ART-VENA: Retinal Vascular Caliber Measurement

Antonio Mosquera1, Raquel Dosil1, Víctor Leborán1, Fernando Pardo1,

Francisco Gómez-Ulla2,3, Bashir Hayik3, Antonio Pose4 and Marta Rodríguez4
1
Grupo de Visión Artificial, Departamento de Electrónica e Computación, Universidade de
Santiago de Compostela. Campus Universitario Sur s/n, 15706 Santiago de Compostela, Spain.
mosquera@dec.usc.es, {rdosil,nelbla,fpardo}@usc.es
2
Departamento de Cirugía, Universidad de Santiago de Compostela.
San Francisco s/n, 15705 Santiago de Compostela, Spain.
ciulla@usc.es
3
Servicio de Oftalmología, Hospital de Conxo, Complejo Hospitalario Universitario de
Santiago de Compostela; Conxo, 15706 Santiago de Compostela, Spain.
4
Servicio de Medicina Interna, Hospital de Conxo, Complejo Hospitalario Universitario de
Santiago de Compostela; Conxo, 15706 Santiago de Compostela, Spain.

Abstract. The size of retinal vascular caliber in eye fundus images is a funda-
mental diagnosis parameter in the study of systemic vascular pathologies, like
arterial hypertension or arteriosclerosis. ART-VENA is a semiautomatic system
to measure the retinal vascular caliber. From the medical point of view, its re-
peatability (coefficients of variation under 1.5%) turns it into a reliable tool to
objectify vascular changes which previously depended on the observer’s sub-
jectivity.

1 Introduction

The automatic computerized analysis of blood vessels from medical images has
played, in the last years, an important role in many clinical and research studies on a
large number of diseases. The presence of noise, the variability of the anatomical
fundus and the low and changing contrast of vessels in many image modalities make
the reliable vessel detection a difficult task. Therefore, automatic detection becomes a
key starting point. Particularly, retinal vascular caliber is a fundamental parameter in
the study of systemic vascular pathologies, like arterial hypertension or arteriosclero-
sis. Vessels present in the eye are the first to manifest the symptoms of the arterial
hypertension [1]. The abnormalities in the retinal vascularization can reflect the de-
gree of damage provoked by these diseases, an aspect that decisively influences the
development of brain and cardiovascular complications.
Before the analysis of the alterations in the retinal vascular caliber due to these sys-
temic diseases, it is necessary to study the variations caused by age in normal patients.
This is a controversial topic in the medical field. Several classic authors have ob-
served that old age is directly related to arteriolar caliber alterations in normal pa-
tients, though there is not an agreement on the causes of this relation [2, 3, 4]. On the
contrary, other authors do not find retinal vascular caliber changes in old patients [5,
6, 7]. It has to be taken into account that these descriptions are highly subjective.
It would be very useful, therefore, to develop an effective and reproducible method
to objectively evaluate the retinal vascular caliber. This objective measure could de-
finitively determine whether observed alterations are related to age in normal patients,
and, in case that this relation exists, it could distinguish between these changes and
those originated by general diseases, like arterial hypertension or arteriosclerosis.

1.1 Retinal Vascular Caliber Measurement

Retinal vascular caliber measurement has been a widely studied subject in the area of
Ophthalmology. In the middle of the 20th century many works begin to be carried out
with the aim of predicting, from these measures, the state and prognosis of diseases
like arterial hypertension or arteriosclerosis.
In a first stage, from the 50’s to the 80’s, vascular width measurement from eye
fundus photographs was manually made by means of retinal image projections [8, 9,
10, 11, 12, 13]. Several works prove that these measures can be performed with an
error inferior to 3µm at the retina level [11, 14]. Obtained results turned out to be
contradictory in many cases, probably as a result of the high subjectivity of the analy-
sis and the lack of statistical studies on the acquired data. Brinchmann-Hansen and
col. realized the first studies statistically analyzed [15]. They found the caliber of three
arteries and three veins, achieving a coefficient of variation of 3.6 %. Posterior studies
reach a decrease in the coefficient of variation down to 2.3% [16, 17].
At the end of the 80’s decade, computerized vessel measurement techniques ap-
peared. In the main, these techniques use the computer just to visualize and store the
data. The spotting of blood vessel boundary points is made manually by the specialist.
Some works reached coefficients of variation between 0.6% and 4.0%, with mean
value around 2.2% [18, 19, 20]. Especially important are the works of Polak and col.
on the reproducibility and sensitivity of the Zeiss’ analyzer of retinal vessels (Zeiss
RVA) [21]. In size measurements of arteries and veins they obtained coefficients of
variation values of 1.3% and 2.6% respectively. These data place the Zeiss’ vessel
analyzer among the reference commercial instruments in this area.

1.2 Vessel Detection

Automatic detection of vessels has involved much research in the field of digital im-
age processing during the last years. Many works in the literature address the problem
of automatic detection of blood vessels in various domains.
Tracking of vessels has been one of the more studied techniques in recent years
[22, 23]. In general, given a starting vessel contour point and one search direction,
vessel tracking involves exploring the image to find other contour points. The crite-
rion to define boundary points is based on the analysis of the pixel gray level values
on a section that is orthogonal to the tracking direction. More sophisticated versions of
this idea transform the problem of vessel contour detection in that of graph optimal
path search [24]. Other artificial intelligence or computer vision methods have been
also proposed to obtain a complete automation of the detection procedure, exploiting
in a more exhaustive way the available anatomical knowledge [25, 26].
The majority of the developed methods adopt boundary detection as the basic tool
for vessel detection. Since vessel contours are usually smooth and fuzzy and the fun-
dus is noisy, a number of authors have applied, as an initial strategy, the maximization
of vessel enhancement. Some approaches are based on rotation invariant operators or
linear morphological operators of suitable size and orientation, which are applied to
the image to minimize noise [25, 27]. Alternative approaches involve image convolu-
tion with multiple filters per model, each of them designed to detect vessels of given
size and orientation [28]. The main drawback to these methods is their high computa-
tional cost, which renders them practically unusable for on-line diagnosis assistance
systems or systems requiring relatively short response times. Zana and Klein devel-
oped a vessel bifurcation detection method as a previous step to image registration
[29]. Quek and Kirbas carried out a method for vessel extraction from angiographies
based on wave propagation [30].

2 Retinal Vascular Caliber Measurement

The measurement method developed in this work is nearly completely automatic. The
operator is initially required just to mark the optic nerve, though high accuracy is not
necessary. Once this initial reference is set, the automatic sequence of detection and
measurement of the vessels on the image starts. The method output is a list of the
structures identified as vessels and their widths, ready to be analyzed by the expert.
As the fundus vascular structures are inherently radial, the width measurement of
retinal vessels will be performed over various circumferences centered on the optic
nerve, at distances that are multiple of the optic disk radius, depicted by the operator
(a common procedure in Ophthalmology). Using a polar coordinate space, a linear
representation of the gray level profiles is obtained for each of the analyzed circum-
ferences, as shown in left side of Fig. 1. The gray level profile along each circumfer-
ence is characterized by the presence of valleys, which are potential vessels. They are
placed over the noise of the anatomical fundus that is nearly uniformly distributed.
Before its detection, a preprocessing stage is needed to minimize noise effects.
The employed filter for noise reduction must not modify the location of the profile
transition regions, which determine vessel widths in the measurement phase. To mini-
mize noise, a median filter has been chosen. It permits to remove gray level values
that differ too much from the surrounding values. This filtering reduces the detection
of valleys caused by noise perturbations induced by the anatomical fundus. The me-
dian filter used in this work has a window size of 11 pixels centered in the pixel under
analysis. One important property of this filter is that it does not alter the magnitude of
the gray level gradient and, therefore, respects the position of transitions. The right
side of Fig. 1 shows the profile obtained after the filtering process.
Although the median filter eliminates many noise valleys, some of them, produced
by illumination variations and the presence of fundus anatomical structures other than
Fig. 1. Left: Gray level profile along one circumference centered at the optic nerve. Right: Gray
level profile after median filter application.

vessels, still remain, together with the blood vessel valleys. For that reason, the simple
detection of filtered sequence minima does not warranty correct identification of ves-
sels. The definitive detection of blood vessels valleys is established as a function of
their 1D and 2D morphology. In this work the hypothesis that the traversal vessel
profiles are deep is assumed (1D model), i.e., their gray levels are much smaller than
the surrounding fundus ones. Consequently, the intensity differences between vessel
valleys and fundus are greater than those of noise valleys. Moreover, the vessel valley
structure is stable in the radial direction regarding the optic disk (2D model).
Fig. 2 shows the 1D vessel model. Points x1, x2 and x3 correspond respectively to
the sites of the previous maximum, the minimum and the posterior maximum on the
valley structure. Parameters y1, y2 and y3 are the profile heights (gray levels) measured
at x1, x2 and x3 respectively. Using the three height parameters, a valley quality meas-
ure reflecting relative depth is established. This measure is given by the expression
2 y2
Q = 1− . (1 )
y1 + y3

Valleys with Q values greater than 0.06 are considered candidates to form blood
vessels. This process is realized individually over 2n+1 circumferences, with n posi-
tive integer and with their radius differing in one pixel, being the innermost circumfer-
ence radius selected by the specialist. The n
value, which can be modified by the operator, is
recommended to be set between 5 and 9.
Smaller values do not warranty the straight
vessel hypothesis which will be applied in the
width measurement stage. These data are the
input to the 2D modeling phase. For a region to
be considered a true vessel, a valley must appear
in 2/3 of the analyzed circumferences. In addi-
tion, the angular distance between valleys asso-
ciated to consecutive circumferences must be
lesser than 0.089 radians. In the top left hand
side of Fig. 3, a sequence of 11 valleys (n=5)
can be observed, which verifies the 1D and 2D
imposed conditions, so that they determine the
existence of a blood. Fig. 2. Vessel model
Fig. 3. Left: Sequence of profiles verifying 1D and 2D modeling conditions. Right: Geometric
calculus of the straight line that is used to determine de the blood vessel width.

When the vessel positions are located, the next step is width measurement. This
task is accomplished by locating the extremes of the gray level derivative along the
angular direction, i.e., along the profiles. Hence, positions of vessel walls are set at the
maximum transition points of the intensity profile. These extreme points are obtained
by means of quadratic interpolation (fit to a parabola) of the derivative values, to
obtain subpixel accuracy. In this way, potential image quantization ambiguities are
avoided and results are better assessed. Assuming the hypothesis of the analyzed ves-
sel segment is straight, the available points can be fitted to a straight line, as shown on
the bottom left hand side and on the right hand side images in Fig. 3. Caliber is finally
estimated over a line perpendicular to the bisector of the lines fitted to both vessel
borders, as can be seen in Fig. 3.

3 Results

In order to validate the developed procedure, a reproducibility/reliability study has

been designed in three stages. These phases are intended to evaluate the reproducibil-
ity/reliability of the method against diverse factors that can influence the study. These
factors include variability on the measurements made by an observer, variability
among different observers and variability due to the patient under study.
Once the goals of this experience were explained to the volunteers, chosen from the
companion of the patients that came to the Servicio de Urgencias de Oftalmología of
the Complejo Hospitalario Universitario de Santiago de Compostela (Spain) and after
they had signed the informed consent to proceed with pupil dilation and an ophthal-
mologic exam, the volunteer pupil was dilated using an instillation of two drips of
tropicamide and two drips of phenylephrine. After a period of 20 to 30 minutes, an
eye fundus photograph was taken from both eyes of each subject at 50 degrees, with
aneritra light (540 nanometers) and centered at the papilla. The camera used to take
the fundus photographs was the Topcon TRC-50 IA connected to the IMAGEnet 1024
Digital Imaging Systems by Topcon Instruments, Paramus, NJ. All captured images
were directly stored in optical disks and displayed on a high resolution monitor
(1280x1024 pixels) for their visualization and posterior analysis. The direct digitaliza-
tion of the captured image was made by means of a digital camera connected to the
aforementioned fundus camera. The elimination of the intermediate step, usually in-
cluded in many other procedures, avoids the introduction of error factors [31].
To achieve the objectives of this work, a basic population sample was formed by
120 normotensive volunteers (240 fundus images), without any known vascular pa-
thologies. A subject is classified as normotensive in the absence of a previous diagno-
sis of arterial hypertension and the presence of PAS/PAD quantities under 140/90,
following the procedures of the American National Committee of HBP [32]. Subject
ages ranged from 10 to 69 years, with a 1:1 ratio between sexes. Each age group (10-
19; 20-29; 30-39; 40-49; 50-59; 60-69) is constituted by 20 individuals, 10 males and
10 females. In all cases the volunteers had transparent media in both eyes and a refrac-
tive defect inferior to 3 diopters of hypermetropia, myopia and astigmatism.
Table 1 shows the results obtained for the coefficient of variation over the individ-
ual measures and the Pearson’s momentum correlation coefficient (p<0.001) over the
global measures. The intraobserver reproducibility indicates agreement in measure-
ments made by the same observer when repeating the measure over the same set of
images. To analyze this factor, one single observer evaluated 40 images corresponding
to 40 eyes of 20 individuals (20 left eyes and 20 right eyes) two times, separated from
each other by an interval of 24 hours, without knowing nor the results neither the
measurement points in the previous attempt. The interobserver reproducibility indi-
cates agreement in measurements realized independently by two different observers.
In this case, two observers evaluated, without information exchange, 30 images corre-
sponding to 30 eyes of 15 individuals (15 left eyes and 15 right eyes). Intraindividual
reproducibility indicates concordance in measurements, realized in a blind fashion, by
the same observer over images of the same subject and same eye taken with a differ-
ence in some time interval. To analyze this factor, one single observer evaluated two
images of the same eye, separated by a time period of one month, repeating the proc-
ess for 30 eyes of 15 individuals (15 left eyes and 15 right eyes), without knowing nor
the results neither the measurement points in the previous trial.

Table 1. Coefficient of variation (CV) and Pearson’s momentum correlation coefficient (r)
obtained in the various planted experiments.

CV (r) C
Intraobserver Interobserver Intraindividual Polak, 2000
Artery 1.10 % (0.96) 1.30 % (0.94) 1.08 % (0.93) 1.3 %
Vein 0.98 % (0.91) 1.47 % (0.86) 1.12 % (0.93) 2.6 %
Arteriovenous index 0.50 % (0.99) 0.99 % (0.95) 0.40 % (0.99)

4 Discussion

Reproducibility of measurement methods is a crucial aspect in Medicine [31]. For this

reason, the first planted goal in this work is to demonstrate that repeated vascular
caliber measurements over the same photograph without knowing the reference points
bring the same results. As a second objective, it has been proposed to demonstrate that
the same thing happens when one certain photograph is measured by two different
observers. And, as a third objective, to demonstrate also that vascular caliber meas-
urements from two photographs of the same individual with a time difference of one
month do not present significant variations.
Intraobserver, interobserver and intraindividual variabilities are small in our
method and the measurement reproducibility in the three groups presents a very high
degree of concordance, with a Pearson’s momentum correlation coefficient (equiva-
lent to the intraclass correlation coefficient) over 0.86 and a coefficient of variation
inferior to 1.47% in all cases, while Polak et col. obtained coefficient of variation
values between 1.3% and 2.6% in their work with the Zeiss’ retinal vessel analyzer
(Zeiss RVA) [21]. In particular, this coefficient of variation superior bound is the best
obtained by the whole of the previously commented methods. These results render the
presented method a secure tool to objectify microvascular changes which previously
depended on the observer’s subjectivity.
The research project “Desarrollo y Validación de un Método de Medida del Calibre
Microvascular Retiniano en la Población Normal” carried out using ART-VENA
(inscribed in the Spanish IP Register, with number 2001/15/934) was prized with the
Premio de Investigación Academia Médico-Quirúrgica /Consellería de Sanidade of
the Xunta de Galicia (Spain).

References

1. EYEfacts, “Systemic Disease and the Eye”, UIC Eye Center, University of Illinois at Chi-
cago (2002)
2. Ballantyne, A.J., Trans. Ophthalmol. Soc. UK., 57, 301-318 (1937)
3. Vogelius, H. and Bechgaard, P., “The Ophthalmoscopical Appearance of the Fundus Oculi
in Elderly Persons with Arteriosclerosis and Normal Blood Pressure”, Br. J. Ophthalmol.,
34, 404-408 (1950)
4. Leishman, R., “The Eye in General Vascular Disease: Hypertension and Arteriosclerosis”,
Br. J. Ophthalmol., 41, 641-701 (1957)
5. Kagan, A., Aurell, E. and Dobree, J.A., “A Note of Signs in the Fundus Oculi and Arterial
Hypertension Conventional Assessment and Significance”, Bull. World Health Organ., 34,
955-960 (1966)
6. Aoki, N., Horibe, H. and col., “Epidemiological Evaluation of Funduscopic Findings in
Cerebrovascular Diseases III: Observer Variability and Reproducibility for Funduscopic
Findings”, Jpn. Circ. J., 41, 11-17 (1977)
7. Dimmitt, S.B., West, J.N. and col., “Usefulness of ophthalmoscopy in mild to moderate
hypertension”, Lancet, 1, 1103-1106 (1989)
8. Boyd, T.A.S. and de Margerie, J., “Caliber of Retinal Arterioles in Hypertension”, Trans.
Can. Ophthalmol. Soc., 23, 65-76 (1960)
9. Hill, D.W. and Dollery, C.T., “Caliber Changes in Retinal Arterioles”, Trans. Ophthal.
Soc. UK., 83, 61-70 (1963)
10.Michaelson, I.C., The Ocular Circulation in Health and Disease, Can. J. S., Ed. Kimpton,
London (1969)
11.Hodge, J.V., Parr, J.C. and Spears, S., “Comparison of Methods of Measuring Vessel
Widths on Retinal Photographs and the Effect of Fluorescein Injection on Apparent Retinal
Vessel Calibers”, Am. J. Ophthalmol., 68, 1060-1068 (1969)
12.Sanchéz-Salorio, M., Retinopatías Vasculares., Ed. Grafinsa (1971)
13.Parr, J.C. and Spears, G.F.S., “General Caliber of the Retinal Arteries Expressed as the
Equivalent Width of the Central Retinal Artery”, Am. J. Ophthalmol., 77, 472-477 (1974)
14.Parr, J. C., “Hypertensive Generalized Narrowing of Retinal Arteries”, Trans. Ophthalmol.
Soc. NZ, 26, 55-60 (1974)
15.Brinchmann-Hansen, O. and Heier, H., “The Apparent and the True Width of the Blood
Column in Retinal Vessels”¸ Acta Ophthalmologica, Supplement 179. 64, 29-32 (1986)
16.Delori, F.C., Fitch, K.A. and col., “Evaluation of Micrometric and Microdensitometric
Methods for Measuring the Width of Retinal Vessels Images on Fundus Photographs”,
Graefe’s Archive for Clinical and Experimental Ophthalmology, 226, 393-399 (1988)
17.Stanton, A.V., Mullaney, P. and col., “A Method of Quantifying Retinal Microvascular
Alterations Associated with Blood Pressure and Age”, J. Hypertens., 13, 41-48 (1995)
18.Meehan, R.T., Taylor, G.R. and col., “An Automated Method of Quantifying Retinal Vascu-
lar Response during Exposure to Novel Environmental Conditions”, Ophthalmology, 97,
875-881 (1990)
19.Rassam, S.M.B., Patel, V. and Col, “Accurate Vessel Width Measurement from Fundus
Photographs: A New Concept”, Br. J. of Ophthalmology, 78, 24-29 (1994)
20.Pedersen, L., Grunkin, M. and col., “Quantitative Measurement of Changes in Retinal
Vessel Diameter in Ocular Fundus Images”, Pattern Recognit. Lett., 21, 1215-1223 (2000)
21.Polak, K., Dorner, G. and col., “Evaluation of the Zeiss Retinal Vessel Analyser”, Br. J.
Ophthalmol., 84, 1285-1290 (2000)
22.Liu, I. and Sun, Y., “Recursive Tracking of Vascular Networks in Angiograms based on the
Detection-Deletion Scheme”, IEEE Trans. Med. Imaging, 12, 2, 334-341 (1993)
23.Miles, F.P. and Nuttall, A.L., “Matched Filter Estimation of Serial Blood Vessels Diame-
ters from Video Images”, IEEE Trans. Med. Imaging, 12, 2, 147-152 (1993)
24.Sonka, M., Wilbricht, C.J. and col., “Simultaneous Detection of both Coronary Borders”,
IEEE Trans. Med. Imaging, 12, 3, 588-599 (1993)
25.Xia, W. and Lü, W., “Correspondence Analysis for Regional Tracking in Coronary Arte-
riograms”, IEEE Trans. Med. Imaging, 11, 2, 153-160 (1992)
26.Coppini, M., Demi, M. and col., “An Artificial Vision System for X-ray Images of Human
Coronary Trees”, IEEE Trans. Pattern Anal. Mach. Intell., 15, 2, 156-162 (1993)
27.Thackray, B.D. and Nelson, A.C., “Semi-automatic Segmentation of Vascular Network
Images using a Rotation Structuring Element (ROSE) with Mathematical Morphology and
Dual Feature Thresholding”, IEEE Trans. Med. Imaging, 12, 3, 835-392 (1993)
28.Gerig, G., Koller, G. Th. and col., “Segmentation and Symbolic Description of Cerebral
Vessel Structure Based on MR Angiography Volume Data”, Computer Assisted Radiology
CAR`93, Lemke, H. U., Inamura, K., Jaffe, C.C. and Félix, R., editors, 359-365 (1993)
29.Zana, F. and Klein, J.C., “A Multimodal Registration Algorithm of Eye Fundus Images
using Vessels Detection and Hough Transform”, IEEE Trans. Med. Imaging, 18, 5, 419-428
(2001)
30.Quek, F.K.H. and Kirbas, C., “Vessel Extraction in Medical Images by Wave-Propagation
and Traceback”, IEEE Trans. Med. Imaging, 20, 2, 117-131 (2001)
31.Couper, D.J., Klein, R. and col., “Reliability of Retinal Photography in the Assessment of
Retinal Microvascular Characteristics: The Arteriosclerosis Risk in Communities Study”,
Am. J. Ophthalmol., 133, 78-88 (2002)
32.Joint National Committee on the Prevention, Detection, Evaluation, and Treatment of High
Blood Pressure, Sixth Report Arch. Intern. Med., 157, 21, 2413-2446 (1997)