Scholarlink

<p>Polyp segmentation plays a pivotal role in early colorectal cancer detection, where accurate and automated delineation of polyps can significantly enhance diagnostic precision. Existing convolutional neural network (CNN)-based approaches exhibit limitations in capturing long-range dependencies, while transformer-based models, though effective in global context modeling, often overlook vital frequency-domain information. Moreover, the highly variable and indistinct boundaries of polyps demand powerful multi-scale representation learning and substantial computational resources, highlighting the necessity of high-performance computing (HPC) support. To address these gaps, we propose DDFFormer, a novel dual-domain fused transformer architecture designed to integrate spatial- and frequency-domain contexts seamlessly. Our key contribution lies in the introduction of the dual-domain collaborative attention (DDCA) mechanism, which synergistically fuses multi-head self-attention (MHSA)-refined spatial-domain tokens and frequency-domain context-enriched tokens, ensuring comprehensive contextual representation. Through a hierarchical framework composed of dual-domain fused transformer blocks (DDFBs) and a mask transformer decoder, DDFFormer progressively refines feature representations across multiple scales. Extensive evaluations on three benchmark datasets—Kvasir-SEG, CVC-ClinicDB, and CVC-ColonDB—demonstrate that DDFFormer consistently outperforms state-of-the-art methods, achieving superior Dice, IoU, sensitivity, and specificity scores. Furthermore, ablation studies highlight the significant contribution of DDCA to performance gains. In addition, we report the computational efficiency of our model and emphasize that both the large-scale training and real-time inference require GPU acceleration and parallel optimization, underscoring the relevance of this work to supercomputing and HPC applications. These findings confirm the robustness and efficacy of DDFFormer, making it a promising solution for real-world clinical applications in gastrointestinal disease detection.</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

DDFformer: a dual-domain fused transformer for polyp segmentation

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • Yao Tong,
  • Xi Yong,
  • Jingchen Liang,
  • Ziqi Chen,
  • Yun Hu,
  • Xin Li,
  • Hongmin Gao,
  • Zuojian Zhou,
  • Kongfa Hu

摘要

Polyp segmentation plays a pivotal role in early colorectal cancer detection, where accurate and automated delineation of polyps can significantly enhance diagnostic precision. Existing convolutional neural network (CNN)-based approaches exhibit limitations in capturing long-range dependencies, while transformer-based models, though effective in global context modeling, often overlook vital frequency-domain information. Moreover, the highly variable and indistinct boundaries of polyps demand powerful multi-scale representation learning and substantial computational resources, highlighting the necessity of high-performance computing (HPC) support. To address these gaps, we propose DDFFormer, a novel dual-domain fused transformer architecture designed to integrate spatial- and frequency-domain contexts seamlessly. Our key contribution lies in the introduction of the dual-domain collaborative attention (DDCA) mechanism, which synergistically fuses multi-head self-attention (MHSA)-refined spatial-domain tokens and frequency-domain context-enriched tokens, ensuring comprehensive contextual representation. Through a hierarchical framework composed of dual-domain fused transformer blocks (DDFBs) and a mask transformer decoder, DDFFormer progressively refines feature representations across multiple scales. Extensive evaluations on three benchmark datasets—Kvasir-SEG, CVC-ClinicDB, and CVC-ColonDB—demonstrate that DDFFormer consistently outperforms state-of-the-art methods, achieving superior Dice, IoU, sensitivity, and specificity scores. Furthermore, ablation studies highlight the significant contribution of DDCA to performance gains. In addition, we report the computational efficiency of our model and emphasize that both the large-scale training and real-time inference require GPU acceleration and parallel optimization, underscoring the relevance of this work to supercomputing and HPC applications. These findings confirm the robustness and efficacy of DDFFormer, making it a promising solution for real-world clinical applications in gastrointestinal disease detection.