Research On The Spike Sorting Method Based On Deep Learning

In the fields of neuroscience and biomedical signal processing,spike sorting is a crucial step to extract information of single neurons from extracellular recordings.The production of spikes is the main way for neurons in the brain to transmit information.The spike is a fast and transient fluctuation in the membrane potential of the nervous cell.Although there are many spike sorting methods,there is still room for improvement in accuracy and robustness.Therefore,a deep learning approach is proposed to more effectively solve the problem of spike sorting.This is of great significance for studying the working mechanism of the brain.Firstly,a deep learning approach based on one-dimensional convolutional neural networks(1D-CNN)is proposed to implement spike sorting.The standard architectures are convolutional layer,pooling layer,and fully connected layer.The simulated database created in “Wave?Clus” is used to verify the proposed model.According to the different proportions of the training data to the total data,six experiments are performed and used to test the model.On the other hand,the 1D-CNN model is also compared with six methods such as “WMsorting” method and a deep-learning-based multilayer perceptron(MLP)model.This model reduces the cost of manual intervention on the premise of improving accuracy.Secondly,the 1D-CNN model is verified using experimental datasets.The data is recorded from primary visual cortex(V1)of a macaque monkey.According to the clusters of spikes in each channel,setting the corresponding number of output nodes in the1D-CNN model.Using different proportions of data to train the model,and then classifying the testing set.As is known,there is always an imbalanced data distribution in experimental datasets;thus,Macro?F,besides accuracy,is also an appropriate criterion to evaluate the classification quality.The results show that the model has high accuracy and strong robustness.Finally,a deep learning approach based on CNN and Long Short Term Memory(LSTM)is proposed to implement the problems of overlapping spike sorting.According to the overlapping degree of the spikes,all overlapping spikes are calibrated.Both simulated database and experimental datasets are employed to evaluate the performance of the CNN+LSTM model.This model is also compared with 1D-CNN method.To evaluate the performance of the model,accuracy is utilized to compute the score of the entire classification.It illustrates the superiority of this model in dealing with the problems of overlapping spike sorting.