A System for Spatio-Temporal Cell Detection and Segmentation in Time-Lapse Microscopy

Cell segmentation is a crucial step for understanding cell mechanisms, behaviors and for analyzing them to improve health and drug discovery. In this work, we propose an automated system for cell segmentation in video-sequences by non-linear diffusion partial differential equations (PDE) modeling extended to the joint spatio-temporal domain and region-based levelset optimization. Our first step is intensity standardization by histogram transformation, to reduce the intensity variability across all the frames. After this step, moving regions are detected in each set of three consecutive frames by solving a non linear PDE in spatio-temporal domain, using the optimal spatial and temporal parameters. We then use watershed segmentation to delineate cells, and apply energy minimizing functions to refine the delineation. Differential motion activity feature maps are computed to detect moving cells. This stage is then followed by cell cluster detection and separation. We validated our method on datasets of image sequences of live cells and reference masks from the Cell Tracking Challenge (CTC) consortium. Our methodology produced promising detection and segmentation results over image sequences varying in size, shape, illumination and dyeing techniques.