DFuse-Net: Disentangled feature fusion with uncertainty-aware learning for reliable multi-modal brain tumor segmentation.

Abstract

Accurate brain tumor segmentation from multi-modal MRI is critical for clinical diagnosis and treatment planning. However, effectively exploiting the complementary information across different modalities remains challenging due to modality-specific noise, semantic inconsistency and inherent model uncertainty. To tackle these issues, we propose a Disentangled Fusion Network named DFuse-Net that integrates disentangled feature fusion with uncertainty-aware learning for reliable multi-modal brain tumor segmentation. Specifically, DFuse-Net explicitly disentangles modality-shared and modality-specific representations, enhancing the discriminability and expressiveness of multi-modal features. Furthermore, a Disentangled Texture Fusion Module (DTFM) and a Disentangled Semantic Fusion Module (DSFM) are designed to effectively integrate texture- and semantic-level information across modalities. In addition, a contrastive-aware learning scheme is proposed to strengthen feature discriminability, while a consistency-aware learning strategy is proposed to enforce structural coherence across modalities. During inference, Monte Carlo dropout is employed to estimate voxel-wise aleatoric and epistemic uncertainties, improving segmentation reliability. Extensive experiments on the BraTS datasets demonstrate that DFuse-Net outperforms state-of-the-art methods, suggesting its potential for reliable clinical application in brain tumor diagnosis and treatment planning.