Page 113 - Computational Retinal Image Analysis

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

annotations. To address this problem, Zhao et al. [93] present a supervised learning
pipeline. The key is the construction of a synthetic retinal image dataset capable
of bridging the gap between an existing reference dataset (where annotated vessel
structures are available) and the new query dataset (where no annotation is available).
Then existing supervised learning segmentation methods are to be engaged to learn
a model dedicated to the set of query images.
Due to the visual similarity between retinal vessels from, for example, fundus
imaging and neurons from, for example, confocal microscope, there are also many
efforts in devising algorithms to address the more general problem of segmenting
such tubular structured objects [85, 94–100], with retinal blood vessel being a special
case. For example, the well-known local Hessian-based vessel enhancement method
developed by Frangi et al. [94] has been widely used for segmenting both 2D and
3D vessels. A mathematical morphology and curvature evaluation-based method
was developed by Zana and Klein [95] to segment the vessel-like structures from
background. Minimal path techniques are considered in a series of closely related
efforts [101–103] to connect vessel segments in 2D or 3D. The work of Benmansour
and Cohen [96] also addresses generic 2D and 3D segmentation of tubular structured
objects in images, with an interactive method using minimal path and anisotropic
enhancement. A multiscale centerline detection method is presented in Ref. [98]
based on learned filters and GradientBoost regression technique. The method of Gu
et al. [100] attempts to capture structural and contextual features from fundus images
through their proposed data-driven feature learning procedure.

4 Vessel tracing

The topological and geometrical properties of retinal vessel trees are vital in screening
for and diagnosing diseases, which call for proper tracing of the individual vessel
trees from fundus images. The problem of vessel tracing is more than segmentation
where vessel pixels are separated from the backgrounds, in that we would like to
first separate artery and vein vessels, as shown in Fig. 1E. An equivalent problem
is to detect junctions and decide as being either branching or crossing. Individual
vessel trees usually could be distinguished from the rim of the optic disk, as shown
in Fig. 1F, thus we answer the question of which pixel belongs to which of the vessel
trees.

4.1 Vascular junction identification
Detecting and categorizing the junction points into either bifurcation or crossing is
useful in vessel tree extraction and classification. Fig. 1B illustrates an exemplar
annotation of such bifurcation and crossing points. In one of the early efforts, a
morphological-type edge detection algorithm is used in Ref. [104] to automate the
recognition of arteriovenous intersections. An image processing-based approach
is considered in Ref. [105] that first extracts the vessels by preprocessing filtering