Fast deep neural correspondence for tracking and identifying neurons in using semi-synthetic training.

Journal: eLife
Published Date:

Abstract

We present an automated method to track and identify neurons in , called 'fast Deep Neural Correspondence' or fDNC, based on the transformer network architecture. The model is trained once on empirically derived semi-synthetic data and then predicts neural correspondence across held-out real animals. The same pre-trained model both tracks neurons across time and identifies corresponding neurons across individuals. Performance is evaluated against hand-annotated datasets, including NeuroPAL (Yemini et al., 2021). Using only position information, the method achieves 79.1% accuracy at tracking neurons within an individual and 64.1% accuracy at identifying neurons across individuals. Accuracy at identifying neurons across individuals is even higher (78.2%) when the model is applied to a dataset published by another group (Chaudhary et al., 2021). Accuracy reaches 74.7% on our dataset when using color information from NeuroPAL. Unlike previous methods, fDNC does not require straightening or transforming the animal into a canonical coordinate system. The method is fast and predicts correspondence in 10 ms making it suitable for future real-time applications.

Authors

  • Xinwei Yu
    Department of Physics, Princeton University, Princeton, United States.
  • Matthew S Creamer
    Interdepartmental Neuroscience Program, Yale University, New Haven, CT 06511, USA.
  • Francesco Randi
    Department of Physics, Princeton University, Princeton, United States.
  • Anuj K Sharma
    Department of Physics, Princeton University, Princeton, United States.
  • Scott W Linderman
    Department of Statistics, Stanford University, Stanford, United States.
  • Andrew M Leifer
    Department of Physics, Princeton University, Princeton, United States.