A Self-supervised Approach for Detecting the Edges of Haustral Folds in Colonoscopy Video

Providing 3D navigation in colonoscopy can help decrease diagnostic miss rates in cancer screening by building a coverage map of the colon as the endoscope navigates the anatomy. However, this task is made challenging by the lack of discriminative localisation landmarks throughout the colon. While standard navigation techniques rely on sparse point landmarks or dense pixel registration, we propose edges as a more natural visual landmark to characterise the haustral folds of the colon anatomy. We propose a self-supervised methodology to train an edge detection method for colonoscopy imaging, demonstrating that it can effectively detect anatomy related edges while ignoring light reflection artifacts abundant in colonoscopy. We also propose a metric to evaluate the temporal consistency of estimated edges in the absence of real groundtruth. We demonstrate our results on video sequences from the public dataset HyperKvazir. Our code and pseudogroundtruth edge labels are available at https://github.com/jwyhhh123/ HaustralFold Edge Detector.