Comparative Analysis of Transfer-Learning Models for Five-Class Diabetic Retinopathy Classification: A Secondary-Data Study Using the APTOS 2019 Dataset

Abstract

Diabetes-related retinal complications are a major cause of preventable visual impairment, with diabetic retinopathy being one of the most significant ocular complications. Recent advances in deep learning and transfer-learning techniques have demonstrated considerable potential for the automated detection and classification of diabetic retinopathy using retinal fundus images. The present study employed a secondary-data-based comparative approach using the publicly available APTOS 2019 dataset, comprising 3,662 images across five DR severity classes. Class distribution and imbalance were analysed, and published performance of commonly used transfer-learning models was compared using accuracy, AUC, precision, recall and F1-score. The dataset showed substantial class imbalance, with No DR accounting for 49.29% of images and Severe DR for 5.27%, resulting in a 9.35:1 largest-to-smallest class ratio. Published results from an APTOS 2019 evaluation reported accuracies of 96.18% for DenseNet121, 96.07% for ResNet50, 95.85% for VGG16, 95.71% for InceptionV3 and 89.31% for MobileNetV2. The findings highlight the importance of considering class distribution and multiple performance measures when comparing transfer-learning models for multi-class DR classification. This secondary-data analysis provides a basis for subsequent computational validation using the same public dataset.