A deep learning approach to improve signal quality: spike denoising for reliable sorting using transformer networks

Abstract

Accurate sorting is critical in neural signal processing. This paper presents a spike denoising method using a transformer network for enhanced spike sorting. Accurate spike sorting involves identifying and isolating signals generated by individual neurons from recordings obtained from multiple neurons. A transformer is a deep learning model that uses self-attention to differentially weight the significance of each part of the input data. Transformer networks consist of two main parts: the Encoder and the Decoder. The Encoder is made up of encoding layers that iteratively map the input to a new space, while the Decoder contains decoding layers that generate a denoised signal from the output of the Encoder. The aim of this paper is to present a supervised sorting model that could isolate the raw spike waveforms to their fired neurons with higher accuracies, leveraging the attention mechanism in the proposed model. The developed approach utilizes the transformer network to increase the accuracy of detecting the main source of spikes. Additionally, this method extracts the main spike source in cases of overlapped spikes. To assess our model’s performance, we tested it on two public datasets. The first dataset was designed to evaluate the model’s ability to handle noise, while the second dataset presented challenges for spike classification, making it a suitable test for our model’s capabilities. Our model outperformed other existing models on both datasets. Our model achieved an average classification accuracy of 99.85% on the simulated dataset and 95.06% on the experimental dataset, outperforming other models. These results suggest that our model is well-suited for handling noise and difficult spike classification scenarios and is superior to other currently available models.