Multimodal medical image fusion aims to combine complementary visual information from different imaging modalities to assist diagnosis and clinical decision-making. Existing methods often struggle to balance global semantic representation and local detail preservation, leading to blurred or incomplete features in salient regions. This work proposes ATDFusion (Adapter-Tuned Dual-Branch Network), featuring a global semantic branch and an auxiliary detail branch to jointly capture high-level context and fine-grained details. A Low-Rank Dynamic Token Adapter (LR-DTA) adaptively fine-tunes intermediate layers of pretrained models based on token number and rank. Additionally, a Fusion Enhancement Guidance (FEG) module imposes explicit spatial supervision via a saliency-aware loss to strictly preserve diagnostically critical regions. On three typical tasks (MRI-CT, MRI-PET, MRI-SPECT), ATDFusion surpasses state-of-the-art methods, improving SSIM and CC by 11.93% and 5.07%, respectively. Other metrics (e.g., EN, PSNR) also achieve leading results, validating its effectiveness. Furthermore, the model demonstrates strong zero-shot generalization on the unseen HECKTOR 2025 dataset. Code is available at https://github.com/pluto628/ATDFusion . • Asymmetric dual-branch framework resolves the semantic-detail trade-off. • Low-Rank Dynamic Token Adapter realizes efficient sample-specific fine-tuning. • Explicit spatial supervision guarantees the retention of metabolic/anatomical hotspots. • Achieves SOTA performance with strong zero-shot generalization on unseen datasets.
Li et al. (2026) studied this question.