Transfer Learning for Diabetic Retinopathy Detection: A Study of Dataset Combination and Model Performance

Abstract

Diabetes’ serious complication, diabetic retinopathy (DR), which can potentially be life-threatening, might result in vision loss in certain situations. Although it has no symptoms in the early stages, this illness is regarded as one of the “silent diseases” that go unnoticed. The fact that various datasets have varied retinal features is one of the significant difficulties in this field of study. This information impacts the models created for this purpose. This study’s method can efficiently learn and classify DR from three diverse datasets. Four models based on transfer learning Convolution Neural Network (CNN)—Visual Geometry Group (VGG) 16, Inception version 3 (InceptionV3), Dense Network (DenseNet) 121, and Mobile Network version 2 (MobileNetV2)—are employed in this work, with evaluation parameters, including loss, accuracy, recall, precision, and specificity. The models are also tested by combining the images from the three datasets. The DenseNet121 model performs better with 98.97% accuracy on the combined image set. The study concludes that combining multiple datasets improves performance compared to individual datasets. The obtained model can be utilized globally to accommodate more tests that clinics perform for diabetic patients to prevent DR. It helps health workers refer patients to ophthalmologists before DR becomes serious.