EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

An enhanced U-shape architecture network with cross-scale attention and frequency-domain perception for robust colorectal polyp segmentation

Rong Gao, Qi Ke, Aiquan Li, Xinning Qin, Sichao Zhao

PLOS One · 20.11.2026 · 21 (11) · e0360491

DOI: 10.1371/journal.pone.0360491 ↗

Kurz gefasst

Fachartikel in der Zeitschrift PLOS One, online erschienen am 20.11.2026. Verfasst von Rong Gao und 4 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Guangxi Higher Education Institutions Young and Middle-aged Teachers’ Basic Research Capacity Enhancement Project, Guangxi First-Class Discipline Statistics Construction Project Fund.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The early detection and removal of colorectal polyps can help reduce colorectal cancer and mortality rates. The segmentation of colorectal polyps in colonoscopy images is difficult due to multi-scale morphology, blurred boundaries, and complex background interference.To solve this problem, we propose a network of cross-scale attention and frequency domain perception of encoder-decoder architecture.The network has two new modules: first, Scale-Aware Cross Attention Skip Connection (SCAS) uses multi-head attention mechanism and multi-scale pyramid fusion to fuse multi-scale features from encoder and decoder to enhance the model to capture polyp features of different sizes.Secondly, Dual-domain fusion (D2F) separates and assigns weights of low-frequency global information and high-frequency detail information of the feature map at bottleneck stage, improving the model to analyze spatial and frequency domains and improve polyp boundaries.Multiple public colorectal polyp segmentation datasets show that our method is competitive and performs better than existing mainstream approaches in terms of mIoU and mDice: On the CVC-ClinicDB dataset our model produces mDice of 0.922 and mIoU of 0.863, outperforming UNet, UNet++, PraNet, ACSNet, DCSAU-Net, HarDNet-CPS, ProMamba, FoBS. These results validate its strong segmentation capability and effectiveness in accurately delineating polyp regions with blurred boundaries, large size variations, and complex background interference. The proposed method provides a reliable and promising tool for clinical endoscopic assistant diagnosis. The source code is publicly available at https://github.com/si-yuan20/Robust-segmentation-of-colon-polyps-Unet .

Förderung

  • Guangxi Higher Education Institutions Young and Middle-aged Teachers’ Basic Research Capacity Enhancement Project
  • Guangxi First-Class Discipline Statistics Construction Project Fund

Zitieren

APA

Gao, R., Ke, Q., Li, A., Qin, X., & Zhao, S. (2026). An enhanced U-shape architecture network with cross-scale attention and frequency-domain perception for robust colorectal polyp segmentation. PLOS One, 21(11), e0360491. https://doi.org/10.1371/journal.pone.0360491

Vancouver

Gao R, Ke Q, Li A, Qin X, Zhao S. An enhanced U-shape architecture network with cross-scale attention and frequency-domain perception for robust colorectal polyp segmentation. PLOS One. 2026;21(11):e0360491. doi:10.1371/journal.pone.0360491

BibTeX
@article{Gao2026,
  author = {Gao, Rong and Ke, Qi and Li, Aiquan and Qin, Xinning and Zhao, Sichao},
  title = {An enhanced U-shape architecture network with cross-scale attention and frequency-domain perception for robust colorectal polyp segmentation},
  journal = {PLOS One},
  year = {2026},
  volume = {21},
  number = {11},
  pages = {e0360491},
  doi = {10.1371/journal.pone.0360491}
}
Kostenlos starten →