De-Seg: Advancing OCTA segmentation with multi-scale feature fusion and edge refinement

Accurate segmentation of retinal vessels in optical coherence tomography angiography (OCTA) is essential for the diagnosis and monitoring of ocular diseases such as diabetic retinopathy and glaucoma. However, this task remains challenging due to the presence of low-contrast microvessels, projection artifacts, and interference from fine vessel structures. To overcome these limitations, we propose De-Seg, a deep learning framework designed to enhance vessel segmentation through multi-scale feature fusion and edge refinement. The model employs a dual-branch encoder that combines a U-Net-style backbone with parallel multi-scale convolutions and a dedicated edge-aware branch utilizing Sobel operators. These branches are effectively fused via skip-connected decoder layers to maintain high spatial precision. The bottleneck is augmented with a multi-layer perceptron (MLP) to capture global contextual information, while a Multi-Scale Feature Extraction Block (MS Block) enlarges the receptive field to better capture faint capillaries. An Edge Block is introduced to enhance boundary clarity and mitigate interference from lesion regions. Extensive experiments demonstrate that De-Seg achieves competitive performance on multiple OCTA benchmarks, yielding Dice scores of 92.13% on OCTA-3M and 94.67% on ROSSA, outperforming recent state-of-the-art methods. Ablation studies validate the necessity of each component in improving feature representation and edge accuracy. The framework exhibits strong generalization capability, showing robust performance even under cross-domain settings. Overall, De-Seg provides a reliable and efficient solution for quantitative OCTA analysis, with potential applicability to other vascular imaging modalities.