Abstract
Purpose :
To develop a deep learning algorithm to automatically diagnose glaucoma using a fundus photograph.
Methods :
Training dataset was consisted of 1,364 color fundus photographs with glaucomatous appearances and 1,768 color fundus photographs without glaucomatous appearances. Testing dataset was consisted of i) 34 eyes of 34 non-highly myopic glaucoma patients (G group), ii) 30 eyes of 30 highly myopic glaucoma patients (mG group), iii) 28 eyes of 28non-highly myopic normative subjects (N group) and iv) 22 eyes of 22 highly myopic normative subjects (mN group). In all eyes, fundus photographs were obtained using the non-mydatric fundus camera (nonmyd WX3D, Kowa ltd). Using the training dataset, a deep learning algorithm of the residual network (resnet) was developed to automatically diagnose glaucoma using a fundus photo. The diagnostic accuracy of the resnet was validated using the testing dataset, using the area under the receiver operating characteristic curve (AROC).
Results :
he AROC in all data was 95.4 %. The sensitivity at the specificity of 95.0 % was 70.3 %. In each testing group, the AROC was i) between G and N groups: 97.2 % and ii) between mG and mN groups: 94.1 %
Conclusions :
A deep learning algorithm was developed to automatically diagnose glaucoma using a fundus photograph. This method had a high diagnostic ability both in non-highly myopic and highly myopic eyes.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.