Behavioral clusters revealed by end-to-end decoding from microendoscopic imaging

In vivo calcium imaging using head-mounted miniature microscopes enables tracking activity from neural populations over weeks in freely behaving animals. Previous studies focus on inferring behavior from a population of neurons, yet it is challenging to extract neuronal signals given out-of-focus fluorescence in endoscopic data. Existing analysis pipelines include regions of interest (ROIs) identification, which might lose relevant information from false negatives or introduce unintended bias from false positives. Moreover, these methods often require prior knowledge for parameter tuning and are time-consuming for implementation. Here, we develop an end-to-end decoder to predict the behavioral variables directly from the raw microendoscopic images. Our framework requires little user input and outperforms existing decoders that need ROI extraction. We show that neuropil/background residuals carry additional behaviorally relevant information. Video analysis further reveals an optimal decoding window and dynamics between residuals and cells. Critically, saliency maps reveal the emergence of video-decomposition across our decoder, and identify distinct clusters representing different behavioral aspects. Together, we present a framework that is efficient for decoding behavior from microendoscopic imaging, and may help discover functional clustering for a variety of imaging studies.