1

Retinal Blood Vessels Classification Into

Arteries And Veins
Gabriel Lopes, João Silveira, Paulo Maia, Tiago Gonçalves

However, these visually descriptive features are usually only

Abstract— The classification of retinal vessels into artery/vein locally valid because it gets harder to discriminate arteries from
(A/V) is an important phase for automating the detection of veins as the distance from the optic disc grows and the vessels
vascular changes, and for the calculation of characteristic signs crossings and bifurcations happen more often, while the vessel
associated with several systemic diseases such as diabetes, width changes, becoming smaller [3]. Moreover, some
hypertension, and other cardiovascular conditions. In this work,
distinctive features can vanish in the outer regions of the image
A/V classification was performed using a Multilayer Perceptron
Classifier. After classification, post-processing was performed to or if the image quality is not good enough, proving that position
increase accuracy. A 68% accuracy at optimal threshold, 75% and image acquisition conditions can influence the extraction
AUC, 67% sensitivity and 68% specificity were achieved on the of meaningful features for the correct vessel classification [3].
test set. Consequently, the acquisition process is a key factor in order to
assure the importance of some features. If the retinal images are
Index Terms— machine learning, computer aided diagnosis non-uniformly illuminated and present local luminosity and
systems, retinal blood vessels classification, biomedical image contrast variability, an intensity based classification method
analysis may fail [4], as vessels in the outer regions of the image can
become very dark to the shading effect [3], making it
I. INTRODUCTION impossible to differentiate both vessel classes. Another
In retinal image analysis, being able to distinguish between pertinent issue concerning the data set of images considered is
arteries and veins is a pertinent issue because it enables the that image resolution and quality is not constant, allowing that
possibility to calculate, from classified retinal vessel trees, the illumination conditions can vary between images, leading
diagnostic indicators as the Arteriolar to Venular Ratio (AVR), to a different absolute blood color among images and across
which is related to some cardiovascular and systemic subjects [5].
pathologies [1]. In fact, in the presence of diseases such as In this paper, a supervised learning approach is proposed for
hypertension or atherosclerosis, the AVR value can decrease classifying retinal blood vessels into arteries and veins, taking
due to retinal vessel diameter changes [2]. On the other hand, a training set of 20 retinal images in which vessels were
diabetes patients may often develop a microvascular correctly identified as being arteries or veins by an expert. The
complication (Diabetic Retinopathy) that is associated to developed model should learn from this training set and predict,
abnormally wide veins. This symptom leads to an unusually using a pixel-wise classification method, which type of vessel
low AVR value [3]. Therefore, retinal vessel morphology is present in a test set of 20 images, given the binary map of its
provides an insight of the entire blood circulatory system and vessels.
enables the detection of cardiovascular diseases.
Several steps are required in order to estimate the AVR value II. LITERATURE REVIEW
from a retinal image. Among other image processing Several approaches have been made regarding this issue.
operations, the most important phases consist of vessel An automatic method for artery/vein classification was
segmentation, precise vessel width calculation and artery/vein proposed by Niemeijer et al [3]: after extracting a set of
classification [4]. centerline pixel features and assigning a soft label to each one
As a result, it is of great importance to pursue an automatic of them representing their probability of being a vein pixel, the
method to determine whether a retinal vessel pixel represents average of the soft labels regarding connected centerline pixels
an artery or a vein pixel, allowing a non-invasive measurement is assigned to each centerline pixel. Then, distinct classifiers
of AVR as well as reducing the time, cost and subjectivity were tested and the one that granted the best results was the k-
associated with current observer-based techniques [4]. nearest neighbor (kNN) classifier.
Subsequently, it becomes relevant to understand how can Zamperini et al [1] investigated the relative importance of
arteries be different from veins and what should be taken into features extracted from vessel pixels, finding that the most
consideration. Assuming that arteries are brighter in color than effective features for this subject were the ones related to color
veins and thinner than neighboring veins, the central reflex is and that some size features were only useful if considered
wider in arteries and smaller in veins (especially using the red together with spatial information. Contrast with the background
channel) and that near the optic disc the two types of vessels was taken into consideration by extracting features related to
alternate with each other before branching out [3], it becomes average and standard deviations of color components and their
easier to perform this task. derivatives in a disc with diameter equal to twice the locally
 2

estimated vessel diameter. They ended up with a set of 16 I(x)=i(x)∙r(x) (1)

features that they found to be the most distinctive ones.
Joshi et al. [6] proposed an automated method for artery and
vein classification based on the conversion of a vessel
segmentation image into a vessel segment map, while
identifying the individual vessel trees by graph search.
Orientation, width and intensity of each vessel segment were Fig. 2. Summary of the operations performed on the Homomorphic Filtering
operation for illumination correction (I’(x,y) is the undegraded image).
used to find the optimal graph of vessel segments. Then, these
separated vessel trees were labeled either as primary vessels or
as branches. Finally, the classification was based on the vessel
color properties in each tree graph.
Dashtbozorg et al. [4] followed a graph-based approach in order
to implement an artery and vein automatic classification
method. First, they extracted a graph from the retinal vascular Fig. 3. (a) Original image (b) Denoised image (c) Image after
tree. Then, a decision was made on the type of each intersection applying homomorphic filter.
point (graph nodes) and one of two labels was assigned to each
vessel segment (graph links) obtained. Finally, the B. Division into regions
classification of the vessel segments as an artery or vein was After obtaining the skeleton of the original vessel tree
based on a combination of the graph-based labeling results with through a thinning operation, regions of interest, namely the
a set of intensity features using a linear classifier. crossings and bifurcations of the tree, were identified. To
achieve this, a hit-miss transform with several structuring
III. METHODOLOGY elements was applied. Afterwards, the intersection points and
its neighbors were removed, and vessel segments were labeled
Our image processing and classification method is divided as
using a region labeling algorithm. The process of this step is
described in Fig. 1. From the 20 training images, 15 were
shown in Fig 4.
randomly selected, and then, 2500-pixel samples from arteries
and 2500 from vessels were extracted. It was followed by
feature selection and determination of the optimal classifier. In
the end, some post processing on the classified image was
performed and the final model was tested with the 5 remaining
images, left on purpose for validation.

Fig. 4. (a) Skeleton of vessels (b) Skeleton with identified crossings (c)
Labeled regions colored.

C. Feature Extraction
Global Features
Fig. 1. Pipeline followed in the work for A/V classification. Beginning in global futures, intensity values for red, green
and blue channels of the RGB image and, on HSV image, the
A. Pre-Processing
hue, saturation and value, were obtained. Also, since usually
Due to the lack of local luminosity and contrast variability in veins have a larger diameter than arteries, vessel diameter was
retinal images [7], it becomes very important to implement a extracted, by using the distance transform on the vessels and
method to normalize luminosity and contrast in retinal images, saving the computed result on the skeleton pixel.
both intra- and inter-image. Local Features
As a result, two different operations were performed on the To compute local features, considering the diameter of the
original image sequentially. The main goal of these operations vessel, and using the approach of Zamperini et al [1], as
was to create an image enhancement method so we could, then, mentioned above, we extracted new features based on statistic
normalize the images as illumination is concerned making it measures on the color components and color derivatives, in
possible to extract more meaningful color features. disks with different diameters (large and small, like shown in
This way, first we applied an image denoising operation in Fig. 5).
order to remove the noise that is associated with any acquisition
system as well as some image artifacts that may exist.
Afterwards, we performed the Homomorphic Filtering
Operation for Illumination Correction so that assuming an
image is non-uniformly illuminated (I(x)), we can obtain an Fig. 5. Small area (p1) vs large area (p2), used for feature extraction [1].
undegraded image (r(x)) while removing the illumination field
(i(x)) that introduces the variability mentioned above. The With this approach, rotation invariant (due to disk-like shape)
equation that represents this operation is presented in (1), the and size invariant (due to adaptive window size) features were
method underlying it is described in Fig. 2, and an example of obtained.
its application on Fig. 3. Besides that, as usually there is an alternation of veins and
arteries near the optic disk [11], the distance to the optic disk
 3

was computed. Optic disk was detected by applying a time it receives an answer, a follow-up question is asked until a
Difference of Gaussian filter, for blob detection. For false conclusion about the class label of the record is reached.
positive reduction, the region with the highest intensity in red Random Forests are an extension of Decision Trees. As the
channel was chosen. Due to luminosity variations, even after name suggests, this algorithm creates a forest - an ensemble of
applying the homomorphic filtering, the wrong region can be decision trees. In general, the more trees in the forest, the more
detected, as shown on Fig. 6, showing a limitation of this robust the classifier looks like. In these two algorithms, variable
algorithm. parameters include the depth, max number of leaves, maximum
leaf nodes, among others.
The k-nearest neighbor (kNN) algorithm is a nonparametric
method based on the similarity of nearby instances and it is a
classifier which does not require a model, and uses the data
directly for classification. The typically variable
hyperparameter is the number of neighbors, which is usually
chosen to be odd to avoid ties.
Fig. 6. (a) Detection of the Optic Disk (b) Failed detection due to luminosity
Logistic regression is a classifier which finds the best fitting
variations. model to describe the relationship between dependent
Distance to image center, which has similar physical meaning (outcome) and independent (predictor/explanatory) variables,
as the distance to optic disk, was also calculated. generating coefficients that maximize the likelihood of
Line Features observing the sample values. The typically variable
Last, the remaining group of features, the line features, hyperparameter is the cost.
contain several statistical measurements regarding vessel Multilayer Perceptron is a class of a feedforward artificial
profile. This was achieved by getting the orientation of each neural network, with nodes that use nonlinear activation
segment (through the fitting of an ellipse), then getting the functions. The tunable hyper parameters chosen were the
perpendicular line centered in the centroid region, discretizing activation function, alpha, which is a regularization parameter
the line and skeletonizing it. From here we could extract the that penalizes weights with large magnitudes, the hidden layer
kurtosis, skewness, mean and standard deviation [8]. size, the learning rate (which can be, for instance, adaptive,
A total of 61 features were obtained. varying in case the loss is constant from iteration to iteration),
and the solver (Adam/SGD/RMSProp).
Feature Selection E. Post-Processing
In this work, recursive feature elimination (RFE) with cross
validation and ANOVA (with different number of chosen After obtaining the first prediction for the full skeleton, each
features) were used for feature selection. Besides that, all the region is iteratively searched, and checked if the number of
features, with some manual feature preprocessing to remove points classified as an artery (with a fixed threshold of 0.5 on
features with a considerable number of zeros, which have no the classification probability) is bigger than the number of
predictive power, were used. For this task, several classifiers, points classified as veins. According to that, the probability of
along with these feature selection techniques, as shown on Fig. the arteries is swapped by the mean probability of the veins in
7, were used. the same segment, or vice-versa.
As a second post-processing step, points which are in an area
closer to the optic disk than 20% of the maximum distance to it
are removed from the skeleton. This is performed because
vessels closer to the optic disk are usually easier to classify, so
their prediction shouldn’t need to be changed.
Fig. 7. Pipeline of feature selection and classification tasks. Afterwards, for every intersection of the skeleton, we get all the
In RFE, all features are chosen on the first step. For every step neighbors in a 5 by 5 window, and see to which region they
or iteration, the worst n number of features (lowest feature belong to. Then, we compute the mean probability of that
importance in case of tree-based models) are eliminated. The segment being a vein, to see the predominant class in the
number of features left at the step which gives the maximum neighborhood. In the regions that doesn’t belong to the
score on the validation data, is considered to be optimal number predominant class, we swap its mean probability by the
of features. maximum probability of the other regions. An example is
The one-way analysis of variance (ANOVA) is used to shown on Fig. 8.
determine whether there are any statistically significant
differences between the means of two or more independent
groups).

D. Classification
Decision Tree classifier applies a straightforward idea to
solve the classification problem, by posing a series of carefully
crafted questions about the attributes of the test record. Each
Fig. 8. Example of the second post processing step, for one region. 0.3 is
changed into 0.7, the maximum between 0.5 and 0.7.
 4

Each region is added to a list of visited regions, so the same

region isn’t changed two times.
Finally, for propagating the labels from the skeleton to the full
vessel, a 1-nearest neighbor classifier, with the number of
classes equal to the number of pixels in the skeleton, and the
features as the position of the skeleton pixels was used. As such,
the pixels that are left to classify are given the same label as the
closest skeleton pixel. An example of application of this
technique is shown on Fig. 9.
Fig. 10. ROC Curve of the test set predictions, showing an AUC of 75%.

V. CONCLUSION
Artery/Vein classification was shown to be a difficult task,
with several challenges in terms of algorithmic approaches and
code implementation.
As possible improvements, more things could be attempted in
the pre-processing step, such as the Retinex algorithm, or an
Fig. 9. Propagation from the skeleton to the vessel. algorithm based on image dehazing to remove shadows, leading
to less dramatic changes on the color of the image than the
IV. EXPERIMENTAL RESULTS homomorphic filtering which was used.
The intersection detection methodology, using hit-miss
A. Model Choice transform, lead to some points which were not intersections
5-fold cross validation was performed for model selection. being detected. Some post-processing which uses the
There was no significant difference between feature selection neighborhood of an intersection can be further tested to see if it
methods, and using all the features, as shown in Fig. 10, so all is a false positive.
features ended up being used, as the computational time wasn’t Regarding feature extraction, more features, such as texture-
significantly different. No pre-processing was used as initial related features, could be extracted from the vessels, as a way
trials showed it led to higher variability. to describe the center reflex. With a higher number of features,
the feature selection methods should lead to a higher difference
in accuracy than what was seen in this work.
As for label propagation, in post-processing, it should be
performed only on smaller vessels, perhaps using a threshold to
distinguish between large and small vessels. This is because
larger vessels are typically well-classified already.
As always, having more data could help our training algorithm
generalize more, and with higher amount of data and data
augmentation, deep learning algorithms can be applied.

VII. REFERENCES
Fig. 10. 5-fold CV results for all the classifiers and all the pathways tested. 1. Zamperini, A., et al. Effective features for artery-vein classification
The best model was shown to be the multilayer perceptron, with in digital fundus images. in Computer-Based Medical Systems
(CBMS), 2012 25th International Symposium on. 2012. IEEE.
a rectified linear unit as an activation function, an alpha of
2. Ikram, M.K., et al., Are retinal arteriolar or venular diameters
0.0001, 100 nodes in one layer, a constant learning rate and associated with markers for cardiovascular disorders? The
Adam as an optimizer. Rotterdam Study. Invest Ophthalmol Vis Sci, 2004. 45(7): p. 2129-
Using the second post processing step, accuracy in both vessel 34.
and skeleton lowered, and standard deviation increased, in the 3. Karssemeijer, N., et al., Automatic classification of retinal vessels
into arteries and veins. 2009. 7260: p. 72601F.
5 validation images, when compared to using only the first step 4. Dashtbozorg, B., A.M. Mendonca, and A. Campilho, An automatic
of post processing. As such, the second post processing step graph-based approach for artery/vein classification in retinal
was not used in the final model. Propagation from skeleton to images. IEEE Trans Image Process, 2014. 23(3): p. 1073-83.
vessel increased average accuracy by around 7%, indicating its 5. Niemeijer, M., et al., Automated measurement of the arteriolar-to-
venular width ratio in digital color fundus photographs. IEEE Trans
successfulness. Med Imaging, 2011. 30(11): p. 1941-50.
6. Joshi, V.S., et al., Automated method for identification and artery-
B. Test Set venous classification of vessel trees in retinal vessel networks. PLoS
The final model was then tested on 20 test images. An area One, 2014. 9(2): p. e88061.
under the curve of 75% was obtained (as shown on Fig. 10), 7. Nieuwenhuis, C. and M. Yan. Knowledge based image enhancement
using neural networks. in Pattern Recognition, 2006. ICPR 2006.
with a 68% accuracy at the optimal threshold. Moreover, the 18th International Conference on. 2006. IEEE.
sensitivity and specificity were shown to be 67% and 68%, 8. Xu, X., et al., An improved arteriovenous classification method for
respectively, showing that the model can classify veins as well the early diagnostics of various diseases in retinal image. Comput
as arteries. Methods Programs Biomed, 2017. 141: p. 3-9.