Abstract
Purpose :
To evaluate the ability of a novel Bayesian deviation map of macular ganglion cell complex (mGCC) thickness to discriminate between healthy and glaucomatous eyes. Our Bayesian modeling improves glaucoma diagnosis by accounting for the known structural relationship between the macula and the circumpapillary retinal nerve fiber layer (cpRNFL).
Methods :
The study included data from 1279 eyes of 393 patients for which mGCC scans were available. The most recent pairs of images for the left and right eye was used per patient for cross-sectional analysis. Macular thickness measurements were analyzed on the 8-by-8 posterior pole grid, while cpRNFL measurements were derived from the 768-point circle scan. A Bayesian linear mixed model was constructed with a novel prior incorporating clinical knowledge of the structural relationship. We compared the Bayesian deviation map to a raw version to assess discrimination between healthy and glaucoma cases. Abnormality was defined as the proportion of negative deviations. Discrimination was measured using both area under the receiver operating characteristic curve (AUC) and partial AUC (pAUC) in the clinically relevant 85-100% specificity range.
Results :
Of the 393 pairs of images , 333 were those of glaucoma patients and 60 of healthy ones. The Bayesian model was trained on a random 80% subset, with 20% held out for testing. Five-fold cross validation (CV) was used for more robust assessments. There was a notable improvement in the average pAUC of the Bayesian deviation map (0.49; 95% confidence interval [CI] 0.32-0.75) compared to the raw map (0.44; CI 0.25-0.79). For 95% specificity, average sensitivity of the Bayesian map was .42 (CI 0.30-0.75) versus .32 (CI 0.23-0.78) for the raw map.
Conclusions :
A novel Bayesian deviation map incorporating structural relationships between macula and peripapillary RNFL performed significantly better to detect glaucomatous damage in the macula.
This is a 2021 ARVO Annual Meeting abstract.