Ocular disease diagnosis from retinal images

• Main Author: Zhang, Jielin
• Other Authors: Yuan, Junsong
• Format: Final Year Project
• Language: English
• Published: 2013
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/52986
• Institution: Nanyang Technological University

Digital color fundus image is a popular imaging modality in the diagnosis of ocular diseases, including age-related macular degeneration (AMD). The presence of large and numerous drusen in the macula is believed to be a common sign of AMD. Therefore, automated methods that can accurately detect the images with drusen are highly desirable in order to design a computer aided diagnosis (CAD) system for AMD. The aim of this project is to test the state-of-the-art drusen image recognition system on a population-based database, build a high-quality AMD database that can be used to train and test computer algorithms, and develop a drusen image detecting algorithm that can improve the detection accuracy compared to existing algorithms. In this work, the state-of-the-art AMD risk index is used to test on a large fundus image database. The results are analyzed against important drusen characteristics such as number of drusen, maximum drusen size and drusen types. The images are then filtered and selected based on the testing result. Finally, a clean database which consists of a normal set and a drusen set is built. To design the drusen recognition algorithm, a multi-feature classification system and a cascaded classification system are proposed. In the multi-feature classification scheme, optimized features from different feature detector-descriptors are combined to train the classifier. In the cascaded classification system, the first level classification contains a number of classifiers that utilize the optimized image features. The outputs of the first level classification are then fed to another classifier to produce the final result. Experimental results show that the proposed methods achieves high detection accuracy and improves significantly compared to existing method, which demonstrates a good potential for these methods to be used in medical applications.