Advanced image analysis techniques for automated glaucoma diagnosis using retinal fundus images

Abstract

Glaucoma, after cataracts, is the second most leading cause of vision loss. It develops due to increased intraocular pressure (IOP) that damages retinal nerve bres (RNFs). Generally, glaucoma does not exhibit any indication of its progression in early stages until it becomes more advanced. Therefore, early diagnosis and routine checkup are required to prevent further vision loss. Ophthalmologists employ certain clinical instruments to diagnose glaucoma. In addition to these clinical examinations, advanced computerised imaging devices are used to detect the presence of glaucoma. These imaging devices generate retinal and optic nerve head (ONH) images along with objective quantitative measures. Experts use these images and measures for further investigation of glaucoma condition. However, the aforementioned clinical methods are manual and require skilled supervision. Moreover, the computerised imaging devices are bulkier, fragile, expensive, require trained professionals, and are generally not available in rural and remote areas. Further, the retinal images obtained using these imaging devices require manual evaluation by quali ed experts. This manual evaluation is subjective and introduces inter and intra-observer variability, which occurs due to inconsistent perception of di erent experts towards the structural and functional damages, which characterises glaucoma, within the eye. Unlike the advanced computerised imaging devices, fundus camera is a basic imaging device without computational setup. Therefore, it is generally portable and economical. Retinal fundus image acquired using fundus camera can be employed to visualise optic cup, optic disk, and blood vessels. In this way, these images help diagnose glaucoma condition. However, as the number of glaucoma cases is increasing every year, it is a time-consuming and challenging task to examine individual retinal images manually. These challenges can be overcome by developing retinal fundus image based computer-aided systems that are fast and accurate. These systems do not involve human intervention and can assist experts in their diagnosis, thereby reducing the burden of mass-screening. In recent years, with the development of advanced image analysis techniques and machine learning algorithms, there exists a huge potential for development of e cient, accurate, and state-of-the-art methods for computer-aided automated diagnosis of glaucoma. These methods can be used to develop prompt, reliable, handy, and cost-e ective glaucoma diagnostic systems. Therefore, this thesis aims to propose computer-aided approaches for automated glaucoma diagnosis based on advanced image analysis techniques and machine learning algorithms. In general, such approaches involve following processing stages. The initial stage involves image preprocessing techniques such as resizing, ltering, etc. In the next stage, meaningful features from the input fundus image are extracted using image analysis techniques. This is followed by ranking and selection of the extracted features. Finally, the features are fed to a classi er to discriminate between normal and glaucoma classes. The rst three approaches presented in the thesis involve the aforementioned processing stages. The rst two approaches for glaucoma diagnosis are based on adaptive non-stationary image analysis techniques. Speci cally, empirical wavelet transform (EWT), and iterative variational mode decomposition (VMD) have been explored. A local bit-level texture descriptor based approach has also been developed for glaucoma diagnosis. Additionally, this thesis explores convolution neural network (CNN) based approach for glaucoma diagnosis. This approach automatically extracts relevant features and classi es the input fundus image. The experimental results presented in this thesis demonstrate that the proposed approaches are e ective for glaucoma diagnosis. Moreover, the proposed approaches achieve state-of-the-art performance on benchmark datasets.