Retinal Disease Classification from OCT Images Using Deep Learning Algorithms

Abstract

Optical Coherence Tomography (OCT) is a noninvasive test that takes cross-section pictures of the retina layer of the eye and allows ophthalmologists to diagnose based on the retina's layers. Therefore, it is an important modality for the detection and quantification of retinal diseases and retinal abnormalities. Since OCT provides several images for each patient, it is a time consuming work for ophthalmologists to analyze the images. This paper proposes deep learning models that categorize patients' OCT images into four categories such as Choroidal neovascularization (CNV), Diabetic macular edema (DME), Drusen, and Normal. Two different models are proposed. One is using three binary Convolutional Neural Network (CNN) classifiers and the other is using four binary CNN classifiers. Several CNNs, such as VGG16, VGG19, ResNet50, ResNet152, DenseNet121, and InceptionV3, are adapted as feature extractors to develop the binary classifiers. Among them, the proposed model using VGG16 for CNV vs. Other classes, VGG16 for DME vs. other classes, VGG19 for Drusen vs. Other classes, and InceptionV3 for Normal vs. other classes shows the best performance with 0.987 accuracy, 0.987 sensitivity, and 0.996 specificity. The binary classifier for Normal class has 0.999 accuracy. These results show their potential to work as a second reader for ophthalmologists.