Evaluating the diagnostic performance with neurovascular segregation of the retinal nerve fibre layer in glaucoma

Abstract

Aims/Purpose: To evaluate the diagnostic performance in detecting eyes with glaucoma by segregating the neuronal and vascular components in circumpapillary optical coherence tomography (OCT) and OCT angiography (OCTA) scans.Methods: In this study, 93 participants with glaucoma were age‐ and gender‐matched with normal controls with no ocular diseases. OCT and OCTA volumetric images of one eye from each participant were acquired using a prototype high resolution swept source OCT imaging system (Zeiss PlexElite) using a 6 mm x 6 mm imaging protocol centered at the optic nerve head. Circumpapillary OCT and OCTA images with a diameter of 3.46 mm were subsequently generated. Information from the circumpapillary and enface OCTA images were used to isolate and segregate the vascular components from the circumpapillary retinal nerve fibre layer (RNFL), which was obtained using a deep learning segmentation approach. Thicknesses of the RNFL with and without the vascular components were measured and diagnostic performance was evaluated using area under the receiving operating characteristic curve (AUC) analysis. Both global and quadrant (superior, inferior, nasal and temporal) measurements around the optic nerve head were considered.Results: Imaging data of one eye from 154 healthy and 143 glaucoma participants were acquired. Using only the global RNFL thickness, the detection of glaucoma achieved an AUC of 0.877 ± 0.021. With the vascular components excluded from the global RNFL thickness, an AUC of 0.890 ± 0.020 was obtained. When the vascular and non‐vascular components of the RNFL in the different quadrants were modelled separately, AUC improved to 0.927 ± 0.015, and was found to be significantly higher than the detection of glaucoma using global measures (p < 0.001).Conclusions: Segregation of vascular and non‐vascular components in the retinal nerve fibre layer led to an improvement in glaucoma detection. Further improvements were observed when the contributions of the components in the RNFL quadrants were modelled separately.