Neuron signal denoising method based on Bi-LSTM and attention mechanism

Abstract

The invention provides a neuron signal denoising method based on Bi-LSTM and an attention mechanism, and the method comprises the steps: actually measuring a neuron signal, carrying out the data preprocessing, synthesizing a data set, carrying out the deep learning of a model, and finally obtaining a denoised neuron signal, the deep learning model is a neural network composed of a shallow auto-encoder, an embedded Bi-LSTM layer and an attention mechanism layer. The method is characterized in that the Bi-LSTM with high time sensitivity and the attention mechanism for enhancing the feature weight are used, compared with a common neuron signal denoising method, the accuracy is high, the original form of the peak of the neuron electric signal can be reserved as far as possible under the conditions that the signal-to-noise ratio of acquisition equipment is low and the number of channels is limited, and the accuracy of the neural signal denoising method is improved. And the method has the remarkable advantages of verification of a synthetic data set and the like.

Description

Technical Field

[0001] The present invention relates to a neuron signal denoising method based on Bi-LSTM and attention mechanism. Background Art

[0002] Microelectrode technology has become one of the most widely used technologies in the field of neuronal signals. It can be used to separate neuronal cells from biological brains for independent signal acquisition (such as patch clamp technology), or to measure voltage differences by placing electrode conductors close to neurons to collect action potential signals emitted by neurons. If there are many active neurons near the electrode, their discharge activities and current transfer can be recorded. The analysis of action potentials usually includes four steps: denoising, spike detection, feature extraction, and spike clustering. First, the neuronal signal must be filtered, but because the signal is easily contaminated by various noises in the environment during the acquisition process, the filtered electrical signal is severely distorted or even submerged. Neuronal signal denoising mainly faces two problems: 1. The pure effective signal is superimposed with noise, resulting in the destruction of the morphology; 2. The effective signal that retains a certain morphology is submerged by large noise, making subsequent spike detection difficult. In order to maintain the original morphology of the spike signal and extract it from the large environmental noise, it is necessary to design an efficient denoising method. Common neuronal signal denoising methods are briefly described as follows:

[0003] PCA-based method: Principal component analysis (PCA) is used to extract the principal components to maximize the explained variance by calculating the covariance matrix between signals from different channels. PCA removes the correlated noise in multi-channel signals by projecting the signals into the principal component space. On this basis, adding the wavelet threshold denoising method can effectively remove independent white noise. However, this method is not effective when the noise components are complex, and it is necessary to combine other denoising methods to improve performance.

[0004] Wavelet methods based on SWT and DWT: Based on the translation-invariant steady-state wavelet transform (SWT) and discrete wavelet transform (DWT), by adjusting different wavelet basis functions and decomposition layers, quantitatively evaluate the effects of various parameter combinations on neuronal signals in terms of denoising and morphology preservation. SWT is translation-invariant and suitable for processing non-stationary signals, while DWT has more advantages in computational complexity. Although these methods can effectively denoise, they will introduce large distortions in the non-peak part of the signal.

[0005] ICA method: Independent component analysis (ICA) is a blind source separation algorithm that decomposes mixed signals into independent components by assuming that the source signals are independent of each other. ICA is widely used in the decomposition of multi-channel EEG signals and the removal of electromyographic interference, but it assumes that the number of separated sources cannot be more than the number of channels, which limits its application in denoising neuronal signals with a small number of channels.

[0006] CCA method: Canonical correlation analysis (CCA) uses second-order statistical information to decompose EEG signals by maximizing the correlation between two sets of multivariate random variables, thereby more effectively removing myoelectric interference. However, CCA assumes that the number of sources cannot be more than the number of channels, so its effect is limited when the number of channels is small.

[0007] Decomposition method based on the inherent characteristics of signal data: Integrated empirical mode decomposition (EEMD) plus PCA and wavelet packet decomposition, using the intrinsic mode decomposition characteristics of the signal to denoise the signal. EEMD reduces the mode aliasing phenomenon by adding noise auxiliary signals to the original signal, decomposing and averaging them multiple times. In the case of a small number of channels, EEMD can achieve good signal-to-noise ratio (SNR) and standard deviation (SD) indicators, but the effect is limited for signals that are severely submerged by noise.

[0008] With the development of deep learning, its ability to understand complex data has been increasingly favored by researchers and has been widely used in the field of physiological signal denoising. 2021 43rd EMBC proposed a fully convolutional autoencoder that can generate clean neuronal activity signals from noisy multi-channel inputs, but it will affect the spike morphology of the neuronal signal. Patent CN109784242A discloses a method for denoising EEG signals based on a one-dimensional residual convolutional neural network, but this method cannot process signals that are completely submerged by noise. To this end, a highly accurate neuronal signal denoising method is needed that can retain the original morphology of electrical signal spikes as much as possible when the acquisition device has a low signal-to-noise ratio and a limited number of channels, and that has been verified by a synthetic data set. Summary of the invention

[0009] The purpose of the present invention is to overcome the deficiencies in the above-mentioned prior art and propose a neuron signal denoising method based on a bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, Bi-LSTM) and an attention mechanism. This method uses a non-contact microelectrode array to collect neuron signals, and the measured neuron signals are processed through data preprocessing, synthetic data sets, and deep learning models to finally obtain denoised neuron signal prediction values. The present invention uses a highly time-sensitive Bi-LSTM and an attention mechanism that enhances feature weights. Compared with common neuron signal denoising methods, it has high accuracy and can retain the original shape of electrical signal spikes as much as possible when the acquisition device has a low signal-to-noise ratio and a limited number of channels, and has significant advantages such as being verified by synthetic data sets. In order to achieve the above-mentioned purpose of the invention, the technical scheme adopted is as follows:

[0010] The neuron signal denoising method based on Bi-LSTM and attention mechanism proposed in the present invention includes the process of measuring neuron signal, data preprocessing, synthesizing data set, deep learning model, and finally obtaining denoised neuron signal; for the measured neuron signal, a non-contact microelectrode array is used to collect the culture solution signal X of the neuron not connected liquid Then the neurons are placed in a culture solution for culture. After the culture is completed, the signal X connected to the neurons is collected. cell ; For data preprocessing, according to the characteristics of neuronal electrical signals and the type of interference signals, X cell With X liquid The Butterworth high-low pass and notch filter preprocessing method is used to obtain an electrical signal waveform with a relatively stable baseline X cell2 With X liquid2 ; For the synthetic data set, generate a section of all zero data as the blank signal X empty , in X cell2 The peak template is selected from the image, its amplitude is scaled and smoothed to different degrees, and then added to X liquid2 With X empty In the dataset X merge ; For the deep learning model, it is composed of an encoder, a decoder, a layer of Bi-LSTM, three layers of unidirectional LSTM and a layer of attention mechanism. The model consists of X merge Pre-trained, then X cell2 As the input data of the model, the output result is the denoised neuron signal.

[0011] Furthermore, the non-contact microelectrode array is a charge-sensing sensor developed independently, and its input impedance is greater than 100 GΩ.

[0012] Furthermore, the preprocessing method includes using a Butterworth filter to remove baseline drift within 5 Hz, power frequency interference at 50 Hz, and high-frequency interference above 1000 Hz.

[0013] Furthermore, the synthetic data set, (1) according to the neuron discharge characteristics, in X cell2 Find and extract at least one set of identifiable spike signals X spike ; (2) Smoothing and filtering are performed to obtain X smooth ; (3) Generate a segment with X liquid2 All zero data with the same length is used as blank signal X empty ; (4) X smooth Randomly multiply different amplitude coefficients A (using a random number generator function, taking a value between 0.6-2), and then randomly add them to X liquid2 As the input signal X in, randomly added to X empty As the signal to be predicted X out (5)X in With X out Together they form the data set X merge .

[0014] Furthermore, in the deep learning model, the encoder consists of two layers of convolution (Conv), the decoder consists of three layers of deconvolution (TransConv), and batch normalization (BN) and linear rectification function (ReLU) are used between layers to stabilize the training process.

[0015] The present invention proposes a neuron signal denoising method, which reasonably predicts the denoised signal based on Bi-LSTM and attention mechanism, and has the following significant advantages:

[0016] (1) The deep learning model suitable for time series signal prediction can preserve the original shape of the electrical signal spike as much as possible when the signal-to-noise ratio of the acquisition device is low and the number of channels is limited.

[0017] (2) It has been verified on a synthetic dataset and compared with four mainstream methods: principal component analysis + wavelet denoising (PCAW), principal component analysis + wavelet denoising + ensemble empirical mode decomposition (PCEW), steady-state wavelet transform (SWT), and fully convolutional autoencoder (FC-DAE). BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart for implementing the method of the present invention.

[0019] Figure 2 The original signal waveforms are (a) solution noise; (b) neuron signal.

[0020] Figure 3 The waveform diagram of the signal after preprocessing: (a) solution noise; (b) neuron signal.

[0021] Figure 4 Schematic diagram of the synthetic dataset.

[0022] Figure 5 This is a structural diagram of the deep learning model.

[0023] Figure 6 Comparison of waveforms before and after denoising: (a) overall comparison; (b) local comparison.

[0024] Figure 7Bar graphs comparing denoising indicators of various methods: (a) signal-to-noise ratio comparison; (b) standard deviation comparison; (c) Pearson correlation coefficient comparison; (d) root mean square error comparison.

[0025] Figure 8 Comparison line graph of spike detection indicators of various methods: (a) accuracy comparison; (b) false alarm rate comparison. DETAILED DESCRIPTION

[0026] To further illustrate the effects and advantages of the present invention in detail, the present invention will be described in detail below in combination with the simulation results of a specific embodiment. All the following results are obtained by simulation using PyCharm 2021.1 and MATLAB R2021b software.

[0027] In order to solve the problem that the neuron signal morphology is destroyed or submerged by noise, and to overcome the drawbacks of traditional mainstream methods in specific usage scenarios, this paper proposes a neuron signal denoising method based on Bi-LSTM and attention mechanism. The algorithm block diagram is shown in the attached figure. Figure 1 As shown, the specific process follows the steps below:

[0028] (1) Measured neuronal signals: A self-made non-contact microelectrode array with a charge-sensing sensor was used with a sampling rate of 39062 Hz to first collect the culture solution signal X without access to the neuronal cells. liquid Then, the neurons of the cerebral cortex of the fetus of C57 female mice were placed in the culture solution for culture, and the neuronal signals were collected after the hatching was completed. cell , attached Figure 2 Shows X cell With X liquid The waveform of X liquid The peak-to-peak value of the baseline noise is about 20μV; cell There are spikes with different shapes and amplitudes in the graph;

[0029] (2) Data preprocessing: According to the characteristics of neuronal electrical signals and the types of interference signals, a filtering preprocessing method is used to obtain an electrical signal waveform with a relatively stable baseline. The preprocessing method includes using a Butterworth high-pass filter to remove baseline drift within 5 Hz, a notch filter to remove power frequency interference at 50 Hz, and a low-pass filter to filter out high-frequency interference above 1000 Hz. Figure 3 The waveform of the pre-processed neuron signal and the culture solution signal is shown. liquid2 With X cell2 The baseline has been reset to zero;

[0030] (3) Synthetic dataset: Based on the cell discharge characteristics, cell2 Find and extract at least one set of identifiable spike signals X spike, and smooth it to get X smooth , generate a separate section with X liquid2 All zero data with the same length is used as blank signal X empty , and finally X smooth Randomly multiply different amplitude coefficients A (using a random number generation function, taking a value between 0.6-2) and randomly add them to X liquid2 As the input signal X in , randomly added to X empty As the signal to be predicted X out , X in With X out Together they form the training set X merge Attached Figure 4 The synthesis process is demonstrated, that is, collecting spike templates from neuronal electrical signals, smoothing and filtering them, and then randomly adding them to the culture medium noise and blank signals. The former is used as the input part of the model, and the latter is used as the part to be predicted by the model.

[0031] (4) Deep learning model: X cell2 As the input of the pre-trained deep learning model, the size of each group of samples is 1×200, and the output of the model is the denoised neuron signal. merge Pre-trained, the training environment is the Pytorch framework of Pycharm software, attached Figure 5 The structure of the deep learning model is shown, and the input and output data sizes are both 1×200. Figure 6 The denoising results of the model are shown, with the noise signal on the left and the denoising result on the right. The attached sub-figure (a) shows the overall denoising performance, and the "A", "B", "C", "D", "E", and "F" within the dashed box of sub-figure (b) show the denoising details of the six spike signals. The noise superimposed on the spike signals of different amplitudes and shapes is removed, and the spike signals submerged in the noise are also exposed;

[0032] (5) Index verification part: In order to verify the performance of the model, the present invention sets up an index verification part in the process. According to the neuron signal simulation method proposed by Musia PG in 2002, the present invention uses MATLAB to synthesize and generate simulated neuron noise signals containing colored noise, correlated noise, and independent noise. Among them, independent white noise is used to simulate thermal noise in the signal acquisition system, correlated noise is used to simulate noise that produces common interference to all electrodes, and colored noise is used to simulate noise with a frequency similar to that of real spikes. The three are then superimposed to form a comprehensive noise. A spike signal with a random amplitude is added to the random position of the simulated noise to obtain a synthetic data set X. noise, use it to denoise and verify the indicators of the deep learning model proposed in this invention, and obtain the denoised results, and compare them with the denoising results of principal component analysis + wavelet (PCAW), principal component analysis + wavelet + ensemble empirical mode decomposition (PCEW), steady-state wavelet transform (SWT), and fully convolutional autoencoder (FC-DAE). The reference indicators are SNR (signal-to-noise ratio), SD (standard deviation), RMSE (root mean square error), and Person (Pearson correlation coefficient). Figure 7 The comparison results are shown, and "Our" is used as the abbreviation of the present invention. In the comparison results, except that RMSE is slightly lower than FC-DAE, the present invention achieves the best in other indicators.

[0033] In order to verify the practical value of the model, the results obtained after denoising by the five methods were subjected to peak detection, and the accuracy (ACC) and false positive rate (FPR) were obtained to further evaluate the advantages and disadvantages of the model. Figure 8 It is demonstrated that the average accuracy of the present invention is always greater than 96% when the noise increases, and the average false alarm rate is 0.0012%, which is better than the other four mainstream methods.

[0034] The above embodiments show that the neuron signal denoising method based on Bi-LSTM and attention mechanism proposed in the present invention is effective.

[0035] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention. Equivalent transformations or replacements made on the basis of the above technical solutions and several improvements such as expansion based on this are all protected by the claims of the present invention.

Claims

1. A neural signal denoising method based on Bi-LSTM and attention mechanism, characterized in that: After measuring neuronal signals, data preprocessing, synthetic data sets, and deep learning models, the de-noised neuronal signals are finally obtained; for the measured neuronal signals, the non-contact microelectrode array is used to collect the culture solution signals without access to neurons. liquid Then the neurons are placed in a culture solution for culture. After the culture is completed, the signal X connected to the neurons is collected. cell ; For data preprocessing, according to the characteristics of neuronal electrical signals and the type of interference signals, X cell With X liquid The Butterworth high-low pass and notch filter preprocessing method is used to obtain an electrical signal waveform with a relatively stable baseline X cell2 With X liquid2 ; For the synthetic data set, generate a section of all zero data as the blank signal X empty , in X cell2 The peak template is selected from the image, its amplitude is scaled and smoothed to different degrees, and then added to X liquid2 With X empty In the dataset X merge ; For the deep learning model, it is composed of an encoder, a decoder, a layer of Bi-LSTM, three layers of unidirectional LSTM and a layer of attention mechanism. The model consists of X merge Pre-trained, then X cell2 As the input data of the model, the output result is the denoised neuron signal.

2. According to claim 1, a neuron signal denoising method based on Bi-LSTM and attention mechanism is characterized in that: The non-contact microelectrode array is a charge-sensing sensor developed independently, and its input impedance is greater than 100 GΩ.

3. According to claim 1, a neuron signal denoising method based on Bi-LSTM and attention mechanism is characterized in that: The preprocessing method includes using a Butterworth high-pass filter to remove baseline drift within 5 Hz, a notch filter to remove power frequency interference at 50 Hz, and a low-pass filter to remove high-frequency interference above 1000 Hz.

4. According to claim 1, a neuron signal denoising method based on Bi-LSTM and attention mechanism is characterized in that: When synthesizing a dataset, (1) according to the neuron discharge characteristics, cell2 Find and extract at least one set of identifiable spike signals X spike ; (2) Smoothing and filtering are performed to obtain X smooth ; (3) Generate a segment with X liquid2 All zero data with the same length is used as blank signal X empty ; (4) X smooth Randomly multiply different amplitude coefficients A (using a random number generator function, taking a value between 0.6-2), and then randomly add them to X liquid2 As the input signal X in , randomly added to X empty As the signal to be predicted X out ; (5)X in With X out Together they form the data set X merge .

5. The method for denoising a neuron signal based on Bi-LSTM and attention mechanism according to claim 1, characterized in that: In the deep learning model, the encoder consists of two layers of convolution (Conv), and the decoder consists of three layers of deconvolution (TransConv). Batch normalization (BN) and linear rectification function (ReLU) are used between layers to stabilize the training process.

Images