1/9 Quick synopsis of our recent double-blind randomized study on computer-aided polyp detection in @LancetGastroHep, which can be found here: thelancet.com/journals/langa… 
2/9 Previous work by our team and others have shown that #AI polyp detection can improve ADR. A major question has been whether use of on-screen computer-aided detection merely increases the vigilance of the endoscopist (vs. specifically helping find more polyps).
3/9 We enrolled 1046 pts. We tried to address operational bias by blinding the endoscopist to whether or not CADe was being used. Blinding is hard. This required development of a 'sham' CADe system by @WisionAI that alerted only to 'polyp-like' findings (bubbles, folds etc).
4/9: While endoscopist performed colo, a 2nd observer viewed real or sham CADe system & used laser pointer on the main endo screen to point out detected areas. Both were blinded to which system was being used. Endoscopist made final call re: whether the area was a polyp or not.
5/9 Key results: ADR in CADe arm was 34% , and ADR in sham system arm was 28% (p=0.03). The vast majority of polyps detected were sessile and small. This latter point should be subject of debate- is it clinically important to find additional tiny adenomas?
6/9 To our knowledge this is the first double-blind RCT for #artificialintelligence in clinical medicine (although @EricTopol or others might know of an #AI study that we are not aware of!)
7/9 Importantly, as with our original study, this was performed in Chengdu, China. There is no national CRC screening in China, and therefore this study was performed on a population of patients undergoing colonoscopy for a variety of indications, with a mean age of 49.
8/9 Generalizability is therefore is a key weakness of our study, and makes it difficult to compare ADR in this study population vs the #colorectalcancer screening populations that are more commonly studied.
9/9 Great editorial by @DrOmerAhmad. To answer question posed by title: 'Deep learning for colorectal polyp detection: time for clinical implementation". My view is the answer is 'yes', but with ongoing careful study of whether benefits are realized in real clinical practice.
