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

Area of Science:

• Neuroscience
• Computational Neuroscience
• Bioengineering

Background:

• Spike sorting is crucial for analyzing extracellular recordings in neuroscience and brain-computer interfaces.
• Large-scale neural recordings present computational challenges and accuracy issues for conventional algorithms due to noise and changing spike patterns.
• Neural networks (NNs) show promise but are limited by manual data labeling and lack of end-to-end frameworks.

Purpose of the Study:

• To develop an end-to-end NN-based spike sorting framework, E-Sort.
• To address limitations of manual labeling and computational intensity in large-scale spike sorting.
• To improve accuracy and efficiency in spike sorting for electrophysiology.

Main Methods:

• Proposed E-Sort, an end-to-end NN framework incorporating transfer learning and parallelizable post-processing.
• Evaluated E-Sort on synthetic datasets to assess training efficiency and accuracy improvements.
• Tested E-Sort on real datasets, using Kilosort4 for initial pre-training of the NN.

Main Results:

• Reduced annotated spike requirements by 44% and increased accuracy by up to 25.7% on synthetic data.
• Achieved accuracy comparable to Kilosort4, sorting 50s of data in 1.32s.
• Pre-training with Kilosort4 improved NN model agreement by approximately 30% on real datasets.

Conclusions:

• E-Sort provides a scalable, efficient, and accurate NN-based solution for large-scale spike sorting.
• Significantly reduces manual labeling effort and processing time in neural data analysis.
• Enables more accessible and rapid analysis of complex electrophysiological data.