Combining Transformer and Reverse Attention Mechanism for Polyp Segmentation

Jianzhuang Lin, Wenzhong Yang, Sixiang Tan

2022

Abstract

The polyp region can be accurately segmented from the images obtained by colonoscopy to assist doctors in diagnosis, which is of great significance for reducing the incidence of colon cancer. For the polyp segmentation problem, a segmentation network combining Transformer and reverse attention (CTRNet) is proposed. First, using the swin Transformer as the backbone network, the polyp image is modeled hierarchically, and long-distance dependencies are obtained, and the receptive field is gradually expanded to obtain more contextual information of the polyp target area; secondly, an inversion attention is proposed. The mechanism is used to mine the polyp area information in the feature map, correct the inconsistent area, establish the relationship between the foreground area and the boundary, and suppress the irrelevant background information through the attention gate, thereby improving the accuracy of the model’s segmentation of the polyp boundary. Extensive experiments are carried out on five challenging datasets. Compared with other existing methods under different evaluation indicators, CTRNet’s performance is better than most of the compared methods, and it can effectively segment polyp regions. Especially on Kvasir and ETIS, STRNet achieves 0.922 and 0.793 mDICE.

Paper Citation

in Harvard Style

Lin J., Yang W. and Tan S. (2022). Combining Transformer and Reverse Attention Mechanism for Polyp Segmentation. In Proceedings of the 4th International Conference on Biotechnology and Biomedicine - Volume 1: ICBB; ISBN 978-989-758-637-8, SciTePress, pages 125-137. DOI: 10.5220/0012014800003633

in Bibtex Style

@conference{icbb22,
author={Jianzhuang Lin and Wenzhong Yang and Sixiang Tan},
title={Combining Transformer and Reverse Attention Mechanism for Polyp Segmentation},
booktitle={Proceedings of the 4th International Conference on Biotechnology and Biomedicine - Volume 1: ICBB},
year={2022},
pages={125-137},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012014800003633},
isbn={978-989-758-637-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 4th International Conference on Biotechnology and Biomedicine - Volume 1: ICBB
TI - Combining Transformer and Reverse Attention Mechanism for Polyp Segmentation
SN - 978-989-758-637-8
AU - Lin J.
AU - Yang W.
AU - Tan S.
PY - 2022
SP - 125
EP - 137
DO - 10.5220/0012014800003633
PB - SciTePress