Applications of deep learning for ophthalmic OCT

Optical coherence tomography (OCT) allows for the cross-sectional visualization of the retinal microvasculature and may benefit clinicians in their management of retinal diseases, such as diabetic retinopathy (DR). However, for DR management, this modality is limited by the field of view (FOV) and patient throughput. This thesis presents machine learning methods to aid in the acceptance of OCT as a secondary modality for DR screening and treatment to improve patient outcomes. First, an increase in FOV reduces the axial resolution. Lower axial resolution was simulated, and a super-resolution generative adversarial network successfully reconstructed lost features. Next, lower lateral resolution was simulated, and the results suggest OCT en face scans may be acquired with 3× fewer lateral scans without affecting a neural network's classification. Finally, machine learning was employed to provide clinicians with an automated classification of DR severity, and federated learning was leveraged to train a more generalizable neural network.