Multilevel space-time aggregation for cell segmentation and trackingpdfauthor

We introduce a multilevel algorithm for the purpose of segmenting brightfield cell images. The V-cycle algorithm has two stages. The first stage takes all pixels in an image and repeatedly groups them into larger overlapping blocks until all blocks become salient. The second stage then takes these segments and determine exactly which pixels belong to each segment. The algorithm is implemented recursively and makes use of algebraic multigrid (AMG) coarsening. At first, using only intensity as the segmentation criterion, the algorithm performs poorly on textured examples. So we include a multilevel isotropic texture measure as a separate criterion, and the result improves greatly. Then, the algorithm is extended for 3D problems and consequently applied to a series of cell images taken over small time intervals. The result is a tube-shaped segment that tracks the cell’s location in space-time.