Artificial intelligence using a deep learning system with transfer learning to predict refractive error and myopic macular degeneration from color fundus photographs

Abstract

Purpose : The global burden of myopia is significant, and increasing. Patients with high myopia are at increased risk of vision-threatening complications, including myopic macular degeneration (MMD). We developed an artificial intelligence (AI) based deep learning system (DLS) using fundus photographs to predict refractive error, high myopia, and MMD.

Methods : The DLS consisted of a convolutional neural network (DenseNet) pre-trained with XGBoost algorithm, developed using retinal images from a multiethnic population-based cohort: the Singapore Epidemiology of Eye Diseases (SEED) cohort, with 15,876 (12,703 for training, 3,173 for testing) and 16,793 (13,433 for training, 3,360 for testing) images used for prediction of refractive error and MMD respectively. Images were classified by professional graders according to the International Meta-Analysis for Pathologic Myopia (META-PM) classification into 5 categories (grades 0 to 4), with MMD defined as category 2 or worse. The DLS was assessed in its ability to (1) predict spherical equivalent (SE), (2) detect high myopia (SE ≤ -6.00D), and (3) detect MMD. The area under the receiver operating characteristic (ROC) curve (AUC) was determined for these classifications.

Results : The DLS was able to predict SE with a mean absolute error (MAE) of 1.20D. The AUC, sensitivity and specificity (at the specified threshold) for detection of high myopia were 0.942, 91.7% and 83.5% respectively. The AUC, sensitivity and specificity for detection of MMD were 0.955, 91.9% and 87.7% respectively. The ROC curves are shown in Figure 1.

Conclusions : Using fundus photographs, the DLS developed was able to predict refractive error, and detect high myopia and MMD with a high degree of accuracy. Application of this DLS in screening programs could allow early identification of patients at high-risk for vision-threatening complications of myopia.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Figure 1. Receiver operating characteristic curves for detection of (A) high myopia and (B) myopic macular degeneration.