Robust Prototypical Networks for Small-Intestine Polyp Recognition in Wireless Capsule Endoscopy Images

Wireless capsule endoscopy (WCE) is a gastrointestinal examination technology, which can help find the polyps in small bowel noninvasively. The computer-aided polyp recognition systems based on deep learning require large amounts of manually annotated data, which is often unavailable for WCE images. Meanwhile, there is a serious imbalance between normal and polypoid samples in WCE image database. We proposed a few-shot learning method of automatic polyp recognition under the circumstance of absolute lack of data, named Robust Prototypical Networks (RPNs), and RPNs made adjustments to polyp position changing. Trained with an imbalanced polypoid WCE image dataset and a polypoid colon endoscopy (CE) image dataset, RPNs can extract their common features by introducing a multi-task learning scheme, separating-diffusing, to overcome imbalance problem. RPNs outperforms the previous work, and the best average AUC-PR score is 0.87.