A method for removing artifacts from electrophysiological signals

Abstract

The present invention discloses a method for removing artifacts from electrophysiological signals: (1) recording the original electrophysiological signals recorded when the tested subject generates neural activity while being scanned in a magnetic resonance imaging device; (2) periodically segmenting the original electrophysiological signals to obtain multiple segments of original electrophysiological signal segments containing periodic noise; (3) preprocessing the original electrophysiological signal segments containing periodic noise to obtain electrophysiological signals after first-stage artifact removal. The method for removing artifacts from electrophysiological signals provided by the present invention can remove the noise caused by radio frequency pulses and gradient changes in the electrophysiological signals recorded simultaneously during functional magnetic resonance imaging (fMRI) scanning, and can make the detected electrophysiological signals clearer and easier to interpret.

Description

Technical Field

[0001] The present invention relates to a method for removing artifacts after acquiring electrophysiological signals under functional magnetic resonance imaging (fMRI) scanning, and in particular to a method for removing artifacts of electrophysiological signals. Background Art

[0002] With the continuous progress of medicine, the relationship between many diseases and electrophysiological signals has been studied in detail. The nervous system includes the brain, spinal cord and peripheral nerves. Due to the unique complexity and functionality of the nervous system, humans have always been eager to understand its operating mechanism in order to protect the brain, treat brain diseases, and even clone the brain to achieve artificial replacement of many brain functions.

[0003] The activity of brain nerve cells is accompanied by electrical activity. EEG signals are obtained by using electrodes to approach or enter the brain. Since functional magnetic resonance imaging (fMRI) is used when detecting brain nerve activity, EEG signals are easily affected by functional magnetic resonance imaging during the acquisition process. In addition, various factors of the subjects themselves and the characteristics of the EEG signals themselves make it very easy to be interfered by neural noise and non-neural noise during the acquisition process. Therefore, it is necessary to remove artifacts through noise removal methods to improve the signal-to-noise ratio of EEG signals, so as to facilitate further interpretation and research. For example, a multi-channel EEG signal electrooculographic artifact automatic removal method and storage medium disclosed in the Chinese patent with publication number CN114176605A, such as a synchronous EEG-fMRI online artifact removal method disclosed in the Chinese patent with publication number CN109222965A.

[0004] Therefore, how to make electrophysiological signals clearer and easier to interpret is currently a research hotspot in this field. Summary of the invention

[0005] The purpose of the present invention is to disclose an electrophysiological signal artifact removal method, which can remove the noise caused by radio frequency pulses and gradient changes in the electrophysiological signals recorded simultaneously during functional magnetic resonance imaging (fMRI) scanning, so that the detected electrophysiological signals can be clearer and easier to interpret.

[0006] The present invention provides the following technical solutions:

[0007] A method for removing artifacts from electrophysiological signals, the method comprising the following steps:

[0008] (1) Recording the original electrophysiological signals of the test subject while generating neural activity while being scanned in a magnetic resonance imaging device;

[0009] (2) segmenting the original electrophysiological signal into periods to obtain a plurality of segments of the original electrophysiological signal containing periodic noise;

[0010] (3) Preprocessing the original electrophysiological signal segment containing periodic noise to obtain the electrophysiological signal after artifact removal in the first stage;

[0011] The pretreatment method is:

[0012] (3-1) averaging the original electrophysiological signal segments containing periodic noise in the complex number to obtain a first average signal segment, and then subtracting the first average signal segment from each original electrophysiological signal segment containing periodic noise in the complex number to obtain a signal segment after the first subtraction of the complex number segments;

[0013] (3-2) summing the absolute values ​​of the differences between each signal segment after the first subtraction and all corresponding data points of other signal segments after the first subtraction to obtain a plurality of first error coefficients; comparing each first error coefficient with a first predetermined value, averaging the first subtraction signal segments corresponding to the first error coefficients less than the first predetermined value as second average signal segments; subtracting the corresponding second average signal segment from each signal segment after the first subtraction to obtain a plurality of signal segments after the second subtraction;

[0014] (3-3) extracting the signal segment after the second subtraction that still contains the noise signal as the second processed segmented signal segment, and obtaining the complex segment signal segment after the third subtraction according to the method in step (3-2);

[0015] (3-4) Overlaying the signal segment after the second subtraction and the signal segment after the third subtraction onto the signal segment at the corresponding time in the original electrophysiological signal in time sequence to obtain the electrophysiological signal after artifact removal in the first stage.

[0016] Preferably, in step (2), the method for obtaining multiple segments of original electrophysiological signal segments containing periodic noise is: obtaining signals greater than 5 times the standard deviation in the original electrophysiological signal and finding the time point of a peak in these signals, and cutting out a complete cycle according to the time point of the peak, which is the electrophysiological signal segment containing periodic noise. The present invention can accurately divide each electrophysiological signal segment through the above method, which is beneficial to the calculation of subsequent analysis.

[0017] Preferably, in step (2), the method for obtaining multiple segments of original electrophysiological signal segments containing periodic noise is as follows: in step (1), the noise signal generated when the subject under test does not generate neural activity but is scanned in the magnetic resonance imaging device is obtained; and the original electrophysiological signal obtained in step (1) is periodically segmented according to the period of the noise signal. The use of periodic segmentation is conducive to the calculation and judgment of subsequent operations.

[0018] The present invention can reduce the amplitude of huge periodic noise by periodically segmenting the original electrophysiological signal, so as to facilitate subsequent neural network processing, which is of great help in judging the read data.

[0019] Preferably, in step (3-2), the first predetermined value is the maximum value among the fifth percentile values ​​in the first error coefficient.

[0020] Preferably, in step (3-3), the absolute values ​​of the differences between each second processed segmented signal segment and all corresponding data points of other second processed segmented signal segments are summed to obtain a plurality of second error coefficients; each second error coefficient is compared with a second predetermined value, and the second processed segmented signal segments corresponding to the second error coefficients less than the second predetermined value are averaged as third average signal segments; each second processed segmented signal segment is subtracted from its corresponding third average signal segment to obtain a signal segment after the third subtraction of the plurality of segments; wherein the second predetermined value is the maximum value among the fifth percentile values ​​of the second error coefficients.

[0021] Preferably, the artifact removal method comprises step (4): further processing the first-stage artifact removal electrophysiological signal to obtain the second-stage artifact removal electrophysiological signal, specifically:

[0022] (4-1) Performing Fast Fourier Transform (FFT) on the electrophysiological signal after artifact removal in the first stage to obtain the spectrum amplitude and spectral phase

[0023] (4-2) The spectrum amplitude The square of the absolute value and the natural logarithm (log) value are obtained As input, the spectrum amplitude is output after calculation by the Fully Convolutional Networks (FCN) model

[0024] (4-3) The spectrum amplitude After the square root of the exponential function As the input of the IFFT model, the IFFT output is the artifact-free spectrum amplitude.

[0025] (4-4) Combined with the spectrum phase and spectrum amplitude The inverse fast Fourier transform method is used to restore the electrophysiological signal after the second stage of artifact removal to obtain the final electrophysiological signal after artifact removal.

[0026] The present invention further processes the electrophysiological signal after artifact removal in the first stage to obtain the electrophysiological signal after artifact removal in the second stage, so that the electrophysiological signal can be restored to be closer to the actual situation.

[0027] In the present invention, the training method of the fully convolutional network model and the long short-term memory model can be: obtaining the pure electrophysiological signal of the tested subject when neural activity is generated in the magnetic resonance imaging device but scanning is not started, and the pure noise signal generated when there is no neural activity in the magnetic resonance imaging device but scanning is performed; then the electrophysiological signal and the noise signal are inspected and preprocessed, divided into two parts, one for training and one for verification, and then two neural network models are trained in MATLAB as input and output; the data source of the input for training is the data obtained by superimposing the above-mentioned "pure electrophysiological signal" and "pure noise signal" for training and undergoing the first stage of preprocessing; the data source of the output for training is the "pure electrophysiological signal" corresponding to the input; the data for verification is then put into the processing flow, and certain parameters in the processing flow are adjusted according to the difference between the result and the clean signal for control.

[0028] Preferably, in step (4-2), the fully convolutional network model includes 6 layers of encoders and 7 layers of decoders, each layer of the encoder includes a convolutional layer, a batch normalization layer and an activation function layer, and each layer of the decoder includes a deconvolution layer, a batch normalization layer and an activation function layer.

[0029] Preferably, the convolution layer of each encoder uses filters, and the number of filters is 40, 20, 20, 20, 40, and 1, respectively, wherein the step of each filter is set to 1, and the padding is set so that the signal after filtering of the convolution layer remains unchanged, and the activation function layer of each encoder uses an exponential linear function; the deconvolution layer of each decoder uses filters, and the number of filters is 1, 40, 20, 20, 20, 40, and 1, respectively, wherein the step of each filter is set to 1, cropping is used to maintain the size of the data, and the activation function layer of each decoder uses an exponential linear function.

[0030] Preferably, in step (4-3), the formula for the square root of the exponential function is

[0031] Preferably, in step (4-4), the long short-term memory model includes a hidden layer, an activation function layer and a fully connected layer; wherein the number of neurons in the hidden layer is 100, the activation function of the activation function layer is an exponential linear function, and the number of neurons in the fully connected layer is 100.

[0032] Preferably, the electrophysiological signal is a neuro-electrophysiological signal.

[0033] Compared with the prior art, the electrophysiological signal artifact removal method of the present invention can remove the noise with huge periodicity and fluctuating amplitude in the electrophysiological signals (such as EEG signals), restore the spike signal and the local field potential (LFP) signal, and can be particularly used to remove the noise caused by the radio frequency pulse and gradient field changes in the electrophysiological signals recorded simultaneously by functional magnetic resonance imaging (fMRI) scanning, which will help to synchronously analyze the performance of a section of neural activity in functional magnetic resonance imaging and in electrophysiology, and then explore the mechanism of neurovascular coupling. In addition, the method provided by the present invention is suitable for detection in an environment of 7T ultra-high field strength. In addition, the method provided by the present invention makes the detected electrophysiological signals clearer and easier to interpret. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The figure is a flowchart of the steps of the method for removing artifacts from electrophysiological signals of the present invention.

[0035] Figure 2 Graph showing the original electrophysiological signal waveform of the embodiment.

[0036] Figure 3 4 is a waveform diagram of a signal segment containing periodic noise according to an embodiment.

[0037] Figure 4 4 is a waveform diagram of an average periodic noise signal segment of an embodiment.

[0038] Figure 5 It is a comparison diagram of the original electrophysiological signal waveform of the embodiment and the waveform after the signal segments after the first subtraction are spliced ​​and backfilled in time sequence.

[0039] Figure 6 It is a partially enlarged waveform diagram of the signal segments after the first subtraction in the embodiment after splicing and backfilling in time sequence.

[0040] Figure 7 It is a comparison diagram of the waveforms after the signal segments after the first subtraction and the signal segments after the second subtraction in the embodiment are spliced ​​and backfilled according to the time sequence.

[0041] Figure 8 It is a partially enlarged waveform diagram of the signal segments after the second subtraction in the embodiment after splicing and backfilling in time sequence.

[0042] Fig. 9 This is a waveform diagram of electrophysiological signals after artifact removal in the first stage of the embodiment.

[0043] Fig.10 FIG. 4 is a schematic diagram of the second stage artifact removal process of the embodiment.

[0044] Fig.11 for Fig.10A zoomed-in image of the Fully Convolutional Networks (FCN) model in Figure 2.

[0045] Fig.12 The figure is a comparison diagram of the preprocessed electrophysiological signal waveform and the artifact-removed electrophysiological signal waveform. DETAILED DESCRIPTION

[0046] The present invention is further explained below in conjunction with examples. These examples are only used for illustration and are not intended to limit the scope of protection of the claims of the present invention.

[0047] like Figure 1 As shown, the method for removing artifacts of electrophysiological signals provided by the present invention comprises the following steps:

[0048] S1. Record the electrophysiological signals of the test subject when the subject generates neural activity in the magnetic resonance imager but the scanning is not started, and the noise signals generated when the subject generates neural activity in the magnetic resonance imager but the scanning is carried out; record the original electrophysiological signals (300-5000Hz) recorded when the subject generates neural activity while being scanned in the magnetic resonance imager.

[0049] S2. Periodically segment the original electrophysiological signal to obtain a plurality of segments of the original electrophysiological signal containing periodic noise.

[0050] This embodiment can segment the original electrophysiological signal according to the period of the noise signal to obtain multiple electrophysiological signal segments containing periodic noise. The periodicity of the noise signal is used to segment the electrical signal, which is beneficial to the subsequent operation of removing artifacts.

[0051] S3. Preprocess the original electrophysiological signal segment containing periodic noise to obtain the electrophysiological signal after first-stage artifact removal.

[0052] Specifically:

[0053] Then proceed with the following steps, including:

[0054] S3-1. The complex segments of original electrophysiological signal containing periodic noise are averaged to obtain a first average signal segment, and then the first average signal segment is subtracted from each complex segment of original electrophysiological signal containing periodic noise to obtain a signal segment after the first subtraction of the complex segments.

[0055] S3-2. Sum the absolute values ​​of the differences between each signal segment after the first subtraction and all corresponding data points of other signal segments after the first subtraction to obtain a plurality of first error coefficients; compare each first error coefficient with a first predetermined value, and average the signal segments after the first subtraction corresponding to the first error coefficients smaller than the first predetermined value as second average signal segments; subtract the corresponding second average signal segment from each signal segment after the first subtraction to obtain a plurality of signal segments after the second subtraction.

[0056] S3-3. Extract the signal segment after the second subtraction that still contains the noise signal as the second processed segmented signal segment, and sum the absolute values ​​of the differences after subtracting each second processed segmented signal segment from all corresponding data points of other second processed segmented signal segments to obtain a plurality of second error coefficients; compare each second error coefficient with a second predetermined value, and average the second processed segmented signal segments corresponding to the second error coefficients less than the second predetermined value as the third average signal segment; subtract the corresponding third average signal segment from each second processed segmented signal segment to obtain a plurality of signal segments after the third subtraction; wherein the second predetermined value is the maximum value among the fifth percentile values ​​of the second error coefficients.

[0057] S3-4, overlaying the signal segments after the second subtraction and the signal segments after the third subtraction onto the signal segments at corresponding times in the original electrophysiological signal in time sequence, to obtain the electrophysiological signal after artifact removal in the first stage.

[0058] Through the above steps S1-S3, the amplitude of the huge periodic noise has been reduced to a sufficiently low level to facilitate the subsequent neural network processing. After the first stage of artifact removal, the electrophysiological signal can be used for the judgment of related physiological signals. This overcomes the disadvantage of the conventional technology that the noise cannot be effectively separated or reduced, resulting in difficulty in judgment.

[0059] This embodiment also includes:

[0060] S4. After the first stage of artifact removal, the electrophysiological signal is subjected to Fast Fourier Transform (FFT) to obtain a spectrum amplitude and a spectrum phase; the absolute value of the spectrum amplitude is squared and the natural logarithm (log) value is taken, and then calculated by a Fully Convolutional Networks (FCN) model, an exponential function square root formula, and a Long Short-Term Memory (LSTM) model, and then combined with the spectrum phase, after Inverse Fast Fourier Transform (IFFT), the electrophysiological signal after artifact removal in the second stage is obtained. Through the above content, the features of the original electrophysiological signal can be effectively extracted and restored, and the offset of the level can be corrected, so that the physiological signal after artifact removal is more accurate.

[0061] To make the above steps easier to understand, they are described below with reference to the drawings.

[0062] S1. The original electrophysiological signal is recorded under functional magnetic resonance imaging (fMRI) scanning, which consists of huge electrophysiological noise caused by radio frequency (RF) pulses and gradient field changes and a spike signal with a firing frequency of 5 Hz (e.g. Figure 2 As shown in the figure, the peak potential signal is invisible due to excessive noise), the original electrophysiological signal X = [x1, x2, x3, ..., x t ], where t is the data length.

[0063] S2. The method of extracting a periodic noise signal segment 1 from the original electrophysiological signal is to detect the signal with a standard deviation greater than 5 times in the electrophysiological signal and find the time point of a peak in these signals. According to the 1000 data points before and 2500 data points after the peak time point, a complete cycle is extracted according to the length of this segment of data, which is the periodic noise signal segment A. i =[x1,x2,x3,…,x 3501 ], i = 1, 2, ..., n, n is the number of periodic noise signal segments (such as Figure 3 shown).

[0064] S3-1. In the steps of this embodiment, the average periodic noise signal segment is calculated. (like Figure 4 As shown), and subtract the average periodic noise signal segment from each periodic noise signal segment, that is, obtain the signal segment after the first subtraction The signal segments after the first subtraction are spliced ​​back in time sequence to form the full-length signal (like Figure 5 The waveform below and Figure 6 shown).

[0065] S3-2, then calculate the two The first error coefficient between

[0066]

[0067] Then for each segment Screening and its first error coefficient is less than the first predetermined value 0.1V 2 The other signal segments after the first subtraction are averaged to obtain The first predetermined value is the maximum value among the fifth percentile values ​​of the plurality of first error coefficients associated with each first subtraction signal segment, and m is the number of other first subtraction signal segments that meet the conditions.

[0068] The first subtraction signal segment of each segment Subtract the corresponding At this time, the signal segment after the second subtraction can be obtained The signal segments after the second subtraction are spliced ​​back in time sequence to form the full-length signal (like Figure 7 The waveform below and Figure 8 shown).

[0069] S3-3, extract the obvious residual periodic noise signal segment in the signal segment after the second subtraction to obtain a set of second processed segmented signal segments Calculate the second error coefficient between the second processed segmented signal segments

[0070] in,

[0071] For each second processed segmented signal segment, a filter is selected whose error coefficient is less than a second predetermined value of 0.001V. 2 The other second process divides the signal into segments and averages them to obtain The second predetermined value is the maximum value among the fifth percentile values ​​of the plurality of second error coefficients associated with each second subtraction signal segment, and m is the number of other second subtraction signal segments that meet the conditions.

[0072] Then the second process splits the signal segments from each segment Subtract the average result of the other second processed segmented signal segments selected for it A set of signal segments after the third subtraction is obtained.

[0073] S3-4. Finally, the signal segments after the second subtraction and the signal segments after the third subtraction are integrated according to the time sequence, and are covered and backfilled to the signal segments of the corresponding time in the original electrophysiological signal to obtain an electrophysiological signal after the first stage of artifact removal. (like Fig. 9 shown).

[0074] S4. Please refer to Fig.10 In the steps of this embodiment, the electrophysiological signal after the first stage of artifact removal is first The spectrum amplitude is obtained by performing a Fast Fourier Transform (FFT) with a window size of 2 milliseconds (ms) and an overlap of 1 millisecond (ms). and spectral phase In addition, if used in the embodiment of the local field potential signal LFP, the window size can be changed to 1 second (s) and the overlap can be changed to 0.5 second (s) before performing fast Fourier transform (FFT).

[0075] In addition, in the steps of this embodiment, first The following line formula performs the conversion:

[0076]

[0077] Please refer to Fig.10 and Fig.11 ,Will The algorithm of the fully convolutional network (FCN) model 1 is put into the algorithm. The fully convolutional network consists of an encoder 11 and a decoder 12, with a total of 13 layers. The first 1 to 6 layers are encoders 11, and each layer includes a convolution layer 111, a batch normalization layer (not shown in the figure) and an activation function layer (not shown in the figure). Each convolution layer uses a filter, and the number of filters is 40, 20, 20, 20, 40, and 1, respectively. The stride of the filter is set to 1 and the padding is set so that the filtered signals of each layer are not changed. The activation function of the activation function layer uses an exponential linear function (ELU). The decoder 12 has a total of 7 layers, each of which includes a deconvolution layer 121, a batch normalization layer (not shown) and an activation function layer (not shown). The filter size used in each deconvolution layer is 3×1 (filter), and the number of filters is 1, 40, 20, 20, 20, 40, 1 in sequence. The step is set to 1 and cropping is used to maintain the size of the data. The activation function uses an exponential linear function (ELU). The spectrum amplitude can be obtained by calculation through the fully convolutional network (FCN) model. And convert it with the following exponential function square root formula for

[0078]

[0079] After calculation by the Long Short-Term Memory (LSTM) model, the model has three layers including hidden layer, activation function layer and full connection layer. The number of neurons in the hidden layer is 100, the activation function is the exponential linear function (ELU), and the number of neurons in the full connection layer is 100. If it is used for local field potential signals (LFP), the number of neurons in the hidden layer is changed to 400 and the number of neurons in the full connection layer is changed to 24414. After calculation by the Long Short-Term Memory (LSTM) model, the denoised spectrum amplitude can be obtained. Finally, combined with the spectrum phase and the spectral amplitude after de-artifacting The inverse fast Fourier transform (IFFT) method is used to restore the electrophysiological signal after the second stage of artifact removal, and finally multiplied by a scaling factor of 0.7 to obtain the final artifact-free electrophysiological signal (such as Fig.12 ). The scaling factor of 0.7 is calculated based on the ratio of the physiological signal before the magnetic resonance imaging scan to the original electrophysiological signal during the magnetic resonance imaging scan.

[0080] In summary, the present invention provides a method for removing artifacts from electrophysiological signals. Figures 1 to 12 It can remove the noise with huge periodicity and fluctuating amplitude in electrophysiological signals (such as EEG signals), restore spike signals and local field potential (LFP) signals, and is particularly used to remove the noise caused by radio frequency pulses and gradient changes in electrophysiological signals recorded simultaneously during functional magnetic resonance imaging (fMRI) scans. It can synchronously analyze the performance of a section of neural activity in functional magnetic resonance imaging (fMRI) and in electrophysiology, and further explore the mechanism of neurovascular coupling. It can also make the detected electrophysiological signals (such as EEG signals) clearer and easier to interpret.

[0081] The above embodiments are only used to explain the inventive concept of the present invention, rather than to limit the protection of the rights of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the essence of the technology and method of the present invention are still within the scope of the technology and method solutions of the present invention.

Claims

1. A method for removing artifacts from electrophysiological signals, characterized in that: The artifact removal method comprises the following steps: (1) Recording the original electrophysiological signals of the test subject while generating neural activity while being scanned in a magnetic resonance imaging device; (2) segmenting the original electrophysiological signal into periods to obtain a plurality of segments of the original electrophysiological signal containing periodic noise; In step (2), the method for obtaining multiple segments of original electrophysiological signal segments containing periodic noise is as follows: obtaining signals with a standard deviation greater than 5 times of the original electrophysiological signal and finding the time point of a peak in these signals, and cutting out a complete cycle according to the time point of the peak, which is the electrophysiological signal segment containing periodic noise; (3) Preprocessing the original electrophysiological signal segment containing periodic noise to obtain the electrophysiological signal after artifact removal in the first stage; The pretreatment method is: (3-1) averaging the original electrophysiological signal segments containing periodic noise in the complex number to obtain a first average signal segment, and then subtracting the first average signal segment from each original electrophysiological signal segment containing periodic noise in the complex number to obtain a signal segment after the first subtraction of the complex number segments; (3-2) summing the absolute values ​​of the differences between each signal segment after the first subtraction and all corresponding data points of other signal segments after the first subtraction to obtain a plurality of first error coefficients; Compare each first error coefficient with a first predetermined value, average the first subtraction signal segments corresponding to the first error coefficients less than the first predetermined value to obtain a second average signal segment; subtract the corresponding second average signal segment from each first subtraction signal segment to obtain a complex second subtraction signal segment; In step (3-2), the first predetermined value is the maximum value among the fifth percentile values ​​in the first error coefficient; (3-3) extracting the signal segment after the second subtraction that still contains the noise signal as the second processed segmented signal segment, and obtaining the complex segment signal segment after the third subtraction according to the method in step (3-2); (3-4) Overlaying the signal segment after the second subtraction and the signal segment after the third subtraction onto the signal segment at the corresponding time in the original electrophysiological signal in time sequence to obtain the electrophysiological signal after artifact removal in the first stage.

2. The method for removing artifacts of electrophysiological signals according to claim 1, characterized in that: In step (3-3), the absolute values ​​of the differences between each second processed segmented signal segment and all corresponding data points of other second processed segmented signal segments are summed to obtain a plurality of second error coefficients; Compare each second error coefficient with the second predetermined value, and average the second processed segmented signal segments corresponding to the second error coefficients smaller than the second predetermined value as the third average signal segment; subtract the corresponding third average signal segment from each second processed segmented signal segment to obtain a signal segment after the third subtraction of the complex segments; wherein the second predetermined value is the maximum value of the fifth percentile value in the second error coefficient.

3. The method for removing artifacts of electrophysiological signals according to claim 1, characterized in that: The artifact removal method comprises step (4): further processing the first-stage artifact removal electrophysiological signal to obtain the second-stage artifact removal electrophysiological signal, specifically: (4-1) Perform fast Fourier transformation on the electrophysiological signal after artifact removal in the first stage to obtain the spectrum amplitude and spectral phase (4-2) The spectrum amplitude The square of the absolute value and the natural logarithm of As input, the full convolutional network model calculates the output spectrum amplitude (4-3) The spectrum amplitude After the square root of the exponential function As the input of the long short-term memory model, the output of the long short-term memory model is the artifact-free spectrum amplitude. (4-4) Combined with the spectrum phase and spectrum amplitude The inverse fast Fourier transform method is used to restore the electrophysiological signal after the second stage of artifact removal to obtain the final electrophysiological signal after artifact removal.

4. The method for removing artifacts of electrophysiological signals according to claim 3, characterized in that: In step (4-2), the fully convolutional network model includes 6 encoder layers and 7 decoder layers, each encoder layer includes a convolution layer, a batch normalization layer and an activation function layer, and each decoder layer includes a deconvolution layer, a batch normalization layer and an activation function layer.

5. The method for removing artifacts of electrophysiological signals according to claim 4, characterized in that: The convolution layer of each encoder uses filters, and the number of filters is 40, 20, 20, 20, 40, and 1, respectively, where the step of each filter is set to 1, and the padding is set so that the signal after the convolution layer filtering remains unchanged, and the activation function layer of each encoder uses an exponential linear function; the deconvolution layer of each decoder uses filters, and the number of filters is 1, 40, 20, 20, 20, 40, and 1, respectively, where the step of each filter is set to 1, cropping is used to maintain the size of the data, and the activation function layer of each decoder uses an exponential linear function.

6. The method for removing artifacts of electrophysiological signals according to claim 3, characterized in that: In step (4-3), the formula for the square root of the exponential function is 7. The method for removing artifacts of electrophysiological signals according to claim 3, characterized in that: In step (4-4), the long short-term memory model includes a hidden layer, an activation function layer and a fully connected layer; wherein the number of neurons in the hidden layer is 100, the activation function of the activation function layer is an exponential linear function, and the number of neurons in the fully connected layer is 100.

8. The method for removing artifacts of electrophysiological signals according to claim 1, characterized in that: The electrophysiological signal is a neuro-electrophysiological signal.

Images