Quality Assurance of Computer-Aided Detection and Diagnosis in Colonoscopy

Abstract

Recent breakthroughs in artificial intelligence (AI), specifically via its emerging sub-field “Deep Learning,” have direct implications for computer-aided detection and diagnosis (CADe/CADx) for colonoscopy. AI is expected to have at least 2 major roles in colonoscopy practice; polyp detection (CADe) and polyp characterization (CADx). CADe has the potential to decrease polyp miss rate, contributing to improving adenoma detection, whereas CADx can improve the accuracy of colorectal polyp optical diagnosis, leading to reduction of unnecessary polypectomy of non-neoplastic lesions, potential implementation of a resect and discard paradigm, and proper application of advanced resection techniques. A growing number of medical-engineering researchers are developing both, CADe and CADx systems, some of which allow real-time recognition of polyps or in vivo identification of adenomas with over 90% accuracy. However, the quality of the developed AI systems as well as that of the study designs vary significantly, hence raising some concerns regarding the generalization of the proposed AI systems. Initial studies were conducted in an exploratory or retrospective fashion using stored images and likely overestimating the results. These drawbacks potentially hinder smooth implementation of this novel technology into colonoscopy practice. The aim of this article is to review both contributions and limitations in recent machine learning based CADe/CADx colonoscopy studies and propose some principles that should underlie system development and clinical testing.