A dynamic texture approach to semi-automatic thrombosis segmentation in in-vivo microscopic video-sequences

The larva of wild-type zebrafish is an excellent model to study clot formation, however, the analysis of respective microscopic image sequences mainly remains manual, thus tedious and error-prone. Despite the obvious benefit of automatic segmentation algorithms for biologists, no satisfactory solution exists yet, which is mainly due to inherent problems like low contrast, motion perturbation, and disturbing edges. In this work we propose a semi-automatic segmentation algorithm, which combines a novel measure derived from dynamic textures, a temporal prior, and an edge-based refinement. With this mixture of time, texture, and gradients, we exploit a large range of temporal and spatial information and thus drive a contour-based segmentation to the accurate solution. Tests on a sequence of hand-labeled microscopic images demonstrate the merit of our approach.