Effective Deep Learning Data Augmentation Techniques for Diabetic Retinopathy Classification

Abstract

Diabetic retinopathy(DR) is an eye disorder that affects 80-85% of diabetic patients, and there is a need to detect it in earlier stages to take precautions. Currently, testing for DR is a time-consuming task and requires qualified human resources. Automating DR detection is time-saving and cost-effective, benefiting both doctors and patients. The proposed work automates the DR image detection using deep learning(DL) by using transfer learning. The proposed work provides better results with multi-class classification and generalization than the previous works. i.e a ResNet-50 DL model is trained with the APTOS dataset and tested using the EyePACS dataset. The low accuracy in multi-class classification and data imbalance problems are overcome by using appropriate image pre-processing and augmentation techniques, which have been found by conducting several experiments. Further, the performance of the DL model is enhanced using Test Time Augmentation(TTA). Finally, the authors are able to achieve a multi-class accuracy of 97.87% and Quadratic weighted kappa(QWK) score of 0.985.