Purchase this article with an account.
Ylenia Giarratano, Alisa Pavel, Jie Lian, Rik Sarkar, Laura Reid, Shareen Forbes, Baljean Dhillon, Tom MacGillivray, Miguel Oscar Bernabeu; Optical coherence tomography angiography (OCTA) analysis of the diabetic eye: network-level vascular changes and patient classification. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4105.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The potential of OCTA for the diagnosis of diabetic eye disease via retinal vascular dysfunction depends on the availability of accurate and reproducible image quantification metrics. In this study, we propose novel vascular metrics based on graph theory and implement machine learning approaches to determine the diabetic retinopathy (DR) status of patients.
OCTA scans (RTVue-XR Avanti) were acquired as part of a longitudinal study that aimed to assess the rate of progression of DR. Baseline data consisted of 58 images from one eye of: 33 controls, 12 diabetic individuals without retinopathy (NoDR), and 13 diabetic individuals with retinopathy. Image segmentation was performed by vessel enhancement (using optimally oriented flux), followed by thresholding. The vascular network was then converted into a graph representation and for each region of interest (foveal, nasal, inferior, temporal and superior) we computed network-based metrics, such as graph density, graph clustering coefficient; coordinate based metrics, such as intercapillary area and perimeter; and topological metrics, graph asymmetry and graphlets distribution. T and Wilcoxon tests (with multiple comparison correction) were used for pairwise vascular metric comparison between groups. Finally, machine learning methods were trained, based on discriminative metrics, to classify patient images into the clinical groups. Performance was evaluated with area under the curve (AUC).
Statistical significant differences in number and size of intercapillary spaces (p=0.036 and p=0.028, respectively), and nestedness of the vasculature (p=0.039) were observed when NoDR and DR patients were compared. Fovea metrics, such as axis ratio (p=0.041), acircularity (p=0.012) were significant for DR and controls comparison. Logistic regression provided the highest AUC score (0.97) in the classification of diabetics and controls. Support vector machines reached the best performance in distinguishing DR from NoDR (AUC=0.82).
Our novel network-level metrics showed promise with regards to characterising diabetic retinal vascular dysfunction and stratifying patient status. Further studies should validate our result on larger, independent cohorts. These data support prospective investigations to determine the potential of OCTA for the prediction of diabetic retinopathy onset and progression.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only