Development and Validation of a Computer-Aided Diagnostic System for Colorectal Lesions Based on the JNET Classification: Supporting the "Resect and Discard" Strategy

Abstract

Objectives: Although the Japan NBI Expert Team (JNET) classification for colorectal lesions is clinically useful, standardization remains challenging owing to endoscopist variabilities. Therefore, we developed a computer-aided diagnosis (CADx) system to support JNET diagnoses, potentially facilitating a "Resect and Discard" strategy. The CADx system's diagnostic performance and its potential to support the "Resect and Discard" strategy were compared with those of endoscopists of varying experience.

Methods: Forty-three patients (60 lesions) who underwent CADx system colonoscopies at Hiroshima University Hospital between September 2022 and July 2024 were evaluated. Endoscopists were grouped by experience: beginners (100–500 cases, n=4), intermediates (1,000–2,000 cases, n=5), and experts (>2,000 cases, n=3). In Study 1, JNET classification concordance between the CADx system and beginner/intermediate endoscopists was compared, using expert diagnosis as the gold standard. Study 2 included 38 <10 mm lesions, with histopathology as the gold standard. Lesions excluding advanced neoplasia were classified as "Resect and Discard". The CADx system and endoscopists were evaluated for identifying JNET Type 1 and 2A lesions as non-advanced neoplasia.

Results: In Study 1, CADx-expert concordance was 83.3% (Kappa=0.75)—significantly higher than beginners (66.7%) and outperforming intermediates (78.0%). In Study 2, the CADx system's accuracy was 92.1%—significantly higher than beginners (70.4%) and outperforming intermediates (82.6%) and experts (78.9%). The CADx system's advanced neoplasia negative predictive value was 100%.

Conclusions: The CADx system showed high concordance with expert diagnoses. For lesions <10 mm, its high advanced neoplasia diagnostic accuracy indicates its potential utility in supporting the "Resect and Discard" strategy.