Abstract
Purpose :
Optical coherence technology (OCT) provides the ability to measure ganglion cell + inner plexiform layer thickness (GCIPL) and visualize the various tissues within the retina. GCIPL thickness is measured as the difference between the inner boundary of the retinal nerve fiber layer (RNFL) and the outer boundary of the inner plexiform layer (IPL). We have developed a multi-retinal layer segmentation algorithm (MLS) that segments the RNFL and IPL as well as other layers. The purpose of this study is to evaluate the performance of the GCIPL using MLS algorithm.
Methods :
Subjects with no retinal disease and with glaucoma were recruited from 2 sites and scanned with the 200x200 and 512x218 macular cube scans over 6x6 mm using the CIRRUSTM HD-OCT 5000 (ZEISS, Dublin, CA). Accuracy of MLS was compared against RNFL and IPL segmentations hand-drawn by a grader trained in segmenting OCT volumetric images (gold standard). The grader drew RNFL and IPL boundaries over three specified B-scans (at 2.1mm, 3 mm, 3.89 mm) per OCT volume (512x218 macular cube scans). Bland Altman and regression plots are reported on a total of 132 B-scans. The repeatability, coefficient of variance (COV) and receiver operating characteristic (ROC) curve for each subfield of the EDTRS grid centered on the fovea is reported from 49 normal subjects (282 cube scans) and 49 glaucoma subjects (244 cube scans).
Results :
In the Bland Altman and regression plots (Fig 1), the overall mean difference between MLS and manual segmentation for RNFL and IPL was found to be less than +/- 3 µm. R2 values show a very high correlation between gold standard and MLS. The 95% limits (+/- 1.96 times standard deviation) of the difference between gold standard and MLS calculated for RNFL and IPL are also reported. The area under the curve (AUC) was calculated in all subfields (Fig 2). The minimum AUC was 0.94 (subfield 3) and the maximum AUC was 0.99 (subfield 6). The repeatability of normal and glaucoma cases were calculated separately for all 6 subfields (Fig 2). The COV for normal cases was less than 2.4%. The largest COV for glaucoma cases was 6% (subfield 4).
Conclusions :
Our proposed segmentation algorithm achieves high accuracy, repeatability and reproducibility for GCIPL thickness analysis. It is therefore a promising tool to detect glaucoma and monitor its progression using OCT imaging.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.