RT-Net: Region-Enhanced Attention Transformer Network for Polyp Segmentation

• Published in: Neural processing letters Vol. 55; no. 9; pp. 11975 - 11991
• Main Authors: Qin, Yilin, Xia, Haiying, Song, Shuxiang
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2023
• Subjects: Artificial Intelligence; Complex Systems; Computational Intelligence; Computer Science; Transformer; Region-enhanced attention; Polyp segmentation; Medical image segmentation
• ISSN: 1370-4621; 1573-773X
• DOI: 10.1007/s11063-023-11405-y

Colonic polyps are highly correlated with colorectal cancer. Prevention of colorectal cancer is the detection and removal of polyps in the early stages of the disease. But the detection process relies on the physician’s experience and is prone to missed diagnoses. The drawbacks motivate us to design an algorithm to automatically assist physicians in detection to reduce the rate of missed polyps. However, polyp segmentation encounters challenges due to the variable appearance and blurred borders with the surrounding mucosa. And it is difficult for existing CNN-based polyp segmentation algorithms to learn long-range dependencies. Therefore, we propose a region-enhanced attention transformer network (RT-Net) for polyp segmentation. Unlike existing CNN-based approaches, it employs a pyramid Transformer encoder to promote the learning ability and robustness of the network. In addition, we introduce three modules, including the residual multiscale (RMS) module, the region-enhanced attention (REA) module and the feature aggregation (FA) module. Specifically, the RMS module learns multiscale information from the features of the encoder. The REA module adopts the prediction maps of each decoder layer to guide the network in building target regions and boundary cues to compensate for the missing local fields of view in the encoder. The role of the FA module is to efficiently aggregate the features from REA with those from the decoding layer to achieve better segmentation performance. RT-Net is evaluated on five benchmark polyp datasets. Extensive experiments demonstrate that our proposed RT-Net exhibits excellent performance compared to other state-of-the-art methods.