An Efficient Deep Learning Framework for Glaucoma Diagnosis Using Convolution Mixed Transformer Network

In this paper, we propose a deep learning based pipeline to predict the presence of Glaucoma, a neuro-degenerative eye condition, by applying object localization, image segmentation, and image classification techniques. The pipeline takes a full fundus image as input and provides Optic Disc and Optic Cup masks along with the likelihood measure to determining Glaucoma. The pipeline uses a lightweight U-Net model called ResNet based Disc Localization U-Net (RDLU-Net) for Optic Disc localization and Convolution mixed SegFormer Glaucoma Detection Network (CSGDN) for segmenting Optic Disc (OD) and Optic Cup (OC) structures and finally predicting the probability of the presence of Glaucoma by using the classification network in CSGDN. Our approach shows dice scores of 0.9871 and 0.9065 on OD and OC segmentation, respectively by using REFUGE dataset. The proposed frame-work obtains accuracy of 95.47% using MobileNet V2 as the detection network when applied on a custom Glaucoma dataset containing data from REFUGE and DRISHTI-GS datasets.