Colonoscopic withdrawal time is an important metric for evaluating colonoscopy quality, but it can be substantially affected by therapeutic maneuvers, impaired visualization, and other interruptions unrelated to mucosal inspection. In this study, we developed an artificial intelligence-based approach to estimate effective withdrawal time (EWT), defined as the periods during which the mucosa or lesions are clearly observed, with the aim of providing a more objective description of inspection time. We developed an InternImage-based model to classify complete colonoscopy withdrawal videos into four categories: Observation, Operation, Obscure, and Outside. The Observation category was defined as EWT. The annotated data were split at the video level into development and independent test sets to avoid data leakage. To explore the potential clinical relevance of EWT, we compared therapeutic, biopsy, polyp, and non-polyp examinations across multiple analyses. The model achieved a classification accuracy of 97.18% (95% CI: 96.66% to 97.39%) on the independent test set. EWT differed between therapeutic and screening-related examinations, whereas the EWT values of the biopsy, polyp, and non-polyp groups were relatively similar despite clear differences in conventional withdrawal time. EWT also varied across hospitals and physician seniority strata, suggesting potential usefulness for cross-context quality assessment. Our model can identify clinically relevant withdrawal phases from complete colonoscopy videos and objectively calculate EWT. These findings support the technical feasibility of EWT as a complementary surrogate metric for colonoscopy quality assessment across different procedural contexts. However, prospective validation with established clinical outcomes, including ADR and APC, is still needed.
Wang et al. (Wed,) studied this question.