Automatic diagnosis of retinal disease with retinal hemorrhage using ultrawide-field fundus ophthalmoscopic images by deep learning

Abstract

Purpose : We evaluated whether retinal diseases images with retinal hemorrhage (branch retinal vein occlusion (BRVO), central retinal vein occlusion (CRVO), diabetic retinopathy (DR) and age-related macular degeneration (AMD)), and normal fundus images of ultrawide-field fundus ophthalmoscopy could be classified by deep learning (DL).

Methods : The study included 2709 images of ultrawide-field fundus ophthalmoscopic images with retinal hemorrhage (BRVO: 203 images, CRVO: 87 images, DR: 2282 images and AMD: 137 images) from 1064 patients (age: 68.3±10.9) and 2093 images of normal ultrawide-field fundus ophthalmoscopic images from 1289 healthy subjects (age: 68.5±10.8). We constructed a convolutional neural network model called Visual Geometry Group-16 and performed binary classification of retinal disease images and normal fundus images. The trained model of DL was diverted from the neural network structure and verified by the K-Fold cross validation (K = 5). Sensitivity and specificity were calculated for the DL model. Furthermore, receiver operating characteristic curve was prepared, and the area under the curve was calculated.

Results : The sensitivity of the DL model for automatic classification of retinal disease was 84.5% (95% confidence interval [CI]: 83.1-85.8%), specificity was 87.8% (95% CI: 86.3-89.1%), and the area under the curve was 0.924 (95 % CI: 0.914-0,935).

Conclusions : DL model can classify retinal hemorrhage disease images with BRVO, CRVO, DR and AMD, and normal fundus images. It may be helpful in diagnosing and screening various diseases associated with retinal hemorrhage.

This is a 2020 ARVO Annual Meeting abstract.