S1559 Reduction of Video Capsule Endoscopy Reading Times Using Artificial Intelligence

Abstract

Introduction: Video capsule endoscopy (VCE) is an innovation that has revolutionized care within the field of gastroenterology, but the time needed to read the studies generated has often been cited as an area for improvement. The purpose of this study is to create a machine learning model capable of significantly reducing capsule endoscopy reading times. Methods: In this study, we have trained a convolutional neural network, ResNet50, using videos from the KID dataset to confidently exclude normal images on VCE, while retaining abnormal ones. The CNN was exposed to 3 full length capsule endoscopy videos A threefold cross-validation scheme was employed whereby the model was trained using 2 of the aforementioned videos, tested on the third and this was repeated for all possible video combinations. We trained our model for 9 epochs and further improved the predictions of our model by adding rotated versions of abnormal segments to our training data. Results: Our study identified abnormal frames in 3 KID videos as outlined in the Table. We were able to reduce the video length by 47%, on average, and captured frames from 118 of the 119 abnormal segments labeled by the expert physician. We were able to develop an algorithm that successfully detected 99% of abnormal segments while reducing the reading time for a physician by over 43%. Conclusion: Our model demonstrated high levels of accuracy with significant reduction in physician reading time. Our results are reassuring and demonstrate the benefit of CNN in processing VCE images. We believe our study lays an excellent foundation for further validation in large multicenter trials. Table 1. - Description of the various characteristics of the of the 3 KID videos in addition to the reduced number of frames produced and abnormal segments detected by the ResNet50 model Video Name Total Number of Frames Number of Abnormal Segments Reduced Number of Frames by ResNet50 Number of Abnormal Segments Detected by ResNet50 KID Video 1 28480 (96 min) 22 14828 (49 min) 22 KID Video 2 117565 (391 min) 86 84672 (282 min) 86 KID Video 3 74762 (249 min) 11 27099 (90 min) 11