Automated Age Related Macular Degeneration Diagnostics via Pre-Trained Deep Learning

Burlina, Philippe ;   Freund, David ;   Kankanahalli, Srihari ;   Neil, Joshi ;   Wolfson, Yulia ;   Bressler, Neil

Our high level goals relate to the development of artificial intelligence methods for automatically detecting and assessing the severity of age-related macular degeneration (AMD) from fundus imagery. Such methods target the development of new automated and intelligent AMD diagnostic tools and applications for point of care/preventive medicine as well as the deployment of self monitoring devices in home, pharmacies, clinical, and/or under-resourced settings. In this paper our efforts are to specifically investigate the possible use of deep convolutional neural networks (DCNNs) for AMD classification. Such approaches are motivated by their recent success and potential to generalize automated classification to large population deployment datasets by allowing for agnostic feature design. Preliminary results regarding feasibility of DCNNs use are very encouraging, with tests performed on NIH AREDS images yielding an accuracy of 89.30%.