Hybrid spatial-channel attention and global-regional context aggregation feature for polyp segmentation

Automatic and accurate polyp segmentation is critically involved in the early diagnosis of colorectal cancer.Despite its importance, achieving accurate polyp segmentation remains a challenging task due to the diverse range of appearances exhibited by the polyps in terms of size, shape, and brightness and the blurred boundary between polyps as well as the blurred boundary that often exists between polyps and their surrounding mucosa.This study proposes an encoder-decoder framework-based depth model for polyp segmentation to overcome these challenges. In order to reweight the encoder features, a Hybrid Spatial-Channel Attention Module (HSCAM) was developed. The segmentation network was induced to place increased emphasis on polyp boundaries while enhancing the key feature channels critical to the segmentation task. Subsequently, a Global-Regional Context Module (GRCM) was designed to extract contextual information applicable to the polyp segmentation of different sizes. Finally, a feature aggregation module was designed to efficiently aggregate features from HSCAM, GRCM, and the previous layer of decoder blocks for better polyp segmentation.The proposed method was rigorously assessed through extensive experiments on three publicly available polyp datasets. The results of the present study showed the superior performance of the proposed method as compared to the other state-of-the-art methods.