Effects of the variability of automated segmentation algorithms on longitudinal analyses: application to the estimation of ganglion cell loss in healthy and glaucoma eyes

Abstract

 

Purpose

 

To estimate and compare the rate of change of ganglion cell layer (GCL) and ganglion cell + internal plexiform layer (GCIPL) thickness in healthy and glaucoma eyes using a novel regression method to reduce the effect of measurement variability.

 

Methods

 

The new Spectralis spectral domain OCT built-in automated segmentation software provides macular GCIPL thickness and GCL thickness. To track changes in OCT measurements over time, the ordinary least squares (OLS) regression is usually used. However, the OLS does not take into account measurement variability, which may affect the estimation of the regression model parameters in particular for the layers where retinal layer segmentation is challenging such as the GCL. We formulated the regression task as a missing data problem where we jointly estimate the slope and intercept and the corrected measurements. To deal with measurement variability, we used a Metropolis-Hasting algorithm to sample the measurements from their posteriori density distribution. Glaucoma and healthy participants from the Diagnostic Innovations in Glaucoma Study (DIGS) with good quality macula volume scans obtained at 3 month intervals were included. Differences between the rates of change (slopes) of GCL and GCIPL global thicknesses in glaucoma and healthy eyes using this approach were compared to those obtained using OLS regression.

 

Results

 

56 eyes of 28 healthy subjects and 73 eyes of 41 glaucoma patients were included. Mean age was 47.3 years (range 24.3-67.0 years) in the healthy group and 68.1 (range 50.8-92.4 years) in the glaucoma group. The median (interquartile range) follow-up time was 2.2 years (2.2-2.3) for healthy eyes and 2.3 years (2.1-2.5) for glaucomatous eyes. The GCL and GCIPL slopes estimated using the OLS method has a wider distribution compared to those obtained with the proposed regression approach (Table 1). Table 2 shows the percentage of glaucoma eyes with a rate of change faster than 95% percentile of normal eyes using the proposed approach and the OLS approach.

 

Conclusions

 

The proposed regression approach reduces the effect of variability in retinal layer segmentation on the estimation of ganglion cell loss in healthy and glaucoma eyes compared to the OLS method. Using this method GCL and GCPIL global thickness identify a similar proportion of progressing glaucoma eyes.