Abstract
Purpose :
Although the determination of the eye’s refractive status is a standard procedure in ophthalmology and optometry, uncorrected refractive errors still represent the most common cause of visual impairment. Therefore, there is a need to develop cost effective, portable, and autonomous systems that measure refractive error. The purpose of this study is to present a new method for determining refractive errors using a retinal pattern projection, an imaging system, and a neural network (NN).
Methods :
This system consisted of a retinal projector display (Brother Air Scouter, Japan) to project patterns on the retina of an eye model, and an ophthalmoscope lens and camera (Basler AG, Germany) to acquire the retinal images. Twenty different patterns were imaged on an eye model’s retina (Ocular Instruments Inc, WA). In addition, to generate the data set with a set of spherical refractive errors, different trial lenses were added to the eye model to recreate the effect of the aberration. Thirty-two trial lenses ranging from -4D to 4D in steps of 0.25D were used to generate 640 images. Data augmentation techniques such as adjusting image rotation and brightness levels was used to increase the data set. A total of 43,560 images were employed to train the VGG16 NN in MATLAB utilizing transfer learning to predict the refractive errors present. 70% of the images were used for training the network and 30% for validating the training.
Results :
The training of the NN using the fundus images of the projected pattern had a percentage accuracy of 99.6%. Validating the network with two different set of images resulted in 84 % and 96% accuracy within one quarter of diopter of the spherical aberrations added to the model.
Conclusions :
Neural networks can be used to predict refractive errors from retinal images. However, the data set needs to be validated and improved. Next steps in the study will be to acquire images in an astigmatic eye model as well as obtaining retinal images of human eyes with different refractive errors. Nevertheless, this study presents the application of retinal projection display technology, fundus imaging and artificial intelligence to predict refractive errors.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.