Page 116 - Computational Retinal Image Analysis

4 Vessel tracing 109

is described in Ref. [133], consisting of segmenting the vessel pixels, thinning to
obtain vessel skeletons, and tracking individual vessel trees. As a result, a number
of geometric and topological properties are subsequently quantified, which include
vessel segment lengths, diameters, branching ratios, and angles. In Ref. [134], the
problem of artery versus vein separation from a vascular graph is cast as a SAT-
problem, where a heuristic AC-3 algorithm is utilized to address a double-layered
constrained search problem. Instead of working with fundus image of a single
wavelength, it is considered in Ref. [135] to work with a specific fundus imaging
set-up that acquires two images simultaneously, at different wavelengths of 570 and
600 nm, respectively. By exploiting the difference between artery and vein vessels
that have distinct central reflex patterns, an SVM classification-based method is
proposed to identify the vessel types. A multistep pipeline is considered in Ref.
[136]. First, a vesselness value is computed for each pixel, which forms a vessel
score image. Then points are sampled from local maxima points of the score image,
and these points are linked into spanning trees by solving the induced k-cardinality
arborescence problem with ant colony optimization-based algorithm. Based on 27
hand-crafted features, a linear discriminant classifier is used in the work of Niemeijer
et al. [33] to classify vessel segments into arteries and veins. In Ref. [26], Hu et al.
use a graph-based metaheuristic algorithm to segment and separate a fundus image
into multiple anatomical trees, while the RITE dataset is also constructed. In Ref.
[27], the classification of artery versus vein vessel segments is carried out based on
the analysis of a graph extracted from the retinal vasculature, where the annotation
of annotate artery and vein vessels is also provided for both DRIVE and INSPIRE-
AVR datasets. For ultra-wide field SLO, Pellegrini et al. [137] also consider a
graph-cut-based approach, where the node and edge-related features are manually
designed features based on local vessel intensity and vascular morphology. A system
is developed in Ref. [138] to segment vessels by Otsu binary thresholding, and obtain
skeleton by applying mathematical morphology. This is followed by localization
of branch points, crossing points, and end points. As a result, a graph structure is
formed, and the Dijkstra algorithm is used to search vessel subtrees by minimizing
the accumulative edge costs. Finally, a k-means clustering algorithm is executed to
separate artery and vein subtrees.
Targeting at wide-FOV images, a planar graph-based method is considered by
Estrada et al. [36] in constructing a likelihood model to take into account the prior
knowledge regarding color, overlap, and local growth aspects of vessel segments at
junctions. A heuristic search is thus carried out for optimal separation of the artery
and vein vessel trees.

4.4 Clinical relevant vessel readouts
Retinal vascular abnormalities, for example widening or narrowing of retina blood
vessels and increased vascular tortuosity, provide significant hints on various
diseases, such as diabetes, hypertension, and coronary heart disease. Measurements
to quantify retinal vascular changes are as follow: