims

1. A method for suppressing the noise of magnetoencephalogram signals, characterized in that, Including: Obtain a target magnetoencephalogram (MEG) signal; Determine the noise in the target MEG signal according to a pre-generated neural network, where the neural network is generated by training with historical MEG signals, and the training cut-off condition of the neural network is that the difference between the noise determined by training and the average value of the pre-determined noise in the historical MEG signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target MEG signal and the output of the neural network; the output of the neural network is the noise determined by training; Suppress the noise of the target MEG according to the average value of the noise.

2. The inhibition method according to claim 1, characterized in that, The fully connected layer structure of the residual attention convolution module of the neural network is one-dimensional convolution; the residual convolution module and the residual attention convolution module are used to learn the noise characteristics in the historical MEG signals.

3. The inhibition method according to claim 1, characterized in that, The historical MEG signals include training data and test data; the average value of the pre-determined noise belongs to the test data; the noise in the training data is uncorrelated with the noise in the test data.

4. The suppression method according to claim 3, wherein The noise in the training data and the noise in the test data follow a zero-mean distribution, and the real MEG signals in the training data are the same as the real MEG signals in the test data.

5. A device for suppressing the noise of magnetoencephalogram signals, characterized in that, Including: A target signal acquisition module, configured to obtain a target MEG signal; A noise extraction module, configured to determine the noise in the target MEG signal according to a pre-generated neural network, where the neural network is generated by training with historical MEG signals, and the training cut-off condition of the neural network is that the difference between the noise determined by training and the average value of the pre-determined noise in the historical MEG signals is less than a preset value; the input of the residual convolution module in the neural network is the difference between the target MEG signal and the output of the neural network; the output of the neural network is the noise determined by training; A noise suppression module, configured to suppress the noise of the target MEG according to the average value of the noise.

6. The suppression device according to claim 5, characterized in that, The fully connected layer structure of the residual attention convolution module of the neural network is one-dimensional convolution; the residual convolution module and the residual attention convolution module are used to learn the noise characteristics in the historical MEG signals.

7. The suppression device according to claim 5, wherein The historical MEG signals include training data and test data; the average value of the pre-determined noise belongs to the test data; the noise in the training data is uncorrelated with the noise in the test data.

8. The suppression device according to claim 7, wherein, The noise in the training data and the noise in the test data follow a zero-mean distribution, and the real MEG signals in the training data are the same as the real MEG signals in the test data.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by a processor, the steps of the method for suppressing the noise of the MEG signal according to any one of claims 1 to 4 are implemented.

10. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, the steps of the method for suppressing the noise of the MEG signal according to any one of claims 1 to 4 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the steps of the method for suppressing the noise of the MEG signal according to any one of claims 1 to 4 are implemented.

Images