O53 AI-assisted detection-characterization-sizing of colorectal polyps. Can AI support non-expert endoscopists to achieve PIVI thresholds?. (19th June 2022)

• Record Type: Journal Article
• Authors: Abdelrahim, Mohamed; Takoh, Kimiyasu; Okuno, Takayuki; Goda, Shimpei; Htet, Hein; Hamson, Joe; Aslam, Shahila; Siggens, Katie; Tanasescu, Andrea; Nair, Sujith Sasidharan; Elias, Mary; Salviato, Antonio; Mohammed, Salma; Parra-Blanco, Adolfo; Ishaq, Sauid; Antonelli, Giulio; Fraile-López, Miguel; Spadaccini, Marco; Subramaniam, Sharmila; Longcroft-Wheaton, Gaius; Alkandari, Asma; Hassan, Cesare; Repici, Alessandro; Bhandari, Pradeep
• Is Part Of: Gut. Volume 71(2022)Supplement 1
• Page Start: A30
• Page End: A31
• Publication Date: 2022-06-19
• Subjects: Gastroenterology -- Periodicals; 616.33
• DOI: 10.1136/gutjnl-2022-BSG.53
• Languages: English
• ISSNs: 0017-5749

Abstract

Abstract : Introduction: Real-time in-vivo characterisation of colorectal polyps remains limited outside expert centres. Data on AI polyp detection and characterisation is promising but accurate sizing remains the missing jigsaw piece. We aimed to study the impact of a novel AI system on non-expert endoscopists detection, characterisation and sizing of colorectal polyps compared to experts. Methods: Prospectively collected endoscopy videos from twelve centres in Europe and Japan were uploaded on a bespoke online platform. All polyps were histologically proven and sized by three experts. The AI model detects polyps and classifies them as neoplastic/non-neoplastic and diminutive/non-diminutive. We asked Six experts to detect, characterise and size polyps without AI support, and Six non-experts to detect polyps assisted by AI, and to characterise and size polyps without and then with AI. Results: 199 videos (100-polyps) were included. On polyp detection, average sensitivity and specificity of non-experts +AI compared to experts was 96.0% and 84.6% compared to 95.7% and 89.9% respectively (p>0.5). Non-experts+AI showed superior sensitivity (95.5% vs 83.3%) and NPV (90.8% vs 70.4%) of characterisation on enhanced imaging compared to non-experts alone (p<0.5). On sizing, non-experts+AI achieved accuracy and sensitivity of 84.0% and 93.6%, respectively. Experts' characterisation and sizing metrics were not significantly different from non-experts+AI. Conclusion: This interim analysis suggests our AI system may support non-experts to perform at experts' level and achieve PIVI-2 threshold (diagnose and leave). Further analysis is underway to understand the impact of the AI system on surveillance interval (PIVI-1). To our knowledge, this is the first report incorporating AI-assisted sizing with detection and characterisation.