.png){kind=icon}
.png){kind=icon}
Return
MFF-Diff: A multi-level feature fusion conditional diffusion model for PET/CT tumor segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asoc.2026.116260.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recently, Denoising Diffusion Probabilistic Models (DDPMs) have shown strong potential for improving the accuracy of PET/CT tumor segmentation. However, existing methods often neglect modality-specific features and introduce conditional cues only through fixed fusion schemes, limiting their ability to fully exploit the complementary strengths of PET and CT throughout the diffusion process. To overcome these limitations, we propose MFF-Diff. Specifically, MFF-Diff introduces a modality-specific feature encoding module to enhance the representation of metabolic hotspots and anatomical boundary details. Building on this, it further employs a two-stage, task-aware conditional fusion strategy that adaptively adjusts cross-modal fusion weights guided by feedback from the primary segmentation loss. Furthermore, we design a global-structure fusion layer to enhance long-range dependency modeling and global structural awareness. Experiments on the HeadNeck and STS datasets demonstrate that MFF-Diff consistently outperforms state-of-the-art methods. Specifically, MFF-Diff achieves Dice scores of 68.05% and 80.31% on the HeadNeck and STS datasets, respectively, outperforming the best-performing competing method by 2.57 and 3.34 percentage points. The corresponding Sensitivity scores are 81.59% and 86.22%, and the Precision scores are 67.54% and 73.44%.
Journal
IF:
6.6
Papers:
1.4W
Citations:
4.8W