SFCFuse: multimodal medical image fusion based on saliency guidance and multi-source feature interaction

Multimodal medical image fusion (MMIF) aims to integrate complementary information from different imaging modalities into a single image, which can facilitate visual inspection and subsequent analysis. Existing deep learning (DL)-based methods mainly focus on bridging inter-modal heterogeneity, while the handling of inter-modal common information is relatively less emphasized. Inadequate treatment of such redundancy may introduce feature interference or blurring artifacts in fused results. In addition, oversimplified multi-source feature aggregation can lead to the loss of fine details. To address these issues, we propose SFCFuse, a saliency-guided fusion framework with multi-source feature interaction. Specifically, we adopt a Swin-Unet backbone for hierarchical multi-scale feature extraction and introduce a Saliency Feature Guidance (SFG) module with gated selection to suppress redundant responses and emphasize informative regions. Furthermore, a Joint Feature Interaction Fusion (JFIF) module is designed to enable attention-based interaction between heterogeneous feature sources, facilitating complementary information exchange. Experimental results on PET-MRI and SPECT-MRI datasets demonstrate that SFCFuse achieves competitive performance against representative MMIF methods in both quantitative metrics and visual comparisons.