WHANet: wavelet and hybrid attention network for vessel segmentation in OCTA fundus images

Shuxin Xue, Zhaohui Zhang, Fen Yan, Fei Ma, Guangmei Jia, Yanfei Guo, Yuefeng Ma, Xiaofei Ai, Jing Meng

Abstract: Optical coherence tomography angiography (OCTA) is a state-of-the-art, non-invasive imaging modality that enables high-resolution visualization of the retinal vasculature, playing a vital role in the early diagnosis of various retinal diseases, including diabetic retinopathy, glaucoma, and choroidal neovascularization. However, traditional OCTA image analysis methods face significant challenges due to the difficulty in detecting retinal microvasculature, the presence of blurred vessel edges, and the complex topological structure of vascular networks. These difficulties are further compounded by limited training data, the inadequate representational capacity of spatial-domain features alone, and the challenges associated with effectively integrating frequency-domain information. To address these limitations, this study introduces WHANet, a novel dual-branch segmentation framework that integrates spatial and frequency-domain feature representations. WHANet comprises two primary modules: the hybrid attention deep convolutional branch (HADCB) and the multi-scale wavelet feature fusion branch (MWFFB). The HADCB enhances spatial feature representation through multi-scale convolutional operations and attention mechanisms, improving the detection of fine vessels, reducing edge blurring and misclassification, and reinforcing local detail perception. In parallel, MWFFB leverages the discrete wavelet transform (DWT) to extract refined vascular structures while preserving global vessel morphology. By complementing spatial-domain features with frequency-domain information, MWFFB strengthens edge representation, enhances the separability of microvessels, and mitigates the effects of image noise. The synergistic integration of these two branches enables comprehensive feature extraction and fusion, significantly boosting the model’s ability to handle complex vascular networks. Extensive experiments conducted on three widely used public datasets demonstrate that WHANet consistently outperforms state-of-the-art methods across multiple evaluation metrics, exhibiting superior robustness and segmentation accuracy.

External IDs:dblp:journals/tjs/XueZYMJGMAM25