Abstract
Purpose :
Occult macular dystrophy (OMD) is a progressive and hereditary type of macular dystrophy, characterized by an amplitude decrease in the macular area, recorded by multifocal ERGs (mfERGs). Adaptive Optics Scanning Laser Ophthalmoscopy (AOSLO) can resolve individual photoreceptors in living human’s eyes. In patients with OMD, distinctive images that reduce cone reflectance at the macular area can be captured, although it is difficult to be classified automatically. Recently, convolutional neural network (CNN) becomes one of main techniques for image recognition and classification because of its overwhelming ability. The present study aimed to classify OMD, retinal diseases other than OMD, and healthy retinas based on AOSLO images using CNN.
Methods :
A total of 36 subjects were enrolled, including 13 healthy individuals (mean ± standard deviation, 28.6 ± 5.1 years), 11 with OMD (50.8 ± 11.8 years), and 13 with retinal diseases other than OMD (44.2 ± 22.2 years ; retinitis pigmentosa, 4; microcystic macular edema, 2; acute zonal occult outer retinopathy, 1; age-related macular degeneration, 1; best vitelliform macular dystrophy, 1; branch retinal vein occlusion, 1; epiretinal membrane, 1; retinal detachment, 1; Vogt-Koyanagi-Harada disease, 1). All patients were diagnosed through complete ophthalmological examinations. Each subject was photographed with a custom built AOSLO (Topcon) with a sampling rate of 30 Hz, at a 21° × 21° view angle centered at the fovea of the right eye. The obtained images were of 512 × 512 pixels with a 12-bit gray scale gradation. In each group, 200 images were obtained for training and 75 were obtained for validation. The CNN model was fine-tuned on the VGG16 model and was the training images augmented from 600 to 3000 using data augmentation. The CNN model then learned 3 classes (classifications) in 60 epochs.
Results :
Training accuracy and validation accuracy were 0.998 and 0.987, respectively. Training loss and validation loss were 0.011 and 0.017, respectively.
Conclusions :
Our results suggest that CNN can discriminate among OMD, retinal diseases other than OMD, and healthy retinas based on AOSLO images.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.