.png){kind=icon}
.png){kind=icon}
Return
Multi-scale shape-aware diffusion models for retinal biomarkers segmentation in OCT images
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bbe.2026.07.001.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Retinal biomarker segmentation in optical coherence tomography (OCT) images can provide critical pathomorphological indicators for ocular and systemic disease progression, aiding clinicians in diagnosis. However, retinal biomarker segmentation remains challenging due to the large variations in biomarker scale and shape, as well as the presence of noise and artifacts in OCT images. To address these challenges, we propose a multi-scale shape-aware network based on a Diffusion model for Retinal Biomarker Segmentation (DiffRBSeg). This framework incorporates three novel components specifically tailored for retinal biomarker segmentation: the Cross-Attention Fusion Module (CAFM), the Wavelet Enhancement Module (WEM), and the Dynamic Deformable Convolution Module (DDCM). Specifically, we first propose CAFM, which employs cross-transformer blocks to establish cross-feature global contextual information. This module is designed to promote global information interaction and fusion between the conditional input and noise representation, thereby guiding the network to focus on key regions. In addition, we design WEM to suppress interference and enhance the texture details of features. By incorporating wavelet transforms, this module interacts with multi-scale wavelet-domain features to refine representations and improve feature expressiveness. Furthermore, the proposed DDCM uses timestep-adaptive dynamic weighting to fuse multi-scale deformable convolution blocks, enabling the capture of diverse morphological characteristics of biomarkers and producing more accurate shapes and boundaries. Comprehensive evaluation on two clinical datasets demonstrates that our method achieves superior performance compared with existing methods on the evaluated datasets.
Journal
IF:
6.6
Papers:
937
Citations:
3.3K