Retinal microvascular changes can occur in retinal diseases including glaucoma, diabetic retinopathies (DR), and radiation retinopathies (RR). For instance, some of these diseases show sparser microvasculature around the foveal area as an early symptom. Therefore, automatically segmenting and analyzing retinal microvasculature may help diagnose and understand the underlying mechanisms of ocular diseases. However, due to the limitations of the image quality and the difficulties of manual tracing, this is not a trivial problem to solve and very few studies address automated segmentations of microvascular networks.
In this thesis, automated approaches with deep learning are developed to analyze microvasculature in fluorescein angiography (FA) and recently developed optical coherence tomography angiography (OCTA). These approaches accurately quantify retinal microvasculature in mice and humans.
First, we develop an automated approach to simultaneously segment retinal structures including the optic nerve head (ONH), main vessels, and the microvascular network in mouse FA images. By combining model-based approaches with deep learning, we can automatically segment these structures and require fewer manual tracings for training.
Then, multiple methods are developed to automatically analyze the microvasculature in human OCTA en-face images. The key novelties in these approaches involve the use of deep learning to not only segment the microvasculature but also directly find regions with different microvascular status. Furthermore, we investigated and analyzed the clinical significance of our region-based OCTA en-face segmentations. In order to achieve this goal, we first analyzed the effectiveness of regional vascular measurements as an indication of RR severity. Additionally, the microvascular differences between affected and unaffected eyes of the RR patients were also studied and compared to healthy subjects. These studies show that the usage of deep-learning networks provides accurate results for the analysis of retinal vasculature.
Overall, the main contributions of this thesis include: 1) combining model-based semi-automated segmentations and a fully automated deep-learning network and developing an approach to simultaneously segment multiple retinal structures in mouse FA images, 2) developing and evaluating three automated approaches to directly segment microvasculature in human OCTA en-face images, 3) developing an alternative deep-learning method to directly segment regions with different vascular states from the OCTA en-face images, and 4) using developed results from automated segmentations to measure the disease severity in radiation retinopathy patients.