Date created: | Last Updated:

Category: Project

Description: Background and Rationale: Real-time in vivo Optical Characterization (OC) of colorectal polyps during colonoscopy has recently gained traction as means to guide ‘resect or discard’ strategies, potentially leading to decreased complications, physician burden, and medical costs. However, this characterization task is challenging and suffers from significant inter- and intra-observer variability, often resulting in community endoscopists performances below accepted thresholds. Several learning-based models have been recently proposed to aid this characterization task. Still, none of them can work on real-time regular colonoscopy videos, as they have either been developed for the characterization of still images or they require human intervention to cope with variations of image quality and polyp appearance. In this study, we intend to assess the performances of a deep learning-based module for real-time characterization of colorectal polyp's histology that can be coupled with any automated polyp detection model in colonoscopy videos (GI Genius CADx). The module consists of a convolutional neural network model which classifies each detected polyp in a single video frame as adenomatous or non-adenomatous polyp and tracks it across all the video frames. This algorithm produces stable, spatio-temporally weighted decisions that are displayed real-time on each frame. The AI model can also abstain from predicting the polyp histology if insufficient confidence was accumulated. The GI Genius CADx was developed to help endoscopists in their clinical practices for polyp characterization. In order to assess GI Genius CADx performances in terms of prediction accuracy, the output of GI Genius CADx (adenoma/non-adenoma/no-prediction) will be evaluated against the histopathology reference standard (ground truth). The performances of GI Genius CADx will be compared against the prediction of the same lesions performed by a pool of endoscopists reviewing the video recording of the procedures blind to the histology results. Objective: To prospectively evaluate if GI Genius CADx accuracy in the automated OC of colorectal polyps in white light is non-inferior to the accuracy of expert endoscopists performing OC (supported by virtual chromoendoscopy), having histopathology as a reference standard. Furthermore, to evaluate if GI Genius CADx accuracy in the automated OC of colorectal polyps in white light is superior to the accuracy of non-expert endoscopists performing OC (supported by virtual chromoendoscopy), having histopathology as a reference standard. Furthermore, to evaluate if GI Genius CADx accuracy in the automated OC of colorectal polyps in white light is non-inferior to GI Genius CADx accuracy in virtual chromoendoscopy, having histopathology as a reference standard.