Efficacy of a Comprehensive Binary Classification Model Using a Deep Convolutional Neural Network for Wireless Capsule Endoscopy
Abstract Manual reading of capsule endoscopy (CE) video is a time-consuming process in diagnosing small bowel diseases. Although many algorithms have been introduced, multi-diagnosis has not been sufficiently validated. They are promising but still premature to be used in clinical practice. Therefore, we developed a practical binary classification model and tested it with unseen cases.400,000 CE images were randomly selected from 84 cases. Among them, 240,000 were used to train an algorithm to categorize images binarily. The remaining images were utilized for validation and internal testing. The algorithm was externally tested with 256,591 unseen images.Diagnostic accuracy was 98.067% when the trained model was applied to the validation set. It was 97.946% when applied to images for internal testing. When the model was applied to a dataset provided by an independent hospital not participated during training, its accuracy was 85.470%. The area under the curve was 0.922.Our binary classification model showed excellent internal test results, and when tested in unseen external cases, misreadings were slightly increased while judging ‘insignificant’ images containing ambiguous substances. When we can get over this problem, CNN-based binary classification will become the most promising candidates for developing clinically ready computer-aided reading methods.