Retinal Images: Blood Vessel Segmentation byThreshold ProbingM. Usman Akram∗ and Aasia Khanum‡

Department of Computer Engineering∗,‡

College of Electrical and Mechanical Engineering∗,‡

National University of Sciences & Technology∗,‡

PakistanEmail: usmakram@gmail.com∗, aasia@ceme.nust.edu.pk‡

Abstract—An automated system for screening and diagnosis ofdiabetic retinopathy should segment blood vessels from coloredretinal image to assist the ophthalmologists. We present amethod for blood vessel enhancement and segmentation. Thispaper proposes a wavelet based method for vessel enhancement,piecewise threshold probing and adaptive thresholding for vessellocalization and segmentation respectively. The method is testedon publicly available DRIVE and STARE databases of manuallylabeled images which has been established to facilitate compar-ative studies on segmentation of blood vessels in retinal images.The proposed method achieves an accuracy of 0.9469 on DRIVEdatabase and of 0.9502 on STARE database.

I. INTRODUCTION

Diabetic retinopathy is the result of microvascular changesin retina [1]. In some patient with diabetic retinopathy, bloodvessels may swell and leak fluid [2]. In other, new abnormalblood vessels grow on the surface of the retina [3] that is whyblood vessel segmentation is an important part of automateddiabetic retinopathy screening system.

A tool which can be used to assist in the diagnosis ofdiabetic retinopathy should automatically detect all retinalimage features such as optic disk, fovea and blood vessel [3],[5], [6] and all abnormalities in retinal image such as microa-neurysms [2], [7], [8], hard exudates and soft exudates [9],[10], hemorrhages, and edema [2]. Illumination equalizationis needed to enhance the image quality as the acquired colorretinal images are normally of different qualities.

Retinal vascular pattern facilitates the physicians for thepurposes of diagnosing eye diseases, patient screening, andclinical study [4]. Inspection of blood vessels provides theinformation regarding pathological changes caused by oculardiseases including diabetes, hypertension, stroke and arte-riosclerosis [11]. The hand mapping of retinal vasculatureis a time consuming process that entails training and skill.Automated segmentation provides consistency and reduces thetime required by a physician or a skilled technician for manuallabeling [1].

Retinal vessel segmentation may be used for automaticgeneration of retinal maps for the treatment of age-relatedmacular degeneration [12], extraction of characteristic pointsof the retinal vasculature for temporal or multimodal imageregistration [13], retinal image mosaic synthesis, identificationof the optic disc position [5], and localization of the fovea [14].

The challenges faced in automated vessel detection includewide range of vessel widths, low contrast with respect withbackground and appearance of variety of structures in theimage including the optic disc, the retinal boundary and otherpathologies [15].

Different approaches for automated vessel segmentationhave been proposed. Methods based on vessel tracking toobtain the vasculature structure, along with vessel diametersand branching points have been proposed by [16]-[21]. Track-ing consists of following vessel center lines guided by localinformation. In [27], ridge detection was used to form lineelements and partition the image into patches belonging toeach line element. Pixel features were then generated based onthis representation. Many features were presented and a featureselection scheme is used to select those which provide the bestclass separability. Papers [22]-[25] used deformable modelsfor vessels segmentation. Chuadhuri et al. [26] proposed atechnique using matched filters to emphasize blood vessels.An improved region based threshold probing of the matchedfilter response technique was used by Hoover et al. [28].

In this paper, we present the retinal image enhancementtechnique that enhances the blood vessels using Gabor wavelet.We also present the colored retinal image vessel localizationand segmentation technique that localizes and segments thevascular pattern using piecewise probing and adaptive thresh-olding respectively.

The paper is organized in four sections. In section II, aschematic overview of our implementation methodology isillustrated. Section II also presents the step by step techniquesrequired for automated vessel enhancement and segmentation.Experimental results of tests on the images of the DRIVE andSTARE databases and their analysis are given in Section IIIfollowed by conclusion in Section IV.

II. PROPOSED SYSTEM

Automatically locating the accurate vascular pattern is veryimportant in implementation of vessel screening system. Ourproposed method segments the blood vessels from retinal im-ages with great accuracy as compared to previous techniques.In proposed method, the monochromatic RGB retinal imageis taken as an input and 2-D Gabor wavelet is used to enhance

2010 IEEE Symposium on Industrial Electronics and Applications (ISIEA 2010), October 3-5, 2010, Penang, Malaysia

978-1-4244-7647-3/10/$26.00 ©2010 IEEE 493

Fig. 1. Complete flow diagram for proposed system

the vascular pattern especially the thin and less visible vesselsare enhanced [28].

Locating of blood vessels is done using piecewise thresholdprobing and vessel segmentation binary mask is created bythresholding the enhanced retinal image. The blood vesselsare marked by the masking procedure which assigns oneto all those pixels which belong to blood vessels and zeroto non vessels pixels. Figure 1 shows the complete flowdiagram for designing an automated vessel screening systemusing proposed blood vessel enhancement and segmentationtechnique.

A. Blood Vessel Enhancement

The monochromatic RGB retinal image is taken as aninput and 2-D Gabor wavelet is used to enhance the vascularpattern especially the thin and less visible vessels are enhancedusing Gabor wavelet [29]. 2-D Gabor wavelet is used dueto its directional selectiveness capability of detecting orientedfeatures and fine tuning to specific frequencies [29], [30].

1) The continuous wavelet transform Tψ(b, θ, a) is definedin terms of the scalar product of f with the transformedwavelet ψb, θ, a using equation 1 [31]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 squareintegrable (i.e., finite energy) function defined over R2

and ψ ∈ L2 be the analyzing wavelet. Cψ , ψ, b, θ and adenote the normalizing constant, analyzing wavelet, thedisplacement vector, the rotation angle, and the dilationparameter respectively.

2) It is easy to implemented wavelet transform using thefast Fourier transform algorithm. Fourier wavelet trans-form is defined using equation 2 [30].

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.3) The 2-D Gabor wavelet is defined as [31]

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 ≥ 1 isa 2 × 2 diagonal matrix that defines the elongation offilter in any desired direction.

4) For each pixel position and considered scale value, theGabor wavelet transform Mψ(b, a) is computed usingequation 4 [30], for θ spanning from 0o up to 170o atsteps of 10o and the maximum is taken.

Mψ(b, a) = maxθ|Tψ(b, θ, a)| (4)

B. Blood Vessel Localization and Segmentation

We present a method for blood vessel localization that com-pliments local vessel attributes with region-based attributes ofthe network structure. A piece of the blood vessel networkis hypothesized by probing an area of the wavelet basedenhanced image, iteratively decreasing the threshold. Pixelsfrom probes that are not classified as vessel are recycled forfurther probing. The strength of this approach is that individual

494

Fig. 2. Flowchart of Threshold Probing

pixel labels are decided using local and region-based propertiesand afterwards adaptive thresholding has been applied.

1) Threshold Probing:1) The basic operation of our algorithm is to probe regions

in a wavelet based enhanced image.2) During each probe, a set of criteria is tested to determine

the threshold of the probe, and ultimately to decide ifthe area being probed (termed a piece) is blood vessel.

3) A queue of points is initialized, each of which will beused for a probe.

4) Upon a probe’s completion, if the piece is determinedto be vessel, then the endpoints of the piece are addedto the queue.

5) In this way, different probes (and thus different thresh-olds) can be applied throughout the image.

A flowchart for vessel probing algorithm is shown in figure 2.This gives us the enhanced vascular pattern for the retinal

image. Histogram for the enhanced retinal image is calculated.Maximum values occur for the grayish background whilethe vessel corresponds to values a slight greater than thebackground values as they are of bright color. An adaptivethresholding technique is used that selects this point whichseparates the vessels from the rest of image. Vessel segmen-

TABLE IVESSEL SEGMENTATION RESULTS (DRIVE DATABASE)

Segmentation Average StandardMethods Accuracy Deviation

2nd. Observer 0.9473 0.0048Staal et al. 0.9441 0.0079

Soares et al. 0.9466 0.0055Bit plane slicing 0.9303 0.0318

Proposed Method 0.9469 0.0053

TABLE IIVESSEL SEGMENTATION RESULTS (STARE DATABASE)

Segmentation Average StandardMethods Accuracy Deviation

2nd Observer 0.9351 0.0171Hoover et al. 0.9275 0.0247Staal et al. 0.9516 0.0329

Soares et al. 0.9480 0.0298Bit plane slicing 0.9367 0.0311

Proposed Method 0.9502 0.0253

tation mask is created by applying this threshold value (Fig.1).

III. EXPERIMENTAL RESULTS

The tests of proposed technique are performed with respectto the vessel segmentation accuracy using publicly availableDRIVE [32] and STARE databases [33]. The DRIVE databaseconsists of 40 RGB color images of the retina. The imagesare of size 768x584 pixels, eight bits per color channel. TheSTARE database consists of 20 RGB color images of theretina. The images are of size 605x700 pixels, 24 bits perpixel (standard RGB). Both retinal image datasets (DRIVE andSTARE) are divided into a test and training set and each onecontains 20 images. The test set is used for measurement ofperformance of the vessel segmentation algorithms. There aretwo hand-labeling available for the 20 images of test set madeby two different human observers. The manually segmentedimages by 1st human observer are used as ground truth andthe segmentations of set B are tested against set A, serving asa human observer reference for performance comparison truth[28], [31]. The true positive fraction is the fraction of numberof true positive (pixels that actually belong to vessels) and totalnumber of vessel pixels in the retinal image. False positivefraction is calculated by dividing false positives (pixels thatdon’t belong to vessels) by total number of non vessel pixelsin the retinal image. We compared the accuracy of proposedtechnique with the accuracies of the methods of Staal et al.[27] and Soares et al. [31]. Table I summarizes the results ofvessel segmentation for DRIVE database. It shows the resultsin terms of average accuracy and their standard deviation fordifferent Segmentation methods and a second human observer.Average accuracy is the fraction of pixels correctly classified.

Same parameters are calculated for different segmentationmethods using STARE database and are summarized in tableII. Figure 3 and 4 illustrates the blood vessel segmentation

495

results for proposed method against manually segmented setA and set B for both DRIVE and STARE databases.

Fig. 3. Proposed technique results and manual segmentation set B forfour images from the DRIVE database. Row 1: Original retinal images fromdataset; Row 2: Enhanced retinal images using gabor wavelet; Row 3: Invertedthreshold probing results; Row 4: Set B manual segmentation results; Row 5:Segmentation results for proposed technique

IV. CONCLUSION

In this paper, vessel enhancement is done in colored retinalimages by using gabor wavelet and then vessels are probedand segmented using piecewise threshold probing and adap-tive thresholding. We have tested our technique on publiclyavailable DRIVE and STARE databases of manually labeledimages and other standard databases. Experimental resultsshow that our method performs well in enhancing, probingand segmenting the vascular pattern.

REFERENCES

[1] E. J. Susman, W. J. Tsiaras, and K. A. Soper: Diagnosis of diabetic eyedisease. JAMA. 247 (1982) 3231-3234.

[2] S. C. Lee, E. T. Lee, R. M. Kingsley, Y.Wang, D. Russell, R. Klein,and A. Wanr: Comparison of diagnosis of early retinal lesions of diabeticretinopathy between a computer system and human experts. Graefes Arch.Clin. Exp. Ophtalmol. 119 (2001) 509-515.

[3] C. Sinthanayothin, J. F. Boyce, H. L. Cook, and T. H.Williamson:Automated localization of the optic disc, fovea and retinal blood vesselsfrom digital color fundus images. Br. J. Opthalmol. 83 (1999) 231-238.

[4] J. KansKy: Clinical Opthalmology. Butterworh-Heinmann. London.(1994).

[5] M. Foracchia, E. Grisan, and A. Ruggeri: Detection of optic disc inretinal images by means of a geometrical model of vessel structure. IEEEtransactions on medical imaging, 23(10) (2004) 1189-1195.

Fig. 4. Proposed technique results and manual segmentations (set A andset B) for four images from the STARE database. Row 1: Retinal imagesfrom STARE data set; Row 2: Set A manual segmentation results; Row 3:Segmentation results for hoover et. al [28] technique; Row 4: Segmentationresults for proposed technique

[6] N. k. M. N. Subhasis Chaudhuri, Shankar Chatterjee and michael gold-baum: Detection of blood vessels in retinal images using two-dimensionalmatched filters. IEEE transactions on medical imaging. 8(3) (1989) 263-269.

[7] T. Spencer, R. P. Phillips, P. F. Sharp, and J. V. Forrester: Automateddetection and quantification of microaneurysms in fluorescein angiogramsin Graefes Arch. Clin. Exp. Ophtalmol. 230 (1991) 36-41.

[8] A. J. Frame, P. E. Undill, M. J. Cree, J. A. Olson, K. C. McHardy,P. F. Sharp, and J. F. Forrester: A comparison of computer basedclassification methods applied to the detection of microaneurysms inophtalmic fluorescein angiograms. Comput. Biol. Med. 28 (1998) 225-238.

[9] A. Osareh, M. Mirmehdi, B. Thomas, and R. Markham: Automaticrecognition of exudative maculopathy using fuzzy c-means clusteringand neural networks. Proc. Medical Image Understanding Analysis Conf.(2001) 49-52.

[10] R. Phillips, J. Forrester, and P. Sharp: Automated detection and quan-tification of retinal exudates. Graefes Arch. Clin. Exp. Ophtalmol. 231(1993) 90-94.

[11] T. Teng, M. Lefley and D. Claremont: Progress towards automateddiabetic ocular screening: A review of image analysis and intelligentsystems for diabetic retinopathy. Med. Biol. Eng. Comput. 40 (2002)2-13.

[12] A. Pinz, S. Bernogger, P. Datlinger and A. Kruger: Mapping the humanretina. IEEE Trans. Med. Imag. 17(4) (1998) 606-619.

[13] C.L. Tsai, C. V. Stewart, H. L. Tanenbaum and B. Roysam: Modelbasedmethod for improving the accuracy and repeatability of estimating vascu-lar bifurcations and crossovers from retinal fundus images. IEEE Trans.Inf. Technol. Biomed. 8(2) (2004) 122-130.

[14] H. Li and O. Chutatape: Automated feature extraction in color retinalimages by a model based Approach. IEEE Trans. Biomed. Eng. 51(2)(2004) 246-254.

[15] A. M. Mendona, A. J. Campilho. Segmentation of Retinal BloodVessels by Combining the Detection of Centerlines and MorphologicalReconstruction. IEEE Trans. Med. Imag. 25(9) (2006) 1200-1213.

[16] I. Liu and Y. Sun. Recursive tracking of vascular networks in angiogramsbased on the detection-deletion scheme. IEEE Trans. Med. Imag. 12(2)(1993) 334-341.

[17] L. Zhou, M. S. Rzeszotarski, L. J. Singerman, and J. M. Chokreff: Thedetection and quantification of retinopathy using digital angiograms. IEEETrans. Med. Imag. 13(4) (1994) 619-626.

[18] O. Chutatape, L. Zheng, and S. M. Krishnan: Retinal blood vesseldetection and tracking by matched Gaussian and Kalman filters. Proc.

496

20th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBS98), 20 (1998)3144-3149.

[19] Y. A. Tolias and S. M. Panas: A fuzzy vessel tracking algorithm forretinal images based on fuzzy clustering. IEEE Trans. Med. Imag. 17(2)(1998) 263-273.

[20] A. Can, H. Shen, J. N. Turner, H. L. Tanenbaum, and B. Roysam: Rapidautomated tracing and feature extraction from retinal fundus images usingdirect exploratory algorithms. IEEE Trans. Inf. Technol. Biomed. 3(2)(1999) 125-138.

[21] M. Lalonde, L. Gagnon, and M.-C. Boucher: Non-recursive pairedtracking for vessel extraction from retinal images. Vision Interface. (2000)61-68.

[22] T. McInerney and D. Terzopoulos: T-snakes: Topology adaptive snakes.Med. Image Anal. 4 (2000) 73-91.

[23] R. Toledo, X. Orriols, X. Binefa, P. Radeva, J. Vitri and J. Villanueva:Tracking of elongated structures using statistical snakes. IEEE Comput.Soc. Conf. Comput. Vision Pattern Recog. (CVPR). (2000).

[24] A. Vasilevskiy and K. Siddiqi: Flux maximizing geometric flows. IEEETrans. Pattern Anal. Mach. Intell. 24(12) (2002) 1565-1578.

[25] D. Nain, A. Yezzi, and G. Turk: Vessel segmentation using a shapedriven flow. Med. Image Comput. Comput.-Assist. Intervention MICCAI.(2004) 51-59.

[26] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum:Detection of blood vessels in retinal images using two-dimensionalmatched filters. IEEE Trans. Med. Imag. (1989) 263-269.

[27] J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever and B.van Ginneken: Ridge-based vessel segmentation in color images of theretina.IEEE Trans. Med. Imag. 23(4) (2004) 501-509.

[28] A. Hoover, V. Kouznetsova and M. Goldbaum: Locating blood vesselsin retinal images by piecewise threshold probing of a matched filterresponse. IEEE Trans. Med. Imag. 19(3) (200) 203-211.

[29] A. Arnodo, N. Decoster and S. G. Roux: A wavelet-based methodfor multifractal image analysis. I. Methodology and test applications onisotropic and anisotropic random rough surfaces. Eur. Phys. J. A. 15(2000) 567-600.

[30] J.P. Antoine, P. Carette, R. Murenzi, and B. Piette: Image analysis withtwo-dimensional continuous wavelet transform. Signal Process. 31 (1993)241-272.

[31] J. V. B. Soares, J. J. G. Leandro, R. M. Cesar, H. F. Jelinek and M.J. Cree. Retinal vessel segmentation using the 2-D gabor wavelet andsupervised classification. IEEE Trans. on Med. Imag. 25(9) (2006) 1214-1222.

[32] Niemeijer and B. van Ginneken: 2002 [Online]. Available:http://www.isi.uu.nl/Reseach/Databases/DRIVE/results.php

[33] Hoover: STARE database [Online]. Availablehttp://www.ces.clemson.edu/ ahoover/stare

497