Automatic Diabetic Retinopathy Classification with EfficientNet
Rachmadio Noval Lazuardi, Nyoman Abiwinanda, Tafwida Hesaputra Suryawan, Muhammad Hanif, Astri Handayani
Abstract
Using the recent known EfficientNet architecture of deep convolutional neural network (CNN), we present an automatic detection of diabetic retinopathy (DR) from given retinal images. We experiment with subsets of the Kaggle diabetic retinopathy dataset consisting of retinal images with varied diagnostic quality. To address the quality variation, we incorporate two preprocessing steps, i.e. contrast limited adaptive histogram equalization (CLAHE) and image central cropping. We trained EfficientNet-B4 and EfficientNet-B5 model on two Kaggle subsets with different class proportions. In this paper, we propose an automatic early diagnosis of diabetic retinopathy which gained 0.7922 / 83.87% and 0.7931 / 83.89% of quadratic weight kappa and accuracy score on EfficientNet-B4 and EfficientNet-B5 respectively.