DR-NET: A Stacked Convolutional Classifier Framework for Detection of Diabetic Retinopathy

Diabetic Retinopathy is the main cause of vision impairment among people suffering from diabetes and often leads to blindness. It has no early warning signs. Hence, it is of utmost importance to detect it as early as possible in order to provide adequate treatment. Most of the current research in this field focuses on a manual process of feature extraction such as annotation of lesions and optic disk segmentation so as to detect the presence of DR. In this paper, a framework DR-NET using stacked convolutional neural networks for diabetic retinopathy detection from digital fundus images is proposed. A network consisting of convolutional layers with different filters stacked in parallel, the output of which is concatenated and global max pooling performed on it, is developed. This architecture helps extract intricate features during the classification task along with minimizing the learnable parameters and reduces overfitting, thus, improving the overall performance of the model. Various preprocessing methods were applied to further improve accuracy. Visualization techniques were also used to gain insights into the learning of the model. The experimental results were performed on about 12,000 images which were an ensemble of various online datasets, yielded an accuracy of 81% and a kappa score of 0.6.