Deep Transfer Learning for Diabetic Retinopathy Detection

It is a difficult and very important job to apply deep learning in medical data. Nevertheless, transfer learning can greatly reduce training costs through pre-trained deep convolutional neural networks. The world is experiencing an increase in the occurrence of blindness due to diabetic retinopathy (DR). Therefore, in this work, we propose to employee deep transfer learning (DTL) approaches for diagnosing DR and its category. To find out the best classification model for DR detection, we used three different deep learning models viz. VGG19, ResNet50, and DenseNet201. The 2019 Kaggle DR detection challenge dataset consisting of 35,120 retinal fundus images was used in the research. It was also employed to scrutinize how well the deep learning models performed with respect to different metrics including precision, recall, F1-rating, accuracy, and AUC. The results are reflective of the excellent performance of the proposed DTL on the dataset.