Combination of white-light imaging-based and narrow-band imaging-based artificial intelligence models during colonoscopy in patients with ulcerative colitis

Abstract

Background and aims: The long-term treat-to-target (T2T) approach in ulcerative colitis (UC) aims for endoscopic remission, but variability among endoscopists and a lack of precision in relapse prediction both limit its clinical usefulness. A recently reported white-light imaging (WLI) artificial intelligence (AI) model helps standardize diagnosis, although challenges remain. Therefore, we attempted to combine a narrow-band imaging (NBI) AI model with the WLI AI model to determine whether these challenges can be overcome.

Methods: This post hoc analysis of a prospective study evaluated the efficacy of combining AI-assisted WLI and NBI models in predicting clinical relapse in patients with UC over a 12-month follow-up period. A total of 102 patients with UC in clinical remission were included, and the combined AI models were used during colonoscopy to assess relapse risk.

Results: The study found that within the same AI-based Mayo endoscopic subscore category, patients with vascular activity were more likely to experience clinical relapse than those with vascular healing. Compared with the WLI model alone, the specificity of the combined method significantly increased from 42.2% (95% confidence interval [CI]: 32.1%-52.9%) to 61.5% (95% CI: 50.7%-71.2%) (P = .013) with its sensitivity being maintained.

Conclusions: The sequential use of WLI and NBI AI models can provide better stratification of relapse risk compared with using either model alone, offering a more accurate and personalized approach to treatment intensification. This dual-model AI approach aligns with the T2T approach in UC management.

Keywords: convolution neural network; mucosal healing; treat-to-target.