Page 117 - Computational Retinal Image Analysis

110 CHAPTER 6 Retinal vascular analysis: Segmentation, tracing, and beyond

Diameter changes. AVR, AVN, and FAN are three quantification measurements
of diameter changes. AVR calculation is restricted to an area of 0.5–1.0 disc
diameters from the OD. An automated pipeline is presented in Ref. [139],
which segments and skeletonizes the vessels, and classifies arterial/venous
vessel segments that are separated at the junction points. The six widest artery
and vein segments are selected for AVR calculation, which is measured by an
iterative algorithm. AVN is a phenomenon where venular caliber decreases
as an arteriole crosses over a venule. A four-level grading approach of AVN
is proposed in Ref. [140]. FAN refers to arterial vascular segment whose
diameter ≥50 μm narrows. Severity degree of FAN is evaluated by the length
of narrowing vessels compared with the diameter of OD.
Tortuosity alteration. This is an early indicator of a number of vascular diseases.
Some of tortuosity quantification of arteries and veins approaches are (1)
tortuous or nontortuous classification; (2) tortuosity ranking of vessel segments
[34]; and (3) tortuosity grading of individual vascular trees [141].

5 Summary and outlook

A significant amount of effor has been devoted to vasculature analysis from retinal
images, which have led to noticeable progress in clinical quantifications to improve
diagnosis and prognosis of related diseases. There are also a number of promising
research directions, some of them discussed in the following sections.

5.1 Vasculature analysis in emerging imaging techniques
We highlight here the emerging retinal imaging techniques referred to as 3D,
multimodal, and mobile imaging.

3D vessel analysis. To date, the majority of existing benchmark datasets and
research efforts are devoted to segmentation in 2D retinal fundus images. Being
a 2D projection of the 3D retinal vasculature, the retinal fundus images contains
necessarily only a partial observation of the underlying 3D vessels. Meanwhile,
with emerging 3D imaging techniques such as spectral-domain OCT (SD OCT
or 3D OCT) and plenoptic ophthalmoscopy [142], we are now capable of
imaging the 3D vasculature volume of the retina. It is thus possible to directly
extract the 3D retinal vasculature volumes.
The endeavors of Haeker et al. [143] and Garvin et al. [144] are among the first
in devising dedicated 3D segmentation techniques for time-domain macular
scans. In terms of spectral-domain OCT volumes, the work of Niemeijer
et al. [145] considers a k-NN pixel classification approach where Gaussian
filter banks are used to produce good features. The results are evaluated on
the macular centered scans as well as the optical nerve head centered scans.
An interactive 3D segmentation approach is developed by Fuller et al. [146].