Multi-class retinal disease detection using VGG-19 with K-Fold cross-validation

Abstract

Early and reliable detection of retinal diseases is crucial for preventing irreversible vision loss and improving clinical decision-making. This paper presents a deep learning-based framework for multi-class retinal disease detection using a VGG-19 convolutional neural network, supported by a comprehensive K-fold cross-validation strategy to enhance robustness and generalization. The methodology integrates a structured preprocessing pipeline that standardizes fundus image resolution, applies color normalization, and incorporates diverse augmentation techniques to preserve diagnostic features while reducing overfitting. The VGG-19 architecture is adapted through transfer learning and fine-tuning, enabling the model to extract deep hierarchical patterns relevant to conditions such as cataract, glaucoma, diabetic retinopathy, and normal retinal states. Cross-validation ensures consistent evaluation across multiple data partitions, allowing the model to learn from varied subsets and reducing sensitivity to class imbalance and dataset bias. The findings indicate that the proposed framework achieves stable classification behavior, improved differentiation between visually similar diseases, enhanced generalization beyond a single train-test split.