Abstract
Purpose :
To develop a deep-leaning framework to estimate the best-corrected visual acuity (BCVA) from optical coherence tomography (OCT) images.
Methods :
Data from 509 patients who visited the Jichi Medical University Hospital from 2004 to 2018 were used. For each eye, a maximum of five OCT (TOPCON Atlantis, Triton; Tokyo, Japan) images including one horizontal and one vertical scans were used. In total, 2,756 OCT images and the BCVA of 809 eyes acquired on the same-day were used for the analysis. To improve accuracy, 1) the last layer of the GoogLeNet deep learning architecture was modified to generate a single outcome from 4 parameters (originally 1,024 parameters) and 2) the Loss Layer was replaced with a Euclidian Loss. 10-fold cross-validation was used to evaluate the model and OCT images from each patient were grouped in separate batches. Standard error and contribution rate of OCT for BCVA were calculated.
Results :
Mean age of patients was 69 years (standard deviation; SD= 12 years), 45% were male (275 eyes) and mean logMAR BCVA was 0.33 (SD= 0.44). The dataset included 139 eyes with age-related macular degeneration, 45 eyes with diabetic retinopathy, 39 eyes with macular hole and epi-retinal membrane, 34 eyes with retinal vein occlusion, 20 eyes with central serous chorioretinopathy, 15 eyes with myopic choroidal neovascularization, 41 eyes with other eye disease, and 476 normal eyes. The mean difference and the standard error of the estimated BVCA from OCT images were -0.02 and 0.35 logMAR, respectively and the R2 was 0.39.
Conclusions :
Deep learning would be useful to estimate visual acuity using OCT images, with about 40% of the best corrected visual acuity to be estimated without other clinical information.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.