COMPARISON OF MATCHED FILTER AND WAVELET BASED BLOODVESSEL ENHANCEMENT

M. Usman AkramComputer Engineering

College of E&ME, NUSTRawalpindi, Pakistan

email: usmakram@gmail.com

Sarwat NasirComputer Sciences & Engineering

Bahria UniversityIslamabad, Pakistan

email: sarwatnasir83@gmail.com

Shoab A. KhanElectrical Engineering

College of E&ME, NUSTRawalpindi, Pakistan

email: shoab@carepvtltd.com

ABSTRACTDiabetic retinopathy occurs when high blood sugar dam-ages small blood vessels in retina so blood vessel seg-mentation is an essential step for the diagnoses of diabeticretinopathy. An automated tool for blood vessel segmen-tation is useful to eye specialists for purpose of patientscreening and clinical study. In this paper, we presenta wavelet based method for enhancing, locating and seg-menting blood vessels in images of retina. This techniqueis tested on publicly available STARE database of hand la-beled images which has been established to facilitate com-parative studies on segmentation of blood vessels in retinalimages. Wavelet based enhancement method is comparedwith matched filter based vessel enhancement method. Thevalidation of our retinal image vessel enhancement tech-nique is supported by results.

KEY WORDSBlood vessels, Matched filter, 2-D wavelet, vessel enhance-ment.

1. Introduction

Diabetic Retinopathy is one of the most serious eye man-ifestation of diabetes and is responsible for most of theblindness caused by diabetes [1]. In this condition, a net-work of small blood vessels, called choroidal neovascular-ization (CNV), arises in the choroid and taking a portion ofthe blood supplying the retina. As the amount of blood sup-plying the retina is decreased, the sight may be degradedand in the severe cases, blindness may occur . The eye isa window to the retinal vascular system which is uniquelyaccessible for the study of a continuous vascular bed in hu-mans [1].

Growth of new vessels leads to intraocular haemor-rhage and possible retinal detachment with profound globalsight loss. The detection and measurement of blood ves-sels can be used to quantify the severity of disease, as partof the process of automated diagnosis of disease or in theassessment of the progression of therapy [2]. Retinal bloodvessels have been shown to have measurable changes in di-ameter, branching angles, length or tortuosity, as a resultof a disease. Thus a reliable method of vessel segmenta-tion would be valuable for the early detection and char-

acterization of changes due to such diseases [2]. Retinalvascular pattern facilitates the physicians for the purposesof diagnosing eye diseases, patient screening, and clinicalstudy [3]. Inspection of blood vessels provides the infor-mation regarding pathological changes caused by oculardiseases including diabetes, hypertension, stroke and arte-riosclerosis [4]. The hand mapping of retinal vasculatureis a time consuming process that entails training and skill.Automated segmentation provides consistency and reducesthe time required by a physician or a skilled technician formanual labeling [5].

Retinal vessel segmentation may be used for auto-matic generation of retinal maps for the treatment of age-related macular degeneration [6], extraction of characteris-tic points of the retinal vasculature for temporal or multi-modal image registration [7], retinal image mosaic synthe-sis, identification of the optic disc position [8], and local-ization of the fovea [9]. The challenges faced in automatedvessel detection include wide range of vessel widths, lowcontrast with respect with background and appearance ofvariety of structures in the image including the optic disc,the retinal boundary and other pathologies [10].

Different approaches for automated vessel segmenta-tion have been proposed. Methods based on vessel trackingto obtain the vasculature structure, along with vessel di-ameters and branching points have been proposed by [12]-[17]. Tracking consists of following vessel center linesguided by local information. In [23], ridge detection wasused to form line elements and partition the image intopatches belonging to each line element. Pixel features werethen generated based on this representation. Many featureswere presented and a feature selection scheme is used toselect those which provide the best class separability. Pa-pers [18]-[21] used deformable models for vessels segmen-tation . Chuadhuri et al. [22] proposed a technique usingmatched filters to emphasize blood vessels. An improvedregion based threshold probing of the matched filter re-sponse (MRF) technique was used by Hoover et al. [24].

In this paper, we present the colored retinal imagevessel enhancement and segmentation technique that en-hances the vascular pattern using 2-D Gabor wavelet. Wecompare our method with region based threshold probingof MFR [24]. The results show the proposed method issignificantly better than MFR method in enhancing and de-

654-095

Proceedings of the IASTED International Conference Signal and Image Processing (SIP 2009) August 17 - 19, 2009 Honolulu, Hawaii, USA

260

tecting the blood vessels. We test our method on publiclyavailable STARE dataset of hand labeled images [11].

The paper is organized in five sections. In Section 2,matched filter response based method for vessel enhance-ment is presented. Section 3 presents proposed waveletbased vessel enhancement and segmentation method. Ex-perimental results of the tests on the images of the STAREdatabase and their analysis are given in Section 4 followedby conclusion in Section 5.

2. Matched Filter for Vessel Enhancement

Hoover et al. [24] proposed a method to segment blood ves-sels that compliments local vessel attributes with region-based attributes of the network structure. In this method, apiece of the blood vessel network is hypothesized by prob-ing an area of the MFR image, iteratively decreasing thethreshold. At each iteration, region-based attributes of thepiece are tested to consider probe continuation, and ulti-mately to decide if the piece is blood vessel or not. Pixelsfrom probes that are not classified as vessel are recycledfor further probing. The strength of this approach was thatindividual pixel labels are decided using local and region-based properties.

Figure 1 show the flowchart of hoover et al. method[24]. Interested researcher can read on this method in detailfrom [24].

Figure 1. Flowchart for MFR based vessel segmentationmethod [24]

3. Wavelet based Vessel Enhancement

We have used 2-D Gabor wavelet to enhance the vascu-lar pattern and thin vessels [25]. 2-D Gabor wavelet isused due to its directional selectiveness capability of detect-ing oriented features and fine tuning to specific frequencies[25], [27].

• The continuous wavelet transform Tψ(b, θ, a) is de-fined in terms of the scalar product of f with the trans-formed wavelet ψb, θ, a using (1) [28]

Tψ(b, θ, a) = C−1/2ψ 〈ψb, θ, a|f〉

= C−1/2ψ a−1

∫ψ∗(a−1r−θ(x− b))f(x)d2x (1)

where f ∈ L2 is an image represented as a square in-tegrable (i.e., finite energy) function defined over R2

and ψ ∈ L2 be the analyzing wavelet. Cψ , ψ, b, θ anda denote the normalizing constant, analyzing wavelet,the displacement vector, the rotation angle, and the di-lation parameter respectively.

• It is easy to implemented wavelet transform usingthe fast Fourier transform algorithm. Fourier wavelettransform is defined using (2) [27].

Tψ(b, θ, a) = C−1/2ψ a

∫exp (jkb)ψ̂∗(ar−θk)f̂(k)d2k

(2)where j =

√−1, and the hat (ψ̂∗ and f̂ ) denotes a

Fourier transform.

• The 2-D Gabor wavelet is defined as [28]

ψG(x) = exp(jk0x) exp(−12|Ax|2) (3)

where k0 is a vector that defines the frequency of thecomplex exponential and A = diag[ε−1/2, 1], ε ≥ 1is a 2 × 2 diagonal matrix that defines the elongationof filter in any desired direction.

• For each pixel position and considered scale value, theGabor wavelet transformMψ(b, a) is computed using(4) [27], for θ spanning from 0o up to 170o at steps of10o and the maximum is taken.

Mψ(b, a) = maxθ|Tψ(b, θ, a)| (4)

This gives us the enhanced vascular pattern for theretinal image. Histogram for the enhanced retinal im-age is calculated. Maximum values occur for the grayishbackground while the vessel corresponds to values a slightgreater than the background values as they are of brightcolor. An adaptive thresholding technique is used that se-lects this point which separates the vessels from the rest of

261

image. Vessel segmentation mask is created by applyingthis threshold value

Fig. 2 shows the blood vessel enhancement and seg-mentation for STARE dataset image. Fig. 2(b) shows theenhanced vascular pattern using Hoover et al. method [24]while fig. 2(c) shows the vascular pattern enhanced usingwavelet method and there is a clear improvement in en-hancement using wavelet method. Fig 2(d) is the vesselsegmentation mask extracted from 2(c) by applying adap-tive thresholding.

Figure 2. (a) Original STARE dataset image, (b) MFRbased vessel enhancement, (c) Wavelet based vessel en-

hancement, (d) Blood vessel segmentation mask

4. Experimental Results

The tests of proposed technique are performed with respectto the vessel segmentation accuracy and their standard de-viation using publicly available STARE database [11]. TheSTARE database consists of 20 RGB color images of theretina. The images are of size 605×700 pixels, 24 bits perpixel (standard RGB). There are two hand-labeling avail-able for the 20 images of made by two different humanobservers. The manually segmented images by 1st humanobserver are used as ground truth and the segmentations ofset B are tested against set A, serving as a human observerreference for performance comparison truth. The segmen-tations of set B are tested against those of A, serving asa human observer reference for performance comparison[24]. The true positive fraction is the fraction of number oftrue positive (pixels that actually belong to vessels) and to-tal number of vessel pixels in the retinal image. False posi-tive fraction is calculated by dividing false positives (pixelsthat don’t belong to vessels) by total number of non ves-sel pixels in the retinal image. We compared the accuracyof proposed technique with the accuracy of the method of

Table 1. Vessel Segmentation Results (STARE Database)

Segmentation Average Standard

Methods Accuracy Deviation

2nd Observer 0.9351 0.0171

Hoover et al. 0.9275 0.0247

Proposed Method 0.9439 0.0253

Hoover et al. [24]. Table 1 summarizes the results of vesselsegmentation for STARE dataset [30]. It shows the resultsin term of average accuracy and their standard deviation.Fig 3. shows the enhanced vascular pattern for MFR andwavelet based methods. Fig. 4 illustrates the blood vesselsegmentation results for MFR based vessel segmentationand vessel segmentation results using proposed method.

5. Conclusion

The problem with retinal images is that the visibility of vas-cular pattern is usually not good. So, it is necessary to en-hance the vascular pattern. In this paper, a wavelet basedcolored retinal image blood vessel enhancement and seg-mentation technique is proposed. The proposed method iscompared with MFR based mathod. We have tested ourtechnique on publicly available STARE database of man-ually labeled images. Experimental results show that ourmethod performs well in enhancing and segmenting thevascular pattern.

References

[1] M. S. Mabrouk, N. H. Solouma and Y. M. Kadah,Survey of retinal image segmentation and registration,GVIP Journal, 6(2), 2006, 1-11.

[2] M. E. Martnez-Prez, A. D. Hughes, S. A. Thom, A.A. Bharath and K. H. Parker, Retinal Blood Ves-sel Segmentation by Means of Scale-Space Analy-sis and Region Growing, Medical Image Comput-ing and Computer-Assisted Intervention MICCAI99,1679, 1999, 90-97.

[3] J. KansKy, Clinical Opthalmology, Butterworh-Heinmann , London, 1994.

[4] T. Teng, M. Lefley and D. Claremont, Progress to-wards automated diabetic ocular screening: A reviewof image analysis and intelligent systems for diabeticretinopathy, Med. Biol. Eng. Comput., 40(1), 2002, 2-13.

[5] E. J. Susman, W. J. Tsiaras, and K. A. Soper, Diagnosisof diabetic eye disease, JAMA, 247(23), 1982, 3231-3234.

262

Figure 3. Vascular pattern enhancement for STARE dataset images: (a) Original retinal images, (b) MFR based enhancedvascular pattern, (c) Wavelet based enhanced vascular pattern

Figure 4. Blood vessel segmentation for STARE dataset images: (a) Original retinal images, (b) MFR based vessel segmenta-tion, (c) Wavelet based segmentation

263

[6] A. Pinz, S. Bernogger, P. Datlinger and A. Kruger,Mapping the human retina, IEEE Trans. Med. Imag.,17(4), Aug. 1998, 606-619.

[7] C.L. Tsai, C. V. Stewart, H. L. Tanenbaum and B.Roysam, Model based method for improving the accu-racy and repeatability of estimating vascular bifurca-tions and crossovers from retinal fundus images, IEEETrans. Inf. Technol. Biomed., 8(2), June 2004, 122-130.

[8] M. Foracchia, E. Grisam, and A. Ruggeri, Detection ofthe optic disc in retinal images by means of a geometri-cal model of vessel structure, IEEE Trans. Med. Imag.,23(10), Oct. 2004, 1189-1195.

[9] H. Li and O. Chutatape, Automated feature extractionin color retinal images by a model based Approach,IEEE Trans. Biomed. Eng., 51(2), Feb. 2004, 246-254.

[10] A. M. Mendona, A. J. Campilho, Segmentation ofRetinal Blood Vessels by Combining the Detection ofCenterlines and Morphological Reconstruction, IEEETrans. Med. Imag., 25(9), Sep. 2006, 1200-1213.

[11] A. Hoover, STARE database [Online]. Available:http://www.ces.clemson.edu/ ahoover/stare.

[12] I. Liu and Y. Sun, Recursive tracking of vascular net-works in angiograms based on the detection-deletionscheme, IEEE Trans. Med. Imag., 12(2), Jun. 1993,334-341.

[13] L. Zhou, M. S. Rzeszotarski, L. J. Singerman, andJ. M. Chokreff, The detection and quantification ofretinopathy using digital angiograms, IEEE Trans.Med. Imag., 13(4), Dec. 1994, 619-626.

[14] O. Chutatape, L. Zheng, and S. M. Krishnan, Retinalblood vessel detection and tracking by matched Gaus-sian and Kalman filters, 20th Annu. Int. Conf. IEEEEng. Med. Biol. Soc. (EMBS98), 20(6), 1998, 3144-3149.

[15] Y. A. Tolias and S. M. Panas, A fuzzy vessel trackingalgorithm for retinal images based on fuzzy clustering,IEEE Trans. Med. Imag., 17(2), Apr. 1998, 263-273.

[16] A. Can, H. Shen, J. N. Turner, H. L. Tanenbaum,and B. Roysam, Rapid automated tracing and fea-ture extraction from retinal fundus images using di-rect exploratory algorithms, IEEE Trans. Inf. Technol.Biomed., 3(2), Jun. 1999, 125-138.

[17] M. Lalonde, L. Gagnon, and M.-C. Boucher, Non-recursive paired tracking for vessel extraction fromretinal images, Vision Interface, 2000, 61-68.

[18] T. McInerney and D. Terzopoulos, T-snakes: Topol-ogy adaptive snakes, Med. Image Anal., 4(2), 2000,73-91.

[19] R. Toledo, X. Orriols, X. Binefa, P. Radeva, J. Vitriand J. Villanueva, Tracking of elongated structures us-ing statistical snakes, IEEE Comput. Soc. Conf. Com-put. Vision Pattern Recog. (CVPR), 1, 2000, 157-162.

[20] A. Vasilevskiy and K. Siddiqi, Flux maximizing geo-metric flows, IEEE Trans. Pattern Anal. Mach. Intell.,24(12), Dec. 2002, 1565-1578.

[21] D. Nain, A. Yezzi, and G. Turk, Vessel segmenta-tion using a shape driven flow, Med. Image Comput.Comput.-Assist. Intervention MICCAI, 2004, 51-59.

[22] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, andM. Goldbaum, Detection of blood vessels in retinalimages using two-dimensional matched filters, IEEETrans. Med. Imag., 8(3), 1989, 263-269.

[23] J. Staal, M. D. Abramoff, M. Niemeijer, M. A.Viergever and B. van Ginneken, Ridge-based vesselsegmentation in color images of the retina, IEEE Trans.Med. Imag., 23(4), Apr. 2004, 501-509.

[24] A. Hoover, V. Kouznetsova and M. Goldbaum, Locat-ing blood vessels in retinal images by piecewise thresh-old probing of a matched filter response, IEEE Trans.Med. Imag., 19(3), Mar. 2000, 203-211.

[25] A. Arnodo, N. Decoster and S. G. Roux, A wavelet-based method for multifractal image analysis. I.Methodology and test applications on isotropic andanisotropic random rough surfaces, Eur. Phys. J. A,15(3), 2000, 567-600.

[26] R. C. Gonzalez and R. E. Woods, Digital image pro-cessing, Prentice hall, second edition, 2002.

[27] J.P. Antoine, P. Carette, R. Murenzi, and B. Piette, Im-age analysis with two-dimensional continuous wavelettransform, Signal Process., 31(3), 1993, 241-272.

[28] J. V. B. Soares, J. J. G. Leandro, R. M. Cesar, H. F.Jelinek and M. J. Cree, Retinal vessel segmentation us-ing the 2-D gabor wavelet and supervised classifica-tion, IEEE Trans. on Med. Imag., 25(9), Sept. 2006,1214-1222.

[29] R. C. Gonzalez, R. E. Woods and S. L. Eddins, Digitalimage processing using matlab, Prentice hall, secondedition, 2007.

[30] M. U. Akram, A. Atzaz, S. F. Aneeque, S. A. Khan,Blood Vessel Enhancement and Segmentation UsingWavelet Transform, IEEE ICDIP, March 2009, 30-34.

264