Complementary Roles of Artificial Intelligence and Disease-Specific Optical Diagnosis During Ulcerative Colitis Surveillance: A Prospective Tandem Colonoscopy Study.

Objectives

Linked-colour imaging/blue-laser imaging (LCI/BLI) and CAD-EYE artificial intelligence have been developed for sporadic colorectal neoplasia, but their role in ulcerative colitis (UC) surveillance remains uncertain. We evaluated CAD-EYE-assisted detection and optical characterization using conventional and disease-specific classifications in UC.

Methods

In this prospective tandem study, patients with UC undergoing surveillance colonoscopy were examined sequentially using white-light imaging (WLI), LCI, and LCI with CAD-EYE. Lesions were characterized using BLI with Kudo, NICE, and Kudo-IBD classifications, followed by CAD-EYE characterization. Detection and characterization performance were assessed using miss rates and diagnostic accuracy.

Results

Among 82 patients, 281 lesions, including 22 neoplastic lesions, were identified. The lesion miss rate decreased from 5.7% with WLI to 2.5% with LCI, while no lesions were missed during the CAD-EYE-assisted withdrawal. However, the fixed examination sequence and absence of prospectively recorded withdrawal times precluded isolation of CAD-EYE's incremental contribution. The neoplasia miss rate was numerically lower with CAD-EYE than with WLI and LCI, although not significantly. Kudo-IBD showed the highest diagnostic performance, with sensitivity, specificity, and positive and negative predictive values of 90.9%, 85.3%, 34.5%, and 99.1%, respectively. Accuracy was significantly higher with KUDO-IBD than with all other methods (all p < 0.001).

Conclusions

In this sequential protocol, CAD-EYE-assisted imaging was associated with complete observed lesion detection, whereas Kudo-IBD provided superior optical characterization. These findings support complementary AI-assisted detection and disease-specific optical assessment. Larger studies with independent examination sequences and prospectively recorded withdrawal times, together with external validation of Kudo-IBD, are needed.