A new convex variational model for liver segmentation

Due to intensity overlapping, blurred edges and complex backgrounds with clutter features, liver segmentation is still a challenging task. In this paper, we address it with a constrained convex variational model, which can definitely avoid leakage through anatomical knowledge from users. A novel heuristic intensity model is proposed to suppress irrelevant strong edges and constrain the segmentation. Both global and local region appearance information are integrated to model higher level features such as local context. As a result, weak liver boundaries and fine structures can be stably delineated according to the information from neighborhood and nearby layers. No precise prior segmentation is needed and few seeds without shape restriction, about three seeds, are adequate to capture fine structures. The initialization is also very easy. Moreover, an accelerated primal-dual algorithm is proposed to efficiently and globally optimize the model. Our method is validated on MICCAI dataset and produces a high score of 80.6. It can be used to segment other abdominal organs.