1191 Use of Artificial Intelligence for Identification of Celiac and Vascular Lesions on Capsule Endoscopy

Abstract

INTRODUCTION: Capsule endoscopy is an important tool for noninvasive identification of gastrointestinal pathology. It requires a physician to review images which can be tedious and time consuming. The use of artificial intelligence has increased in popularity for colonoscopy for identification of polyps and vascular lesions. We aim to use computer-assisted image analysis using convolutional neural networks (CNNs) for the identification of inflammatory and vascular lesions on capsule endoscopy. METHODS: We examined a total of 2371 publicly available wireless capsule endoscopy (WCE) images obtained using MiroCam® (IntroMedic Co, Seoul, Korea) capsule endoscopes. The capsule images illustrated normal and assorted small bowel findings including polypoid, vascular and, inflammatory lesions and were notated with each finding. Images were divided into “normal” (725 images), “inflammatory” (225 images), and “vascular” (300 images) for the purposes of this analysis. Pre-processing was performed on these images. Augmentation of the image was performed by flipping and rotating the image to obtain multiple views of pathologic and normal images. The machine learning algorithm was trained on the original and augmented images, testing was only performed on original images. RESULTS: Using this machine learning architecture, a total testing accuracy of 73.7% was able to be achieved for the differentiation of normal versus inflammatory and a total testing accuracy of 70.2% was able to be achieved for the differentiation of normal versus vascular capsule endoscopy images. In a set of 2371 capsule endoscopy images, the CNN identified both vascular and inflammatory images using a 5-Fold Cross validation with an accuracy of 73.7% for normal compared to inflammatory images (Figure 1) and 70.20% for normal versus vascular images. The area under the receiver operating characteristic curve (ROC) for identification of inflammatory images was 0.70 (95% CI 0.664-0.736) compared to 0.68 (95% CI 0.646-0.714) for vascular lesions. CONCLUSION: This system demonstrates that AI can be used to find more subtle inflammatory and vascular lesions. The CNN system detected and localized polyps well within real-time constraints using an ordinary desktop machine with a contemporary graphics-processing unit. It can increase the findings of pathology and decrease time needed to review capsule endoscopy studies, but requires further validation in large multicenter trials.