See more, detect more: investigating colonic inspection techniques and polyp detection using machine learning

Abstract

Abstract Background Colorectal cancer (CRC) is a leading cause of cancer mortality despite widespread colonoscopy screening. Colonoscopy effectiveness, which can reduce CRC risk by 90%, depends largely on adenoma detection rate (ADR), a metric strongly influenced by mucosal inspection quality during withdrawal. However, ADR miss rates remain as high as 26% due to incomplete inspection. Although simulation-based training (SBT) improves procedural skills, physical simulators lack objective assessment and feedback on inspection quality. This study aimed to develop an automated machine-learning model (ML) to objectively assess colonoscopy inspection performance and identify feedback to improve ADR. Methods A Vision Transformer (ViT)-based ML model was developed to automatically recognize three inspection techniques during SBT required for effective inspection: clear lumen view, lateral wall, and obscured (fold) inspection. Model performance was evaluated using accuracy, precision, recall, and F1 score. The validated model was applied to videos from experts ( n = 5), intermediate ( n = 5), and novice residents ( n = 7) performing withdrawal on a physical simulator embedded with silicone polyps. The model quantified time spent in each technique and surrounding polyp detection. Polyp detection rate (PDR) was quantified based on number of polyps detected. Results The model achieved high accuracy, precision, recall, and F1 score. Greater inspection time in obscured regions was significantly associated with higher PDR ( r = 0.50, p = 0.041). Compared with novices, experts spent more time inspecting with a clear lumen view ( p = 0.039) and deeper obscured views ( p = 0.017), demonstrated higher PDR, and devoted more time to obscured regions before or after polyp detection. Conclusion The ViT ML model accurately recognizes key inspection techniques and objectively differentiates inspection performance across expertise levels using time-based metrics and PDR. This enables objective assessment and targeted feedback—such as highlighting uninspected, obscured regions within physical SBT, with the potential to improve inspection performance and ADR.