Deep Classification of Fundus Images Using Semi Supervised GAN

In ophthalmology, fundus image analysis is an efficient way to avoid blindness. Existing deep learning methods with fundus images fail to attain high classification performance because there are considerable numbers of pixel wise annotated data are used for training. The proposed work examines the performance of semi-supervised generative adversarial network from labeled ODIR dataset and from private dataset available in hospital. Training process is enhanced by using unlabeled dataset at various levels, weights which are updated frequently in training phase and finally test phase with labeled ODIR dataset improves classification accuracy. The optimized loss function is used to update weight parameters of discriminator and generator. The wide-ranging research shows that our model achieves state-of-art retinal classification accuracy by using ODIR and hospital dataset.