Author’s Disclosure Block: Charley Cai, none; Hoyoung Jung, none; Yudan Chen, none; Tiffany Tse, none; Jiwon Hwang, none; Myeong Jin Ju, none; Zaid Mammo, none

Abstract Body
Purpose:The vascular theory of glaucoma posits that dysregulated ocular blood flow leads to ischemic optic nerve damage. We compare a novel voxel-based 3D algorithm with our previous 2D en-face method, evaluating its ability to capture vascular perfusion heterogeneity in serial OCTA of glaucoma and control subjects with higher spatial fidelity.Design: Cross sectional imaging study. Methods: UBC research ethics approval and signed informed consent from all subjects were obtained. Ten serial OCTA acquisitions were captured for each subject using the Zeiss PlexElite9000 OCTA system, focused on a 3x3mm area centered at the fovea. Subjects included healthy controls and patients with open-angle or normal-tension glaucoma (OAG/NTG) with no vasoactive medications or comorbidities. The 2D algorithm calculates a per-pixel Coefficient of Variation (CoV), defined as the ratio of the standard deviation to the mean intensity, across the ten serial 2D en-face projections. The serial acquisitions are aligned spatiotemporally to counter motion artifacts between scans and retinal vasculature was segmented using a Deep Neural Network. Our new 3D volumetric algorithm calculates CoV on a per-voxel basis before the en-face projection, preserving depth information to better stratify vessel layers via an additional alignment step in the z-axis applied across B-scans to correct depth discrepancies. Paired student t-tests were used to compare the mean and standard deviation (SD) of CoV calculations and to evaluate the differences in normalized CoV residuals to assess their single gaussian goodness-of-fit. Results: 6 patients (2 controls, 2 OAG, 2 NTG) with mean age 61 (range 33-85) (2M, 4F) were included. The mean CoV value (mean (SD)) across patients for the 2D and 3D algorithms was 0.23 (0.029) and 0.38 (0.032) respectively (p < 0.001), with the mean CoV SD at 0.095 (0.012) and 0.12 (0.011) (p <0.001). Our new method also demonstrates excellent internal validity, with highly repeatable mean CoV variability below 4.9% (0.03) across two control subjects imaged in quadruplicate. Analysis of the CoV distribution using the new algorithm revealed a notable divergence from a single Gaussian model with significantly larger normalized residuals at 0.14 (0.038) and 0.28 (0.052) respectively (p = 0.01). Conclusion: Our 3D algorithm advances the assessment of perfusion heterogeneity by distinctly revealing superficial and deep retinal vessels, overcoming limitations of prior 2D methods. Its proficiency in identifying a non-gaussian CoV distribution in complex vessel structures enables tracing flow changes to specific capillary network layers. This advancement creates new opportunities to investigate how such heterogeneity might indicate critical retinal vascular phenomena, enhancing the diagnosis and staging of ocular conditions.