Medical Image Analysis 11 (2007) 47–61
M. Elena Martinez-Perez , Alun D. Hughes , Simon A. Thom , Anil A.
Bharath , Kim H. Parker
黃銘哲 2010/11/9
1
Outline
 1.Introduction
 2.Method
 3.Result
 4.Validation
 5.Conclusions
2
Introduction
 The eye is a window to the retinal vascular system which
is uniquely accessible for the non-invasive, in vivo study
of a continuous vascular bed in humans.
 Retinal blood vessels have been shown to change in
diameter, branching angles or tortuosity, as a result of a
disease, such as :
 Hypertension
 diabetes mellitus
 retinopathy of prematurity (ROP)
3
Introduction
 Green band is chosen in the present work because it is known
to show the improved visibility of the retinal blood vessels.
4
Multiscale
 Retinal blood vessels have a range of different sizes.
 Multiscale techniques have been developed to provide a
way to isolate information about objects in an image by
looking for geometric features at different scales.
5
Multiscale
 The effect of convolving an image with a Gaussian kernel
is to suppress most of the structures in the image with a
characteristic length less than s.
6
Feature extraction
 Gradient magnitude - The magnitude of the gradient
represents the slope of the image intensity for a particular
value of the scale parameter s.
7
Feature extraction
 Principal curvature - Since vessels appear as ridge-like
structures in the images, we look for pixels where the
intensity image has a local maximum in the direction for
which the gradient of the image undergoes the largest
change (largest concavity).
8
Feature extraction
 The second derivative information is derived from the
Hessian of the intensity image I(x,y):
 The eigenvalues, λ+and λ-, where we take λ+ ≥ λ-, measure
convexity and concavity in the corresponding
eigendirections.
9
Feature extraction
10
Feature extraction
 In order to analyse both red-free and fluorescein images
with the same algorithm, we define
λ1 = min(|λ+|, |λ-|) and λ2 = max(|λ+|, |λ-|).
 The maximum eigenvalue, λ2 , corresponds to the
maximum principal curvature of the Hessian tensor, which
we will refer to as maximum principal curvature.
11
Multiscale integration
 This might be expected, particularly for maximum
principal curvature, since the vessels are approximately
cylindrical so that the total amount of blood in the light
path corresponding to each pixel is larger in large vessels.
 Vessels with diameter d ≈ 2s are most strongly detected
when the scale factor is s, we normalised each feature
along scales by d and then kept the local maxima over
scales:
12
Multiscale integration
13
Region growing
 The region growing algorithm we use is based on an
iterative relaxation technique, using:
 Histograms of the extracted features, primarily upon the
maximum principal curvature κ.
 The classification of the eight-neighbouring pixels
14
Region growing
 In first stage, classes grow initially in regions with low
gradient magnitude, γ, allowing a relatively broad and fast
classification while suppressing classification in the edge
regions where the gradients are large.
 In second stage, the classification constraint is relaxed
and classes grow based solely upon κ to allow the
definition of borders between regions.
15
Region growing
16
Result
17
Comparison with average diameter
18
Conclusions
 MS doesn’t perform better than other method. However,
we concluded that an analysis of true and false positive
rate values does not give us the information needed about
the accuracy of segmenting vessels when geometric
measurements such as vessel widths are of interest.
 And from the third validation, we conclude that the
approach presented gives comparable estimates of
diameter and branching angles in both red-free and
fluorescein image.
19