Part-Boundary-Aware Networks for Surgical Instrument Parsing

Jiaqi Liu, Yu Qiao, Jie Yang, Yun Gu

Abstract: Semantic parsing of surgical instruments provides essential information for further control. Due to the visual variations in surgery scenes, it challenging for automate segmentation task of instruments. In this paper, we proposed PBANet, which is short for Part-Boundary-Aware Networks, decomposing the instrument segmentation into two sub-tasks. An encoder-decoder architecture is adopted to predict the part-aware distance map that highlights the spatial structure of instruments. The segmentation mask is then obtained via the sigmoid function. We further propose to use a multi-scale dilation loss to reduce the boundary confusion. Empirical evaluations are performed on EndoVis2017 sub-challenge, demonstrating that the proposed method achieves superior performance compared to baseline methods.