Automated Detection and Classification of Glaucoma and Macular Degeneration Using Convolutional Neural Network

Abstract

Eye is the most complex and sensitive organs of our body which enables us to visualize things. Globally, around 1.3 billion people will live with visual impairment. Most eye diseases may gradually lead to permanent vision loss if not detected and treated early. Glaucoma is a chronic eye disease often called “silent thief of sight” as it has no symptoms. If it’s not detected at an early stage, it may cause permanent blindness. Age-related macular degeneration is another chronic eye disease that causes blurred or reduced central vision leading to blindness. Regular eye health checkup is required, especially for the elders, to avoid blindness. Tonometry, ophthalmoscopy, visual field test, and optical coherent tomography are used to diagnose glaucoma and age-related macular degeneration. Most of the modern imaging devices are expensive and unaffordable for health facilities in low resource setting. Also, diagnostic methods are manual, time-consuming and exposed to intra/inter observer variability and require expert supervision. Most works are focused on binary classification of eye disease. However, grading of the disease is also an important diagnosis factor that is used to identify the disease progression. Segmenting cup and disk areas are one of challenging task in the area. Segmentation of OC and OD using traditional machine learning approach is difficult for extraction of deep feature and learning hidden pattern by itself which reduce computation time, reliability and performance of the system. This thesis presented automated glaucoma and AMD detection using a convolutional neural networks approach from fundus image. Glaucoma detection and grading have been performed based on the cup to disk ratio calculation after the segmentation of OC and OD features. The experiments shows that Unet++ achieves state-of-the-art OD and OC segmentation results compared to Unet. The detection accuracy was found to be 91%. AMD classification has also been done with 97% accuracy using the VGG16 model.