A pilot evaluation of the artificial intelligence system CAD-EYE to optically characterise lesions in inflammatory bowel disease surveillance

Abstract

Background: Patients with inflammatory bowel disease (IBD) have an increased risk of colorectal cancer. Endoscopic surveillance is recommended but is challenging due to the presence of active inflammation, flat dysplasia and inflammatory pseudopolyposis. CAD-EYE, an artificial intelligence (AI) powered endoscopic module by FUJIFILM, optically characterises lesions in real time. The aim of this study was to evaluate the accuracy of CAD-EYE in IBD surveillance. Methods: Ninety-seven lesions were identified from 38 patients with IBD, undergoing surveillance at a single centre. Non-magnified, still images of lesions identified during the procedure were captured, followed by characterisation by CAD-EYE as neoplastic or hyperplastic (non-neoplastic) prior to resection. Inflammatory pseudopolyps were imaged and only resected based on the physician’s discretion. Images of lesions identified were characterised by two expert IBD clinicians. The accuracy of CAD-EYE was assessed for all lesions (composite of histology for resected lesions and expert-verified non-resected pseudopolyps). For the resected lesions, the accuracy of CAD-EYE was compared to expert characterisation. Results: CAD-EYE correctly characterised 92/97 lesions (94.8%) with a sensitivity of 80.0%, specificity of 97.6%, positive predictive value of 85.7% and negative predictive value of 96.4% for neoplastic lesions. For resected lesions, diagnostic accuracy was similar between CAD-EYE (93.0%) and expert characterisation (88.4%), with no statistically significant differences in sensitivity. Conclusion: CAD-EYE demonstrated its utility in IBD surveillance with excellent accuracy in the characterisation of lesions, including inflammatory pseudopolyps. Larger studies are required to confirm these findings, particularly for flat dysplasia.