Abstract
BACKGROUND: Optical Coherence Tomography Angiography (OCTA) provides high-resolution visualization of retinal microvasculature, with vascular density (VD) serving as a one of the key quantitative metrics. However, VD measurements are highly sensitive to image binarization step, and no standardized approach exists. METHODS: We analyzed 51 OCTA scans (human and porcine) using 29 binarization algorithms, including global and local thresholding techniques from ImageJ and DoxaPy, as well as Random Walker segmentation. VD was calculated for each binarization algorithm and compared against Optovue-generated values (ground truth). Results were evaluated using hierarchical clustering and agreement between them was determined by Bland-Altman analysis. RESULTS: Wolf algorithm was found to exhibit least deviation from mean Optovue VD values for human SCP layer (46.5 ± 1.2% vs. 48.3 ± 1.4%; p = 3.62 x 10 (-5) ); however, there is not significant difference between VDs from Optovue and Wolf algorithms from porcine SCP layer (46.2 ± 1.8 % vs 46.3 ± 1.4% ; p =0.74). For DCP layer, Phansalkar algorithm exhibited least VD variability (50.7 ± 2.0% vs. 51.9 ± 1.7%; p =2.53 x 10 (-4) ) in the human cohort. Whereas Percentile algorithm exhibited least, non-significant variations in the porcine DCP layer VD (50.0 ± 1.4 % vs. 50.3 ± 1.4%; p =0.75). DISCUSSION: Each binarization technique evaluated in this study impacts OCTA-derived VD measurements differently. Local adaptive algorithms collectively outperform global methods, particularly for SCP analysis. Standardization of image processing pipelines and layer-specific optimization are essential to improve reproducibility and clinical consistency.