Optical coherence tomography and deep learning for ophthalmology

Robust quantitative tools require large data sets for testing efficacy and accuracy, which is especially true when using machine learning and neural networks. However, large datasets with corresponding manual annotations are uncommon with state-of-the-art imaging systems, particularly in the medical field. Ophthalmology is one such field, for which recent imaging advances allow clinicians to use multiple imaging modalities to diagnose and monitor patients. Optical coherence tomography (OCT) has become an integral imaging modality in ophthalmic clinics due to its non-invasive nature and ability to acquire micrometer scale sub-surface images of ophthalmic tissue. In this thesis, several different techniques to mitigate the need for large annotated datasets when translating machine learning tools to an ophthalmic clinic are evaluated. First, the concept of transfer learning is assessed through fine-tuning networks trained on a different domain (adaptive optics scanning laser ophthalmoscopy) to the domain of interest (adaptive optics OCT) to detect cone photoreceptors. Second, both adversarial and semi-supervised learning are investigated which allow for unlabelled data to be used in the training process. Finally, the more challenging task of diagnostics with limited data was investigated using diabetic retinopathy OCT Angiography data and an ensemble of networks. Through these investigations, the utility of transfer learning, adversarial and semi-supervised learning, and ensembling is shown for small ophthalmic datasets.