Quantifying Phenotypic Traits in Retinal Coronary Angiography: Automated Extraction of Retinal Vascular Networks and Localization of Optic Discs in Fundus Images
Numerous retinopathies are related to the dysfunction of retinal vasculature, especially micro-vessels. Extensive research in ophthalmology has singled out critical roles of vascular morphology, and the functional dynamics of blood flow in diseases. Advances in angiography has yielded a myriad of applications for computational methods that design efficient tools to complement retinal imaging and microscopy in analytic ophthalmology. In this paper, we propose a novel mathematical approach for the design of quantitative tools that enable researchers, as well as automated vision-based systems, to perform pattern recognition, and feature extraction in retinal vasculature. The present feasibility-stage implementation of these new algorithms demonstrates the power and versatility of the set of tools we provide for the detection of morphological pathology, as well as the theoretical study of retinal neurovasculature anatomy when regarded as a complex (dynamic) system. In contrast to current state-of-the-art methods that rely on bottom-up algorithms to deal with noise and trace the vessels, we propose a top-down scheme to overcome noise and capture morphological features such as center-lines, radii, and the edge locations of circulatory blood vessels. This approach is comprised of three components. First, the algorithms for detection and measurement of the vasculature morphological structures in two-dimensional fundus images are implemented. These algorithms combine advanced kernel-based methods to extract blood vessels, and are further enhanced by variants of Canny Edge Detection algorithms. Second, a fully automated approach is provided to identify the optic disc in healthy/diseased fundus images, eliminating current bottle-necks requiring extensive human expertise. Third, we construct a hierarchical network of geometric (topological) structures of the extracted vessels, rooted in the optic disc. A notable application of our methods is to capture complex vasculature structures in noisy, blurred, and light-reflecting fundus images. Another advantage of our approach is the automation of in vivo quantification of complex phenotypic traits of retinal neurovasculature, which are expected to play an important role in emerging computational models for mapping genotype-phenotype relations and personalized medicine.
展开▼
