Scholarlink

Addressing limited generalization in MRI reconstruction caused by cross-device variations and heterogeneous sampling patterns, we propose KFDualUNet: a dual-domain collaborative network integrating UNet and Transformer architectures. Our approach introduces three key innovations: (1) The KS-MFE module extracts discriminative features from undersampled k-space data and sampling masks via multi-scale convolutional layers, enabling adaptive acquisition-specific feature perception; (2) The MS-ADC mechanism dynamically integrates k-space information during image-domain reconstruction; (3) KGCM and KWM loss functions enforce physical constraints on reconstructed outputs. Experimental results demonstrate superior artifact suppression and high-frequency detail preservation compared to existing methods, establishing a novel framework for MRI cross-domain joint optimization.

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

KFDualUNet: Enhancing MRI Reconstruction Generalization Through Cross-Domain Synergistic Learning with Dual Cascaded UNets and K-Space Physical Constraints

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ye Wang,
  • Rongjun Ge

摘要

Addressing limited generalization in MRI reconstruction caused by cross-device variations and heterogeneous sampling patterns, we propose KFDualUNet: a dual-domain collaborative network integrating UNet and Transformer architectures. Our approach introduces three key innovations: (1) The KS-MFE module extracts discriminative features from undersampled k-space data and sampling masks via multi-scale convolutional layers, enabling adaptive acquisition-specific feature perception; (2) The MS-ADC mechanism dynamically integrates k-space information during image-domain reconstruction; (3) KGCM and KWM loss functions enforce physical constraints on reconstructed outputs. Experimental results demonstrate superior artifact suppression and high-frequency detail preservation compared to existing methods, establishing a novel framework for MRI cross-domain joint optimization.