Machine Learning Based End-to-End Pipeline for Optical Coherence Tomography Angiography of Diabetic Retinopathy

Abstract

Purpose : Machine learning for binary classification of diabetic retinopathy (DR) in fundus images has achieved high validation accuracies, however classification results of multiple severity levels are generally less impressive. The use of optical coherence tomography angiography (OCTA) has potential to improve classification. We investigate an end-to-end machine learning based pipeline for the volumetric OCTA images consisting of cascaded Deep Neural Networks (DNNs) for the segmentation of five retinal layers, intraretinal fluid, microvasculature, the foveal avascular zone (FAZ), and DR severity grading.

Methods : 3x3mm volumetric scans centered on the fovea were used from two commercial OCTA systems: AngioPlex (Cirrus 5000 HD-OCT; Carl Zeiss Meditec), and PlexElite (Carl Zeiss Meditec). The first DNN, which segmented 5 retinal layers and intraretinal fluid, was trained on 14,700 OCT Bscans. The vessel and FAZ segmentation DNN was trained on 80 en face OCTA images. Both segmentation DNNs were modified UNET architectures. Seven FAZ metrics were calculated (area, perimeter, minimum and maximum diameter, axis ratio, eccentricity, acircularity index) along with four vascular parameters (density, skeleton density, fractal dimension, and average vessel diameter). Classification of DR severity (no DR, non-proliferative DR, proliferative DR) was performed separately.

Results : Figure 1 shows representative results from the stages of the pipeline. Figure 1 (a-b) is the en facesuperficial and deep OCTA, (c) a representative Bscan with the layer and fluid segmentations, (d-e) superficial and deep vessel segmentation maps with FAZ min (red) and max (green) diameter, and perimeter (yellow), (f) FAZ segmentation, (g) inner ring density, (h) central circle density, (i) en face fluid segmentation map.
The proposed layer and fluid segmentation DNN achieved a Dice index of 0.942 for RNFL, 0.970 for GCL, 0.957 for INL, 0.968 for OPL, 0.959 for the outer retina, and 0.786 for fluid. For the vessel and FAZ segmentation, the Dice index was 0.836 for the microvasculature and 0.957 for the FAZ. All segmentation performances were comparable to a second manual rater. The classification DNN achieved an accuracy of 88%.

Conclusions : Machine learning based segmentation enables reliable quantification of retinal OCTA scans. The results are encouraging for improving the classification towards a clinical severity scale.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.