A Framework for Diabetic Retinopathy Detection using Transfer Learning and Data Fusion

Diabetic retinopathy stands as a significant concern for individuals managing diabetes. It is a severe eye condition that targets the delicate blood vessels within the retina. As it advances, it can inflict severe vision impairment or complete blindness in extreme cases. Regular eye examinations are vital for individuals with diabetes to detect abnormalities early. Detection of diabetic retinopathy is challenging and a time-consuming process, but deep learning and transfer learning techniques offer vital support by automating the process, providing accurate predictions, and simplifying diagnostic procedures for healthcare professionals. This study introduces a multi-classification framework for grading diabetic retinopathy into five classes using Transfer Learning and data fusion. The objective is to develop a robust, automated model for diabetic retinopathy detection to enhance the diagnostic process for healthcare professionals. We fused two distinct datasets, APTOS and IDRiD, which resulted in a total of 4178 fundus images. The merged dataset underwent preprocessing to enhance image quality and to remove unwanted regions, noise and artifacts from the fundus images. The pre-processed dataset is then resized and a balancing technique called SMOTE is applied to it due to uneven class distribution present among classes. To increase diversity and size of the dataset, data augmentation techniques including flipping, brightness adjustment and contrast adjustment are applied. The dataset is split into 80:10:10 ratios for training, validation, and testing. Two pre-trained models, EfficientNetB5 and DenseNet121, are fine-tuned and training parameters like batch size, number of epochs, learning rate etc. are adjusted. The results demonstrate the highest test accuracy of 96.06% is achieved by using EfficientNetB5 model followed by 91.40% test accuracy using DenseNet121 model. The performance of our best model i.e. EfficientNetB5, is compared with several state-of-the-art approaches, including DenseNet-169, Hybrid models and ResNet-50 where our model outperformed these methodologies in terms of test accuracy.