Automahout: an autonomous image analysis system for cell tracking

Abstract

Cell tracking in time-lapse images is complex image analysis that involves detection, linking and proofreading, often requiring the algorithm to be tuned for every new dataset. We present Automahout, an autonomous agent for the Mastodon/ELEPHANT platform that carries out these operations. It looks at the data and tries existing models; where none fits, it annotates reference cells, trains a detector from those sparse annotations, reviews its output against them and feeds the corrections back until its own stopping criteria are met. It then links, checks the lineage and moves on. On three public datasets, a 2D fluorescence culture, a 2D phase-contrast culture and a 3D embryo, it produced lineages without human input and reached TRA scores of 0.95-0.96.