Diagnostic performance of real-time characterization in artificial intelligence-assisted colonoscopy

BACKGROUND AND STUDY AIMS: Because most colorectal polyps are diminutive and carry minimal cancer risk, artificial intelligence (AI) might enable diagnostic strategies such as leave-in-situ and resect-and-discard, provided it meets predefined quality benchmarks. We assessed diagnostic performance of a computer-aided polyp characterization system (CADx) in predicting histology of diminutive rectosigmoid polyps during colonoscopy in a Danish population.

PATIENTS AND METHODS: We conducted a prospective diagnostic accuracy study across four endoscopy centers. Adults referred for colonoscopy due to a positive fecal immunochemical test (FIT), surveillance, or other diagnostic indications were included. Patients with inadequate bowel preparation were excluded. Histopathology served as the reference standard and all polyps were categorized as adenomas or non-adenomas. Several performance metrics are reported.

RESULTS: We included 278 patients and a total of 772 polypectomies for analysis. For diminutive rectosigmoid polyps (n = 184), the CADx-system achieved a sensitivity of 93% (95% confidence interval [CI] 87%-97%) and a specificity of 33% (95% CI 22%-46%). Positive and negative predictive values were 70% (95% CI 62%-77%) and 74% (95% CI 55%-88%), respectively, with an overall accuracy of 71% (95% CI 64%-77%) and a diagnostic odds ratio of 6.69.

CONCLUSIONS: The CADx-system demonstrated high sensitivity, but significantly lower specificity compared with prior studies, driven by a high false-positive rate. Inconsistent findings across studies highlight the challenges in standardizing AI-based characterization systems. Our results indicates that CADx is a promising future adjunct to colonoscopy, but its current performance is insufficient for reliable implementation in resect-and-discard or leave-in-situ strategies.