Abstract
Purpose :
Testing visual performance especially for different optical designs of presbyopia correction require long lasting clinical studies. The purpose of the current study was to train and validate a convolutional neuronal network (CNN) approach combined with psychophysical methods to predict subjective visual acuity (VA) from optical wavefront errors.
Methods :
A two-layer convolutional neural network (CNN) was trained to classify the gap orientations of Landolt rings with 14.000 artificially blurred images from acuity values of -0.2 to 1.2 logMAR. Eye’s optical errors (LOA and HOA) were applied using a Fourier approach for a 4mm pupil and a reference wavelength of 530nm. The network was trained using the machine learning toolbox in Matlab. The simulation combines the feature extraction abilities of a CNN with psychophysical methods of VA testing using a BestPEST staircase algorithm which is comparable to subjective procedures. The simulation was validated with subjectively determined VA assessed under defocus from ±1.5D collected from 39 normal eyes and on measurements with low contrast (0 % to 100 %, step size: 1 %) optotypes. Agreement between the simulation and the subjective measurement was calculated using a Bland-Altman analysis.
Results :
The simulation with additional spherical defocus resulted in an offset of 0.20 logMAR ±0.035 logMAR between the subjectively measured data and the predicted values, which was shown to be constant over the simulated spherical blur. Correcting this offset, the limits of agreement were -0.08 logMAR and 0.07 logMAR. Low contrast acuity with a maximum error of 0.09 logMAR for contrast levels greater than 35%, however revealed an increased error for lower contrast levels.
Conclusions :
Combined algorithmic approach utilizing convolutional neuronal networks and psychophysical staircase methods can be used to predict Landolt C visual acuity. The proposed approach can be applied to test variety of optical corrections for presbyopia prior to a costly and time-consuming clinical evaluation of such.
This is a 2020 ARVO Annual Meeting abstract.