Retinal Fundus Image Analysis for Early Glaucoma Detection Using Deep Learning Models

This study investigates the use of retinal fundus pictures and deep learning techniques for the early identification of glaucoma. The dataset, sourced from the Kaggle repository, is divided into training and validation sets and categorised into referable glaucoma (RG) and non-referable glaucoma (NRG). Two deep learning models, EfficientNetB0 and MobileNetV2, were implemented to classify the images. Both models were trained and evaluated over 20 epochs. EfficientNetB0 outperformed MobileNetV2 with a validation accuracy of 95.04% compared to 92.91%. It also demonstrated superior precision, recall, and F1-score performance across both classes, although MobileNetV2 showed slightly better recall for the RG class. External unseen data was used to assess model confidence and generalizability. Overall, EfficientNetB0 provided a better balance and more robust classification. Future work aims to explore more complex architectures and diverse datasets to further enhance diagnostic performance in clinical settings. Adding explainability features would allow doctors to see why the AI flagged a case as glaucoma, helping them trust and rely on its assessments.