SurgicalSemiSeg: A Semi-Supervised Framework for Laparoscopic Image Segmentation

Abstract

Deep learning applications in surgery are heavily reliant on large-scale datasets with high-quality annotations, which are costly and time-consuming to obtain. Self-supervised learning (SSL) has shown significant potential for reducing reliance on labelled data. This work investigates the use of SSL for semantic segmentation in laparoscopic cholecystectomy (LC) surgery. Through evaluation of existing SSL methods, we find that pixel-level objectives enable the most effective representation learning for laparoscopic imaging, characterised by highly variable and deformable anatomy. Building on this insight, we develop a tailored masked denoising autoencoder with a carefully optimised masking ratio and patch size for semantic segmentation. Our method achieves state-of-the-art performance across three LC datasets. Of note, it significantly improves segmentation accuracy for critical anatomical structures that are under-represented in training datasets. Furthermore, our approach achieves generalisability, with pre-trained representations performing effectively across fine-tuning datasets from different LC datasets.