Deep-Learning-Based Early Detection of Diabetic Retinopathy on Fundus Photography Using EfficientNet

Abstract

Diabetic retinopathy, which is the leading cause of blindness among working aged adults, is a serious health problem worldwide. As millions of people suffer from diabetic retinopathy, the need for an automated method of diabetic retinopathy detection to prevent lifelong blindness has long been recognized. Therefore, it's a key challenge to build an automated diagnosis system to detect such disease early and efficiently. For that purpose, we propose a deep-learning-based method to early detect diabetic retinopathy on fundus photography in this paper. The dataset of fundus images is provided by Aravind Eye Hospital in India through a featured competition on the Kaggle platform. Deep convolutional neural networks based on EfficientNet-B3 are utilized to simultaneously extract features from fundus images, where the severity level of diabetic retinopathy is subsequently identified through a regression method. The results show that our deep learning model can achieve an expert-level performance on diagnosing the severity level of diabetic retinopathy, with a quadratic weighted kappa score 0.935 on the private dataset. Such performance locates at the top 1% among all teams and can get a gold medal prize in the Kaggle competition.