Dual-level Dynamic Imbalance Learning Network for Multimodal Medical Image Fusion

Multimodal medical image fusion seeks to integrate complementary information from various imaging sensors to produce diagnostically informative representations. However, inherent differences in imaging mechanisms result in substantial variations in structural, textural, and functional characteristics across sensors. These discrepancies often cause fusion networks to prioritize high-energy or high-contrast modalities, thereby suppressing critical information from lower-intensity modalities and resulting in modality imbalance. To address these challenges, we propose a dual-level dynamic imbalance learning fusion network, which adaptively balances and integrates features at both inter-modal and intra-modal levels. Specifically, at the inter-modal level, a dynamic adaptive modal reweighting mechanism is introduced to adjust the contribution of each modality based on reliability estimation and gradient feedback, ensuring a balanced and stable flow of information across heterogeneous sources. At the intra-modal level, a dynamic modal balancing module is developed to enable spatially adaptive expert fusion, thereby achieving fine-grained feature equilibrium through the selective enhancement of structural, contextual, functional, and differential cues. Additionally, a rank-reduced window-based frequency transformer is incorporated to enhance the shared feature representation through rank-decomposed attention and multi-domain collaborative modeling, thereby enabling efficient capture of spatial-frequency dependencies. Comprehensive experiments on multiple clinical sensor datasets demonstrate that the proposed framework achieves strong overall performance across complementary evaluation criteria. In particular, on SPECT-MRI fusion, DDILFusion ranks first in SD, SF, SCD, AG, and SSIM and second in MS-SSIM, while its FMI and VIF values are competitive but not best-in-class. The results indicate that it effectively mitigates modality imbalance, enhances structural and functional fidelity, and produces perceptually consistent and clinically reliable fusion outcomes. These findings highlight the promising potential of the proposed framework in clinical decision support and advanced multimodal medical image analysis.