Automated diagnosis system for endocytoscopy with narrow-band imaging provides accurate characterization for colorectal lesions

Background and Aim:Endocytoscopy (EC) can evaluate not only cell nuclei but microvessels in vivo. We reported the efficacy of endocytoscopic vascular (ECV) pattern observed by EC with narrow-band imaging for diagnosing colorectal lesion. However, interpretation of ECV is difficult for novices. The aim of this study was to develop the computer aided diagnosis (CAD) system for ECV.Methods:The algorithm of CAD system for ECV (ECV-CAD) was programmed based on texture analysis ECV-CAD provided 2-class diagnosis (neoplastic or non-neoplastic) with its probability. To validate the diagnostic ability of ECV-CAD, 173 randomly selected EC images (non-neoplasm, 49; neoplasm, 124) were evaluated by ECV-CAD. To compare with the ability of endoscopist, we select 4 expert endoscopists (experienced >200 EC cases), and 3 trainees (experienced <20 EC cases). EC images were randomly allocated to assessors. Assessors recorded their diagnosis (non-neoplasm or neoplasm) with its confidence level (high or low). The overall accuracy to distinguish neoplasms from non-neoplasms were calculated. Results: The overall accuracy of ECV-CAD, experts and trainees were 87.8%, 84.2% and 63.4% respectively, whereas with regard to high confidence cases these were 93.5%, 90.8% and 71.7% respectively. Conclusions:The diagnostic ability of ECV-CAD was better than that of trainee and comparable to that of experts. Thus ECV-CAD could be a powerful decision making tool for less-experienced endoscopist.