Learning to segment mouse embryo cells

Recent advances in microscopy enable the capture of temporal sequences during cell development stages. However, the study of such sequences is a complex and time consuming task. In this paper we propose an automatic strategy to address the problem of semantic and instance segmentation of mouse embryos using NYU’s Mouse Embryo Tracking Database. We obtain our instance proposals as refined predictions from the generalized Hough transform, using prior knowledge of the embryo’s locations and their current cell stage. We use two main approaches to learn the priors: Hand-crafted features and learned features. Our strategy increases the baseline jaccard index from 0.69 to 0.71 using learned features.