Page 109 - Computational Retinal Image Analysis

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

scenarios when no reference is available. This is tackled by, for example, the paper of
Galdran et al. [45], where a similarity metric is presented to quantify the segmentation
performance in the absence of reference annotation. A web-based client-server
system is described in Ref. [46], which consists of a component for vessel tracing, an
interactive editing interface for refining parameters and fixing prediction errors, and
a component responsible for outputting the clinical indexes.

3 Vessel segmentation
It is often of foremost interest to segment the retinal blood vessels pixel by pixel in a
retinal image, corresponding to the blood column within the vessels. An example of
vessel segmentation on CF images is shown in Fig. 1D.
Existing methods can be roughly divided into two categories based on whether
an annotated training set is required: unsupervised and supervised methods.
While supervised methods learn their models based on a set of training examples,
unsupervised methods do not require a training set.

3.1 Unsupervised segmentation
As mentioned, arguably the first research effort of retinal segmentation is by Matsui
et al. [10], 45 years ago, where a mathematical morphology approach is taken toward
the problem of vessel segmentation.
Matched filtering and mathematical morphology-based techniques were
popular choices in the early days. In Ref. [47], Chaudhuri et al. approximate the
vessel segment profiles by a set of 1D Gaussian shaped filters over various spatial
directions. As part of an initiative at the Vienna Eye Clinic for diagnosis and treatment
of AMD, a dedicated pipeline is presented [48] to analyze retinal images based on
SLO. In addition to separate modules for detecting the optic disk, fovea, and scotoma
locations, its vessel segmentation module operates by matching the prototype zero-
crossing of the gradient profiles and applying grouping. Hoover et al. [25] discuss the
application of 12 Gaussian shape-matched filters in 2D, which produces a response
image recording the highest response at each pixel. This is followed by a sequence
of carefully designed threshold probing steps to produce the predicted segmentation.
The STARE dataset is introduced in this paper. A second-order Gaussian matched
filtering approach is considered by Gang et al. [49] to detect vessels. The work of
Staal et al. [24] focuses on extracting the centerlines or ridges of the 2D vessels,
which is achieved by means of a k-nearest neighbor classifier and sequential forward
feature selection-based method. Importantly, it contributes the widely used DRIVE
dataset. A set of three vessel detection variants are studied in Ref. [50] to account for
noisy input images. The methods are derived from the likelihood ratio test in making
local decisions, and model selection for choosing the optimal intrinsic parameters.
A supervised likelihood ratio test is developed in Ref. [51] that combines matched-
filter responses, and customized measures of confidence and vessel boundary-ness.