DKNET: A Hybrid Model with Enhanced Attention and Multi-level Fusion for Accurate Early Polyp Segmentation.

Saved in:

Bibliographic Details
Title:	DKNET: A Hybrid Model with Enhanced Attention and Multi-level Fusion for Accurate Early Polyp Segmentation.
Authors:	Dehong Kong¹ 1839079041@qq.com, Jie Liu² 17359111@qq.com, Xiang Zhou³ 702603395@qq.com, Wenyu Zhang⁴ zhangwenyu8518@126.com, Yixiao Sun³ robinsunyixiao@163.com, Jieyi Xu³ 232081203006@stu.ustl.edu.cn, Qinghong Zeng⁵ 1600250563@qq.com
Source:	Engineering Letters. Dec2025, Vol. 33 Issue 12, p5003-5014. 12p.
Subjects:	Image segmentation, Colon polyps, Artificial neural networks, Cancer prevention, Deep learning
Abstract:	Colorectal polyps are gastrointestinal growths that may develop into colorectal cancer, so the early detection and removal of these polyps are crucial for cancer prevention. In recent years, deep-learning techniques have made significant progress in colorectal polyp image segmentation, significantly enhancing segmentation accuracy and automation. However, early-stage polyps are often small, irregularly shaped, and typically hidden within larger organs or complex backgrounds, making them difficult to detect and segment. Existing segmentation models face challenges in effectively handling such polyps due to the lack of specialized structures for small-scale targets and the inability to integrate local detail information with global context effectively. In this paper, we propose the DKNET model, which leverages an enhanced attention mechanism (the DCE module) and a multi-level fusion strategy (the K-Fusion module) to tackle the challenges posed by small-scale polyps and complex segmentation tasks. Experimental results demonstrate that the proposed model performs well across multiple polyp datasets. Notably, on datasets with a high prevalence of small-scale polyps, such as CVC-ColonDB and Kvasir, the model achieves DICE scores of 82.68% and 92.39%, and mIoU scores of 74.75% and 87.27%, respectively, showcasing its effectiveness in handling challenging cases with small-scale targets. Compared to the state-of-the-art segmentation model PVT-GCASCAD, the proposed model achieves notable gains of 1.2% and 0.34% in DICE scores, and 1.1% and 0.36% in mIoU scores on CVC-ColonDB and Kvasir datasets, respectively. Additionally, its strong generalization capability is demonstrated by favorable performance on other datasets within this domain. [ABSTRACT FROM AUTHOR]
	Copyright of Engineering Letters is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	Colorectal polyps are gastrointestinal growths that may develop into colorectal cancer, so the early detection and removal of these polyps are crucial for cancer prevention. In recent years, deep-learning techniques have made significant progress in colorectal polyp image segmentation, significantly enhancing segmentation accuracy and automation. However, early-stage polyps are often small, irregularly shaped, and typically hidden within larger organs or complex backgrounds, making them difficult to detect and segment. Existing segmentation models face challenges in effectively handling such polyps due to the lack of specialized structures for small-scale targets and the inability to integrate local detail information with global context effectively. In this paper, we propose the DKNET model, which leverages an enhanced attention mechanism (the DCE module) and a multi-level fusion strategy (the K-Fusion module) to tackle the challenges posed by small-scale polyps and complex segmentation tasks. Experimental results demonstrate that the proposed model performs well across multiple polyp datasets. Notably, on datasets with a high prevalence of small-scale polyps, such as CVC-ColonDB and Kvasir, the model achieves DICE scores of 82.68% and 92.39%, and mIoU scores of 74.75% and 87.27%, respectively, showcasing its effectiveness in handling challenging cases with small-scale targets. Compared to the state-of-the-art segmentation model PVT-GCASCAD, the proposed model achieves notable gains of 1.2% and 0.34% in DICE scores, and 1.1% and 0.36% in mIoU scores on CVC-ColonDB and Kvasir datasets, respectively. Additionally, its strong generalization capability is demonstrated by favorable performance on other datasets within this domain. [ABSTRACT FROM AUTHOR]
ISSN:	1816093X