E-Sort: empowering end-to-end neural network for multi-channel spike sorting with transfer learning and fast...

Objective.Spike sorting, which involves detecting and attributing spikes to their putative neurons from extracellular recordings, is a common process in electrophysiology and brain-computer interface systems. Recent advances in large-scale neural recording technologies are challenging the conventional algorithms because of the intensive computational workloads required and the accuracy degradation suffered from time-variant spike patterns and significant levels of noise. Neural networks (NNs) have demonstrated promising performance in processing these large-scale neural recordings. However, their applications are constrained by the labor-intensive data labeling and the lack of fully vectorized frameworks with end-to-end NNs.Approach.We propose E-Sort, an end-to-end NN-based spike sorter with transfer learning and parallelizable post-processing to address both obstacles.Main results.We examined our framework in both synthetic and real datasets. The results of the processing of the synthetic datasets show that our approach can reduce the number of annotated spikes required for training by 44% compared to training from scratch, achieving up to 25.7% higher accuracy. We evaluated E-Sort on various probe geometries, noise levels, and drift patterns, which demonstrates that our design can achieve an accuracy that is comparable with Kilosort4 while sorting 50 s of data in only 1.32 s. To test with real datasets, we first sorted the spikes using Kilosort4 and used the sorted spikes at the initial period to pre-train the NN; then we compared and measured the agreement between the results from the trained model and those from Kilosort4. On average the pre-training process improved the result agreement by 30% approximately.Significance.E-Sort offers a scalable, efficient, and accurate NN-based framework for large-scale spike sorting, significantly reducing manual labeling effort and processing time.