Two stage-network: Automatic localization of Optic Disc (OD) and classification of glaucoma in fundus images using deep learning techniques

Glaucoma is an ophthalmic disorder which results in permanent vision loss because high intraocular pressure damages the optic nerve in the eye. This paper proposes a two-stage network for automated glaucoma identification utilizing fundus images. In the first stage, Yolo-v4 is used to locate and extract the optic disc from a retinal fundus image, and ResNet-101 is used in the second stage to determine whether the retrieved disc is glaucomatous or healthy. Unfortunately, none of the publicly accessible retinal fundus image datasets contain the necessary bounding box ground truth for disc localization. In this regard, a semi-automatic ground truth creation strategy has been proposed that gives the essential annotations enabling the training of the Yolo-v4 based model for autonomous disc localization. The proposed method is evaluated on ORIGA publicly available dataset. The proposed automated OD localization showed good results with 87.4% accuracy, 89.79% precision and 88.7% recall. Whereas the proposed glaucoma diagnosis module attained good results with 88.5% accuracy and AUC of .920.