Quantitative Analysis of Deep Learning Techniques in Retinal Disease Classification

Retinal diseases such as diabetic retinopathy (DR) and age-related macular degeneration (AMD) are among the leading causes of vision loss worldwide. Early detection is critical, and deep learning has emerged as a powerful tool for automating retinal image classification. In this study, we present a quantitative comparison of state-of-the-art deep learning architectures for retinal disease classification, focusing on both performance and reliability. We evaluate convolutional neural networks (ResNet, DenseNet, EfficientNet, ConvNeXt) and transformer-based models (Vision Transformer, Swin Transformer) using large-scale public datasets including EyePACS, Messidor, and APTOS. A standardized pipeline was applied across all models, consisting of fundus image preprocessing, class imbalance handling with focal loss, and data augmentation to enhance generalization. Performance was assessed using AUROC, AUPRC, quadratic kappa, sensitivity at fixed specificity, and calibration error. Robustness was further evaluated through external dataset testing and statistical significance was measured using bootstrap confidence intervals, DeLong's test, and McNemar's test. Results show that transformer-based models generally achieve higher AUROC and better generalization than CNNs, while ensembles provide the most stable performance across datasets. However, calibration analysis revealed that most models remain overconfident, which could limit their safe use in clinical decision-making. The aim is to establish a benchmark for retinal disease classification, emphasizing not only accuracy but also reliability, calibration, and cross-dataset robustness-factors essential for deploying deep learning systems in real-world ophthalmic screening.