Polyp Detection and False Alarm Rates by Computer-Aided Analysis of Videos of Withdrawal Phase of Colonoscopy in a Randomized Controlled Trial of Water Exchange Versus Air Insufflation

Worldwide colorectal cancer (CRC) is the second most common cancer in women and the third in men. Early detection and removal of the colon polyps (cancer precursors) reduce the incidence of CRC. However, interval colon cancers still occur within 3-5 years after colonoscopy among patients of colonoscopists with low adenoma detection rate (ADR), defined as the proportion of patients with at least one adenoma. ADR was widely variable, suggesting that some adenomas were missed. Twenty six percent of adenomas were missed during tandem examination reported in a recent meta-analysis. Missed adenomas accounted for about 58% of interval cancers. Adenomas are more likely to be missed in the right colon than in other segments because of their flat morphology and hiding behind the accentuated folds and curvatures. Innovations in colonoscopy to increase ADR and decrease adenoma miss rate (AMR) hold the potential to reduce interval cancers.

The consensus statements in a recent modified Delphi review confirmed water exchange (WE) as a standardized insertion method produced less insertion pain, better bowel cleanliness and higher ADR than gas insufflation. It is characterized by infusing water to guide insertion in an airless lumen and almost simultaneous suctioning of the infused water during insertion, aiming at near-complete removal of the infused water and debris upon cecal intubation. Although an RCT with tandem examination showed WE significantly decreased right colon adenoma miss rate (rAMR) compared with CO2 insufflation (18.0% [33/183] vs. 34.6% [62/179], P = 0.0025), a considerable percentage of polyps in the right colon were still overlooked.

In recent years, the field of machine learning and artificial intelligence has made remarkable progress, and an increasing number of publications showed improved polyp detection rate (PDR) and ADR using computer-aided detection (CADe). CADe can detect polyps overlooked by the colonoscopist due to human limitations of inattention or inexperience. However, one major drawback of current CADe systems is false alarms (FAs), or false positives (FPs). Usually triggered by bubbles and fecal debris, FAs might distract the endoscopists with potential unfavorable effect on ADR. One study reported a FP rate of up to 60%.

In an overview on applying deep learning algorithms and WE in colonoscopy to improve adenoma detection, the authors noted that WE could enhance the performance of artificial intelligence (CADe) by improving bowel cleanliness and thus the exposure of polyps. In a follow-up review, the authors reported that artificial intelligence might mitigate operator-dependent factors that limited the potential of WE, while WE might provide the platform to optimize the performance of artificial intelligence by increasing bowel cleanliness and improving visualization, Therefore, the strengths of WE and artificial intelligence may complement the weaknesses of each other to maximize adenoma detection.

One of our recently completed studies compared right colon ADR evaluated by a blinded endoscopist using either air insufflation or WE for insertion, with all the colonoscopies video recorded (NCT02737514). We developed and applied a CADe system to detect the polyps in the videos. The current report is a proof of principle study to test the hypothesis that WE could yield a significantly higher additional PDR (APDR) and reduce false alarms rate (FAR) as compared to air insufflation in the right colon.