Performance Evaluation of ResNet-50 and Inception-V3 for Diabetic Retinopathy Detection on Retinal Images

Abstract

Diabetic retinopathy (DR), a leading cause of blindness among diabetics, poses significant diagnostic challenges due to subtle early-stage symptoms like microaneurysms, often undetectable through conventional manual fundus examinations, which are time-consuming and inaccessible in remote areas. This study evaluates the performance of ResNet-50 and Inception-V3 deep learning models in detecting DR using the APTOS 2019 dataset, preprocessed with Gaussian filtering to reduce noise. Binary classification (DR/No_DR) was implemented to address extreme class imbalance in the original dataset (No_DR: 1,750 images; Severe DR: 150 images). Methodology included data augmentation (rotation ±20°, 20% horizontal/vertical shifts, 20% zoom, horizontal flip) and evaluation using accuracy, precision, recall, and F1-score. Results demonstrated that Inception-V3 achieved the highest validation accuracy of 97.64% with augmentation, outperforming ResNet-50 (91.82%). Inception-V3 was also 40% more computationally efficient, requiring only 9,959.88 seconds compared to ResNet-50’s 16,673.65 seconds. Augmentation improved model generalization, particularly for ResNet-50 (+2.37% accuracy). Inception-V3 achieved an optimal F1-score of 0.9856, balancing precision (99.27%) and recall (97.85%). These findings recommend Inception-V3 as an accurate and efficient solution for automated DR screening, especially in resource-constrained settings.