Multi-Stage Attention-Unet for Wireless Capsule Endoscopy Image Bleeding Area Segmentation

Bleeding in gastrointestinal tract (GI) is caused by many different diseases and may lead to serious consequence if not being treated correctly. The most effective method to locate bleeding area is using wireless capsule endoscopy (WCE). However, the amount of WCE images of one patient is so large that even a professional physician may take a long time on analyzing them, and human errors may also occur with work time increasing. Thus the computer-aided diagnosis methods become more attractive. In this paper, we propose a novel deep learning based method for WCE images bleeding area segmentation. We design multi-stage architecture and attention blocks to deal with small areas segmentation. The latter stages' input is the combination of feature maps transferred from former stages so that even very deep layers can obtain small areas features, and attention blocks help shallow layers to better extract small areas features thus they can transfer more useful information to latter stages. Extensive experiments are conducted on public available WCE image dataset to show the effect of multi-stage architecture and attention blocks. Compared with other bleeding areas segmentation methods, our approach achieves state-of-the-art performance with 98.5% overall accuracy, 90.1% mean accuracy and 86.3% mean intersection over union (IoU). There is a 10.7% improvement on mean IoU compared with the previous state-of-the-art WCE bleeding segmentation method.