EUVIMEDOpdag EUVIMED
Open access · CC BY 4.0

CAM-PoEm: Contour-guided state-space modeling with multi-scale positional encoding for gastrointestinal polyp segmentation

Fabliha Afaf Sarwar, Abdullah All Jilan, Zaid Rehman, Md. Farhadul Islam, Jannatun Noor Mukta

Array · 2026-11 · 101419

DOI: 10.1016/j.array.2026.101419 ↗

At a glance

Journal article in Array, 2026-11. Written by Fabliha Afaf Sarwar and 4 co-authors. 97 references. The full text is freely available (open access, CC BY 4.0). Funded by: United International University.

AbstractBibliografiske oplysningerCiteRelated studies

Abstract (original language)

Accurate segmentation of gastrointestinal polyps is vital for early detection of colorectal cancer, yet remains challenging due to ambiguous boundaries, heterogeneous textures, and high morphological variability. While convolutional neural networks (CNNs) effectively capture local features, they often struggle with long-range dependencies, and transformer-based methods though effective at capturing long-range dependencies, incur prohibitive quadratic complexity. Recent state-space models (SSMs), such as Mamba, enable global modeling with linear-time complexity but remain insensitive to boundary structures that are critical for precise segmentation. In this work, we propose CAM-PoEm, a segmentation framework that integrates global context modeling with explicit boundary awareness. The proposed Contour-Aware Mamba (CAM) bottleneck incorporates learnable contour cues into Mamba’s state-to-output readout, enhancing morphology-aware global feature modeling. To complement this, we introduce a Multi-Kernel Positional Embedding (MKPE) encoder that captures spatial information across multiple receptive fields, improving local representation of diverse polyp structures. Experiments on four benchmark datasets show that CAM-PoEm achieves Dice scores of 0.9161 on Kvasir-SEG (Jha et al., 2019), 0.8827 on PolypGen2021 (Ali et al., 2021), 0.9150 on CVC-ClinicDB (Bernal et al., 2015), and 0.8831 on CVC-ColonDB (Bernal et al., 2012), with only 4.96M parameters . The results indicate that combining contour-aware global modeling with multi-scale spatial encoding enhances both boundary delineation and overall segmentation consistency. The model implementation, experimental configurations, and evaluation scripts are provided as supplementary materials and complete source code will be made publicly available upon acceptance of the manuscript.

Funding

  • United International University

Cite

APA

Sarwar, F. A., Jilan, A. A., Rehman, Z., Islam, M. F., & Mukta, J. N. (2026). CAM-PoEm: Contour-guided state-space modeling with multi-scale positional encoding for gastrointestinal polyp segmentation. Array, 101419. https://doi.org/10.1016/j.array.2026.101419

Vancouver

Sarwar FA, Jilan AA, Rehman Z, Islam MF, Mukta JN. CAM-PoEm: Contour-guided state-space modeling with multi-scale positional encoding for gastrointestinal polyp segmentation. Array. 2026:101419. doi:10.1016/j.array.2026.101419

BibTeX
@article{Sarwar2026,
  author = {Sarwar, Fabliha Afaf and Jilan, Abdullah All and Rehman, Zaid and Islam, Md. Farhadul and Mukta, Jannatun Noor},
  title = {CAM-PoEm: Contour-guided state-space modeling with multi-scale positional encoding for gastrointestinal polyp segmentation},
  journal = {Array},
  year = {2026},
  pages = {101419},
  doi = {10.1016/j.array.2026.101419}
}
Start for free →