Hybrid CNN-based Framework for Enhanced Glaucoma Detection in Fundus Images

Glaucoma is the most common irreversible cause of blindness. For early detection, there is a growing need to prevent vision loss. A deep learning-based framework for improved glaucoma detection using multimodal fundus images from the Drishti-GS dataset is presented here. With the aim of enhancing the classification performance of one strategy, three state-of-the-art convolutional neural networks-DenseNet, ResNet50, and InceptionV3-are used together in an ensemble approach. Here, the model trains on 80% of the dataset and uses 20% for testing. All the models mentioned take generalization through data augmentation using rotation, zoom, and horizontal flip. Then, Adam optimizer is used with an initial learning rate of 0.001 and then tuned down to 0.00001 for the best output of the model. Ensemble model has impressed accuracy of 97.5% in the test set as compared to that of individual models. This establishes new benchmarks for performance and demonstrates the ability of hybrid deep learning models to be useful in glaucoma detection. The framework is scalable and reliable for real-time automated glaucoma screening, which may be ready to be employed into clinical practice.