A Transfer Learning Based Framework for Diabetic Retinopathy Detection Using Data Fusion

Diabetic retinopathy is a condition associated with diabetes that damages the blood vessels within the retina resulting in vision impairment or even blindness. Early detection and classification of DR enables timely intervention, which is crucial for preventing vision loss and blindness in diabetic patients. This research presents a framework for binary classification, employing transfer learning to identify diabetic retinopathy in individuals with diabetes. APTOS19 and IDRiD, two datasets containing fundus images, are merged together for training the transfer learning models to predict the presence or absence of the disease. Many preprocessing techniques have been applied to these images like resizing, Gaussian filtering, and dataset splitting. After the split, training set is augmented using zooming, rotation, flipping etc. to increase diversity. The transfer learning models used are: ResNet50 and DenseNet121. These models are fine tuned for classification. The results highlight that the DenseNet121 model achieved a superior test accuracy of 97.22% as compared to ResNet50.