Purchase this article with an account.
Philipp M Prahs, Caroline Brandl, Christian Mayer, Yordan Cvetkov, Viola Radeck, Horst Helbig, David Märker; Deep Learning for the Prediction of Visual Function using Macular OCT Scans. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1509.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Optical coherence tomography (OCT) scans of the central retina provide detailed morphological information and are widely used by clinicians. The degree of correlation between retinal morphology and visual function is often used in practice as a measure of certainty by which a reduction of visual acuity (VA) is explained and whether additional diagnostics should be performed. To follow this thought in an automated system, we trained a deep convolutional artificial neural network to predict VA from the retinal anatomy provided by OCT.
We identified medical reports written at our institution in the years 2008 to 2016 that referred to patient visits were OCT images were acquired. After extraction of VA information and OCT images, we divided the available data into training and test datasets. We trained a deep convolutional neural network of the Inception V4 type using the Google TensorFlow framework by minimizing the mean squared difference of documented VA and neural network output. A retinal specialist decided whether or not each individual B-scan subjectively explains the amount of observed visual impairment for a small sample (n=100) of images with both low and high prediction errors in the test set. A chi squared test was performed to test for a difference in frequencies.
A total of 72,282 individual medical reports with accompanying OCT B-scans were identified. The neural network was successfully trained using 276,930 individual B-scan images. After training predictions were tested on 43,368 individual OCT B-scans in the test dataset by assessing the difference of actual visual acuity and predicted values. The median of the observed difference was 0.15 decimal units, 75% of the test dataset had a prediction error of less than 0.27 decimal units. The retinal specialist rated 84% vs. 14% of images as sufficiently explaining VA from the groups with low and high prediction errors. The difference was statistically significant (p<0.001).
After training with historical clinical data, artificial intelligence systems are able to predict visual function from retinal morphology acquired by OCT within limits. Cases with high prediction errors are likely to exhibit a low amount of correlation between visual function and retinal anatomy. An automated system that makes this information available to the treating physician might prove useful in practice.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only