Machine Learning-assisted Automated Quantification of Foveal Avascular Zone Parameters and Perifoveal Capillary Density of Prototype and Commercial Optical Coherence Tomography Angiography (OCT-A) Platforms in Healthy and Diabetic Eyes

Abstract

Purpose : To demonstrate the utility of an automated method in quantifying foveal avascular zone (FAZ) parameters and perifoveal capillary density of OCT-A images in healthy eyes and eyes with diabetic retinopathy (DR).

Methods : 74 OCT-A images of the foveal region were acquired using either a 1060nm Swept-Source OCT prototype or the commercially available RTVue XR Avanti system (Optovue, Inc). 21 healthy eyes and 9 eyes with DR were imaged on the prototype system, while 32 healthy and 9 with DR were imaged on the commerical system. The microvasculatures in the enface angiograms were then manually segmented by two blinded, trained raters. Automated segmentation classified each pixel as a vessel or non-vessel class using deep neural networks (DNNs). FAZ morphometric parameters (area, min/max diameter, and eccentricity) and perifoveal capillary density were used as outcome measures for the image processing pipeline.

Results : The accuracy (healthy: 0.80, DR: 0.83), sensitivity (healthy: 0.76, DR: 0.76) and specificity (healthy: 0.87, DR: 0.91) of the algorithm on the commercial system was comparable to the accuracy (healthy: 0.80, DR: 0.82), sensitivity (healthy: 0.81, DR: 0.72) and specificity (healthy: 0.79, DR: 0.87) on the prototype system (Figure 1). Comparing manual and automated segmentations, no statistically significant difference existed between the means of any FAZ morphometric parameters or perifoveal capillary density in either system. Correlation of these clinical measures between automated and manual segmentations was strong for the commercial system (r>0.7, p<0.01) but poorer for the prototype. Eyes with DR had significantly lower perifoveal capillary density (p<0.01), greater maximum diameter (commercial p=0.031, prototype p<0.01), and greater eccentricity (p<0.01) compared to healthy eyes.

Conclusions : The DNN based automated segmentation of OCT-A may be suitable for both commercial and research purposes for better characterization of the FAZ and quantification of the retinal capillary density in healthy subjects and in subjects with retinal vascular disease.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.

 

Figure 1. Samples images of healthy eyes from the prototype and commercial OCT-A systems, with corresponding manual and automated vessel segmentations. Automatic segmentations had 80-83% accuracy.