Real-Time Lumen Detection for Autonomous Colonoscopy

Lumen detection and tracking in the large bowel is a key prerequisite step for autonomous navigation of endorobots for colonoscopy. Attempts at detecting and tracking the lumen so far have been made using optical flow and shape-from-shading techniques. In general, these methods are computationally expensive, and most are either not real-time nor tested on real devices. To this end, we present a deep learning-based approach for lumen localisation from colonoscopy videos. We avoid the need for extensive, costly annotations with a semi-supervised learning and a self-training scheme, whereby only a small subset of video frames is annotated. We develop an end-to-end pseudo-labelling semi-supervised approach incorporating a self-training scheme for colon lumen detection. Our approach reveals a competitive performance to the supervised baseline model with both objective and subjective evaluation metrics, while saving heavy labelling costs in terms of clinicians' time. Our method for lumen detection runs at 60 ms per frame during the inference phase. Our experiments demonstrate the potential of our system in real-time environments, which contributes towards improving the automation of robotics colonoscopy.