Glaucoma Detection using Deep Learning on Retinal Fundus Images

Glaucoma is one of the most prevalent permanent blindness in the world, and it may be hard to detect at the initial stages or even be asymptomatic because of slow progression. Retinal fundus imaging has been used as a popular technique for detecting the retinal optic nerve head and nearby structures in a very accurate manner; it has become a widely used tool due to its timely diagnosis. The publicly available datasets, namely, ACRIMA and ORIGA, and a hybrid dataset aimed at enhancing the model robustness with respect to different image sources, the given research is a deep learning-based blind glaucoma detection model. The model used transfer learning based on VGG16 on models trained on ACRIMA and EfficientNetB0, augmented with contrast preprocessing on ORIGA, trained with data augmentation on class imbalance and overfitting. The experimental assessment revealed that the custom CNN achieved training accuracy of 96.28% and validation accuracy of 93.57% on ACRIMA, and the transfer learning model trained on the combined dataset gained through training accuracy of 95.71% on the ACRIMA test set, which demonstrates that generalization is strong across datasets. Findings demonstrate that the suggested approach is correct and could be used in early glaucoma detection as well as in large-scale clinical screening.