Performance Analysis of InceptionV3, ResNet50 and VGG16 for Diabetic Retinopathy Detection

Abstract

 Diabetic retinopathy (DR) is an illness that has the potential to cause vision impairment in people who are diabetic by affecting the retina’s blood vessels. The identification of DR using color fundus images typically relies on the expertise of trained medical professionals to manually detect lesions is labor-intensive and costly. This paper proposed a model based on InceptionV3, ResNet50, and VGG16 to detect and classify DR based on its severity level. The model incorporates the synthetic minority oversampling technique to address class imbalance issues commonly encountered in medical image analysis. Training and testing are conducted using the APTOS2019 and Messidor datasets, resulting in a validation accuracy of 73%, 60% and 52% for InceptionV3 and ResNet50 while VGG16 obtained a validation accuracy of 52%. Despite these promising results, the study identifies areas for improvement, such as mitigating underfitting for the VGG16 model.  Strategies to enhance model generalization, optimize hyperparameters and diversify datasets should be prioritized to facilitate broader applicability across diverse clinical settings.