Surgivisor: Transformer-based semi-supervised instrument segmentation for endoscopic surgery

Abstract

Precise instrument segmentation helps tracking of instruments in surgery. The most of the existing instrument segmentation methods are fully supervised, which are based on 100% labeled data. However, the annotation for instrument segmentation are really expensive, which need the skilled professionals who can identify the parts and types of the surgical instruments. In this work, we propose a transformer-based semi-supervised instrument segmentation for endoscopic surgery, called Surgivisor. First, we present a data augmentation technique to generate synthetic data from endoscopic images to overcome the complex background and instrument collision problem, by fully using the information of unlabeled data and pseudo labels. Second, we propose a mutual prototype loss and a dual structural similarity loss to address illumination reflection and bloody condition issues in the training phase. With the two improvements, the effectiveness of proposed method is validated by the experiments on EndoVis Challenges. It exceeds the state-of-the-art results on the sub-tasks of binary, part, and type.