扩散
材料科学
物理
噪音(视频)
声学
机械
数学
计算机科学
数学模型
核磁共振
作者
Yi Zhao,Qirui Fan,Xiaohui Duan,Donglong Chen
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-21
卷期号:: 8157-8161
标识
DOI:10.1109/icassp55912.2026.11464098
摘要
Recent advances in diffusion models have greatly improved medical image generation, yet enforcing precise anatomical control in multi-region brain tumors remains challenging. We propose SegMaDiff, a segmentation-guided diffusion framework inspired by Mamba state-space modeling. By integrating tumor subregion masks at each denoising step and introducing the Mamba Dynamic Former—combining Omni Scan recurrence, Gaussian high-pass filtering, and Dynamic Tanh activation—our method captures complex spatial dependencies while preserving fine structures. Experiments on the BRATS-2023 and BRATS-2024 datasets confirm that SegMaDiff achieves SOTA performance, exhibiting the lowest FID scores and the highest segmentation accuracy for the tumor subregion across existing methods. Specifically, SegMaDiff attains an FID of 0.0031 on BRATS-2024 and 0.0037 on BRATS-2023. In addition, our model has restoration capability: SegMaDiff can reconstruct corrupted MRI slices using only segmentation masks as guidance. This work highlights the potential of combining anatomical priors with dynamically inspired generative modeling for advancing medical image synthesis.
科研通智能强力驱动
Strongly Powered by AbleSci AI