BH-CDiff: Boundary-aware and heterogeneous feature synergy cold diffusion for medical image segmentation.
Authors
Affiliations (4)
Affiliations (4)
- School of Computer and Artificial Intelligence, Beijing Technology and Business University, Beijing, 100048, China. Electronic address: [email protected].
- School of Computer and Artificial Intelligence, Beijing Technology and Business University, Beijing, 100048, China. Electronic address: [email protected].
- School of Computer and Artificial Intelligence, Beijing Technology and Business University, Beijing, 100048, China. Electronic address: [email protected].
- School of Computer and Artificial Intelligence, Beijing Technology and Business University, Beijing, 100048, China. Electronic address: [email protected].
Abstract
Recently, diffusion models have shown significant potential for medical image segmentation, providing improved interpretability through their iterative optimization process. To address the uncertainty introduced by Gaussian noise in traditional diffusion methods, the cold diffusion paradigm replaces stochastic noise with task-relevant perturbations, enabling more controllable mask generation. However, existing cold diffusion-based segmentation methods still face challenges in recovering boundary structures during iterative restoration and effectively exploiting dynamically evolving heterogeneous representations. This paper proposes a boundary-aware and heterogeneous feature synergy cold diffusion network (BH-CDiff). Specifically, we introduce a boundary-aware modeling mechanism through the boundary-aware feature modulation (BAFM) module, which uses boundary priors to guide channel-spatial attention and effectively enhance boundary information. The heterogeneous feature synergy mechanism is realized through the multi-domain representation synergy (MDRS) module and the dual-source feature fusion (DSFF) module. The former jointly models frequency-domain and spatial-domain features for synergistic multi-domain representation, while the latter fuses segmentation features with conditional priors to promote complementary enhancement of dual-source information. Extensive experiments on the ISIC 2018, CVC-ClinicDB, and REFUGE datasets demonstrate that the proposed method attains state-of-the-art results.