Efficient Classification of Diabetic Retinopathy using Binary CNN

Abstract

Diabetic Retinopathy (DR) is a fastly spreading disease that may lead to loss of vision if not quickly detected and treated. Early-stage detection is beneficial to restrict the progress of disease and reduces the recovery expenditure. The current detection process of DR heavily depends on domain experts. Machine-dependent approaches are gain attention with large-scale fundus image repositories to overcome this difficulty. Recent techniques with deep learning are successful in getting noticeable results with pre-trained networks. However, the increase of memory occupancy and runtime with existing models is the bottleneck. We propose Binary Convolutional Neural Networks (BCNN), which drastically reduces memory consumption and faster the execution process to combat this problem. Our model is hardware friendly and efficient in DR classification with large scale fundus images. Experiments conducted using the Kaggle dataset reduce memory consumption by 37% and increase runtime by 49% compared to the base model.