Improved Transfer Learning Approach for Diabetic Retinopathy Screening

Diabetic retinopathy (DR) is a frequent consequence of diabetes mellitus. DR damages the retina’s blood vessel network and can be a significant vision-threatening disorder if not diagnosed early. Using fundus imaging to manually diagnose DR is costly and time-consuming because it calls for trained medical professionals. Automated DR detection is therefore a necessary endeavor. In this paper, we assess and modify multiple Transfer Learning models-VGG, ResNet, and DenseNet models with globalaverage pooling and dropout layers to learn the features from the images with reduced overfitting. After that, the models are trained to adjust to the specialized features. The robustness of the presented models is assessed using accuracy, kappa score, precision, recall, and f-measure. The experimental analysis shows that the proposed approach significantly boosts the performance of DR detection on the APTOS dataset and achieved accuracy of 97.49% and 97.16% and kappa score of 95.56% and 94.76% for multiclass and binary classification, respectively.