Annotation-free 3D reconstruction and quantification of retinal microvasculature by RADAR

Three-dimensional mapping of retinal microvasculature is essential for monitoring systemic vascular health. Existing methods rely heavily on manual annotation or training-intensive deep learning models that lack generalizability. Here we report RADAR. This is an annotation-free computational framework for the 3D segmentation and quantification of optical coherence tomography angiography data. The pipeline integrates adaptive physics-aware denoising with topology-preserving centerline extraction to reconstruct complex networks without manual labeling. We validated the framework in healthy individuals and patients with early-stage diabetic retinopathy. The method outperformed standard segmentation tools and resolved layer-specific morphological alterations obscured in conventional two-dimensional projections. Quantitative analysis revealed distinct patterns of compensatory remodeling and increased tortuosity in diabetic eyes. RADAR enables precise extraction of volumetric biomarkers including vessel length and branching complexity. It provides a scalable tool for early detection and longitudinal assessment of ocular and systemic vascular diseases.