The process for training a classifier in image segmen- tation for computer vision, relies highly on how appropri- ate the image feature extraction methods used to train these classifier are. Although one could implement as many fea- tures available as possible. But there is no doubt some of them have a much higher contribution in producing the final probability map. In this work, we implemented a so called auto-context feature extraction algorithm on a biomedical image analysis problem by using Beiers et al method of fusion moves for solving the NP-hard problem of optimizing the MAP.