.png){kind=icon}
Return
SFA-DiffNet: Spatial-frequency aware diffusion network for robust medical image segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.133529.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diffusion models have demonstrated significant potential in medical image segmentation, yet they remain challenged in capturing complex fine-grained structures. Existing spatial-domain backbones exhibit bias towards low-frequency components, while the denoising process often leads to over-smoothing of high-frequency details, compromising boundary precision and cross-domain generalization. To overcome these limitations, we propose a Spatial-Frequency Aware Diffusion Network (SFA-DiffNet) that jointly models spatial and frequency representations, augmented with conditional guidance from foundation models. This design improves detail preservation and boundary accuracy in our experiments. Specifically, we introduce a Spatial-Frequency Aware Encoder (SFAE) that decomposes features into complementary spatial and frequency representations. It consists of a Spatial Aware Block (SAB) for contextual information learning, a Frequency Aware Block (FAB) for spectral cues exploration, and an Adaptive Fusion Block (AFB) for complementary fusion. Moreover, we develop a foundation model guidance mechanism that dynamically injects semantic priors via an Adaptive Feature Projector (AFP) and Semantic Alignment Fusion Block (SAFB), ensuring alignment between conditional and denoising features. Extensive experiments on multi-modal medical imaging datasets show that SFA-DiffNet achieves competitive performance against existing methods and shows potential for cross-domain generalization. The source code is available at GitHub.
Keywords:
Medical image segmentation
Diffusion model
Deep learning
Wavelet transform
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
Organization
View more organizations
Cited Papers
No cited papers available