RetiNet — Feature Extractor for Learning Patterns of Diabetic Retinopathy and Age-Related Macular Degeneration from Publicly Available Datasets

Diabetic Retinopathy (DR) and Age-related Macular Degeneration (AMD) are two common vision threatening eye conditions. In a large-scale screening environment DR and AMD can be assessed by detecting specific retinal findings in fundus images. In this paper, we introduce a new deep learning based feature extractor for automatic classification of DR and AMD from fundus images. We used a small dataset containing 60000 images with four severity levels of DR and two classes of AMD to design and fine-tune a deep learning model called RetiNet. This dataset, which consisted of two publicly available datasets (MESSIDOR and Kaggle), was augmented and employed to evaluate RetiNet. RetiNet can achieve diagnosis performance comparable to retina experts on the MESSIDOR dataset with cross-dataset testing (i.e., the feature extractor was trained on an independent dataset and tested on MESSIDOR). Our algorithm obtained an average accuracy of 88% on the validation set.