Frequency-augmented and uncertainty-aware learning for polyp segmentation

We propose FULNet for polyp segmentation, combining Frequency-Augmented Learning (FAL) to enrich data-style diversity while preserving lesion morphology and Uncertainty-Aware Learning (UAL) to improve segmentation on complex and ambiguous regions, validated on five datasets with SegNeXt and PraNet baselines.

摘要

Polyp segmentation plays a crucial role in the early diagnosis of colorectal cancer, yet it still faces two major challenges in practical applications. First, the scarcity of annotated data limits the model's ability to learn the diverse characteristics of polyps, while existing data augmentation methods often distort key lesion structures during sample expansion. Second, current algorithms remain insufficient in handling complex and ambiguous regions. To address these issues, we propose the Frequency-Augmented and Uncertainty-Aware Learning Network (FULNet), which incorporates Frequency-Augmented Learning (FAL) to enhance data-style diversity while preserving lesion morphology, and Uncertainty-Aware Learning (UAL) to improve segmentation performance in challenging regions. We validated our proposed method on five datasets using SegNeXt and PraNet as baselines to demonstrate the generalization ability. On two unseen datasets, SegNeXt achieved 2.5% and 4.4% Dice improvement on CVC-ColonDB and ETIS, while PraNet improved by 3.2% and 6.2% respectively.

主要结果

BibTeX

@article{zhai2026fulnet,
  author  = {Zhai, Ying and Li, Huiqian and Han, Longfei and Li, Haisheng},
  title   = {Frequency-augmented and uncertainty-aware learning for polyp segmentation},
  journal = {Biomedical Signal Processing and Control},
  year    = {2026},
  doi     = {10.1016/j.bspc.2026.110346}
}