Abstract
Purpose :
To develop a neural network-based method for adjusting of the dependence of the peripapillary retinal nerve fiber layer (RNFL) profile on age, gender, and ocular biometric parameters, and to evaluate its performance in glaucoma diagnosis, especially in high myopia.
Methods :
Participants were randomly selected from the population-based Beijing Eye Study 2011. A total of 2477 non-glaucomatous participants and 254 glaucoma patients were included. A detailed ophthalmic and systematic examination included RNFL profile measurement using spectral-domain optical coherence tomography, with 768 points at circumferential 3.4mm. In a test group of 2223 non-glaucomatous eyes, a fully connected network and radial basis function network were trained for the vertical (scaling) and horizontal (shift) transformation of the RNFL thickness profile with adjustment for age, axial length (AL), disc-fovea angle and disc-fovea distance. The performance of the RNFL thickness compensation was evaluated in an independent validation group of 254 glaucoma patients and 254 non-glaucomatous participants.
Results :
Applying the RNFL compensation algorithm improved the area under the receiver operative characteristic curve in detecting glaucoma from 0.703 to 0.842, from 0.748 to 0.889, from 0.767 to 0.891, and from 0.779 to 0.867, for eyes in the highest 10% (mean: 26.0±0.9mm), 20% (25.3±1.0mm), and highest 30% (24.9±1.0mm) percentile subgroup of the AL distribution, and in the eyes of any AL (23.5±1.2mm), as compared with the unadjusted RNFL data, respectively. The difference between the uncompensated versus compensated RNFL thickness values expressed in relative percentage points increased with longer axial length: it increased by 19.8%, 18.9% and 16.2% in the highest 10%, 20% and 30% percentile subgroups, respectively.
Conclusions :
In a population-based study sample, an algorithm-based adjustment for age, gender, and ocular biometric parameters improved the diagnostic precision of the peripapillary RNFL thickness profile for glaucoma detection in particular in myopic and highly myopic eyes.
This is a 2020 ARVO Annual Meeting abstract.