Measurements of retinal microvasculature in mice and humans with deep learning

Abstract

Retinal diseases such as diabetic retinopathies (DR) and radiation retinopathies (RR) show changes in the microvasculature as early symptoms. 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, as well as the difficulties of evaluation, very few studies address automated segmentations of microvascular networks. A commonly used imaging modality to record retinal vasculature is fluorescein angiography (FA). It is an invasive technique and uses fluorescein dyes to provide information of blood flow in the retina. Another relevant technology is optical coherence tomography angiography (OCTA). It is a newly developed non-invasive technique that captures regions with blood flow in the retina in 3D. Better ways to generate automated quantification for these modalities could potentially enhance clinical studies and help in the diagnosis and understanding of retinal diseases. In this work, automated approaches are developed to accurately quantify retinal microvasculature in mouse FA and human OCTA images. 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 traditional approaches and deep learning, we can automatically segment these structures with minimal human effort. Then, multiple methods are developed to automatically analyze the microvasculature in human OCTA projection images. The key novelties in these approaches are by using deep-learning based approaches, we can 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 projection segmentations. In order to achieve this goal, we first analyzed the effectiveness of regional vascular measurements as an indication of RR disease. Additionally, the microvascular differences between affected and unaffected eyes of the RR patients were also studied and compared to v healthy subjects. The proposed approaches are expected to provide more accurate and reliable automated measurements to analyze retinal microvasculature. vi

Dictionary-based Method for Vascular Segmentation for OCTA Images

“…Most studies obtain a segmentation through thresholding and filtering schemes [1,13]. A few studies utilize manually annotated data to create segmentation models, such as probabilistic models [8], convolutional neural networks [15], and Hessian-and deep learning-based methods [6] to segment all vessels. A single study [6] automatically segments main vessels and capillaries separately in retinal images using deep learning.…”