Comparison of Transfer Learning Strategies for Diabetic Retinopathy Detection

Diabetic Retinopathy (DR) is a common complication of diabetes, one of the leading causes of blindness worldwide. Diabetes and DR cases have been increasing at an alarming rate in recent years. However, early diagnosis of DR is difficult and the diagnostic procedure is time-consuming. Due to the importance of early detection of DR in terms of treatment, many approaches have been applied to DR detection in the last two decades. Recent research shows that deep learning-based convolutional neural network (CNN) structure and transfer learning are the most used approaches in DR detection. Accordingly, in the present study, DR images were classified considering their severity levels using the Messidor-2 + EyePac Balanced data set in Kaggle using two different transfer learning strategies: training from scratch and fine tuning. The results obtained using two different approaches were compared in terms of accuracy and training time. In the proposed study, AlexNet, VGG-16, DenseNet121, and ResNet50 architectures were used to compare the performance of these two approaches. According to the results, it was determined that the fine tuning approach for the classification of DR images performs better than CNN training from scratch.