DualSort: online spike sorting with a running neural network

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Spike sorting, the process of identifying and separating neuronal action potentials, is a critical yet challenging step in brain activity analysis.
• Existing neural network (NN) approaches often focus on individual components of the spike sorting pipeline, requiring complex architectures.
• There is a need for efficient, low-complexity methods for real-time spike sorting with reduced manual intervention.

Purpose of the Study:

• To introduce DualSort, a simple NN combined with post-processing for efficient and real-time spike sorting.
• To demonstrate that high performance in spike detection and sorting can be achieved without complex NN architectures, even under high noise.
• To reduce the need for extensive manual labeling through data augmentation techniques.

Main Methods:

• DualSort, a simple neural network (NN), was trained and evaluated using synthetic and experimental single-channel extracellular recordings.
• The NN detects and categorizes spike waveforms in unison by classifying spikes iteratively across the signal.
• A downstream post-processing algorithm refines the NN output into precise spike trains, enhancing overall system robustness.

Main Results:

• DualSort successfully detected, distinguished, and separated different neuronal spike waveforms from background noise.
• The integrated post-processing significantly improved the model's performance and robustness.
• DualSort demonstrates competitive performance against state-of-the-art methods that target specific sub-problems.

Conclusions:

• Simple neural networks, like DualSort, coupled with post-processing, are sufficient for high-performance spike sorting, challenging the need for complex architectures.
• The framework enables reduced manual labeling through data augmentation and can operate autonomously with unsupervised pseudo-labeling.
• DualSort's low complexity facilitates efficient real-time processing on basic hardware and shows potential for analyzing other biosignals like EEG.