MDF-Net: A Decoupled Multimodal Fusion Network with Contrastive Learning for Reliable Brain Tumor Segmentation on Multi-sequence MRI

Accurate segmentation of brain tumors from multimodal magnetic resonance imaging (MRI) is critical for clinical diagnosis, treatment planning, and longitudinal disease monitoring. Multimodal MRI provides complementary information across sequences, yet this complementarity is entangled with substantial cross-modal redundancy. Prevailing coupled fusion schemes merge modalities into a single shared space, where discriminative tumor-specific cues and redundant shared content are blended indiscriminately, an effect further aggravated by cross-scale feature inconsistency and by the difficulty of delineating small enhancing lesions. To address these issues, we recast multimodal segmentation as a decoupled representation learning problem and propose a Multimodal Decoupled Fusion Network (MDF-Net) that explicitly disentangles the fused representation into complementary and non-redundant components. At its core, a Tri-Branch Decoupling (TBD) module separates the fused features into spatial, channel, and instance components, and a Feature Contrastive Learning (FCL) objective enforces their complementarity while suppressing inter-branch redundancy. To provide scale-aware and locally coherent features for decoupling, an Atrous Spatial Pyramid Pooling (ASPP) module and a Local Window Self-Attention (LWSA) module jointly model multi-scale context and local dependencies, while a dedicated Small Lesion Refinement (SLR) module improves the delineation of small and ambiguous enhancing tumors through multi-scale detection and boundary-aware refinement. Extensive experiments on two publicly available benchmarks, BraTS 2023 and FeTS 2022, demonstrate that MDF-Net achieves competitive and well-balanced performance relative to recent state-of-the-art methods across tumor sub-regions and multiple evaluation metrics, with particularly accurate whole-tumor boundary delineation. Ablation studies and feature-space visualizations further confirm that explicit decoupling yields a more structured embedding space and that each component makes a complementary contribution to segmentation performance.