Optic Disc Detection from Fundus Images for Glaucoma – Using Deep Learning
Brown Bag Lecture by Ben Kussmaul | 8/7/2018 11:00AM – 12PM | 7th Floor Conference Room, Bldg 38A
Glaucoma is an eye disease which damages the optic nerve, resulting in vision loss and a reduced field of view. It is the second leading cause of blindness worldwide, and while preventative measures may delay its effects, the damage it causes is irreversible. Therefore, early detection is extremely important. A major warning sign for glaucoma is a large ‘cup’ (white area in the center of the optic disc) in relation to the optic disc, when the size ratio exceeds 30%. Currently, this requires a trained ophthalmologist to annotate a fundus image (inner eye scan). However, recent improvements in deep learning models and image analysis suggest that automated detection of the cup and optic disc may be possible. We implement RetinaNet, a state-of-the-art deep learning model, for this purpose. RetinaNet has a faster processing time than Faster RCNN and Mask RCNN, and has similar accuracy. The training and testing images were taken from NEI (AREDS) and several publicly available fundus image databases. The results are promising, over 99% accuracy and less than a fifth of a second processing time per image.
Ben Kussmaul is a summer researcher at the Communications Engineering Branch, Lister Hill National Center for Biomedical Communications, working with Dr. Jongwoo Kim. He is a rising junior at Swarthmore College in Pennsylvania, and is majoring in Computer Science and Engineering.