Decision support system for retinal health with digital fundus images

• Main Author: Koh, Joel En Wei
• Other Authors: Ng Yin Kwee, Eddie
• Format: Theses and Dissertations
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Bioengineering
• Online Access: http://hdl.handle.net/10356/74636
• Institution: Nanyang Technological University

Background: Many health-related problems arise with aging. One of the diseases that are prevalent among the elderly is the loss of sight. Various eye diseases namely age-related macular degeneration, diabetic retinopathy, and glaucoma are the prime causes of vision loss as people grow old. Nevertheless, early detection of such eye diseases can impede the progression of this problem. Therefore, the elderly is encouraged to attend regular eye checkups for early detection of eye diseases. However, it is time-consuming and laborious to conduct a mass eye screening session frequently. This thesis proposes a retinal screening system to automatically differentiate normal image from abnormal (AMD, DR, and glaucoma) fundus images. Methods: To achieve these, two methods will be used. The first method uses the combination of pyramid histogram of oriented gradients (PHOG) and speeded up robust features (SURF) technique. Then, the extracted data are subjected to adaptive synthetic sampling to balance the number of data in the two classes. Subsequently, the canonical correlation analysis approach to fuse the highly-correlated features extracted from the two (PHOG and SURF) descriptors was used. The second method applies pyramid histogram of visual words (PHOW) for the detection of retinal health with digital fundus images using Fisher vector and visual vocabularies. The algorithm can discriminate four classes. As the aim is to get the best performance with the least number of features, the second method outperforms the first method. Results: An average accuracy, sensitivity, and specificity of 96.21%, 95.00%, and 97.42% respectively is obtained using the first method for the classification of normal and abnormal classes using ten-fold cross validation. An average accuracy, sensitivity, and specificity of 96.79%, 96.73%, and 96.96% respectively is obtained using the second method for the classification of normal, AMD, DR and glaucoma classes using ten-fold cross validation. This work has high potential in the diagnosis of normal eye during the mass eye screening session or in polyclinics quickly and reliably. Hence, the patients having abnormal eye can be sent to the main hospitals which will reduce the workload for the ophthalmologists.