Early Detection of Glaucoma from Cropped Fundus Images Using Transfer-Learned Convolutional Neural Network

Glaucoma, a major contributor to irreversible blind-ness, which requires early diagnosis and appropriate treatment. In this work, we use deep learning techniques to automate the diagnosis of glaucoma from retinal fundus images. For increased efficiency, we use a cropped fundus images instead of analyzing entire fundus photos. We use VGG-19, ResNet-50, and EfficientNet B3 for binary classification on the ACRIMA dataset. Notably, ResNet-50 achieves 100% accuracy. Next, we combine glaucoma suspect samples from publicly accessible datasets with ACRIMA and DRISHTI-GS1 to create “Glaucoma suspect” class in a custom multiclass dataset. Using this dataset, EfficientNet-B3 and ResNet-50 produce 97% and 98% accuracy rates, respectively. Both models perform flawlessly when it comes to categorizing the suspect class & negative class with 100% sensitivity. Our results highlight the effectiveness of deep learning in glaucoma detection and provide interesting directions for enhancing diagnostic precision and facilitating early intervention.