Diabetic Retinopathy Classification Using Binary CNN and Data Augmentation

Abstract

Diabetic Retinopathy (DR) is a rapidly spreading disease that can lead to blindness. Early detection can help to limit disease progression and minimize treatment costs. The process of finding a real DR is very much dependent on the clinical experts. The computer-aided software approach in solving this problem gain attention with large scale datasets. Modern techniques of Deep Learning can achieve extraordinary results in DR classification. However, with the current models, memory usage and increased runtime is a significant problem. To solve the problem, we recommend Binary Convolutional Neural Networks (BCNN), which significantly reduce memory usage and speed up the execution process. We augmented the dataset in preprocessing to train our model effectively. Our hardware friendly model outperforms in restricted memory environments. Our experiments using the Kaggle dataset reduced the memory usage and increased execution speed compared to the base model.