Ice Finder: Few-Shot Learning for Non-Vitrified Ice Segmentation

Abstract

This study introduces Ice Finder, a novel tool for quantifying crystalline ice in tomography, filling a crucial gap in existing methodologies. We establish the first application of the meta-learning paradigm to tomography, demonstrating that various tomographic tasks across datasets can be unified under a single meta-learning framework. Our approach utilizes few-shot learning to enhance domain generalization and adaptability to domain shifts, facilitating rapid adaptation to new datasets with minimal examples. Ice Finders performance is evaluated on a comprehensive set of in situ datasets from EMPIAR, proving its ease of use and fast processing capabilities, with inference times in the milliseconds. This tool not only accelerates workflows but also enhances the precision of structural studies in structural biology.