Needles in the Haystack-Rapid Error Detection for Cell Tracking Solutions

Extracting complete cell lineages or trajectories from time-lapse microscopy images, or cell tracking, is a necessary pre-cursor for many bioimaging analyses. Powered in part by extremely accurate segmentations produced by deep learning methods, automated tracking tools make fewer mistakes than ever. Still, no method is perfect, and remaining errors must be corrected for biologists to draw meaningful conclusions from the data. While some interfaces exist for exploring and editing tracks, this is an entirely unguided process, with users having to skim through thousands of mostly correct tracks to find the rare error. In this paper, we present an edge-ranking method for proofreading auto-generated ground truth or finding remaining errors in a set of tracks. Using the Cell Tracking Challenge benchmark datasets and their ground truth annotations to simulate human proofreading, we show that through our method, users can correct 50% of remaining errors in a tracking solution after viewing a median of just 2.6% of edges.