MambaFusion: Registration-assisted lumbar CT-MRI image fusion via Mamba and dynamic feature modeling

Medical images play a pivotal role in clinical diagnosis, yet a single imaging modality often fails to provide complete anatomical and pathological information. Multimodal image fusion addresses this limitation by integrating complementary information from different modalities. In this study, a CT-MRI registration and preprocessing procedure was first introduced to reduce spatial mismatch between lumbar CT and MRI images. The MRI volume was registered and resampled into the CT coordinate space, ensuring spatial consistency before image fusion. Landmark-based registration evaluation across 30 patients showed that the mean target registration error (TRE) decreased from 26.725 ± 0.847 mm after geometry-based initialization to 2.808 ± 0.983 mm after final rigid registration, corresponding to an 89.49% reduction. Based on the registered CT-MRI image pairs, this study proposes MambaFusion, a novel multimodal medical image fusion model that integrates Mamba sequence modeling with dynamic convolution for enhanced feature extraction and reconstruction. In addition, a Dynamic Feature Fusion Module (DFFM) is designed to improve cross-modal representation learning by enhancing texture details and cross-modal correlations. Experiments on registered lumbar CT-MRI images demonstrate that MambaFusion achieves competitive visual quality and quantitative fusion performance compared with existing methods. The fused images preserve CT-derived vertebral bone structures while retaining MRI-derived soft-tissue information, providing a complementary multimodal representation for lumbar spine imaging. The code is available at https://github.com/Marcusmakusi0612/MambaMIF .