Scholarlink

<p>Accurate medical image segmentation is limited by unreliable skip-feature fusion and upsampling artifacts, degrading boundary precision for small/complex structures. We propose MAGE-UNet with a Progressive Refinement Attention Module (PRAM) and Artifact-Aware Upsampling (AAU). PRAM shifts from coupled fusion to decoupled progressive refinement: intra-scale purification (spatial low-pass filtering to suppress noise) followed by cross-scale semantic alignment for decoder compatibility. AAU replaces transposed convolutions with interpolation-based depthwise separable convolution, mitigating checkerboard artifacts and enhancing boundary fidelity. On Synapse multi-organ CT, MAGE-UNet achieves mean Dice 80.29% and HD95 21.06 mm, outperforming TransUNet by +3.70% Dice and -10.44 mm HD95. On ACDC MRI under zero-shot hyperparameter transfer, it achieves 86.96% mean Dice. Results confirm exceptional accuracy, boundary reliability, and robustness.</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

MAGE-UNet: Progressive skip-feature refinement for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Yuhang Wang,
  • Shuoyang Wang,
  • Chengze Lv,
  • Gongrui Wang,
  • Zicheng Lin,
  • Zhengxiang Ma,
  • Weichao Pan

摘要

Accurate medical image segmentation is limited by unreliable skip-feature fusion and upsampling artifacts, degrading boundary precision for small/complex structures. We propose MAGE-UNet with a Progressive Refinement Attention Module (PRAM) and Artifact-Aware Upsampling (AAU). PRAM shifts from coupled fusion to decoupled progressive refinement: intra-scale purification (spatial low-pass filtering to suppress noise) followed by cross-scale semantic alignment for decoder compatibility. AAU replaces transposed convolutions with interpolation-based depthwise separable convolution, mitigating checkerboard artifacts and enhancing boundary fidelity. On Synapse multi-organ CT, MAGE-UNet achieves mean Dice 80.29% and HD95 21.06 mm, outperforming TransUNet by +3.70% Dice and -10.44 mm HD95. On ACDC MRI under zero-shot hyperparameter transfer, it achieves 86.96% mean Dice. Results confirm exceptional accuracy, boundary reliability, and robustness.