Abstract
Purpose :
Chronic kidney disease (CKD) is associated with long-term morbidity and mortality. Kidney donation is also now recognised to be associated with long-term risk, albeit to a lesser extent. Previous studies have shown reduced retinal vascular density in CKD. To date, there are no studies of retinal vascular density following kidney donation. Using OCTA, we explored changes in the retinal vasculature in health, CKD and following kidney donation.
Methods :
We recruited 30 healthy volunteers, 18 CKD patients and 24 kidney donors to a prospective cross-sectional study. Image segmentation was performed using UNet, a deep learning architecture designed specifically for medical images. The binary image was converted into a graph object, which was divided into five regions of interest: foveal, nasal, inferior, temporal, and superior. For each region, we computed graph-based metrics (including network density and branching points). Metrics associated with the hierarchical structure and nestedness of the network were calculated from a binary tree representation of the graph. For the foveal region, in addition we calculated, among others, area, perimeter and axis ratio of the foveal avascular zone (FAZ). After removal of correlated features, different machine learning approaches were investigated to attempt automated classification of subjects into groups using area under the curve (AUC) as the evaluation metric.
Results :
Compared to health, CKD and kidney donors had reduced vascular density and branching points (p<0.05 for both). Metrics of network nesting were significant between healthy subjects and kidney donors (p<0.01). FAZ measures did not differ between groups. A Naïve Bayes classifier best discriminated healthy subjects from CKD (AUC 0.76). A decision tree in combination with linear discriminant analysis best discriminated health from kidney donors (AUC 0.80).
Conclusions :
As shown previously, we report a reduced vascular density in CKD. Interestingly, we show a previously unreported comparable reduction in kidney donors. Our analyses show that retinal imaging is able to discriminate healthy volunteers from both CKD patients and kidney donors. Ongoing studies will determine the predictive value of these OCTA metrics.
This is a 2020 ARVO Annual Meeting abstract.