Context-Aware Deep Network With Cross-Layer Feature Fusion for Polyp Segmentation

Abstract

Accurate polyp segmentation in colonoscopy images is crucial for early diagnosis and timely treatment of colorectal cancer. Recently, deep learning-based methods have exhibited distinct advantages in polyp segmentation. However, most existing methods often obtain ordinary performance due to the limited ability to extract context information and foreground cues. To overcome these drawbacks, we propose a context-aware deep network (CADNet) with cross-layer feature fusion for polyp segmentation in this article. Specifically, CADNet is designed with an encoder-decoder framework. In the encoder stage, a group guidance context module (GGCM) is proposed for the top-three highest layers to make the network focus more on target regions by aggregating multiscale context information with the prior knowledge from the adjacent high layer. In the decoder stage, a cross-layer feature fusion module (CLFFM) is proposed to obtain rich foreground cues by adaptively fusing low-level spatial details and high-level semantic concepts with the assistance of an attention mechanism. After that, the foreground cues serve as the input for the subsequent decoding stage. Considering that there is a lack of two-modal datasets of polyp segmentation, we construct a new dataset with 1200 images. Extensive experiments on our dataset and three public datasets demonstrate that our CADNet has considerable generalization ability across two-modal data and cross-center data and obtains comparable results compared with the state-of-the-art methods.