Scholarlink

<p>Multimodal medical image fusion integrates structural information from high-resolution anatomical images with physiological and metabolic information from low-spatial-resolution functional images, thereby providing more comprehensive imaging information for clinical diagnosis and treatment. However, existing methods still struggle to properly model cross-modal discrepancies induced by inherent modality heterogeneity, achieve precise feature alignment, and capture long-range global dependencies, which may hinder the full exploitation of complementary information. To address these issues, we propose a dual-discrepancy feature extraction and adaptive weighted fusion network, termed DDAFusion. Specifically, we first develop a dual-discrepancy feature extractor to explicitly model modality-specific characteristics and capture complementary yet heterogeneous features from anatomical and functional images. An adaptive weighted fusion module is then introduced to enhance feature alignment and establish global dependencies, enabling more effective cross-modal interaction. In addition, a feature filtering and channel reconstruction module is designed to suppress redundant information and strengthen diagnostically relevant features, leading to enhanced feature representations. These components collectively contribute to improving both structural detail preservation and functional semantic representation. The training and optimization of our proposed DDAFusion model fully demonstrate the essential role of high-performance computing (HPC) in enabling efficient development and deployment of complex deep learning models for medical image analysis. Accordingly, we leverage HPC resources to accelerate the model optimization process and improve its scalability for efficient training and practical clinical deployment. Extensive experiments on PET-MRI and SPECT-MRI datasets demonstrate that our proposed method achieves superior performance compared with existing mainstream approaches, with consistent improvements observed in both qualitative and quantitative evaluations.</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

DDAFusion: dual-discrepancy feature extraction and adaptive weighted fusion network for multimodal medical image fusion

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Jiayi Wang,
  • Lei Yu,
  • Huiqi Wang

摘要

Multimodal medical image fusion integrates structural information from high-resolution anatomical images with physiological and metabolic information from low-spatial-resolution functional images, thereby providing more comprehensive imaging information for clinical diagnosis and treatment. However, existing methods still struggle to properly model cross-modal discrepancies induced by inherent modality heterogeneity, achieve precise feature alignment, and capture long-range global dependencies, which may hinder the full exploitation of complementary information. To address these issues, we propose a dual-discrepancy feature extraction and adaptive weighted fusion network, termed DDAFusion. Specifically, we first develop a dual-discrepancy feature extractor to explicitly model modality-specific characteristics and capture complementary yet heterogeneous features from anatomical and functional images. An adaptive weighted fusion module is then introduced to enhance feature alignment and establish global dependencies, enabling more effective cross-modal interaction. In addition, a feature filtering and channel reconstruction module is designed to suppress redundant information and strengthen diagnostically relevant features, leading to enhanced feature representations. These components collectively contribute to improving both structural detail preservation and functional semantic representation. The training and optimization of our proposed DDAFusion model fully demonstrate the essential role of high-performance computing (HPC) in enabling efficient development and deployment of complex deep learning models for medical image analysis. Accordingly, we leverage HPC resources to accelerate the model optimization process and improve its scalability for efficient training and practical clinical deployment. Extensive experiments on PET-MRI and SPECT-MRI datasets demonstrate that our proposed method achieves superior performance compared with existing mainstream approaches, with consistent improvements observed in both qualitative and quantitative evaluations.