P282 Artificial intelligence for real-time optical diagnosis of neoplastic polyps during colonoscopy

Background

Current colonoscopy practice allows endoscopists to leave tiny hyperplastic polyps in-situ in the rectosigmoid colon if endoscopic diagnosis is done with a high level of accuracy and confidence. Artificial intelligence (AI) is expected to further facilitate this process, however, there are no studies that have evaluated the additional value of using AI in polyp assessment.

Methods

This is a prospective, open-label, comparative, multicenter trial. We enrolled patients scheduled for screening, surveillance, diagnostic, or therapeutic colonoscopy at three hospitals in Norway, the United Kingdom and Japan. All patients with colorectal polyps ≤ 5 mm in rectosigmoid colon underwent visual inspection to assess polyp histology (adenoma vs non-adenoma). This assessment was done firstly by endoscopists alone followed by an inspection with the aid of the AI device for polyp classification. The primary endpoint was an increased sensitivity to differentiate adenomas with the use of AI. We also evaluated the specificity to differentiate adenomas and confidence level of endoscopists’ prediction as secondary outcome measures. McNemar’s test was used for the analyses.

Results

Out of the 1,289 recruited participants, we enrolled 531 participants (64% men with average age 67 years, 36% women with average age 65 years) who had colorectal polyps ≤ 5 mm in rectosigmoid colon. A total of 894 eligible polyps (359 adenomas and 535 non-adenomas) of these patients were assessed through visual inspection. There is no significant difference in the sensitivity to differentiate adenomas between endoscopists’ inspection alone and inspection with the aid of AI (88.9% [95% confidence interval, 85.1-91.9] versus 90.5% [87.0-93.4]; Absolute difference 1.7% (95%CI: -1.4-4.7), P=0.33). The specificity to identify adenomas increased with the use of AI (83.6% [80.1-86.6] versus 86.4% [83.2-89.1]; absolute difference 2.8% [95% CI, 0.2-5.4], P=0.03). The confidence level of endoscopists’ prediction also increased from 65.7% to 86.9%.

Conclusions

Use of AI for polyp classification did not improve the sensitivity to differentiate adenomas in the rectosigmoid colon. However, higher specificity and confidence level in polyp classification under the use of AI may contribute to broader use of optical assessment of polyps and reduction of unnecessary polyp removal and the associated costs.