Comparative Evaluation of Diabetic Retinopathy Detection Using VGG-16 and ResNet-50 Models: Insights from Matthew’s Correlation Coefficient and Cohen’s Kappa Metrics

Abstract

The diagnosis of diabetic retinopathy utilizing artificial intelligence is a subject of debate, particularly involvement of convolutional neural networks (CNN).  This article employs VGG-16 and ResNet-50 models for the diagnosis of degrees of diabetic retinopathy by analyzing the common retinal features of diabetic retinopathy, which are microaneurysms, hemorrhages, macula oedema, and exudates. The datasets from the APTOS 2019 Blindness Detection fundus images, which were resized to 224×224 pixels, are used for model training to prevent overfitting and processed via yCbCr color filter to emphasis on the image brightness for ease of diagnosis and evaluation. The evaluation results were determined by the metrics of Matthew’s Correlation Coefficient (MCC) and Cohen’s Kappa, both capable of handling imbalanced data, ensure reliable agreement with true labels, and provide balanced insights into the model’s predictive accuracy. The findings revealed that the VGG-16 performed better than ResNet-50 via measurement of Matthew’s Correlation Coefficient in classifying images in a balanced manner, while ResNet-50 performs better than VGG-16 according to values of Cohen’s Kappa thanks to the deep layers of the model.