HC-Net: Hierarchical Context integration Network for medical image segmentation

Abstract

Accurate medical image segmentation is essential to achieve precise medical image analysis, e.g., blood vessel detection or lung segmentation, which has attracted researchers’ attention. So far, state-of-the-art techniques improving segmentation results have extracted the multi-scale contextual features with dilated convolution. However, these techniques cannot capture sufficient scale information due to only using several parallel independent branches, which result in leaving a gap between existed and ideal segmentation results.To address this problem, we design a hierarchical context integration network (HC-Net) which includes encoder module, hierarchical context integration module (HCM), and decoder module. Our HCM can not only extract the scale features in a single branch, but also learn the scale correlation among different branches. Meanwhile, self-attention mechanism therein is utilized to integrate context features and global dependencies adaptively for decoder Module. The proposed HC-Net has been applied two popular datasets, DRIVE and LUNA, and the experimental results show that our method outperforms state-of-the-art ones.