Automated Arteriole and Venule Recognition in Retinal Images using

Ensemble Classification

M. M. Fraz

, A. R. Rudnicka

, C. G. Owen

, D. P. Strachan

and S. A. Barman

School of Computing and Information Systems, Faculty of Science Engineering and Computing,

Kingston University London, London, U.K

Division of Population Health Sciences and Education, St. George’s, University of London, London, U.K

Keywords: Medical Image Analysis, Retinal Image Processing, Artery Vein Classification, Ensemble Learning.

Abstract: The shape and size of retinal vessels have been prospectively associated with cardiovascular outcomes in

adult life, and with cardiovascular precursors in early life, suggesting life course patterning of vascular

development. However, the shape and size of arterioles and venules may show similar or opposing

associations with disease precursors / outcomes. Hence accurate detection of vessel type is important when

considering cardio-metabolic influences on vascular health. This paper presents an automated method of

identifying arterioles and venules, based on colour features using the ensemble classifier of boot strapped

decision trees. The classifier utilizes pixel based features, vessel profile based features and vessel segment

based features from both RGB and HIS colour spaces. To the best of our knowledge, the decision trees

based ensemble classifier has been used for the first time for arteriole/venule classification. The

classification is performed across the entire image, including the optic disc. The methodology is evaluated

on 3149 vessel segments from 40 colour fundus images acquired from an adult population based study in

the UK (EPIC Norfolk), resulting in 83% detection rate. This methodology can be further developed into an

automated system for measurement of arterio-venous ratio and quantification of arterio-venous nicking in

retinal images, which may be of use in identifying those at high risk of cardiovascular events, in need of

early intervention.

1 INTRODUCTION

With the development of digital imaging and

computational efficiency, image processing, analysis

and modeling techniques are increasingly used in all

fields of medical sciences, particularly in

ophthalmology (Abràmoff, Garvin et al. 2010).

Automated detection of micro-vascular disease such

as diabetic retinopathy in the retinal image using

digital image analysis methods has huge potential

benefits in screening programs for early detection of

disease (Fraz, Remagnino et al. 2012). The blood

vessel structure in retinal images is unique in the

sense that it is the only part of the blood circulation

system that can be directly observed non-invasively,

can be easily imaged using Fundus cameras.

Morphological characteristics of retinal blood

vessels (particularly width) have been prospectively

associated with cardiovascular outcomes in adult life

(Wong, Klein et al. 2001), and with cardio-

metabolic risk factors in early life (Owen, Rudnicka

et al. 2011). Associations between retinal vessel

morphology and disease precursors / outcomes may

be similar or opposing for arterioles and venules. For

instance, hypertension and atherosclerosis may have

different effects in retinal arterioles and venuels,

resulting in a decreased arteriole to venule width

ratio (AVR) (Jack J. Kanski and Brad Bowling

2011). Retinal arteriovenous nicking, a

pathognomonic sign of hypertension, is another

retinal feature worthy of study, characterized by a

decrease in the venular calibre at both sides of an

artery-vein crossing (Jack J. Kanski and Brad

Bowling 2011). However, more subtle changes in

arteriole / venular morphology may be an early

physio-maker of vascular health, which might

predict those at high risk of disease in middle and

later life. However, identifying small changes in

retinal arterioles and venules is a difficult task to

perform manually, as it is subjective, open to

measurement error, and time consuming, limiting its

use in large population based studies. Automated

194

M. Fraz M., R. Rudincka A., G. Owen C., P. Strachan D. and A. Barman S..

Automated Arteriole and Venule Recognition in Retinal Images using Ensemble Classiﬁcation.

DOI: 10.5220/0004733701940202

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 194-202

ISBN: 978-989-758-009-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

segregation of retinal arterioles and venules could be

used to assist with this task, which would be a pre-

requisite for the development of a computer assisted

tool for use in large populations to identify those at

high risk of disease.

The appearance of arterioles and venules in

retinal images are similar. The general assumption is

that there is a difference in the colour and size of the

venules and arterioles; the later one appears to be

thinner, brighter and present more frequently with a

central light reflex. However, there are some

challenges in building a robust vessel classification

system. There is intra-image and intra-subject

variance in the blood colour. The size and colour of

similar blood vessels changes as they move away

from the optic disc. In the periphery vessels become

so this they are almost indistinguishable. The

context based features may also fail at these

locations due to vessel crossings and branching. In

addition, the curved shape of the retina and non-

uniform illumination add complexity to the

automated vessel classification task.

A number of methods have been reported in

literature for retinal arteriole/venule (a/v)

classification, which can be divided into two broad

categories; automated and semi-automated methods.

In automated methods (Niemeijer, Xiayu et al. 2011;

Huang, Zhang et al. 2012; Dashtbozorg, Mendonca

et al. 2013; Nguyen, Bhuiyan et al. 2013), the vessel

centerline pixels forming the vascular skeleton are

extracted from the segmented vascular tree, followed

by the calculation of various distinguishable features

for each centerline pixel and finally each pixel is

assigned as an arteriole or venule by a classifier. In

semi-automated methods (Rothaus, Jiang et al. 2009;

Vázquez, Cancela et al. 2013), the initial pixels on

the main vessels are marked as arteriole or venule by

an expert, and then these labels are propagated

across the vascular network through vessel tracking

using the structural characteristics and connectivity

information.

Grisan’s method (Grisan and Ruggeri 2003) was

amongst the first to propose automatic a/v

separation. The main idea was to divide the optic

disc centered images into four quadrants with the

assumption that each quadrant will contain

approximately the same number of arterioles and

venules with significant differences in the features.

The variance of the red channel and mean of the hue

are used as vessel features, fuzzy clustering is

applied to each partition independently. In another

method (Saez, González-Vázquez et al. 2012), the

quadrants are rotated in steps of 20 degrees with the

aim of fulfilling the assumption that each quadrant

should contain at least one venule and one arteriole.

K-Means clustering is used to classify the vessels in

two concentric circumferences around the optic disc.

The quadrant-wise classification enforces a

condition to have at least one arteriole and one

venule per quadrant and it seems more suitable for

optic disc centered images rather than macula

centered images. Also, basic K-Means clustering is

sensitive to the initialization and may often become

stuck at a local optimal.

In this paper we have presented an automated

method for retinal a/v classification utilizing an

ensemble classifer of boot strapped decision trees.

The classifier based on the boot strapped decision

trees is a classic ensemble classifier, which has been

broadly applied in many application areas of image

analysis (Fraz, Remagnino et al. 2012), but has not

been extensively utilized for retinal vessel

classification. To our knowledge, this is the first use

of a decision trees based ensemble method for a/v

classification. An important feature of the bagged

ensemble is that the classification accuracy can be

estimated during the training phase, without

supplying the classifier with test data. Moreover, the

importance of each feature in classification can also

be predicted during the training phase, which helps

in identifying the most relevant features used in a/v

classification thus automatically reducing the

dimensionality of the feature vector and boosting

computational speed. The method is validated on 40

macula centered fundus photographs acquired from

20 middle-aged and elderly adults examined as part

of the latest phase of the European Investigation into

Cancer in Norfolk study (EPIC-Norfolk 2013). The

classification is not only performed near the optic

disc but across the entire image. The proposed

method achieves a high classification rate without

increasing the training samples or adding many

features.

The organization of the paper is as follows.

Section 2, presents the methodology for automated

segmentation of retinal blood vessels. Next, the

vessel classification methodology is explained in

section 3. Section 4 presents the validation

methodology and experimental results. Finally, the

discussion and conclusions are given in Section 5.

2 THE METHODOLOGY

The vascular network is segmented from the

coloured retinal image and the vascular skeleton

consisting of centerline pixels is constructed. Vessel

segments are generated by search and removal of

AutomatedArterioleandVenuleRecognitioninRetinalImagesusingEnsembleClassification

195

bifurcations crossing points. For each centerline

pixel in the vessel segment, the feature vector is

computed using pixel based features, profile based

features and vessel segment based features of the

RGB and HSI colour spaces, and finally each

centerline pixel is assigned an artery or vein label by

a decision tree based ensemble classifer.

2.1

Retinal Vessel Segmentation

The retinal vasculature is composed of arterioles and

venules, appearing as piecewise linear features, with

variation in width and their branches visible within

the retinal image (Fraz, Barman et al. 2012). We

have computed a measure of vessel-ness for each

pixel in the retinal image by combining a multi-scale

line detection which is calculated from the inverted

green plane of the coloured retinal images. The

average pixel intensity is measured along lines of a

particular length passing through the pixel under

consideration at 12 different orientations spaced by

15 degrees each. The line with the highest average

pixel intensity is selected. The line strength of a

pixel is calculated by computing the difference in

the average grey values of a square sub-window

centred at the target pixel with the average intensity

of the selected line. This concept was first

introduced by (Ricci and Perfetti 2007) and has also

been employed elsewhere (Fraz, Remagnino et al.

2012). We have used a generalized multi-scale line

detector (Nguyen, Bhuiyan et al. 2012), which uses a

variable length of aligned lines in a fixed square sub-

window, for calculating the line strength measures

for the pixels in the images containing a central

vessel reflex. In the line strength image (LSI), each

value corresponds to the confidence measure of each

pixel to be a part of the vessel or not. The LSI, as

illustrated in Figure 1(b), is often considered as a

greyscale image, where bright pixels indicate a

higher probability of being a vessel pixel.

In order to obtain a vessel binary segmentation, a

hysteresis thresholding based morphological

reconstruction is applied. Hysteresis thresholding

employs a bi-threshold procedure such that the

intensity image is thresholded for two ranges of grey

values, one being included in the other. The image is

first segmented by a narrow threshold range which

concedes only high confidence object pixels and

thus also contains many false negatives. This image

is termed a marker image. The mask image is

generated by applying a wide threshold range to the

greyscale image. These threshold values are derived

from the intensity histogram of the non-null pixels;

each one of these thresholds; T

for the marker

image and T

for the mask image, is defined as the

highest intensity value such that the number of

pixels with intensities above this limit is greater or

equal to a predefined percentage. This percentage

value is empirically selected for T

and T

as 90%

and 95% respectively.

(a) (b) (c)

(d) (e) (f)

Figure 1: (a) Coloured retinal image, (b) Line strength

image, (c) Marker Image, (d) Mask Image, (e) Segmented

vascular, (f) Extracted Vessel segments.

The image is first segmented by a narrow

threshold range which concedes only high

confidence object pixels and thus also contains many

false negatives. This image is termed a marker

image. The mask image is generated by applying a

wide threshold range to the greyscale image. These

threshold values are derived from the intensity

histogram of the non-null pixels; each one of these

thresholds; T

for the marker image and T

for the

mask image, is defined as the highest intensity value

such that the number of pixels with intensities above

this limit is greater or equal to a predefined

percentage. This percentage value is empirically

selected for T

and T

as 90% and 95% respectively.

The marker image is used as a seed for the

morphological reconstruction using the mask image.

Figure 1(c-e) shows the marker, mask and

segmented vessels image respectively.

2.2

Vessel Segment Extraction

The skeletonization is applied to the binary

segmented vasculature image which reduces all the

vessels to a single centreline one pixel wide. The

bifurcation points and crossing points are detected in

the vessel centreline image by counting the

neighbourhood of each pixel. The vessel centreline

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

196

image is scanned to analyse the neighbours of each

pixel for eight-connectivity. This way, the

intersection number,

(, )In x y

is calculated for each

pixel

(, )

of the centreline image, as shown in

equation (2).

(, ) 0.5* | (, ) (, )|

nxy N xy N xy













(1)

where

(, )

Nxy

are the neighbours of the

analysed point,

(, )

According to its intersection number In

each

pixel p(x,y) will be marked as vessel end point if

In(x,y)=1, vessel internal point if In(x,y)=2,

bifurcation if In(x,y)=3 and crossover if In(x,y)= 4

The detected bifurcation points and crossing

points are deleted from the vessel centreline image

and the retinal vasculature is cut into the vessel

segments as shown in Figure 1(f).

2.3 Image Normalization and Shade

Correction

The absolute colour of the blood in the vessels varies

between the images and across the subjects. To

overcome this variability some authors (Grisan and

Ruggeri 2003; Saez, González-Vázquez et al. 2012)

have used a quadrant based approach which is

computationally intense, also it enforces a condition

to have at least one arteriole and one venule per

quadrant, and it seems more suitable for optic disc

centered images rather than macula centered images.

We have used a simple linear transformation for

shade correction and variability normalization. For

this purpose, the estimate of the background o is

obtained by applying a filtering operation with a

large arithmetic mean kernel. The size of the filter

kernel is not a critical parameter as long as it is large

enough to ensure the blurred image contains no

visible structures such as vessels. In this work, we

use a 121×121 pixel size kernel. Then the difference

between the morphologically opened image I

and

the estimated background I

is then computed for

each pixel to obtain a normalized image I

(, ) (, ) (, )

nOBE

Ixy I xy I xy

(2)

Likewise, when the fluctuation in background

intensity of retinal images is examined, there can be

significant variation in intensities between images

due to different illumination conditions in the

acquisition process. Therefore, a shade corrected

image is obtained by applying a global

transformation with the purpose of reducing the

intensity variation and contrast enhancement. For

this purpose, the pixel intensities are modified

according to the following global linear

transformation function,

0, ( , ) 0

( , ) 1, ( , ) 1

(, ),

if I x y

Ixy ifIxy

x y otherwise

















(, ) (, ) 0.5

n MAX PIXEL

pxy I xy val





(3)

where, I

(x,y) is the homogenized image, I

(x,y)

is the normalized image shown in (7), val

MAX_PIXEL

the intensity value presenting the highest number of

pixels in the normalized image I

(x,y). The pixels

with intensity value equal to val

MAX_PIXEL

belong to

the background of the retinal image. This global

transformation will set them to 0.5 and will

standardize the intensity around this value of those

background pixels with different illumination

conditions. In order to find differences between

arterioles and venules, we have analysed RGB and

HIS colour spaces. Figure 2 shows the red, green

and blue components of the image shown in Figure

1(a) along with respective shade corrected images.

(a) (b) (c)

(d) (e) (f)

Figure 2: The colour spaces; (a-c) Red, Green and Blue

channels of RGB. (d-f) the shade corrected images of Red

green and blue channels respectively.

2.4 Feature Extraction

In previous work (Fraz, Remagnino et al. 2013), we

have computed the local orientation angle of vessel

segment and the width of vessel for each centreline

pixel. Based on this information we have extracted

the vessel profile for each centreline pixel. The

vessel profile is perpendicular to vessel direction,

and the length of vessel profile is equal to the vessel

width. The centreline, edges and the vessel segment

AutomatedArterioleandVenuleRecognitioninRetinalImagesusingEnsembleClassification

197

profiles are shown in Figure 3(a). We define three

types of features extracted from the vessel segments

for each centreline pixel: pixel based features,

profile based features and vessel segment based

features.

The pixel based features are the centreline pixel

intensity values taken from the respective colour

channel. The profile based features are the mean and

variance of the intensity values across a vessel

profile for each centreline. The vessel segment based

features are calculated in two ways for each

centreline pixel from the respective colour channel.

First, the mean and variance of the pixel intensities

are calculated for the entire vessel segment. Second,

the relatively longer vessel segments are divided into

smaller vessel parts of length of almost 50 pixels,

and then the mean and variance of the pixel

intensities are calculated with in these vessel . We

have tried and tested the different size of vessel sub-

segments ranging from 15 to 75 pixels and achieve

the best results using 50 pixels.

Figure 3(b) illustrates a vessel segment where

the edges are marked with blue lines and the

centreline pixels (Ci) are shown with green circles

on the black line. The vessel profile (Pi) consists of

the pixels in the line drawn perpendicular to the

local vessel direction, equal to the approximate

vessel width, and is shown in white. VPi are the

vessel segment parts. Table 1 shows the complete set

of features extracted for each centreline pixel.

Figure 3: (a) Vessel profiles; (b) Vessel Features for A/V

classification.

The feature set (fv) can be summarized as

follows.

( ) [ ]

() ( [ ])

() ( [ (, )])

( ) ( [ ( , )])

()

fv n X C C vessel segment

fv n X P P vessel segment

fv n X VP x y VP vessel segment

fv n











( [(, )]) (, )

() ( [(, )]) (, )

XIxy pixelsxy invesselsegment

v n X I x y pixels x y in vessel segment







Table 1: Complete set of features extracted for each

centreline pixel.

No. Feature description Type

1-6

The centreline pixel intensity values from each

component of RGB and HSI colour space

Pixel

based

7-12

Mean of pixel intensities of vessel part VP

from

each colour component of RGB and HSI.

Vessel segment

part based

13-18

Standard deviation of pixel intensities of vessel

part VP

from each colour component of RGB and

HSI.

19-22

Min and Max values of pixel intensities of vessel

segment part VP

from each R and G component

23-25

Mean of pixel intensities of vessel profiles P

from

each colour component of RGB

Vessel profile

based features

26-31

Standard deviation of pixel intensities of vessel

profiles P

from each colour component of RGB

and HSI.

32-35

Min and Max values of pixel intensities of vessel

profiles P

from R and G component.

36-41

Mean of pixel intensities of complete vessel

segment from each component of RGB and HSI.

Vessel segment

based features

42-47

Standard deviation of pixel intensities of complete

vessel segment from each colour component of

RGB and HSI.

48-51

Min and Max values of pixel intensities of

complete vessel segment from each R and G

component from RGB.

where, n is the number of features, X[C

] is the value

of the colour channel X at point C

in the image, X ɛ

{R,G,B,H,S,I}. For instance, R[C

] is the value of

the centreline pixels in the R colour component,

µ(G[P

]) is the mean of the vessel profile pixels in

the green channel.

Figure 4(a) shows the graphs for feature

importance index calculated from classifiers created

with 30 decision trees and trained with 15000

training samples. Based on the information shown in

the graph, five sets of features have been created by

selecting those features, where the importance index

is more than the following values {0.75, 1.0, 1.5,

2.0,2.75}. Five classifiers have been trained with the

reduced feature sets and the out-of-bag (OOB)

classification error is computed for each of the

classifiers and is plotted in Figure 4(b). It has been

observed that the performance of the classifier

trained with 16 most significant features is best

among all; therefore 16-D is utilized for vessel

classification.

2.5 Vessel Labelling

The final labelling of vessel pixels is obtained by

employing supervised classification with an

ensemble classifier of bagged decision trees. We

assume that all the pixels in the vessel segment are

either in an arteriole or venule. The classification of

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

198

(a)

(b)

Figure 4: (a) Feature Importance Index, (b) Out-of-Bag

classification error with reduced feature sets.

the retinal vessel is a two class classification

problem where each pixel in the image either

belongs to an artery (C

) or to a vein (C

In Ensemble classification (Polikar 2006),

multiple classifiers or models are tactically

generated and combined in order to give the solution

to a machine learning problem; with the goal of

obtaining better predictive performance than could

be obtained from any of the constituent classifiers /

models. This process is used to abbreviate the

likelihood of inadequate or unfortunate selection

while improving the classification or prediction

performance of the classifier. We use this strategy

instinctively in our day to day activities, where we

consider the opinion from several experts, evaluate

and merge their recommendations for establishing a

well optimized and well-versed conclusion. In the

same manner, the ensemble methods utilize multiple

classifiers/models to accomplish gain in

classification performance by mixing/aggregating

the outcomes from several weak learners into one

high-class classifier, with the goal of reducing the

variance and amplifying the confidence in the

decision. In the present work, the decision trees have

been used as the component classifier of the

ensemble system and which is created by employing

boot strapped aggregation

Let us consider a set of observations “x

” from

the feature vector with a known class label “y” as a

training set, where y ∈{C

}. The objective is to

predict the class label “y” for the given observations.

The classifier assigns soft labels to the centreline

pixel labels, which can be regarded as a vote for the

label of the complete vessel segment, and the mean

of these votes is assigned as the label for the entire

vessel segment.

3 RESULTS

3.1

Evaluation Criteria and

Performance Measures

We have tested this methodology on retinal images

obtained from EPIC Norfolk study (EPIC-Norfolk

2013). The dataset contains 40 macula centered

retinal images from both of the eyes of 20 study

participants, with the vessel types manually labelled

by trained observers. The images were captured with

non-mydratic fundus cameras and saved in 24-bit

JPEG format with a resolution of 3000x2000 pixels.

The performance measure are obtained for the

centreline pixels in the entire image and evaluated

separately for arterioles and venules. For each of the

vessel type, the True Positives (TP), False Positives

(FP), True Negatives (TN) and False Negatives (FN)

are calculated in the same way as in (Saez,

González-Vázquez et al. 2012), and tabulated in

Table 2.

Table 2: Vessel classification (Class ∈ {a,v}).

Observer Identify

pixel ∈ Class

Observer

identify pixel

∉Class

System Identify

pixel ∈ Class

True Positive

(TP

Clas

)

False Positive

(FP

Class

)

System Identify

pixel ∉ Class

False Negative

(FN

Class

)

True Negative

(TN

Class

)

The algorithm is evaluated in terms of Detection

Rate / Sensitivity (SN

a|v

), Specificity (SP

a|v

Classification Accuracy (ACC

a|v

), Classification

Error Rate (CER

a|v

), Positive Predictive Value

(PPV

a|v

), Negative Predictive Value (NPV

a|v

) and the

Positive and Negative Likelihood Ratios (PLR

a|v

and

NLR

a|v

). The ACC

a|v

is measured by the ratio of the

total number of correctly classified pixels (sum of

true positives and true negatives) by the number of

pixels under consideration in the image. SN

a|v

reflects the ability of an algorithm to detect the true

positives. SP

a|v

measures the proportion of negatives

that are correctly identified. PPV

a|v

or precession rate

gives the proportion of vessel pixels with correctly

identified positive test results and NPV

a|v

is the

proportion of vessel pixels with negative test results

AutomatedArterioleandVenuleRecognitioninRetinalImagesusingEnsembleClassification

199

that are correctly identified. The predictive values

depends on the percentage of a/v in the retina

(prevalence), therefore the likelihood ratios (PLR

a|v

and NLR

a|v

) are also computed which are not

dependent on prevalence. These metrics are

illustrated in Table 3, based on the terms defined in

Table 2.

Table 3: Performance metrics for vessel classification.

Measure Description

a|v

/(TP

a|v

+FN

a|v

)

a|v

/(TN

a|v

+FP

a|v

)

ACC

a|v

(TP

a|v

+TN

a|v

)/(TP

a|v

+FP

a|v

+TN

a|v

+FN

a|v

)

CER

a|v

(FP

a|v

+FN

a|v

) / (TP

a|v

+FP

a|v

+TN

a|v

+FN

a|v

)

PPV

a|v

/(TP

a|v

+FP

a|v

)

NPV

a|v

/ (TN

a|v

+FN

a|v

)

PLR

a|v

/ (1- Sp

a|v

)

NLR

a|v

(1-SN

a|v

)/ SP

a|v

3.2 Experimental Results

We have analysed the methodology by using 3149

vessel segments from 40 colour fundus images from

20 EPIC Norfolk participants. The algorithm is

evaluated by using a two-fold validation

methodology. The images of the right eye are

assigned to set S

and the left eye images are

allocated to the set S

. The classifier is then trained

on S

and tested on S

, followed by training on S

and testing on S

. The performance metrics are

computed separately for arterioles and venules and

presented in Table 4.

Table 4: Vessel classification performance metrics.

Measure Arterioles Venules

Tested

on S1

Tested

on S2

Tested

on S1

Tested

on S2

a|v

0.9067 0.8795 0.7658 0.7652

a|v

0.7658 0.7836 0.9067 0.8804

ACC

a|v

0.8387 0.8298 0.8327 0.8261

CER

a|v

0.1612 0.1701 0.1672 0.1738

PPV

a|v

0.8162 0.8170 0.8757 0.8513

NPV

a|v

0.8757 0.8513 0.8162 0.8170

PLR

a|v

3.6183 5.7391 7.9255 8.8678

NLR

a|v

0.1261 0.1623 0.2763 0.2505

The similarity in the performance metrics

obtained for the sets S

and S

indicates the

repeatability of the methodology in classification of

vessels. The sensitivity in the set S

for the arteries is

0.9067 and for the veins is 0.7658. This in turn

indicates that the probability of incorrect

classification (i.e., false positives) for arterioles and

venules is 9.33% and 23.42% respectively. Some

results of the presented A/V classification

methodology are illustrated in Figure 5.

Figure 5: A/V classification result.

A comparison of the proposed method’s

performance metrics with the recently published

methods is shown in Table 5. The sensitivity of

proposed method in identifying arterioles is higher

than the other methods, which suggests that our

method has lower probability of incorrect

classification. The higher specificity of our

algorithm in identifying veins also indicates fewer

false positives compared with other methods. We

have analysed 40 coloured fundus images compared

to 35, 58 and 35 images analysed in (Grisan and

Ruggeri 2003), (Saez, González-Vázquez et al.

2012) and (Relan, MacGillivray et al. 2013),

respectively. Also the resolution of our images are

3000 x 2000 pixels which is greater than 1300x1000

and 786x567 used in previous studies (Grisan and

Ruggeri 2003) (Saez, González-Vázquez et al.

2012). It should be noted that the resolution of our

test images is higher than those used to test other

methods, so the performance metrics for our method

may differ with a different image-set. The

classification accuracy of the method is dependent

upon the vessel segmentation results and the

extraction of vessel centreline pixels. Moreover, the

choice of a different classifier, feature set and retinal

image zone is likely to have an impact of the

method’s performance.

Table 5: Results of our method compared with recently

published methods.

Method SN SP PPV NPV

Arteries

(Saez, González-Vázquez et al.

2012)

0.78 0.89 0.88 0.80

(Relan, MacGillivray et al. 2013) 0.81 0.89 0.90 0.80

Proposed Method 0.90 0.76 0.81 0.87

Veins

(Saez, González-Vázquez et al.

2012)

0.87 0.79 0.81 0.86

(Relan, MacGillivray et al. 2013) 0.76 0.95 0.94 0.83

Proposed Method 0.76 0.90 0.87 0.81

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

200

4 CONCLUSIONS

An automated method for A/V classification in

retinal vasculature based on colour features utilizing

the ensemble classifer of boot strapped decision

trees is presented. To the best of our knowledge, this

is the first time the decision trees based ensemble

classifier has been used for A/V classification.

An application of image processing algorithms

for computer assisted analysis of digital fundus

images offers a number of advantages over a manual

system, including fast, timely and reliable

quantification of abnormalities. The presented

methodology will be incorporated in to a software

package QUARTZ (QUantitative Analysis of

Retinal vessel Topology and siZe). The QUARTZ

software will assist in examining arterio-venous

morphological associations with cardiovascular risk

factors and outcomes in large population based

studies, furthering our understanding of the vascular

changes / consequences associated with the

development of disease.

In future we aim to extend the QUARTZ

software to incorporate the analysis of other retinal

vessel features pathognomonic of cardiovascular

disease, including measurement of arterio-venous

ratio, identification of venous beading and

quantification of arterio-venous nicking.

ACKNOWLEDGEMENTS

The authors would like to thank Professor Paul

Foster and the European Investigation into Cancer in

Norfolk (EPIC Norfolk) study for providing the

retinal images used in this analysis. The EPIC

Norfolk study is supported by grants from the

Medical Research Council, Cancer Research UK and

Research into Ageing.

REFERENCES

Abràmoff, M. D., M. K. Garvin, et al. (2010). "Retinal

Imaging and Image Analysis." Biomedical

Engineering, IEEE Reviews in 3: 169-208.

Dashtbozorg, B., A. M. Mendonca, et al. (2013). "An

Automatic Graph-based Approach for Artery/Vein

Classification in Retinal Images." Image Processing,

IEEE Transactions on PP(99): 1-1.

EPIC-Norfolk. (2013). "European Prospective

Investigation of Cancer (EPIC)." Retrieved September,

2013, from http://www.srl.cam.ac.uk/epic/.

Fraz, M. M., S. A. Barman, et al. (2012). "An approach to

localize the retinal blood vessels using bit planes and

centerline detection." Computer methods and

programs in biomedicine 108(2): 600-616.

Fraz, M. M., P. Remagnino, et al. (2013). "Quantification

of blood vessel calibre in retinal images of multi-

ethnic school children using a model based approach."

Computerized Medical Imaging and Graphics 37(1):

60-72.

Fraz, M. M., P. Remagnino, et al. (2012). "Blood vessel

segmentation methodologies in retinal images – A

survey." Computer methods and programs in

biomedicine 108(1): 407-433.

Fraz, M. M., P. Remagnino, et al. (2012). "An Ensemble

Classification-Based Approach Applied to Retinal

Blood Vessel Segmentation." Biomedical Engineering,

IEEE Transactions on 59(9): 2538-2548.

Grisan, E. and A. Ruggeri (2003). A divide et impera

strategy for automatic classification of retinal vessels

into arteries and veins. Engineering in Medicine and

Biology Society, 2003. Proceedings of the 25th

Annual International Conference of the IEEE.

Huang, Y., J. Zhang, et al. (2012). "An automated

computational framework for retinal vascular network

labeling and branching order analysis." Microvascular

Research 84(2): 169-177.

Jack J. Kanski and Brad Bowling (2011). Clinical

Ophthalmology: A Systematic Approach. London,

Elsevier Health Sciences (UK).

Nguyen, U., A. Bhuiyan, et al. (2013). "An Automated

Method for Retinal Arteriovenous Nicking

Quantification from Colour Fundus Images."

Biomedical Engineering, IEEE Transactions on

PP(99): 1-1.

Nguyen, U. T. V., A. Bhuiyan, et al. (2012). "An effective

retinal blood vessel segmentation method using multi-

scale line detection." Pattern Recognition(0).

Niemeijer, M., X. Xiayu, et al. (2011). "Automated

Measurement of the Arteriolar-to-Venular Width Ratio

in Digital Color Fundus Photographs." Medical

Imaging, IEEE Transactions on 30(11): 1941-1950.

Owen, C. G., A. R. Rudnicka, et al. (2011). "Retinal

Arteriolar Tortuosity and Cardiovascular Risk Factors

in a Multi-Ethnic Population Study of 10-Year-Old

Children; the Child Heart and Health Study in England

(CHASE)." Arteriosclerosis, Thrombosis, and

Vascular Biology 31(8): 1933-1938.

Polikar, R. (2006). "Ensemble Based Systems in Decision

Making." IEEE Circuits and Systems Magazine 6(3):

21-45.

Relan, D., T. MacGillivray, et al. (2013). Retinal vessel

classification: sorting arteries and veins. 35th Annual

International Conference of the IEEE EMBS

Engineering in Medicine and Biology Society

(EMBC), Osaka, Japan, IEEE.

Ricci, E. and R. Perfetti (2007). "Retinal Blood Vessel

Segmentation Using Line Operators and Support

Vector Classification." Medical Imaging, IEEE

Transactions on 26(10): 1357-1365.

Rothaus, K., X. Jiang, et al. (2009). "Separation of the

retinal vascular graph in arteries and veins based upon

AutomatedArterioleandVenuleRecognitioninRetinalImagesusingEnsembleClassification

201

structural knowledge." Image and Vision Computing

27(7): 864-875.

Saez, M., S. González-Vázquez, et al. (2012).

"Development of an automated system to classify

retinal vessels into arteries and veins." Computer

methods and programs in biomedicine 108(1): 367-

376.

Vázquez, S. G., B. Cancela, et al. (2013). "Improving

retinal artery and vein classification by means of a

minimal path approach." Machine Vision and

Applications 24(5): 919-930.

Wong, T. Y., R. Klein, et al. (2001). "Retinal

microvascular abnormalities and incident stroke: the

Atherosclerosis Risk in Communities Study." The

Lancet 358(9288): 1134-1140.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

202