In Vivo Computer-aided Prediction of Polyp Histology on White Light Colonoscopy

Overview

Our group, prior to the present study, developed a handcrafted predictive model based on the extraction of surface patterns (textons) with a diagnostic accuracy of over 90%24. This method was validated in a small dataset containing only high-quality images. Artificial intelligence is expected to improve the accuracy of colorectal polyp optical diagnosis. We propose a hybrid approach combining a Deep learning (DL) system with polyp features indicated by clinicians (HybridAI). A pilot in vivo experiment will carried out.

Optical diagnosis aims to predict the histology of a polyp based on its endoscopic features. This practice could avoid histopathological analysis and reduce the derived costs. Under this premise, the American Society of Gastrointestinal Endoscopy (ASGE), in its Preservation and Incorporation of Valuable endoscopic Innovations (PIVI) statement, established a diagnostic threshold for real-time endoscopic assessment of diminutive polyps. The rationale for its implementation is that the prevalence of advanced histology in polyps \< 5mm is very low (0.5%). Several studies have demonstrated that optical diagnosis of small polyps is safe and feasible in clinical practice and comparable to the current gold standard, histopathology. However, the accuracy of optical diagnosis has been shown to be insufficient in community-based practices or in non-expert hands and the diagnosis is even more difficult in diminutive polyps \< 3 mm in which the discrepancy between the endoscopic and pathological diagnosis is about 15%. Artificial Intelligence (AI) has emerged as a help tool for polyp characterization. Aiming to improve optical diagnosis using AI methods, we propose a hybrid approach that combines DL with characteristics of polyps manually indicated by endoscopists (HybridAI).

Conditions