Comparison of Transfer Learning Approaches with Multiple Data Sets in Early Stage Detection of Diabetic Retinopathy

Diabetic Retinopathy (DR) is one of the leading causes of blindness today, therefore early diagnosis of the disease is crucial to preserve patients' vision. This study aims to develop a decision support model that can achieve high accuracy and discrimination in the early-stage diagnosis of DR patients. Convolutional Neural Networks (CNNs), a deep learning approach, have been used with transfer learning methods from well-known models in the literature to create this decision support model. The models that were trained with transfer learning were also used as feature extractors and were trained with different machine learning classifiers such as Support Vector Machines, K-nearest neighbors, and so on. Contrary to the common practice in the literature, different datasets were combined for training to improve the generalization ability of the created model. The most commonly used and publicly available datasets in the literature, namely, Kaggle EyePACS, DDR, Kaggle APTOS2019, and IDRiD, were chosen for this study. The training was conducted on images obtained from the combined dataset for three classes: no diabetic retinopathy (0), mild and moderate non-proliferative diabetic retinopathy (1+2), and severe non-proliferative and proliferative diabetic retinopathy (3+4). The proposed study achieved the highest accuracy and discrimination by classifying with a mediumdepth (two-layer) artificial neural network using the Xception model.