Endoscopy 2022; 54(05): 473-474
DOI: 10.1055/a-1669-8814

Missed lesions and artificial intelligence during colonoscopy: the tireless working expert in the room

Referring to Zippelius et al. p. 465–472
Timo Rath
Department of Medicine, Ludwig Demling Endoscopy Center of Excellence, University Hospital Erlangen, Friedrich-Alexander-University, Erlangen-Nuernberg, Germany
› Author Affiliations

Although colonoscopy is generally considered to be the most accurate screening modality for detecting polyps, a substantial number of polyps are still missed. Recent data show a worrying proportion of missed lesions during screening or surveillance colonoscopy. In a meta-analysis of 43 publications and more than 15 000 tandem colonoscopies, miss rates of 26 % for adenomas, 9 % for advanced adenomas, and 27 % for serrated polyps were calculated [1]. Needless to say, missed lesions, especially missed advanced adenomas, may have the potential to develop into cancer, and it is thought that at least 50 % of all interval colorectal cancers arise from lesions missed during colonoscopy [2].

“Computer-aided detection systems can be subject to bias, with preferential detection of those lesions that were contained within the training set.”

This begs the question of why we miss lesions during colonoscopy. Clearly, several factors contribute to missed lesions and the past decade has shown that use of auxiliary devices and techniques that increase mucosal exposure, such as distal attachment devices (caps, endorings, endocuff), balloon-assisted colonoscopy, or behind-fold visualizing technologies, can increase detection or reduce miss rates [3] [4] [5].

However, even with optimized mucosal exposure and a polyp being within the visual field, a lesion can still be missed by the endoscopist and several hypotheses have been offered to explain this. The first involves perceptual phenomena called “inattentional blindness,” where an observer experiences a lapse of focus (e. g. distraction), and “change blindness,” where a change in a visual stimulus is introduced (e. g. a polyp entering the visual field) but goes unnoticed by the endoscopist. Fatigue or emotional states may also contribute to missed lesions.

Awareness of these factors and the scale of the problem has led to the development of artificial intelligence (AI) systems to assist the endoscopist in polyp detection. Such computer-aided detection (CADe) systems can detect, highlight, and alert the endoscopist, often using an acoustic signal, to the presence of a polyp within the visual field. Thus, lesions that might otherwise be missed, for whatever reason, will be detected constantly and consistently during colonoscopy examinations.

During the past 2 years, several studies, including randomized controlled trails (RCTs), have analyzed detection rates using different CADe systems, so it is tempting to question why we need to focus on miss rates under AI, especially in studies that are single centered and nonrandomized. However, the adenoma miss rate (AMR) serves a valuable function. First, owing to the nature of tandem studies, AMR is a more direct and representative parameter than the adenoma detection rate (ADR) for analysis of individual endoscopist performance, either with or without CADe. Second, the AMR, and especially the miss rate of advanced adenomas, directly translates into clinical relevance. Advanced adenomas are estimated to progress to colorectal cancer at the rate of 1 % per year, and based on the prevalence of advanced adenomas and the number of screening and surveillance colonoscopies performed in the USA, it is estimated that 1700–9200 interval cancers resulting from missed advanced adenomas will occur within the next 10 years [6] [7].

Given this association and the fact that various CADe systems have entered the market in the past 2 years, the scarcity of data on the influence of AI on AMR is surprising. The elegant prospective trial by Zippelius et al. [8] in this issue of Endoscopy is one of the first real-world datasets on miss rates with the use of AI. In this trial, 150 patients undergoing routine colonoscopy at a single endoscopy center in Germany were examined in parallel by an endoscopist and the commercially available CADe system GI Genius (Medtronic, Minneapolis, Minnesota, USA) using two opposing screens, with the AI screen observed by a second observer and not visible to the endoscopist. Using this approach, the authors found neither significant nor clinically relevant differences in the AMR between the CADe system and conventional colonoscopy performed by experienced endoscopists. Furthermore, the polyp miss rate as a secondary endpoint was virtually identical between the CADe system and conventional colonoscopy. Importantly, the ADR, which was already high during conventional colonoscopy (52.0 %), did not increase with the use of CADe (50.7 %).

Despite these “negative” results, the study by Zippelius et al. adds an important piece to the puzzle when all aspects are considered. Although the study did not show any decrease in AMR with the use of CADe, it did demonstrate that miss rates under AI were similarly low to those of experienced endoscopists participating in the study. Of note, pooled AMR for all endoscopists was 2 %, and almost 60 % of the endoscopists exhibited an ADR > 50 %, indicating a high level of colonoscopy expertise in this study.

In 2020, Wang et al. analyzed the influence of a different CADe system on AMR in a randomized setting and found significantly lower polyp and adenoma miss rates with CADe colonoscopy compared with routine colonoscopy [9]. While these results seem contradictory to those presented by Zippelius et al., it is worth pointing out that the AMR of 40 % in the Wang et al., RCT was remarkably high for conventional colonoscopy, compared with the very low AMR of 2 % in the Zippelius et al. study. Furthermore, any AI system can only be as good as the data used to train it. CADe systems, including the GI Genius system used in the current study, are usually developed using training sets comprising images that have been manually annotated by experts. CADe systems can therefore be subject to bias, with preferential detection of those lesions that were contained within the training set. With such potential variation in the images used to train CADe systems, it is clear that different systems cannot be readily compared. Finally, it is known that the prevalence of colorectal adenomas is lower in mainland China compared with Western countries and other differences, including genetic, dietary or lifestyle factors, are likely to have also contributed to different outcomes between these two studies. However, even with different effects of AI on AMR between these two studies, it appears to be clear that use of AI has at least a similar AMR compared with experienced endoscopists.

The use of AI during colonoscopy has the advantage of offering an additional pair of “computer eyes” that are tirelessly examining and detecting lesions, and the current data by Zippelius et al. impressively and elegantly show that AI systems are comparable to experienced endoscopists in terms of lesions detected (and missed). In light of these fascinating results, it is tempting to speculate that application of CADe systems in routine colonoscopy could decrease performance variability between endoscopists, increase detection, and reduce the number of lesions missed by less experienced endoscopists. With this, AI is here, and it is here to stay.

Publication History

Publication Date:
14 December 2021 (online)

© 2021. Thieme. All rights reserved.

Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany