Self-Paced Learning for Automatic Prostate Segmentation on MR Images with Hierarchical Boundary Sensitive Network
Wenjian Qin, Zhibo Xiao, Yaoqin Xie, Yixuan Yuan · 2020
Accurate segmentation of Magnetic Resonance (MR) on prostate is an essential step for robotics surgery in prostate cancer treatment planning. This paper proposes a Hierarchical Boundary Sensitive Residual U-net (HBS-RUnet) model with self-paced learning strategy for prostate segmentation in MR image. Instead of regarding the segmentation task independently, our network consists of two branches: one segmentation branch detects the prostate region and the boundary branch finds prostate shape. The outputs of boundary branch are employed to refine the HBS-RUnet model by adding a boundary regularization, which helps to find desirable and spatially consistent prostate region. Moreover, a hierarchical dynamic self-paced learning strategy is proposed to measure the difficulty for each prostate image and gradually select the relatively simpler samples for model training. Such a simple-to-complex learning strategy could robustly learn image features and enable the robust prostate segmentation. We applied 66 cases from the PROSTATEx Challenge to evaluate the robustness and effectiveness of the proposed HBS-RUnet, and our fully automatic segmentation results demonstrate high consistency (DSC 87.1%) with the manual segmentation results by experienced physicians.