Presentation
Cross-dataset Evaluation of Lumen Centering Methods for Real-Time Colonoscopy Navigation
SessionWednesday Poster Session
DescriptionColorectal cancer (CRC) is the second leading cause of cancer death worldwide. Colonoscopy is widely used for CRC screening and diagnosis, yet procedure quality depends on effective navigation and complete mucosal inspection. A major challenge is the loss of spatial orientation, leading to uncertain scope advancement and interrupted inspection. Accurate lumen-centering supports high-quality colonoscopy by helping endoscopists maintain visual orientation, advance the scope effectively, and support complete mucosal inspection. Several lumen-centering methods include traditional image-processing approaches, such as K-means clustering, Ellipse Fitting, and Hybrid methods, while deep learning-based YOLO models have recently been used for real-time lumen detection and localization. However, their robustness across heterogeneous colonoscopy datasets has not been systematically evaluated. This study evaluated traditional and deep learning-based lumen-centering methods across diverse colonoscopy datasets selected based on sample size, image quality, visual conditions, and ground-truth annotations. Two publicly available datasets were selected, and one simulated dataset was developed to represent controlled navigation scenarios. Traditional methods showed limited robustness under variable illumination, deformable lumen shapes, and occlusions. In contrast, YOLOv11s with optimal hyperparameters achieved best performance. This study contributes to a systematic cross-dataset evaluation of lumen-centering approaches and identifies a robust YOLO model for supporting navigation guidance in simulation-based training.
Event Type
Poster
TimeWednesday, October 21st5:30pm - 6:30pm PDT
Location