Data Fusion Based Ensemble Transfer Learning Approach to Detect Diabetic Retinopathy

Diabetic retinopathy is an eye disease damaging the blood vessels of retina as a result of long term diabetes. In this research, an ensemble classification system is proposed by combining two transfer learning models to identify diabetic retinopathy among diabetic patients. A fusion of two datasets containing fundus images is used in this study: 1) APTOS, which contains 3662 images and 2) IDRiD, which contains 516 images. For binary classification, the dataset is divided into two classes: DR and No DR. Preprocessing steps are applied on the dataset such as resizing the images to 150x150x3, applying Gaussian filtering, balancing minority classes using SMOTE and splitting the dataset to 80:20 ratios. Data augmentation techniques like zooming, rotation etc. are also used to augment the images. Two transfer learning models, Xception and EfficientNetB3, are used for classification after fine-tuning. An ensemble of these models is built which achieved the highest test accuracy of 97.47% outperforming individual models.