Assessing quality indicators of screening colonoscopy using an artificial intelligence system

The efficacy of screening colonoscopy is contingent upon adherence to established quality indicators, including cecal intubation (CIR) and adenoma detection rate (ADR). Artificial intelligence (AI) presents a viable means of automating the assessment of these metrics. Objective. To evaluate the feasibility of an AI system for automated assessment of CIR and ADR with preliminary estimation of invasive growth. Materials and Methods. A prospective single-center study was conducted using photo and video materials from colonoscopies performed at the Yaroslavl Regional Clinical Oncology Hospital. High-quality visualization of the colonic mucosa was ensured through a bowel preparation regimen using low-volume trisulfate solution (Eziclen) in split dose combined with Meteospazmil. The AI model was trained on endoscopic images of the appendiceal orifice, colorectal neoplasms, and normal mucosa. Results. The AI algorithm was integrated into the endoscopy processor. It provided real-time visual assistance, displaying a blue perimeter frame upon cecal intubation confirmation. The model demonstrated high performance: AUC 0.97 (F1=0.85) on the validation dataset and AUC 0.95 (F1=0.90) on the test dataset. The mean processing time per frame (29 ms) enabled real-time operationFor detected adenomas, the system annotated the lesion with a yellow bounding box and provided a probability for superficial versus deep submucosal invasion. The high detection rate was facilitated by optimal bowel preparation quality. Conclusion. This AI system provides objective, real-time quality assurance by verifying procedural completeness and detecting adenomas. Its application facilitates standardized auditing, automated documentation, and enhanced endoscopic training, addressing the growing demand on endoscopy services.