Polyp Detection at Colonoscopy – is AI better than 2 I’s?
Review by Prof. Eamonn Quigley (USA)
Study Summary
Adenoma detection rate is the Holy Grail when it comes to the assessment of colonoscopy quality in colon cancer screening programs and a variety of endoscope attachments, colonoscopist maneuvers and imaging modifications have been recommended to optimize polyp yield. Inevitably, artificial intelligence (AI) has now been applied to polyp detection based on algorithms trained to predict histologic features characteristic of adenomas. Two papers in the same edition of Annals of Internal Medicine provide updates on the performance of AI in colonoscopy. The first of these, a multicenter study involving 6 centers, 49 endoscopists and 1,252 consecutive patients across the US compared the performance of the endoscopist in the histologic diagnosis of polyps with and without assistance from one particular AI system.1 Using white light followed, if needed, by narrow band imaging (NBI) the endoscopist made a prediction of whether a diminutive polyp (≤ 5 mm diameter) was an adenoma or not. Once this decision had been made the AI system was activated and its prediction recorded. Polyp histology provided confirmatory diagnosis. Sensitivity at 90.8% and 90.7% was similar but specificity higher at 64.7% and 59.5%, for AI vs non-AI approaches, respectively. This increase in specificity was considered small of minimal clinical impact and suboptimal.
The second paper performed a systematic review and meta-analysis on the same issue based on 10 studies including 3,620 patients with 4,103 small rectosigmoid polyps and again found no difference between AI-assisted and the non-AI approach in accuracy in predicting polyp histology.2 Overall, AI-assistance showed a sensitivity and specificity for predicting neoplastic change of 87% and 89%, respectively. The proportions of polyps that were predicted to be non-neoplastic (mostly hyperplastic, one assumes) were 58% and 55% for AI vs non-AI.
Commentary
These studies set out not to detect the accuracy of AI assistance in detecting polyps in the first place, but rather to define whether AI enhanced the accuracy of prediction of small polyp histology over that provided by white light and NBI. The very practical implication here is that if one can be certain that a polyp is non-neoplastic one can choose to ignore it (the so called “do not resect” strategy) or, if a diminutive lesion is deemed to be neoplastic (i.e. an adenoma), one could resect and discard it (the “resect and discard” strategy) with resultant significant reductions in pathology costs. As borne out in the meta-analysis, the issue of separating adenomas from non-adenomas is most relevant to the rectosigmoid region where hyperplastic polyps are so common (almost 50% of all polyps in this area in the Rex et al study and 66% of all polyps in the Hassan et al study). What are the takeaway messages? First, general endoscopists (at least in the setting of a clinical study) are very good at identifying adenomatous polyps (as reflected by a low false negative rate) but may overall call adenomas (relatively high false positive rate). Second, the AI system tested in these studies and, it must be emphasized, only this particular system did not do much better than human eyes (I’s). I am sure that there will be further developments in the area.
Citation
1Rex DK, et al. Artificial Intelligence for real-time prediction of the histology of colorectal polyps by general endoscopists. Ann Intern Med 2024;177:911-8.
2Hassan C, et al. Computer-aided diagnosis for leaving colorectal polyps in situ. A systematic review and meta-analysis. Ann Intern Med 2024;177:919-28.
View past News You Can Use commentaries
