Artificial Intelligence automated assessment of colonoscopy key performance measures

Aims Colorectal cancer is the third most frequent cancer and is an important cause of morbidity and mortality. A high-quality colonoscopy is the most, and only, effective method for the screening and early treatment of this disease. The key performance measures (KPM) that define a high-quality colonoscopy include the completeness of the examination, adequate intestinal cleansing, and retraction technique (evaluated through the retraction time) as established by the ESGE. The use of a single AI software can potentially make simpler, faster, and more objective the evaluation of the aforesaid KPM.

Methods A total of 127 fully anonymized videos of screening or diagnostic colonoscopies (excluding subjects undergoing surgery or suffering from IBD) were retrospectively annotated both in the intubation and retraction phases. At each stage of the colonoscopy the different colic segments (left, transverse and right colon, terminal ileum), the main anatomical landmarks (appendicular orifice, cecum, ileocaecal valve, ascending colon, hepatic flexure, transverse colon, splenic flexure, descending colon, sigma, rectum) and any endoscopic instruments used (tongs, loops, etc.) have been noted. Each entry was subsequently validated by two endoscopist experts with more than 10 years of experience. Bowel preparation was evaluated using the Boston Bowel Preparation Score during the withdrawal phase, as per guidelines. Retraction time has been estimated by subtracting the time of intubation of the terminal ileum and the time of using any endoscopic devices from the time elapsed between the intubation of the cecum and the end of the procedure.

A deep learning model for the assessment of cecal intubation, bowel preparation score and withdrawal time was generated using 70 videos for training, 28 for validation and 29 for testing.

Results The accuracy of cecal intubation was around 75%, with 79% for the appendiceal orifice. For the assessment of the bowel cleansing a regression model was developed with excellent results in all the colonic segments. The software was able to identify the endoscopic devices with a 72.3% accuracy. This has guaranteed an accurate estimate of the retraction times with an absolute error of 12s±4s.

Conclusions The preliminary results of our project support the idea that artificial intelligence is an important tool to evaluate objectively and quickly the main KPM of colonoscopy, with a view to a fully automated reporting of the procedure.

Publication History

Article published online:
15 April 2024