计算机科学
图像融合
融合
人工智能
算法
可扩展性
计算机视觉
扩散
图像(数学)
图像处理
图像分割
模式识别(心理学)
传感器融合
可视化
数据挖掘
图像配准
理论计算机科学
作者
Xinyi Shi,Weisheng Li,Dan He,Yucheng Shu,Feiyan Li,Yidong Peng,Xiang Li
标识
DOI:10.1109/tcsvt.2026.3740876
摘要
Multimodal medical image fusion (MMIF) aims to integrate complementary anatomical and functional information from heterogeneous modalities for clinical diagnosis. Although diffusion models have shown strong potential for MMIF, most existing methods treat various modalities symmetrically despite their visual disparities, causing anatomical details to be over-whelmed by functionally salient regions. In addition, direct multi-scale feature aggregation is vulnerable to noisy shallow features in the early denoising stages, leading to anatomical distortion and detail loss. To address these issues, we propose AD-cDiff, an asymmetric dynamic conditional diffusion framework. Anatomical images dominate the diffusion process through the content branch, while the conditional branch dynamically injects stage-specific information from functional modalities through timestep conditioning and frequency adaptation. We further develop a progressive self-distillation alignment strategy that aligns shallow detail features under deep semantic guidance before multi-scale fusion. Finally, leveraging the independence of the conditional branch, we design a common-difference integration module to support trimodal fusion without incurring the multiplicative computational overhead caused by modality scaling. Extensive experiments on three public datasets show that AD-cDiff better preserves complex lesion contours and achieves state-of-the-art fusion performance, while downstream BraTS 2023 segmentation further demonstrates the effectiveness of fused images. The code is available at https://github.com/SXY-F/AD-cDiff.
科研通智能强力驱动
Strongly Powered by AbleSci AI