Retinal layer segmentation results differ between two generations of OCT devices

Abstract

Purpose : As spectral domain optical coherence tomography (SD-OCT) has become routine in clinical practice and research, and newer OCT devices are replacing first generation OCTs, it is essential to verify their output. We studied retinal layer thickness in healthy eyes of two comparable populations obtained with a first and a second generation OCT device. Additionally, we quantified differences in image quality between the two systems.

Methods : OCT data were obtained in the population-based Rotterdam Study (RS) with the TOPCON 3D-OCT1000 in 2007 (N=2785, age range: 47-96 yrs, mean 67.8; mean axial length 23.6mm; 56% ♀), and the TOPCON 3D-OCT2000 from 2008 onwards (N=1047 , age range: 52-99 yrs, mean 68.7; mean axial length 23.6mm; 54% ♀). Scans were segmented with the validated Iowa Reference Algorithms 3.6 from the University of Iowa. A dataset in which eyes from each subject were scanned using both OCT systems was collected in the Beaver Dam Eye Study (University of Wisconsin). We studied differences in image quality between the two systems by comparing pixel intensity histograms using the variational distance metric.

Results : Systematic differences in average retinal layer thickness were observed between the two OCT systems. Sign and magnitude of this bias was layer dependent. The ganglion cell layer measured 32.5 ± 5.8 µm (mean±SD) with the OCT1000 and 36.2 ± 5.3 µm with the OCT2000, outer nuclear layer 87.2 ± 8.8 µm vs. 90.5 ± 8.2 µm, and the RPE measured 19.5 ± 1.6 µm vs. 16.9 ± 0.9 µm (all P<<0.001). Biases of similar magnitude were found in the BDES dataset, although layer thickness correlated relatively well between the two OCT systems. One explanation for the observed biases could be a difference in image quality between OCT devices.Comparing pixel intensity distributions indeed showed qualitative differences, with the second generation OCT system having less variance in dark region intensities and fatter high intensity tails. The retinal layer segmentation algorithm was robust to histogram rescaling and white noise image degrading procedures, applied to make image quality more uniform.

Conclusions : Replacing OCT devices with newer versions could introduce unexpected differences in the estimated thickness of retinal layers, even when segmentation is performed with the same robust algorithm. Automated thickness parameters should therefore be interpreted with caution when comparing data from different devices.

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