Article Preview
Top1. Introduction
Glaucoma is a chronic eye disorder and one of the world’s most prominent causes of irreversible vision loss. (Anil et al., 2018; Chethan Kumar et al., 2017). Glaucoma is mainly characterized by damaging the optic nerve and progressively results the vision loss. The increasing degeneration of optic nerve affects directly structural changes in the “Optic Disc (OD). The Optic Disc is the entry point for outgoing vessels into the retina, it looks like a bright circle, and it varies from one person to another. OD consists of two main regions: (i) a bright central part called Optic Cup (OC) and (ii) retinal vessels that carry blood to each part of an eye (Tugashetti et al., 2017: Dey et al., 2012). In normal patient, the size of the Optic Cup is one-third of the Optic Disc but when the size increases, it leads to glaucoma (Touahri et al, 2018).
Since this disease is incurable, proper treatment in time can prolong its development. Therefore, early diagnosis and timely management of this ocular disease may prevent total vision loss. OC and OD are the key factors for the specialists in the detection of glaucoma. This manual identification is time consuming and highly subjective task, thus there is a need for the automatic and fast segmentation of OC and OD. Various computer aided diagnosis systems (CAD) were investigated to automatically localize and diagnose glaucoma. Deep learning (DL) based approach has efficiently shown its capability in the automatic identification of retinal disease last few years (Shriranjani et al. 2018; Benzebouchi et al. 2019).
The motivation of this paper is to propsoe a predictive CAD system towards glaucoma localization and classification. Leveraging the powerful design of deep learning method, this work proposes an accurate segmentation of the OC and OD parameters based on deep learning model that has proven its capabilities in medical image processing. Indeed, the U-Net model will be analyzed and designed to extract local intensity and textural features to perform the semantic segmentation. However, this model presents some deficiencies when analyzing the thin and extensive shapes of retinal vessels. The naive use of this architecture on retinal segmentation gives good results but shows numerous classification incoherencies.
The presented work highlights the impact of zooming stage in the OD zone to expand the rate of classification and confine the researcher’s diagnosis to the aimed zone. The novelty of this study lies in localizing and cutting the region of interest ROI to eliminate the unnecessary data which in its turn increases the accuracy of the classification, in less time and better precision most wished by the patient. More importantly, the aided diagnosis systems-based U-Net architecture offer qualified results with lesser costs.
To tackle the above challenges, a two-stage segmentation of the optic disc and optic cup was proposed. Firstly, the extraction of the region of interest (ROI) will be established from the fundus images by localizing and cutting the optic disc zone. Secondly, a deep learning model based on U-Net architecture will be built in order to obtain the refined segmentation. For the classification step, a Deep Convolutional Neural Network (CNN) was employed to classify the segmented disc into normal or glaucomatous.
The rest of the paper is organized as follows: Section 2 summarizes a brief review of approaches for Optic Disc and Optic Cup segmentation based on U-Net architectures. Section 3 provides some preliminaries used to achieve our purpose. Section 4 presents the comprehensive details of the whole methodology. Used data set and obtained results and discussion of the proposed system are presented in Section 5.