Abstract

Info

Title:: Colorectal polyp segmentation algorithm integrating Transformer and convolution

Author(s):: LIU Hongbin; GU De; School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, China

Keywords:: Keywords: polyp segmentation feature fusion Transformer convolution

Abstract:: Abstract: In response to the challenges of varied sizes and diverse shapes of colorectal polyps, especially with blurred boundaries that often complicates localization and smaller polyps being particularly prone to oversight, a colorectal polyp segmentation algorithm integrating Transformer and convolution is proposed. Transformer is employed to extract global features from images for ensuring the networks capability for global modeling and improving the localization capability for both main polyp regions and vague boundaries. Subsequently, convolution is introduced to augment the networks ability to process polyp details, refining boundary segmentation and enhancing the capture capability for small-sized polyps. Finally, a deep fusion of the features extracted by Transformer and convolution is carried out to realize feature complementarity. The experimental evaluation using CVC-ClinicDB and Kvasir-SEG datasets show that the algorithm has similarity coefficients of 95.4% and 93.2%, and mean intersection over union of 91.3% and 88.6%, respectively. Further tests on the generalization capability of the algorithm are conducted on CVC-ColonDB, CVC-T, and ETIS datasets, in which similarity coefficients of 81.3%, 90.9% and 80.1% are obtained. The results indicate a notable improvement in the accuracy of polyp segmentation achieved by the proposed algorithm.