MAC-S³ Robo: Mamba-Driven Contrastive Semi-Supervised Surgical Robot Segmentation for Minimally Invasive Surgery

Accurate and reliable surgical robot segmentation during endoscopic surgery is critical for improving robotic perception, task automation, and patient safety. Current deep learning approaches rely heavily on supervised training, requiring extensive, manually annotated datasets, which are costly and difficult to obtain in surgical contexts.To address this challenge, MAC-S$^{3}$Robo adopts a semi-supervised cross-supervision scheme in which two branches exchange hard pseudo-labels on unlabelled frames and are jointly optimised with labelled supervision under a unified objective. Within this framework, a Mamba-based U-shaped encoder-decoder replaces U-Net convolutional blocks with visual state-space (VSS) blocks while preserving the multi-scale design and identity skip connections, enabling stronger modelling of long-range dependencies and global context. Contrastive self-supervision is integrated into the same objective to strengthen pixel-level representations and stabilise features, benefiting the semi-supervised setting. Extensive experiments on the EndoVis 2017 and 2018 benchmarks demonstrate that MAC-S$^{3}$Robo achieves the best mean Dice among semi-supervised baselines at 10% labelled data on both datasets and maintains consistent gains as the labelled ratio increases. The code will be available at https://github.com/ziyangwang007/CV-SSL-Robot