Improving Semi-Supervised Surgical Tools Segmentation with Contrastive Dual Attention

Abstract

Accurate surgical tool segmentation is crucial for robot-assisted surgery. However, the scarcity of labelled pixel-level data hinders effective neural network training for precise instrument localization in surgical areas. This paper introduces a novel semi-supervised segmentation network with Convolutional Block Attention Module (CBAM) and contrastive learning. By employing classifiers and projectors, positive and negative feature sets are constructed, framing the learning process as a contrastive problem. Evaluated on the Kvasir instrument dataset, our approach consistently outperforms state-of-the-art semi-supervised and fully supervised segmentation models, demonstrating effective segmentation of previously unseen surgical equipment with accurate predictions.