Second glance framework (secG): enhanced ulcer detection with deep learning on a large wireless capsule endoscopy dataset

Abstract

Wireless Capsule Endoscopy (WCE) enables physicians to examine gastrointestinal (GI) tract without surgery. It has become a widely used diagnostic technique while the huge image data brings heavy burden to doctors. As a result, computer-aided diagnosis systems that can assist doctors as a second observer gain great research interest. In this paper, we aim to demonstrate the feasibility of deep learning for lesion recognition. We propose a Second Glance framework for ulcer detection and verified its effectiveness and robustness on a large ulcer WCE dataset (largest one to our knowledge for this problem) which consists of 1,504 independent WCE videos. The performance of our method is compared with off-the-shelf detection frameworks. Our framework achieves the best ROC-AUC of 0.9235 and outperforms the results of RetinaNet (0.8901), Faster-RCNN(0.9038) and SSD-300 (0.8355).