Purchase this article with an account.
Julian Lo, Morgan Heisler, Donghuan Lu, Sonja Karst, Vinicius Vanzan, Eduardo V. Navajas, Marinko V. Sarunic; Deep Neural Network Segmentation of Optical Coherence Tomography Angiography for Diabetic Retinopathy. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB0113.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Diabetic retinopathy (DR) is a microvascular complication of diabetes manifesting in abnormal retinal circulation. The purpose of this study is to investigate the viability of machine learning algorithms in the automatic segmentation and quantification of retinal microvasculature and inter-capillary spaces in optical coherence tomography angiography (OCTA) images in clinical diabetic populations.
A convolutional neural network (CNN) was used to generate a binarized segmentation of the retinal microvasculature, followed by additional post-processing to analyze inter-capillary spaces. OCTA images of both healthy and diabetic patients across all levels of DR severity were acquired using three commercial systems (PLEX Elite 9000, Zeiss, Inc.; AngioPlex, Zeiss, Inc.; AngioVue, Optovue, Inc.), as well as one prototype swept-source OCT system. The retinal layers were segmented to separate the superficial and deep capillary complex. Manual segmentation of the microvasculature was performed on 80 images, of which 60 were used for training, and 20 for validation and evaluation. Inter-capillary spaces were identified based on two metrics: the area of the region, as well as the distance from the centroid to the nearest capillary. Results were compared to a reference set of 10 healthy subjects and labeled in standard deviation maps.
Across OCTA images from both healthy and diabetic patients, the CNN segmented the microvasculature with an accuracy of 86.77%, with a Dice similarity index of 83.94% to the manual segmentations. Representative performance of the automated segmentations is shown in Fig. 1 (a)-(d). Fig. 1 (a) shows the original OCTA image, (b) the microvasculature segmentation, and (c) and (d) the standard deviation map based on area and the distance to the nearest capillary, respectively.
The CNN-generated segmentations of the microvasculature had a high accuracy and similarity to manual segmentations across all OCTA systems. Automated quantification of inter-capillary spaces allows for accurate analysis of retinal circulation which is a biomarker for DR. Although the area measurements are useful in identifying large inter-capillary spaces, the distance to the closest capillary may present the highest clinical utility due to its direct correlation to the amount of circulation received.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.
This PDF is available to Subscribers Only