Download Advances in Natural and Applied Sciences

Advances in Natural and Applied Sciences, 9(6) Special 2015, Pages: 600-606
AENSI Journals
Advances in Natural and Applied Sciences
ISSN:1995-0772 EISSN: 1998-1090
Journal home page: www.aensiweb.com/ANAS
Automatic Retinal Vessel Segmentation through Gabor Filtering, Principal Component
Analysis and Ensemble of Classifiers (C4.5 with Bagging)
1
R. Geetha Ramani and 2Lakshmi Balasubramanian
Associate Professor, Department of Information Science and Technology, College of Engineering, Guindy, Anna University, Chennai –
600025, India.
2
Research Scholar, Department of Information Science and Technology, College of Engineering, Guindy, Anna University, Chennai –
600025, India.
1
ARTICLE INFO
Article history:
Received 12 October 2014
Received in revised form 26 December
2014
Accepted 1 January 2015
Available online 25 February 2015
Keywords:
Retinal Image Analysis
Blood vessel segmentation
Classification
Gabor Filtering
ABSTRACT
Background: Computational techniques have been exploited in the domain of
ophthalmology for better retinal fundus image analysis. Blood vessels, one of the most
important retinal anatomical structures are analyzed for diagnosis of many diseases like
retinopathy, occlusion and much other vision threatening diseases. Objective: In this
paper, retinal blood vessel segmentation is attempted through sequential application of
image pre-processing data mining and image post-processing techniques. The image
pre-processing phase comprises of image cropping, color space conversion and color
channel extraction, contrast enhancement, Gabor filtering and half wave rectification.
Image pre-processing is followed by the data mining phase which includes feature
vector construction through principal component analysis and classification through
C4.5 with bagging. The output of the data mining phase is post processed through
connected component analysis. Results: The proposed methodology is evaluated on a
publicly available DRIVE database. The results reported 95.08% accuracy on the test
set of the dataset. A comparison of these results with the existing techniques is also
reported. Conclusion: This methodology can help ophthalmologists in better and faster
retinal analysis and hence earlier treatment to the patients.
© 2015 AENSI Publisher All rights reserved.
To Cite This Article: R. Geetha Ramani and Lakshmi Balasubramanian., Automatic Retinal Vessel Segmentation through Gabor Filtering,
Principal Component Analysis and Ensemble of Classifiers (C4.5 with Bagging). Adv. in Nat. Appl. Sci., 9(6): 600-606, 2015
INTRODUCTION
Retinal image analysis has been an interesting area of research in the recent times. Fundus imaging of the
retina reveals its anatomical structures viz., blood vessels, optic disc, macula, fovea and any other abnormal
structures such as exudates and microneurysms if present (Niall Patton et al., 2006). Changes in the properties
associated with the blood vasculature aids in identification of various retinal diseases such as Diabetic
Retinopathy, Retinal artery and vein occlusion, Hypertensive Retinopathy etc. Disorders such as Hypertension,
Diabetes, Cardiovascular abnormalities and Stroke also cause alterations in the blood vessels of the retina.
Hence, analysis of the retinal blood vessel positions itself as a core area in retinal image analysis. Manual
segmentation of blood vessels requires high expertise. It is also a time consuming task. Hence, employing
computational techniques for automatic retinal blood vessel segmentation would highly simplify the task of the
ophthalmologists.
In this work, automatic retinal blood vessel segmentation is attempted through image pre-processing
followed by data mining techniques and subsequently image post processing.
Subsequent sections present the literature work on automatic blood vessel segmentation. the proposed
methodology and the experimental results.
1. Related Work:
Modifications in retinal blood vessels reveal many retinal and other disorders. Hence automatic blood
vessel segmentation helps in efficient disease diagnosis. A few works have been previously attempted in this
area. The works which report results on DRIVE database are concisely presented below.
Corresponding Author: R. Geetha Ramani, Associate Professor, Department of Information Science and Technology,
College of Engineering, Guindy, Anna University, Chennai-600025, India.
601
R. Geetha Ramani and Lakshmi Balasubramanian, 2015
Advances in Natural and Applied Sciences, 9(6) Special 2015, Pages: 600-606
In 2004, supervised classification through KNN classifier on 31 features including the green intensity plane,
Gaussian and its derivative at various scales resulted in an accuracy of 94.16% (Niemeijer et al., 2004). In 2006,
retinal blood vessel segmentation was attempted through extraction of vessel centerlines followed by vessel
filling through region growing achieving an accuracy of 94.63% (Mendonca and Campilho, 2006). In 2006, yet
another attempt to classify vessel pixels through Bayes classifier was employed on feature vector formed with
pixel intensities and Gabor wavelet reponses at various scales, yielding an accuracy of 94.66% (Soares et al.,
2006). In 2007, a multi-pass region growing algorithm using local maxima over scales of magnitude of gradient
and maximum principal curvature of hessian tensor obtained an accuracy of 93.44% (Martinez-Perez et al.,
2007). In 2010, a local adaptive thresholding was applied to form a binary image, in which the largest
component was considered as large vessels and SVM was applied to the residual segments followed by
application of track growing to grow the small vessels (Xu and Luo, 2010). This methodology yielded 93.28%
accuracy. In 2011, neural network based supervised classification was applied on a 7-D feature vector
comprising of gray level and moment invariant features providing an accuracy of 94.52% (Diego Marin et al.,
2011). In 2012, Isotropic Undecimated Wavelet Transform was applied for vessel segmentation resulting in an
accuracy of 93.71% (Bankheard et al., 2012). In 2013, another methodology with Gabor wavelet for vessel
enhancement and multi-layered thresholding for vessel segmentation attained an accuracy of 94.69% (Usman
Akram and Shoab, 2013). In 2014, an approach to segment vessels using combination of bee colony
optimization for identifying cluster centers followed by fuzzy c-means clustering achieved an accuracy of 93.9%
(Eid et al., 2014). Yet another framework which comprised of bias correction and adaptive histogram
equalization for vessel enhancement and probabilistic modeling optimized by expectation maximization
algorithm for vessel segmentation obtained an accuracy of 94.10% (Djibril Kaba et al., 2014). In the same year,
an enhanced multi-scale operator was proposed, which combines the responses of line operator at various scales
by assigning appropriate weights to each scale, accomplishing an accuracy of 94.15% (Yanli Hou, 2014). One
more attempt in 2014, employed multilayer perceptron on RGB color channels to segment the blood vessels and
this technique achieved an accuracy of 95.03% (Wilfred and Edward, 2014).
In this article, a methodology to segment retinal blood vessels through sequential application of image preprocessing, data mining and image post processing techniques to the fundus image is proposed. The following
section describes the proposed methodology.
2. Proposed Methodology:
Retinal vessel segmentation through application of computational intelligence would be highly appreciated
by the ophthalmologists. The proposed framework is depicted in fig, 1.
Fig. 1: Proposed Framework for Automatic Retinal Vessel Segmentation.
602
R. Geetha Ramani and Lakshmi Balasubramanian, 2015
Advances in Natural and Applied Sciences, 9(6) Special 2015, Pages: 600-606
The proposed framework to segment the retinal blood vessels includes sequential application of image preprocessing, data mining techniques and post processing of the classified pixels. Image pre-processing comprises
of image cropping, color space conversion and color channel extraction, contrast enhancement, application of
Gabor filtering and half wave rectification done in the specified order. Then, a feature vector representation
including the pixel intensities and Gabor responses for every pixel is formed which serves as input to the data
mining phase. The data mining phase includes application of principal component analysis to the feature vector
followed by classification through the ensemble classification model. Each pixel is classified either as a vessel
or a non-vessel pixel. The classified image hence represents s binary vessel segmented image. Now, this image
is post processed using connected component analysis, yielding to final segmented vessel image.
The entire framework is described in detail in the following sub-sections.
Image Pre-Processing:
The fundus image is pre-processed to make it suitable for further processing. The image pre-processing
phase includes image cropping, color space conversion and color channel extraction, contrast enhancement,
Gabor filtering and half wave rectification.
Image Cropping:
The image is cropped such that border of the image touches the field of view in all the sides. This reduces
the processing area of the image thus reducing the computational complexity. The cropped area will not contain
any vessel pixels as it is the area out of the region of field of view. Hence the pixels discarded in the step
(approximately 40,000 pixels; varies with every image) would be included during performance evaluation as
correctly identified non-vessel pixels.
Color Space Conversion and Color Channel Extraction:
Followed by image cropping, the RGB color model is converted into YCbCr, L*a*b and Gaussian models
(Guesebroek et al., 2001). Since the blood vessels are well depicted at high contrast, the color channel
displaying the highest contrast is chosen from every color model and considered further. Hence Green channel
from RGB; Y Channel from YCbCr; L channel from L*a*b model and G1 channel from Gaussian color model
is considered for further processing.
Contrast Enhancement:
The four channel images thus obtained are contrast enhanced using Contrast Limited Adaptive Histogram
Equalization (CLAHE) (Pizer et al., 1987). The CLAHE technique is adopted so that the thin vessels are also
enhanced.
Gabor Filtering:
2-D Gabor filters (Fogel and Sagi, 1989), which are sinusoidally modulated Gaussian functions, have been
used to further enhance the blood vessels in the contrast enhanced images. The Gabor responses obtained at
wavelength 9, 10 and 11 are investigated further. Since this step is applied to four images, twelve resultant
images are obtained.
Half wave rectification:
The process of half wave Rectification operates on the Gabor response images based on a percentage value
(10 in this case) of the maximum intensity of the image. Pixels with Gabor response value lesser than the
specified percentage will be set to zero.
After the image pre-processing phase, a feature vector representing each pixel is formed with the twelve
half wave rectified Gabor responses and the pixel intensities of the Green channel image. The feature vector is
given as input to the data mining phase
Data Mining:
The image pre-processing phase is followed by the application of data mining techniques (GeethaRamani,
Lakshmi and Shomona, 2012a) to the feature vector formed from the previous phase. The data mining phase
comprises of feature vector formulation through principal component analysis and classification through
ensemble of classifiers.
Feature Contraction through Principal Component Analysis:
Principal Component Analysis (Jollife, 1986) has always been applied to analyse patterns in high
dimensional data. A new set of more representative features is constructed from the feature vector through
application of Principal Component Analysis. This new set of features is considered for further processing.
603
R. Geetha Ramani and Lakshmi Balasubramanian, 2015
Advances in Natural and Applied Sciences, 9(6) Special 2015, Pages: 600-606
Classification through Ensemble of Classifiers:
Supervised Classification techniques ( Shomona Gracia Jacob and Geetha Ramani, 2012) can be applied to
the feature vector formed from the preceding step. The segmentation of vessels is viewed here as a binary
classification problem i.e., the pixel belongs to either a vessel class or non-vessel class. The supervised
classification technique demands training data to form the decision rules (Shomona Gracia Jacob,
GeethaRamani and Nancy, 2012). In this work, one of the images is taken as reference and used for training.
The performance of the classifier is always increased through a meta-learning procedure. In this study, bagging
technique (Breiman 1996) is applied to the classifier model to achieve increased performance.
C4.5 (Steven, 1994) with Bagging is applied on the data to build the classification model to classify the
vessel and non-vessel pixels. The output of this stage would be a binary image consisting of vessel pixels
represented as white and non-vessel pixels represented as black. This image is further post processed to
eliminate false vessels and false non-vessels.
Image Post-Processing:
The result from the data mining phase is a classified binary image. This image contained a few white pixels
hanging all over the image. Since, even small vessels would occupy greater than 10 pixels, all components
lesser than 10 pixels were eliminated. The entire process comprising of image pre-processing, data mining and
image post-processing is illustrated on a sample fundus image in Fig. 2.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
(m)
(n)
(o)
(p)
(q)
(v)
(r)
(w)
(s)
(x)
(t)
(u)
(y)
Fig. 2: Sample illustration of the proposed framework on image 20 of the DRIVE dataset. a) Input Retinal
fundus image b) cropped image discarding 49259 pixels c) green channel image d) Y channel image e)
L channel image f) G1 channel image g-j) Contrast Enhanced images of Green, Y, Land G1 channel
images k-m) Half wave rectified Gabor responses of Green channel image at wavelength 9, 10 and 11 np) Half wave rectified Gabor responses of Y channel image at wavelength 9, 10 and 11 q-s) Half wave
rectified Gabor responses of L channel image at wavelength 9, 10 and 11 t-v) Half wave rectified Gabor
responses of G1 channel image at wavelength 9, 10 and 11 w) Data Mining result x) post-processed
result y) ground truth.
Experimental results of the proposed methodology are discussed in the next section.
RESULTS AND DISCUSSION
The proposed methodology is validated on the fundus images from the publicly available database DRIVE
(Niemeijer et al., 2004). DRIVE database is the widely used database for assessing the performance of the
vessel segmentation algorithms. The database consists of forty images, 20 of which are packed as training and
the remaining 20 as test set. The results reported here are the average performance measures for the twenty
604
R. Geetha Ramani and Lakshmi Balasubramanian, 2015
Advances in Natural and Applied Sciences, 9(6) Special 2015, Pages: 600-606
images in the test set of the database. A manual graded ground truth is available for these images and a second
human observer prediction is also provided.
The proposed methodology is implanted using Matlab r2008a and weka 3.6.11. Matlab r2008a is utilised
for implementing the image pre-processing and image post processing phases. Weka, an open source data
mining tool is used for implementing the data mining phases. Performance comparison is facilitated through
average accuracy of the segmentation algorithm on all the twenty images.
Various classification algorithms (GeethaRamani and Jacob, 2013) were attempted for the purpose of
classification of vessels and non-vessels. C4.5 with bagging was observed to yield the best results. C4,5
algorithm utilises gain ratio for the purpose of evaluation of the attributes to form the decision tree. As
supervised classification algorithms demands training data, one of the images from the training set is taken as
reference and considered for training (first image of the training set in this case). The classification prediction
from the C4.5 algorithm yields a binary segmented image, which is post processed for enhanced performance.
Accuracy (GeethaRamani, Lakshmi and Shomona, 2012b, 2013)(GeethaRamani and Lakshmi, 2013) is
calculated as the ratio of sum of correctly classified non-vessel pixels and correctly classified vessel pixels to the
total number of image pixels. Sensitivity is defined as the ratio of correctly identified vessel pixels to the total
number of actual vessel pixels. Specificity is defined as the ratio of correctly identified non-vessel pixels to the
total number of non vessel pixels. In all these calculations, the total numbers of non-vessel pixels indicate the
sum of pixels discarded during image cropping and the correctly identified non-vessel pixels in the current step.
Table 1 presents the accuracy of the proposed algorithm in every image.
Table 1: Performance of the proposed methodology on test set of DRIVE database.
Image ID
Accuracy (%)
Image ID
1
95.58
11
2
95.84
12
3
94.19
13
4
95.31
14
5
94.85
15
6
95.00
16
7
94.78
17
8
94.78
18
9
95.56
19
10
95.21
20
Accuracy (%)
94.94
95.46
94.97
95.13
91.74
95.63
95.55
95.91
95.16
95.96
The proposed technique thus achieved an average accuracy, sensitivity and specificity of 95.08%, 68.20%
and 97.70% respectively. Performance comparison of the presented technique with the existing techniques is
presented in Table 2.
Table 2: Performance Comparison with the existing methodologies.
Technique
Proposed technique
Wilfred and Edward, 2014
Second Observer
Usman Akram and Shoab, 2013
Soares et al., 2006
Mendonca and Campilho, 2006
Diego Marin et al., 2011
Niemwijer et al., 2004
Yanli Hou, 2014
Djibril Kaba et al.. 2014
Eid et al., 2014
Bankhead et al., 2012
Martinez Perez et al., 2007
Xu and Luo, 2010
Accuracy (%)
95.08
95.03
94.73
94.69
94.66
94.63
94.52
94.16
94.15
94.10
93.90
93.71
93.44
93.28
The proposed methodology is thus efficient in segmenting blood vessels from the retinal fundus images. It
can be of great use to the ophthalmologist society in analyzing the patient’s retina.
4. Conclusion:
Computational techniques viz. image processing and data mining are exploited for the purpose of retinal
image analysis. Changes in the properties associated with the retinal blood vessels are the main indicator to
diagnose many sight threatening diseases. In this paper, segmentation of retinal blood vessel is attempted
through sequential application of image cropping, color space conversion and color channel extraction, Contrast
Enhancement, Gabor Filtering, Half wave rectification, Principal Component Analysis, Classification through
C4,5 with Bagging and image post-processing through connected component analysis.. Results reported
outperform the existing retinal segmentation algorithms. An average accuracy of 95.08% is accomplished on the
605
R. Geetha Ramani and Lakshmi Balasubramanian, 2015
Advances in Natural and Applied Sciences, 9(6) Special 2015, Pages: 600-606
test set of the DRIVE dataset. The insight for future directions includes building new classification models, and
enhanced post processing techniques for better segmentation. The proposed framework can assist the medical
practitioners in better retinal analysis, earlier diagnosis and hence better treatment to patients.
REFERENCES
Bankheard, P., C.N. Scholfield, J.G. McGeown and T.M. Curtis, 2012. Fast Retinal Vessel Detection and
Measurement Using Wavelets and Edge Location Refinement, PLoS One, 7(3): e32435.
Breiman, Leo, 1996. Bagging predictors, Machine Learning, 24(2): 123-140.
Diego Marin, Aquino, A., M.E. Gegundwz-Arias and J.M. Barvo, 2011. A new supervised method for
blood vessel segmentaion in retinal imaes by using Gray-Level and Moments Invariants-based features, IEEE
Transactions on Medical Imaging, 30(1): 146-158.
Djibril Kaba, Chaung Wang, Yongmin Li, et al., 2014. Retinal blood vessel extraction using probabilistic
modelling, Health Information Science and Systems, 2(2): 1-10.
Eid Emary, Zawbaa, H.M., A.E. Hassanien and G. Schaefer, 2014. Retinal blood vessel segmentation using
bee colony optimisation and pattern search, In Proceedings of International Joint Conference on Neural
Networks, 1001-1006.
Fogel, I. and D. Sagi, 1989. Gabor filters as texture discriminator, Biological Cybernetics, 61(2).
Geetha Ramani, R. and Lakshmi Balsubrmanian, 2013. Multi-Class Classification for Prediction of Retinal
Diseases (Retinopathy and Occlusion) from Fundus Images, Proceedings of ICKM’, 13: 122-134.
Geetha Ramani, R., B. Lakshmi and Shomona Gracia Jabob, 2012a. Automatic Prediction of Diabetic
Retinopathy and Glaucoma through Image processing and Data Mining Techniques, Proc. of Int. Conf. on
Machine Vision and Image Processing, 163-167.
Geetha Ramani, R., B. Lakshmi and Shomona Gracia Jacob, 2012b. Data Mining Method of Evaluating
Classifier Prediction Accuracy in Retinal Data, Proc. of IEEE Int. Conf. on Computational Intelligence and
Computing Research, 426-429.
Geetha Ramani, R., B. Lakshmi and Shomona Gracia Jacob, 2013. ROC Analysis of Classifiers in
Automatic Detection of Diabetic Retinopathy using Shape Features of Fundus Images, Proc. of Int. Conf. on
Advances in Computing, Communications and Informatics, 66-72.
Geetha Ramani, R. and S.G. Jacob, 2013. Improved Classification of Lung Cancer Tumors based on
Structural and Physicochemical Properties of Proteins Using Data Mining Models. PLoS One, 8(3): e58772.
Geusebroek, J., R. Van Den Boomgaard, A.W.M. Smeulders and H. Gearts, 2001. Color Invariance, IEEE
Transactions on Pattern Analysis and Machine Intelligence, 23(2): 1338-1350.
Jolliffe, I.T., 1986. Principal Component Analysis, Springer-Verlag, 487; ISBN 978-0-387-95442-4.
Martinez-Perez, M.E., A.D. Hughes, S.A. Thom, et al., 2007. Segmentation of blood vessels from red-free
and fluoroscein retinal images, Medical Image Analysis, 11: 47-61.
Mendonca, A.M. and A. Campilho, 2006. Segmentaion of retinal blood vessels by cobining the detection of
centerlines and morphological reconstruction, IEEE Transactiona on Medical Imaging, 25: 1200-1213.
Niall Patton, Aslam, T.M., T. MacGillivray, et al., 2006. Retinal image analysis: Concepts, applications and
potential, Progress in Retinal and Eye Research, 25: 99-127.
Niemeijer, M., J.J. Staal, B. Van Ginnekan, et al., 2004. Comparative study on retinal vessel segmentation
methods on a new publicly available database, SPIE, 648-656.
Pizer, S.M., E. Philipi Amburn, J.D. Austin, et al., 1987. Adaptive Histogram Equalization and Its
Variations, Computer Vision, Graphics, and Image Processing, 39: 355-368.
Shomona Gracia Jacob and R. Geetha Ramani, 2012. Evolving Efficient Classification Rules from
Cardiotocography Data through Data Mining Methods and Techniques, European Journal of Scientific
Research, 78(3): 468-480.
Shomona Gracia Jacob, R. Geetha Ramani and P. Nancy, 2012. Efficient Classifier for Classification of
Hepatitis C Virus Clinical Data through Data Mining Algorithms and Techniques. in Proc. of the International
Conference on Computer Applications, Techno Forum Group, Puducherry, India, 27-31.
Soares, J.V.B., J.J.G. Leandro, R.M. Caesar, et al., 2006. Retinal vessel segmentation using the 2-D Gabor
wavelet and supervised classification, IEEE Transactions on Medical Imaging, 25: 1214-1222.
Steven L. Salzberg, 1994. C4.5: Programs for Machine Learning by J. Ross Quinlan, Morgan Kaufmann
Publishers, Inc. 1993, Machine Learning, 16(3): 235-240.
Usman Akram, M. and Shoab A. Khan, 2013. Multilayered Thresholding- based blood vessel segmentation
for screening of diabetic retinopathy, Engineering with Computers, 29(2): 165-173.
Xu L. and S. Luo, 2010. A novel method for blood vessel detection from retinal images, BioMedical
Engineering Online, 9(14).
Yanli Hou, 2014. Automatic segmentation of retinal blood vessels based on improved multiscale line
detection, Journal of Computing Science and Engineering, 8(2): 119- 128.
606
R. Geetha Ramani and Lakshmi Balasubramanian, 2015
Advances in Natural and Applied Sciences, 9(6) Special 2015, Pages: 600-606
Wilfred Franklin, S. and S. Edward Rajan, 2014. Compuerized Screening of diabetic retinopathy employing
blood vessel segmentation in retinal images, Biocybernetics and Biomedical Engineering, 34: 117-124.