Research On Key Technologies Of Neural Spike Sorting

The extracellular recording of neural signals and the spike sorting methods are fundamental to the study of neurological synergism.Unlike the intracellularly recorded neural action potential,the extracellularly recorded spikes have several components,which is owing to so many neurons around the tip,the locations between each neuron and the recording tip are different,and the issuing times of each neuron are various also,etc.So,processing them will confront a lot of problems.They are including sorting the overlapping spikes,low signal-to-noise ratio,the unknown number of neurons,and so on.These problems pose challenges to the spike sorting methods.How does it extract the most discriminative spikes' feature?And how does it perform high-accuracy clustering for the extracellularly recorded spikes?To meet these challenges,from the view of feature extraction and clustering,this paper carries out the following three studies of spike sorting.1.Spike feature extraction method based on Wavelet packet decomposition and mutual information.The main contribution of this work focuses on extracting discriminative features of spikes.Specifically,by performing wavelet packet decomposition on each detected spike,the original time-frequency feature set of the spikes is obtained firstly.These time-frequency features got from the wavelet packet decomposition are discriminative for the overlapping spikes.Then,given a small number of training spike samples,the mutual information and conditional mutual information between the time-frequency features and the samples' label could be calculated.By taking these pieces of mutual information as criteria,the final representative features could be well selected.This method could solve the redundancy problem that existed in the feature selection procedure and select the most discriminative time-frequency features of spikes.Simulation results confirmed that the wavelet packet decomposition and mutual information-based method could extract more discriminative spike features.And it also performs well in the tasks such as overlapping spike sorting and noise immunity testing compared with the widely used features got from spike's waveform or wavelet coefficients decomposed from the spikes signal over the low-frequency band.2.A robust spike sorting method based on low-rank and sparse representation.The main contribution of this work focuses on the study of the robust clustering of spikes.Specifically,the detected spikes are first modeled as a cumulative sum of the clean spikes and the noise.Then,by forcing the spike matrix with a low-rank structure and the noise matrix with a sparse structure,an affinity matrix is obtained.This affinity matrix can robustly characterize the correlation between the spikes.Further,the number of neurons is estimated based on this affinity matrix by employing matrix decomposition theory.Follow that,the affinity matrix and the estimated number of neurons are assigned as the inputs of the spectral clustering algorithm to achieve an automatic spike sorting process finally.The numerical simulation results show that the proposed robust spike sorting method based on the low-rank and sparse representation has better anti-noise performance compared with the recently emerging methods,which model the spike sorting as one sparse coding problem.Besides,the proposed method also shows considerable performance in solving the problem of overlapping spike sorting as well as the problem of automatic neuron number estimation.3.An automatic spike sorting method based on a joint model of feature extraction and clustering.The main contributions of this work focus on jointly modeling the feature extraction and clustering stages in spike sorting and efficiently solving the model.More concretely,by analyzing the internal relationship between the feature extraction and the clustering stages in spike sorting,and considering the feedback from the clustering stage in the feature extraction stage,a model for unifying the spike feature extraction and clustering process is proposed.In finding the solution of the model,the original separately performed feature extraction and clustering are alternately iterated towards increasing the value of the objective function.Finally,it converged with stable and highly accurate sorting results output.Besides,by adopting the K-menas++initialization and a comparative updating strategy in finding the solution stage,the drawbacks of the clustering method,such as initializing sensitivity and easily trapped into a local optimum,are well solved.Meanwhile,by embedding the existing clustering validity index in the process of initializing the indication matrix,which is associated with the clustering results,the number of neurons is effectively estimated.Finally,this paper gets an automatic spike sorting method.Experimental results on synthetic and real-world datasets show that the automatic spike sorting method with unified feature extraction and clustering achieves stable and highly accurate results based on the automatic estimation of the number of neurons,robustness to noise,and well classifying the overlapping spikes.In summary,this paper focuses on feature extraction and clustering-related techniques in spike sorting.The experimental validations are conducted on the extracellular spikes recorded by single-channel electrodes and four-channel tetrode.These methods and experimental results are of great practical importance for the acquisition of neural signals with single-neuron resolution or single-action-potential resolution.Especially,they are of high academic significance for understanding the coordination of the nervous system in large animals or humans.