Scholarlink

<p>Precise polyp segmentation holds significant clinical value for the early screening and intervention of colorectal cancer. Despite advancements in spatial domain segmentation precision achieved through multi scale and edge feature modeling, existing methodologies often overlook the intricate detail information residing within high frequency components of the frequency domain. To address these limitations, we propose FMU-Net. We innovatively introduce a Fourier Domain Augmentation (FDAug) strategy to efficiently expand datasets by reconfiguring amplitude information while strictly preserving structural semantics. To capture fine-grained features often lost in spatial processing, we design a Detail Boosting Wavelet (DBW) module that effectively mines high frequency components. These details are integrated into the decoder via a Detail-Aware Mamba (DAM) module, utilizing a cross-domain state-aware mechanism for dynamic feature fusion. Furthermore, an Uncertainty-Aware Foreground Reweighting Module (FRM) is designed to optimize segmentation along ambiguous boundaries. Extensive experiments demonstrate FMU-Net’s superior performance across five public datasets, notably achieving mDice/mIoU scores of 93.1%/87.7% on Kvasir-SEG and 94.9%/90.2% on CVC-ClinicDB.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FMU-Net: Mamba-driven frequency and uncertainty aware network for polyp segmentation

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Lingyun Zhu,
  • Hongyu Yi

摘要

Precise polyp segmentation holds significant clinical value for the early screening and intervention of colorectal cancer. Despite advancements in spatial domain segmentation precision achieved through multi scale and edge feature modeling, existing methodologies often overlook the intricate detail information residing within high frequency components of the frequency domain. To address these limitations, we propose FMU-Net. We innovatively introduce a Fourier Domain Augmentation (FDAug) strategy to efficiently expand datasets by reconfiguring amplitude information while strictly preserving structural semantics. To capture fine-grained features often lost in spatial processing, we design a Detail Boosting Wavelet (DBW) module that effectively mines high frequency components. These details are integrated into the decoder via a Detail-Aware Mamba (DAM) module, utilizing a cross-domain state-aware mechanism for dynamic feature fusion. Furthermore, an Uncertainty-Aware Foreground Reweighting Module (FRM) is designed to optimize segmentation along ambiguous boundaries. Extensive experiments demonstrate FMU-Net’s superior performance across five public datasets, notably achieving mDice/mIoU scores of 93.1%/87.7% on Kvasir-SEG and 94.9%/90.2% on CVC-ClinicDB.