Effective 3D Boundary Learning Via A Nonlocal Deformable Network

Due to the unbalance between the boundary pixels and regional pixels, the accuracy of boundary prediction is a challenging issue for learning-based medical segmentation approaches. In this paper, we propose a two-stage segmentation method to identify and refine the object boundary accordingly. By modeling the boundary by the signed distance function, we develop a nonlocal deformable convolutional network to accurately predict the local geometry of boundaries. We also introduce an efficient loss function to enhance the learning ability in the boundary area. Experiments on two public spleen datasets can evidence the superior performance of the proposed model compared to the existing 2D, 3D, and boundary-based learning methods.