SAAIL Lab SAAIL Lab
Expert Systems with Applications · 2025 SCI Q1 · IF 7.5 Published · Vol. 299, 129998 · IF 7.5

BAFNet: Deep contour-aware features for colorectal polyps segmentation

Dibin Zhou 1 , Ni Chen 1 , Yueping Zhu 1 , Xueyan Zhang 2 , Innocent Nyalala 3 , Jiayu Zhang 4 , Junfeng Gao 5
1School of Information Science and Technology, Hangzhou Normal University, Hangzhou, China
2Women's Hospital, School of Medicine, Zhejiang University, Hangzhou, China
3School of Engineering and Science, Indian Institute of Technology Madras Zanzibar, Tanzania
4Xuzhou University of Technology, Xuzhou, China
5Department of Computer Science, University of Aberdeen, Aberdeen, UK
Expert Systems with Applications (Elsevier) (Oct 2025)
BAFNet: Deep contour-aware features for colorectal polyps segmentation

Summary: BAFNet couples a boundary-aware gating branch (four Conv-BN blocks) with Multi-Scale Feature Aggregation (MFAM) and Dynamic Global Attention (DGAM) to achieve 0.914 Dice on EndoScene in 5.98M parameters (3.59 GFLOPs).

Abstract

Automated segmentation of colorectal polyps remains challenging due to low tissue contrast and blurred mucosal boundaries. We propose BAFNet, a boundary-aware feature fusion network incorporating gating mechanisms that selectively preserve sharp contour cues alongside multi-scale semantic representations, advancing polyp segmentation accuracy across clinical colonoscopy benchmarks.

Keywords: BAFNetcolorectal polypsmedical image segmentationcontour-awareExpert Systems with Applications

Key Findings & Contributions

0.914
EndoScene Dice
0.863 IoU · 0.905 F-measure
0.808
ColonDB & ETIS Dice
Consistent generalization
5.98M
Model Parameters
3.59 GFLOPs footprint
IF 7.5
Expert Syst. Appl.
Elsevier Q1 Journal
  1. Designed a boundary-aware gating branch integrating four Conv-BN blocks to suppress background mucosal texture and preserve sharp colorectal polyp contours.
  2. Developed the Multi-Scale Feature Aggregation Module (MFAM) and Dynamic Global Attention Module (DGAM) to fuse encoder-decoder semantic cues across lesion scales.
  3. Outperformed five prior colonoscopy baselines (ASCNet, MSNet, TransFuse, C2F-Net, PraNet) on CVC-ColonDB, ETIS, and EndoScene at only 5.98M parameters and 3.59 GFLOPs.

Benchmark Results

BAFNet Colorectal Polyp Segmentation Performance (5.98M Params, 3.59 GFLOPs)

Benchmark DatasetDice ScoreIoU ScoreWeighted F-Measure
EndoScene0.9140.8630.905
CVC-ColonDB0.8080.7220.782
ETIS-LaribPolypDB0.8080.7320.774

BibTeX

@article{zhou2025bafnet,
  title={BAFNet: Deep contour-aware features for colorectal polyps segmentation},
  author={Zhou, Dibin and Chen, Ni and Zhu, Yueping and Zhang, Xueyan and Nyalala, Innocent and Zhang, Jiayu and Gao, Junfeng},
  journal={Expert Systems with Applications},
  year={2025},
  doi={10.1016/j.eswa.2025.129998},
  publisher={Elsevier}
}
APA Citation

Zhou, D., Chen, N., Zhu, Y., Zhang, X., Nyalala, I., Zhang, J., & Gao, J. (2025). BAFNet: Deep contour-aware features for colorectal polyps segmentation. Expert Systems with Applications. https://doi.org/10.1016/j.eswa.2025.129998