Hierarchical Disentanglement Guided Diffusion for Multimodal Brain Tumor Segmentation.

Brain tumor segmentation from magnetic resonance imaging (MRI) is crucial for diagnosis and treatment. However, it faces the challenge of multi-level se mantic perception. The challenge includes local modality specificity and global subregion heterogeneity. Diffusion based segmentation methods can potentially address this challenge, but they are hindered by two critical issues: in adequate modeling of modality interactions and inappropri ate coupling of structural conditions with denoising stages. In this paper, we propose a hierarchical disentanglement guided diffusion framework (HD-Diff) for multimodal brain tumor segmentation. We introduce a modality-aware en coder to harmonize and integrate multimodal features. A dual-stream feature fusion module is employed to integrate extracted multimodal features across multiple scales. This module eliminates semantic misalignment and spatial in consistencies between modalities caused by differences in imaging principles. We utilize boundary and core enhanced conditioners to explicitly encode structural prior information about the tumor boundary and core regions. Thenboundaryand corefeatures are hierarchically injected into the denoising process. This strategy achieves independent optimization of both the overall tumor topology and local edge details. Extensive experiments are conducted on multiple public brain tumor segmentation datasets. Across these five benchmarks, HD-Diff achieves strong overallseg mentation performance, with highly competitive Dice and HD95 results. The consistent results further demonstrate its effectiveness across diverse brain tumor datasets.