A unified graphical models framework for automated human embryo tracking in time lapse microscopy

Time lapse microscopy has emerged as an important modality for studying early human embryo development. Detection of certain events can provide insight into embryo health and fate. Embryo tracking is challenged by a high dimensional search space, weak features, outliers, occlusions, missing data, multiple interacting deformable targets, changing topology, and a weak motion model. We address these with a data driven approach that uses a rich set of discriminative image and geometric features and their spatiotemporal context. We pose the mitosis detection problem as augmented simultaneous segmentation and classification in a conditional random field framework that combines tracking based and tracking free elements. For 275 clinical image sequences we measured division events during the first 48 hours of embryo development to within 30 minutes resulting in an improvement of 24.2% over a tracking-based approach and a 35.7% improvement over a tracking-free approach, and more than an order of magnitude improvement over a traditional particle filter, demonstrating the success of our framework.