FreqPolyp: Single-Source Domain Generalization for Polyp Segmentation via Amplitude Decoupling and Wavelet Cross-Attention

Abstract

Accurate segmentation of colorectal polyps is important for colorectal cancer prevention. However, domain shifts across medical centers and blurred boundaries of diminutive lesions remain major challenges for robust polyp segmentation. We propose FreqPolyp, a frequency-domain polyp segmentation framework designed for single-source domain generalization. At the input stage, the frequency domain augmentation module perturbs amplitude-related appearance statistics while preserving phase-related structural information, thereby reducing dependence on source-domain photometric styles without using target-domain data. At the decoding stage, the wavelet dual-stream fusion module uses discrete wavelet transform to extract high-frequency boundary cues and semantic-guided cross-attention to refine low-contrast lesion boundaries. In addition, a hybrid spatial-frequency supervision strategy combines amplitude-domain spectral consistency with phase-aware structural consistency. Experiments on five public benchmarks show that FreqPolyp achieves mDice/mIoU scores of 0.849/0.738 on ColonDB and 0.831/0.711 on ETIS, with an average unseen-domain mDice/mIoU of 0.865/0.765 across ColonDB, ETIS, and CVC-300. Under the tested RTX 3090 setting, FreqPolyp contains 26.2 M parameters and reaches 58 FPS. These results suggest that FreqPolyp provides a favorable accuracy–efficiency trade-off under public benchmark and controlled GPU evaluation settings.