Hybrid CNN Approach for Automated Glaucoma Detection and Severity Grading from Fundus Image

This study presents a deep learning-based approach for automated detection and grading of glaucoma using retinal fundus images. Leveraging a hybrid model built on EfficientNetB3 as a feature extraction backbone, the system is trained to classify fundus images across different severity levels of glaucoma. Our methodology incorporates a custom convolutional neural network (CNN) architecture that includes fully connected layers and dropout for regularization, ensuring robust feature learning and improved accuracy. Data pre- processing and augmentation techniques such as rotation, horizontal flip, and adjustment of brightness are used to increase the generalization of the model on different image inputs. Stratified train-validation-test split is used to balance the classes; early stopping and model checkpointing are used to prevent overfitting and retain the best model state. In the experiments, the model could reach a training accuracy of around 96% and validation accuracy of about 94%, which may ensure reliable detection and grading in clinical settings. This approach underscores the effectiveness of hybrid deep learning architectures in ophthalmic disease classification, with future work aimed at enhancing diagnostic accuracy and expanding classification granularity.