Data Balancing Techniques Evaluation on Convolutional Neural Network to Classify the Diabetic Retinopathy of Fundus Image

Diabetic retinopathy (DR) is a common complication diabetic patients that causes impaired vision, and may even lead to blindness. Several studies on the DR diagnosis based on Computer-Aided Diagnosis (CAD) had been conducted. The method used various feature extraction modules and a particular classifier. However, this method required a long step. In a different circumstance, deep neural networks had been successfully applied in various fields and showing good performance. For this reason, we proposed a classification system for DR based on Convolutional Neural Networks (CNN). In this study, we used retina images dataset from the Asia-Pacific Tele-Ophthalmology Society (APTOS) to train CNN under three different conditions. Sequentially is imbalanced, balanced by undersampling, and balanced by oversampling. The best results were obtained in the third condition, with an accuracy of 73.64%, precision 59.01%, sensitivity 60.69%, and specificity 93.49%. The classification method in the proposed study should be realized in clinical use.