A.Eye: Advanced Eye Imaging - Enhancing OCT and OCTA through Machine Learning Techniques

Optical Coherence Tomography (OCT) and Optical Coherence Tomography Angiography (OCTA) are emerging imaging techniques owing to their ability to noninvasively generate two-dimensional (2-D) and three-dimensional (3-D) in vivo crosssections of the human retina with a micron scale resolution. Due to their increasing use in ophthalmology, a growing amount of data is generated in the clinical routine that needs to be reviewed and processed. Developing automatic methods in order to help clinicians and researchers can be extremely useful for them to handle this amount of data. In the course of this thesis, several fully automatic algorithms for different steps in the image processing pipeline are developed, ranging from image enhancement via denoising to the segmentation of retinal structures to the classification and interpretation of diseases. One important step in the pipeline is the enhancement of image quality. OCT and OCTA images are degraded by speckle noise which renders the interpretation of fine details in these images a challenging task. However speckle also contains information about moving particles that is used to generate OCTA images. For the task of OCT B-scan denoising a new unsupervised loss is introduced that incorporates OCT as well as OCTA information which enables the algorithm to differentiate between the speckle information used to compute OCTA images and the undesired speckle noise and only remove the latter. Another important step in the pipeline is processing the generated data to make it usable for subsequent tasks. One challenge here is the reliable segmentation of retinal layers which can serve as boundaries for the computation of en face projections used in subsequent analysis steps. For this task a fully automated graph-based segmentation algorithm that jointly uses OCT and OCTA data to segment Bruch’s Membrane (BM) is presented. This combination of OCT and OCTA information can be more robust, compared to using only OCT or OCTA as input, in cases of pathology where the spatial relationships between the retinal layers can be distorted. A final step in the pipeline is to analyse and interpret the resulting data. In this thesis a network is trained for the first time to distinguish glaucoma patients from healthy controls using only OCTA data. An interpretation of attention maps leads to the conclusion that the network also focuses on lower perfused areas, like commonly used biomarkers, but performs better. Because of these observations, two follow-up works are conducted. First, a novel biomarker is introduced in order to quantify the areas highlighted by the attention maps providing an explainable measure instead of the black-box approach of network. Second, a comprehensive study is conducted to compare all combinations of acquisition areas and plexuses of OCTA scans in order to obtain insightful information on the best acquisition setup.