A Severity Grading Framework for Diabetic Retinopathy Detection using Transfer Learning

Diabetic retinopathy is one of the primary reasons of blindness among individuals. It is a threatening visual disorder that affects delicate blood vessels present within the retina. Early identification of this condition is beneficial for preserving vision and managing it effectively. Detecting diabetic retinopathy manually takes great amount of time and is susceptible to errors. This research aims to propose an automated and efficient framework for grading the severity of diabetic retinopathy with higher accuracy using transfer learning techniques. For this purpose, we have used 3 distinct diabetic retinopathy datasets; Messidor-1, Messidor-2 and IDRiD, containing fundus images. Preprocessing techniques including cropping, denoising, CLAHE, and image resizing are applied to each dataset. Data augmentation is used to tackle class imbalance issues in order to avoid biased results. The dataset was split into three subsets for training, validation, and testing, using a split of 70:10:20. Pre-trained model EfficientNetB5 is fine-tuned and hyper-parameters are adjusted. This model is used to train and test each dataset separately. Results show that our model attained test accuracies of 98.43% for Messidor-1, 97.36% for Messidor-2, and 97.67% for the IDRiD dataset. On comparing the best results of this research with previous state of the art techniques, EfficientNetB5 outperformed those existing methods.