Transfer Learning based Classification of Diabetic Retinopathy Stages

Diabetic Retinopathy is caused due to the damage of the blood vessels located at the back of the eye (retina). Symptoms such as blurriness, dark spots, blindness, etc. can be seen due to diabetic retinopathy. In this paper, we perform the analysis of diabetic retinopathy detection using three pre-trained networks - VGG16, InceptionV3, and ResNet50 on Kaggle Diabetic Retinopathy dataset. The dataset consists of five classes of images and is very disproportionately balanced. The bulk of the images belong to class 0 (73.48%) and the rest is divided among class 1 to class 4. So, in order to mitigate the problem, we augment new images for the inconsistent classes. Then we perform image preprocessing, use different evaluation metrics and finally compare the results. Out of these three models, VGG16 achieves the highest accuracy of 78% and Kappa score of 0.721. Another observation we were able to distinguish is that, even after using different techniques such as augmentation and preprocessing, class 2 has the lowest precision, recall and F1-score values out of the five classes of images.