BarlowTrack: A Self-Supervised Framework for Zero-Shot Multi-Object Cell Tracking

Recent advances in neuroscience have made it possible to image large brain regions at single-cell resolution. However, classic methods for processing these videos into neuronal time series fail in the presence of large and nonrigid deformations, in particular for freely moving animals. Several successful algorithms have been proposed to solve this problem in moving C. elegans, but they are highly specific to the conditions of a single research setting. We propose a tracking pipeline based on self-supervised learning that achieves a high level of zero-shot accuracy across conditions, and, for the first time, independent laboratories. We contribute a novel term in the Barlow Twins loss function to encourage decorrelation of features across detected instances at the same time point. To encourage broad adoption, we use the standardized Neurodata Without Borders (NWB) format and we provide a GUI for visualization of the final results and neuronal time series. Finally, we provide a benchmark of datasets with ground truth annotations in the NWB format for further algorithmic development.