Adaptive optics retinal imaging in patients with retinal disease

High-resolution in vivo retinal imaging with an adaptive optics scanning laser ophthalmoscope (AOSLO) allows visualisation of degenerative changes at the cellular level. The improvement in resolution with this type of imaging when compared to conventional retinal imaging technologies has made the use of cell-specific imaging biomarkers possible. With the development of new therapies for retinal disease, cell-specific biomarkers of disease are becoming increasingly important for monitoring treatment outcomes. This thesis aims to investigate the AOSLO as a clinical tool and to develop fully automated image processing and analysis pipelines for AOSLO images from patients with retinal disease. Participants with Stargardt disease were investigated using multimodal conventional clinical imaging and high-resolution in vivo retinal imaging. Hyper-reflective spots, that were termed discrete hyper-reflective foci, were identified in AOSLO images in areas of retinal degeneration. We hypothesize that they may be cellular in nature and could represent cellular debris related to ongoing retinal degeneration. To optimise AOSLO image quality in patients with retinal disease, recommendations for future participant selection criteria and AOSLO image acquisition protocols are described. Reference frames are used to measure and correct intraframe eye motion in order to produce averaged AOSLO images with a high signal to noise ratio. A fully automated algorithm for reference frame selection was developed and validated using synthetic data simulating normal retinae. Although the results did not show a significant difference between selected reference frames, averaged images generated using the developed algorithm were significantly better when compared to the default method used for reference frame selection with the Oxford AOSLO. We hypothesise that the improvement in averaged images may be due to clustering of spatially overlapping frames that occurs in the developed algorithm, which may lead to an improvement in the subsequent step of eye motion correction. A deep learning approach for automated cone identification in AOSLO images of the photoreceptor mosaic was developed and validated. There was good agreement between the developed convolutional neural network (CNN), trained with synthetic and real data, and the current gold standard of manual marking. The developed CNN also marginally outperformed the results of current state-of-the-art automated cone detection algorithms. Although high-resolution in vivo retinal imaging provides substantial advantages over conventional clinical imaging, a number of technical, image processing and image analysis related challenges need to be overcome to promote its widespread clinical application.