Potential of artificial intelligence‐assisted colonoscopy using an endocytoscope (with video)

Colonoscopic polypectomy for all adenomatous polyps is considered to contribute to the reduction of both the incidence and mortality of colorectal cancers. However, a substantial number of unnecessary polypectomies are carried out for non-neoplastic polyps because of endoscopists’ misdiagnoses, resulting in considerable financial concerns.1 Thus, increasing attention is being given to artificial intelligence (AI) designed to assist accurate optical biopsy of colorectal polyps.2, 3 Artificial intelligence-assisted endocytoscopy is a recently developed technique that supports optical biopsy. In this AI system, images acquired by an endocytoscope (CF-H290ECI; Olympus Corp., Tokyo, Japan), which is an ultra-magnifying endoscope allowing cellular observation, are analyzed in real time during colonoscopy. The predicted histology of the target polyps is output as either neoplastic or non-neoplastic with a probability (Video S1).4 The algorithm of this AI system is based on automatic image feature extraction followed by machine learning. We conducted a retrospective analysis to evaluate the performance of AI-assisted endocytoscopy in the clinical setting. In total, 144 diminutive polyps (≤5 mm) were assessed using this AI system from January to May 2017 at Showa University Northern Yokohama Hospital. The diagnostic performance of differentiation between neoplastic and non-neoplastic lesions was calculated with reference to the pathology of the resected specimen as the gold standard. Table 1 shows that the sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of the AI system were 98%, 71%, 81%, 67%, and 98%, respectively. The AI model showed an excellent sensitivity and negative predictive value that potentially met the threshold required for optical biopsy.5 To confirm this result with a higher level of evidence, a prospective trial evaluating this AI model that is powered with a higher number of learning images is ongoing (UMIN000027360). We thank Dr Masashi Misawa from Showa University and Dr Hayato Itoh from Nagoya University for their invaluable support. We also thank Angela Morben, DVM, ELS, from Edanz Group (www.edanzediting.com/ac) for editing a draft of this manuscript. This research was supported by Grants-in-Aid for Scientific Research (Number 17H05305) from the Japan Society for the Promotion of Science. Author S.K. received endocytoscope from (Olympus Corp.). Author K.M. received research grants from (Cybernet System Corp.) Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.