A Robust Glaucoma Screening Method for Fundus Images Using Deep Learning Technique

In this paper, the optic disc and optic cup are segmented for a cup to disc ratio (CDR) based glaucoma diagnosis. For this purpose, segmentation is implemented by a modified U-Net architecture employing the pre-trained SE-ResNet50 as its downsampling layers. Finally, due to cup and disc areas obtained from the proposed segmentation step, CDR is evaluated. This model is trained on Drishti-GS1 and RIM-ONE v3 databases and is tested on test images of the Drishti-GS1 database. Additionally, to demonstrate the robustness of the proposed method across different datasets the test phase is performed on validation images of the REFUGE database. In terms of F1-score criteria, segmentation results of the optic cup and optic disc are respectively 0.926 and 0.977 for the Drishti-GS1 database and 0.79 and 0.91 for the REFUGE database. Also, the correlation coefficient between the proposed method CDR and the ground truth CDR is 0.94 for the Drishti-GS1 database and is 0.81 for the REFUGE database. Finally, the AUC value is obtained 0.94 and 0.939 for Drishti-GS1 and REFUGE databases, respectively, where the latter result shows the robustness of the proposed diagnosis model.