A Saliency-Aware Hybrid Dense Network for Bleeding Detection in Wireless Capsule Endoscopy Images

Wireless Capsule Endoscopy (WCE) has been widely used for the screening of Gastrointestinal (GI) diseases. However, manually reviewing the huge number of WCE images is time-consuming and error-prone, thus a computer-aided diagnosis (CAD) system is highly demanded in clinical practice. In this paper, we propose a novel Saliency-aware Hybrid Network (SHNet) for automatic GI bleeding detection. Specifically, the SHNet consists of two densely connected convolutional networks (DenseNets) that are respectively considered as global image stream and saliency-aware stream. Moreover, we introduce polar transformation to reduce the noise from the background and highlight the image information. Finally, we employ the ensemble learning strategy to jointly optimize the final diagnosis result. Extensive experiments on the clinical WCE dataset illustrate that our proposed method outperforms the-state-of-the-art algorithms in GI bleeding detection with the F1 score of 0.959.