Scholarlink

<p>The encoder-decoder framework with a U-shaped design has revolutionized the segmentation of medical image but remains hindered by semantic gaps caused by mismatched feature hierarchies between shallow and deep layers. To resolve this, we suggest HAT-UNet, a dual-encoder framework that synergizes residual learning, transformer enhanced attention, and multi-source feature fusion. The architecture features a primary encoder with HAT blocks, novel modules integrating channel attention, spatial attention, and transformer layers to prioritize diagnostically salient features while modeling long-range dependencies. A secondary residual encoder preserves gradient flow and mitigates information loss during downsampling. The decoder employs a multi-source fusion block that dynamically reconciles features from both encoders. Evaluated across six public datasets from diverse modalities embracing dermoscopy, colonoscopy, endoscopy, and retinal fundus images, HAT-UNet attains cutting-edge performance, demonstrating high dice values of 97.50% for skin cancer, 95.11% for polyps abnormalities segmentation, and 95.37% for endoscopic polyps delineation. However, challenges persist in fine structure segmentation, highlighting limitations under data scarcity. HAT-UNet establishes a versatile framework for multi-modal medical image analysis, bridging the gap between architectural innovation and clinical applicability.</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

HAT-UNet: Hybrid attention transformer-based U-Net with multi-source fusion for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Noura Bentaher,
  • Younes Kabbadj,
  • Samira Lafraxo,
  • Mohamed Ben Salah

摘要

The encoder-decoder framework with a U-shaped design has revolutionized the segmentation of medical image but remains hindered by semantic gaps caused by mismatched feature hierarchies between shallow and deep layers. To resolve this, we suggest HAT-UNet, a dual-encoder framework that synergizes residual learning, transformer enhanced attention, and multi-source feature fusion. The architecture features a primary encoder with HAT blocks, novel modules integrating channel attention, spatial attention, and transformer layers to prioritize diagnostically salient features while modeling long-range dependencies. A secondary residual encoder preserves gradient flow and mitigates information loss during downsampling. The decoder employs a multi-source fusion block that dynamically reconciles features from both encoders. Evaluated across six public datasets from diverse modalities embracing dermoscopy, colonoscopy, endoscopy, and retinal fundus images, HAT-UNet attains cutting-edge performance, demonstrating high dice values of 97.50% for skin cancer, 95.11% for polyps abnormalities segmentation, and 95.37% for endoscopic polyps delineation. However, challenges persist in fine structure segmentation, highlighting limitations under data scarcity. HAT-UNet establishes a versatile framework for multi-modal medical image analysis, bridging the gap between architectural innovation and clinical applicability.