A software approach for the densification of sparse OCTA volumes to achieve 3D visualization of retinal vasculature
We developed a cost-effective, software-based densification pipeline to overcome the diagnostic limitations of standard 2D OCTA projections, enabling the comprehensive assessment of 3D retinal vascular structures without the need for expensive and inaccessible hardware-based 3D upgrades. We developed a software-based pipeline for 3D densification and reconstruction of retinal vasculature from sparse OCTA B-scan volumes, involving Hue, Saturation, Value (HSV)-based thresholding for layer and vessel segmentation, 5th-order spline interpolation along the y-axis, standard deviation projection for en face views, and affine registration. The method was applied to 304 B-scans per volume from 16 eyes (9 healthy controls, mean age 32 ± 5 years; 7 with diabetic retinopathy, mean age 48 ± 7 years, mild-to-severe NPDR) acquired using a 70 kHz SD-OCT system (RTVue-XR, Optovue, CA). Densification and reconstruction accuracy was evaluated against device-generated 2D en face projections using Dice scores and vessel density metrics, with layer-specific analysis for superficial, deep, and full retina. Post-interpolation, the method achieved mean Dice scores of 0.8321 ± 0.0148 (full retina), 0.7993 ± 0.0309 (superficial), and 0.6871 ± 0.0624 (deep), markedly improving from pre-interpolation values (e.g., full retina: 0.4028 ± 0.0161; p < 0.01). 3D vessel density more than doubled post-interpolation (e.g., full retina: from 0.0778 ± 0.0125 to 0.2095 ± 0.0331), resolving capillary discontinuities. Pathologic eyes showed a trend toward reduced 3D vessel density, particularly in the deep layer (healthy: 0.1803 ± 0.0446 vs. pathologic: 0.1584 ± 0.0330; p = 0.0703), consistent with microvascular dropout in diabetic retinopathy. However, these differences did not reach statistical significance (p > 0.05), likely due to the modest sample size. This hardware-independent software method enables detailed layer-specific 3D visualization of retinal vasculature from sparse OCTA B-scan volumes. By providing enhanced spatial context, it has the potential to support the clinical evaluation of vascular abnormalities. An interactive tool further allows real-time exploration, serving as a cost-effective bridge for 3D volumetric analysis in resource-limited clinics. However, because this approach currently relies on a manual data extraction workaround rather than fully automated workflow integration, and because its direct impact on diagnostic accuracy remains unproven, larger cohort validation and formal clinical diagnostic studies are warranted.