Automatic Glaucoma Detection from Fundus Images Using Deep Convolutional Neural Networks and Exploring Networks Behaviour Using Visualization Techniques

Glaucoma is an irreversible eye disease due to increased intraocular pressure that damages the optic nerve in the eye. It does not initially exhibit any symptoms. The effect is so gradual that one may not detect a change in vision until the condition has advanced, and it may lead to permanent vision loss. Therefore, early detection and appropriate routine eye screening that evaluates pressure in the eye is essential to prevent and reduce vision loss. In this work, we developed a model for automatic glaucoma detection in fundus images using three deep convolutional neural networks (CNNs): Resnet101, Nasnet_mobile, and Nasnet_large, and tested the model on five publicly available fundus image datasets: ACRIMA, RIMONE-v2, Drishti-GS, FTVD, and the Harvard Dataset (HVD). Model performance metrics such as area under the curve (AUC), accuracy (Acc), sensitivity (Sen), specificity (Spe), etc. are evaluated for each dataset in particular, we achieved an AUC of 1, Acc of 99.43%, Sen of 98.99%, and Spe of 100% for the ACRIMA dataset. The results of our method show that our solution outperforms state-of-the-art methods for glaucoma diagnosis in fundus images. Understanding model output includes attribution-based methods such as activations and class activation maps using gradients (G-CAM), as well as perturbation-based approaches such as locally interpretable model diagnostic explanations (LIME) and occlusion sensitivity (OS), which generate heat maps of different image sections for model prediction. © 2023, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.