Diabetic retinopathy classification using deeply supervised ResNet

Diabetic retinopathy (DR) is the leading cause of blindness in the working-Age population all over the world. Classifying DR is a time-consuming work, which requires an experienced ophthalmologist to evaluate color fundus photographs of retina. In this paper, we propose a deeply supervised ResNet approach to classify the severity of DR automatically. In this new convolutional neural networks (CNN) architecture, we add three sets of additional side-output layers to intermediate hidden layers of a 11-layer ResNet. By introducing these deeply supervised layers, we can provide additional regularization during training network. More importantly, we can perform multi-scale learning by leveraging the predictions of intermediate supervised layers, thus improving the final performance. Furthermore, to combat the issue of class-imbalance in the dataset, we adopt cost sensitive learning and an oversampling method of cropping images. We train and evaluate our network on the publicly available Kaggle dataset, and the results show that our method outperforms the state-of-The-Art method.