tbme-01423-2014 1 iterative vessel segmentation of … ieee papers/matlab(image... · iterative...

0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TBME.2015.2403295, IEEE Transactions on Biomedical Engineering

TBME-01423-2014 1

Iterative Vessel Segmentation of Fundus ImagesSohini Roychowdhury, Member, IEEE, Dara D. Koozekanani, Member, IEEE and Keshab K. Parhi, Fellow, IEEE

Abstract—This paper presents a novel unsupervised iterativeblood vessel segmentation algorithm using fundus images. First,a vessel enhanced image is generated by tophat reconstructionof the negative green plane image. An initial estimate of thesegmented vasculature is extracted by global thresholding thevessel enhanced image. Next, new vessel pixels are identifiediteratively by adaptive thresholding of the residual image gen-erated by masking out the existing segmented vessel estimatefrom the vessel enhanced image. The new vessel pixels are thenregion grown into the existing vessel, thereby resulting in aniterative enhancement of the segmented vessel structure. As theiterations progress, the number of false edge pixels identified asnew vessel pixels increases compared to the number of actualvessel pixels. A key contribution of this paper is a novel stoppingcriterion that terminates the iterative process leading to highervessel segmentation accuracy. This iterative algorithm is robustto the rate of new vessel pixel addition since it achieves 93.2%-95.35% vessel segmentation accuracy with 0.9577-0.9638 areaunder ROC curve (AUC) on abnormal retinal images from theSTARE data set. The proposed algorithm is computationallyefficient and consistent in vessel segmentation performancefor retinal images with variations due to pathology, unevenillumination, pigmentation and fields of view since it achievesa vessel segmentation accuracy of about 95% in an averagetime of 2.45 seconds, 3.95 seconds and 8 seconds on imagesfrom three public data sets DRIVE, STARE, and CHASE DB1,respectively. Additionally, the proposed algorithm has more than90% segmentation accuracy for segmenting peripapillary bloodvessels in the images from the DRIVE and CHASE DB1 datasets.

Index Terms—Vessel segmentation, fundus image, morpholog-ical reconstruction, stopping criterion, accuracy, computationalcomplexity, iterative algorithm.

I. INTRODUCTION

Retinal blood vessel (vasculature) segmentation using fun-dus photographs has played a vital role in assessing theseverity of retinal pathologies that can lead to acquiredblindness such as retinopathy of prematurity [1], glaucoma,vein occlusions and diabetic retinopathy (DR) [2]. Accordingto research statistics in the year 2011, retinal pathologiessuch as DR affect over 4.2 million Americans each year,while glaucoma affects about 2.3 million Americans annually[3]. Automated blood vessel segmentation algorithms canbe very useful in screening patients that are affected bysuch retinal complications and require follow-up [4]. Also,

Manuscript received November 20, 2014; revised December 24, 2014;accepted January 19, 2015. This research was supported in part by a grantfrom the Institute for Engineering and Medicine at the University of Minnesotaand in part by an unrestricted departmental grant from Research to PreventBlindness Inc, New York, NY.

Sohini Roychowdhury (email: [email protected]) and Keshab K. Parhi (email:[email protected]) are with the Department of Electrical and Computer En-gineering. Dara D. Koozekanani (email: [email protected]) is with theDepartment of Ophthalmology. Keshab K. Parhi and Dara D. Koozekanani areaffiliated to University of Minnesota and Sohini Roychowdhury is affiliatedwith the University of Washington, Bothell.

automated blood vessel segmentation systems can be usefulin determining variations in the blood vessels based on thevessel branching patterns, vessel width, tortuosity and vesseldensity as the pathology progresses in patients [5]. Suchevaluations will help to enhance the resourcefulness of thepresent-day retinal therapeutics and guide research towardsanalyzing patients for hypertension [6], variability in retinalvessel diameters due to a history of cold hands and feet [7],and flicker responses [8]. Some existing automated detectionsystems for non-proliferative DR detection, such as [9] [10],require masking of the blood vessels to ensure that the bloodvessels are not mistaken for red lesions that are caused byDR. Additionally, proliferative DR detection systems [11]require analysis of the vessel parameters for detecting neovas-cularization, vessel beading and intra retinal micro-vascularabnormalities (IRMA). Thus, an accurate vessel segmentationalgorithm that is robust to image variability and that has lowcomputational complexity is desirable for such automated real-time detection and screening systems.

All existing algorithms for automated segmentation of bloodvessels using fundus images can be broadly categorized as su-pervised and unsupervised methods. A comprehensive surveyon existing retinal vessel segmentation algorithms and publiclyavailable data sets has been presented in [12]. Also, a com-parative analysis of the two categories of vessel segmentationalgorithms has been presented in [13]. In the supervised cate-gory of algorithms, classifiers such as the k-Nearest Neighbor[14], Gaussian Mixture Model (GMM) [15], support vectormachine (SVM) [16], neural networks [17], decision trees [18]and AdaBoost [19] have been applied to classify vessel pixelsfrom the non-vessels. The unsupervised algorithms mostlyapply matched filtering [20], line detectors [21], morphologicaltransformations [22] [23] [24], model-based methods [25] [26][27], or multi-scale vessel segmentation methods [28] [29][30]. While most supervised vessel classification methods aredependent on the training data and sensitive to false edges, theexisting unsupervised methods are computationally complexand hence they are not viable for real-time portable DRscreening systems such as [9]. Most of the existing approachessuch as the ones in [15], [21], and [26] perform well on healthyretinal images but have low segmentation accuracy in imageswith pathology. Thus, there is a need for a general method withlow computational complexity and high segmentation accuracyfor normal as well as pathological fundus images. In thiswork, we propose an iterative vessel segmentation algorithmthat segments the major vessels first, followed by additionof finer vessel branches by adaptive thresholding in iterativesteps. This iterative approach has high segmentation accuracyfor vasculature in normal and abnormal retinal images. Also,the proposed method has lower computational complexity thanmost existing supervised and unsupervised approaches.



TBME-01423-2014 2

Key contributions of the paper include a novel iterativeblood vessel segmentation algorithm, where vasculature es-timates are identified by adding new pixels iteratively usingadaptive global thresholding, and a novel stopping criterionto terminate the iterative vessel addition process therebyreducing the number of false edge pixels in the segmentedimage. The proposed vessel segmentation algorithm is robustto both image variability and inter-observer variability in theground-truth. For healthy and pathological images from theSTARE data set [20], the proposed method achieves an averagesegmentation accuracy in the range of 93.32%-95.6% forthree sets of manually marked ground-truth. The analysis ofautomated vessel segmentation performance with respect tomultiple ground-truth sets has not been presented in any ofthe existing works so far. Additionally, we identify vesselsegmentation accuracy in the peripapillary region as anotherimportant metric; this metric is well suited for images withabnormalities such as proliferative DR, glaucoma and retinalocclusions. The peripapillary blood vessels contain vesselswithin 1-optic disc diameter centered at the papilla [11]. Itis shown that the proposed approach provides robust and highvessel segmentation accuracy in the peripapillary region whencompared to well-known existing algorithms.

The organization of this paper is as follows. The proposedmethod and materials are described in Section II. In sectionIII, the experimental results are presented and compared withexisting methods, for normal and abnormal retinal images,and for peripapillary vessel extraction. Finally, in Section IV,discussion and significant conclusions are presented.

II. METHOD AND MATERIALS

The principal idea behind iterative vessel segmentationis that in a vessel enhanced image, the bright and largevessels overshadow the thinner fine vessel branches. In sucha situation, global thresholding would extract only the largeprominent vessels while the finer vessel branches would re-main unsegmented. Thus, to include these fine vessels intothe segmented vasculature estimate, iterative adaptive globalthresholding is proposed.

For every color fundus photograph, a vessel enhanced imageis obtained by morphological tophat reconstruction of thenegative green plane image. The brightest pixels from thisvessel enhanced image are extracted as the major portionsof the vasculature, or major vessels. A residual image isthen generated by masking these major vessels from thevessel enhanced image followed by contrast adjustment andadaptive thresholding in an iterative manner to uncover newvessel pixels and add them to the existing vessel estimate.This method of iterative vessel addition and a novel stoppingcriterion for estimating the best segmented vasculature ispresented in this section. The segmentation performance of theproposed algorithm is evaluated using three publicly availabledata sets with different resolutions and FOVs. The proposedalgorithm has been implemented in MATLAB environment ona Laptop with Intel Core i3 processor, 2.6GHz and 2GB RAM.

A. Data

The following data sets have been manually annotated forthe blood vessel regions for analyzing the performance ofblood vessel segmentation algorithms.

• STARE [20] data set contains 20 images with 35o FOV ofsize [605x700] pixels that are manually annotated by twoindependent human observers. Here, 10 images representpatients with retinal abnormalities (STARE Abnormal).The other 10 images represent normal retina (STARENormal).

• DRIVE [31] data set contains 40 images with 45o FOVof size [584x565] pixels. This data set is separated by itsauthors into a training set (DRIVE Train) and a test set(DRIVE Test) with 20 images in each set. The DRIVETrain set of images are annotated by one human observerwhile the DRIVE Test data set is annotated by twoindependent human observers.

• CHASE DB1 [32] data set contains 28 images with30o FOV of size [960x999] pixels corresponding to twoimages per patient (one image per eye) for 14 children.Each image is annotated by two independent humanobservers [18].

The automated vessel segmentation performance is evaluatedon the images from DRIVE Test, STARE and CHASE DB1data sets with respect to three sets of manually marked ground-truths. The ground-truth O1 and O2 are referred to the manualannotations by first and second human observer, respectively.The annotations by the second human observer contain morefine vessels than the first human observer. Additionally, toanalyze inter-observer variability, ground-truth O3 = O1 ∪O2

is formed as the union of all regions in the ground-truth imagesof O1 and O2.

B. Proposed Method

As a pre-processing stage, the green plane of each fundusimage is scaled in [0,1] (I). A fundus mask (g) is utilized toremove the dark background region from the photographs tofocus attention on the retinal region only. The fundus masksfor DRIVE, STARE and CHASE DB1 data sets are generatedas centered elliptical regions with diameters of [521x520],[550x650] and [860x869] pixels, respectively.

In the scaled green plane image (I), the red regions corre-sponding to the blood vessel segments appear as dark pixelswith intensities close to 0. To focus attention on the bloodvessel regions, image I is inverted to make the red regionsappear the brightest, followed by superposition of the mask g,resulting in image Iv . Iv is then subjected to contrast enhance-ment followed by morphological tophat transformation. 12linear structuring elements each of length 21 pixels and 1 pixelwidth and angles incremented by 15o from 0 through 180o

are used to generate tophat reconstructions from Iv [22] [18].The length of 21 pixels for the linear structuring element ischosen to approximately fit the diameter of the biggest vesselsin the images [18]. For each pixel location, the reconstructedpixel with the highest intensity is selected, thereby resultingin tophat reconstructed vessel enhanced image T .



TBME-01423-2014 3

The major vessels (V0) are extracted by thresholding imageT for pixels greater than ‘p’ : p ∈ [0, 1], and retaining theregions with area greater than 200 pixels. An optimal value of‘p = 0.5’ is selected to minimize error in the final segmentedvessel. The process of optimally estimating the major vesselsis presented in Section II-C. The major vessels (V0) is thesegmented vessel estimate (Vt) at iteration t = 0. Next, thefollowing steps are iteratively performed:

For each iteration t, the pixels from the existing vesselestimate Vt are removed from image T and the remainingimage is contrast enhanced resulting in residual image Rt.This image Rt is thresholded at pixel value φ1(t) to extract abinary image VRt

containing new vessel regions that have anarea of at least 10 pixels. Here, the pixel threshold value φ1(t)varies with ‘t’, the iteration number. However, it is desirablefor the segmented vessel to have a continuous structure andhence it is imperative to fill any gaps or holes between theexisting vessel estimate Vt and the newly identified vesselpixels in VRt

. Thus, the pixels in Vt and VRtare added to

vessel enhanced image T , followed by rescaling the pixels in[0,255], resulting in base image Bt. This image Bt is thenregion-grown with a threshold pixel value (φ2(t)) that is afunction of the iteration number. The image obtained at theend of the regiongrow operation [33] is the new iterative vesselestimate (Vt+1).

The iterative addition of the newly identified vessel regionsto the existing vessel estimate is continued till a stoppingcriterion is met. The notations and their definitions used inthis iterative algorithm are defined in Table I.

TABLE IDEFINITION OF NOTATION.

Notation MeaningV Manually marked vessel image.Vf Final segmented vasculature.T Tophat transformed vessel enhanced image.p Threshold pixel intensity in [0,1] to extract major vessels.Vt Binary image representing vasculature estimated in iteration t.

V0 represents the major vessels at t = 0.Rt Residual image obtained by removing the pixels in the

existing vessel estimate (Vt) from T .VRt Thresholded residual image from Rt using threshold φ1(t).

This contains the new vessel pixels identified in iteration t.Bt Base image to fill gaps between the new vessel pixels and

the existing vessel estimate. It is the union of vessel pixelsin VRt and Vt added to image T .

Vt+1 Vasculature at iteration t+ 1 obtained from region grownbase image (Bt) with an adaptive threshold φ2(t).

Ct Change in segmented vessel estimate in iteration ‘t’ definedas the difference between number of pixels in Vt and V0divided by the total number of image pixels.

Cmt For m = [1, 2, 3], the first, second and third order derivatives

of Ct computed using backward difference method.Et Error in number of vessel pixels between Vt and manually

marked vessels V divided by the total number of pixels.Em

t For m = [1, 2, 3], the first, second and third order derivativesof Et computed using backward difference method.

Qt Quality of the segmented vessel estimate Vt in iteration‘t’ defined as the ratio between the number of actualvessel pixels and the number of false edge pixels.

For an image from the DRIVE data set, the iterativevessel estimates are shown in Fig. 1 for 10 iterations, i.e.,

t = [1, 2...10]. The proposed iterative segmentation algorithmis given in Algorithm 1. For the example shown in Fig. 1, thebest segmented vasculature is estimated at iteration t = 4.

For post-processing, the regions in the best vessel estimatewith area less than 20 pixels are discarded, and the remainingvessel regions are median filtered with a [3x3] filter.

Algorithm 1 Iterative Vessel Segmentation (I, g)Input: I ∈ [0, 1], g ∈ [0, 1]Output: VfPre-Processing:

Iv ← contrast adjust((1− I) g)T ← tophat reconstruction(Iv)V0 ← [(T > p), and area > 200]

Iterations:t = 0, stop← ‘No’, stop ∈ [‘Yes’,‘No’],

while [stop← ‘No’ ] do∀(i, j), Rt(i, j) = [T (i, j)− Vt(i, j)]|Rt(i, j) ≥ 0.VRt← [Rt > φ1(t), and area > 10]

Bt ← scale (VRt∪ Vt+ T ∈ [0, 255])

Vt+1 ← regiongrow(Bt, seed=255, threshold=φ2(t))t← t+ 1stop ← Stopping criterion (Vt, Vt−1)

end whiletf ← tVf ← Vtfreturn Vf

It is important to note that Algorithm 1 relies on twothreshold parameters that impact the accuracy of the finalvessel estimate. These two threshold parameters φ1(t) andφ2(t) affect the quality of new vessel pixels that are addedto the existing vessel estimate. The other important feature inthe iterative vessel segmentation algorithm is the estimation ofthe iteration number (tf ) at which the best vessel estimate withthe highest segmentation accuracy occurs. To determine tf , astopping criterion is introduced. The methods for estimatingthe adaptive thresholds φ1(t) and φ2(t), and stopping criterionare given below.

1) Thresholds for vessel addition: In each iteration step‘t’, new vessels are identified from the iterative residualimage in VRt

using pixel threshold φ1(t). Here, we initializeφ1(t) = 1−(0.05∗t). As the algorithm progresses to combinethe new vessel pixels (VRt ) to the existing vessel estimate(Vt), a regiongrow operation [33] is performed by scaling theimage (VRt

∪ Vt) + T in [0,255], using a seed pixel value of255, and a threshold pixel value φ2(t). The threshold functionφ2(t) that would result in significantly high accuracy for vesselsegmentation is obtained by one-time training using the 20images from the DRIVE Train data set. We consider φ2(t)given by:

φ2(t) = 205+α∗(t−1)k, α ∈ [1, 2...6], k ∈ [0, 0.2, ...3] (1)

Each of the 20 images were subjected to vessel segmenta-tion for all the 6x16 combinations of ‘α’ and ‘k’ values in (1)and [t = 1, 2..10] iterations. For every image (l = [1...20]),



TBME-01423-2014 4

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j)

Fig. 1. The iterative vessel segmentation algorithm on an image with 45o

FOV. (a) Green plane image (I). (b) Vessel enhanced image by tophatreconstruction (T ). (c) Major vessels extracted from T (V0). (d) Residualimage R0 with pixels from V0 removed from image T . (e) New vesselpixels identified by thresholding R0 (VR0

). (f) Base image B0 obtained bycombining pixels in VR0 and V0 on image T . (g) V1 extracted after region-growing. (h) Final vasculature estimate obtained after 4 iterations (Vf ). (i)Manually marked vasculature (V ). (j) Vessel estimates extracted after eachiteration t = 3 to t = 10 by repeating steps (d) to (g) iteratively. A stoppingcriterion is required to stop the iteration at t = 4 to prevent over-segmentation.

and for each combination of [α, k], the highest segmentationaccuracy (ACCl) was achieved at iteration number tl. Themean ACCl and the corresponding mean tl versus [α, k] areshown in Fig. 2 (a) and Fig. 2 (b), respectively. Since, alow mean iteration number implies higher segmentation speedfor the algorithm, the goal was to select a threshold functionthat generates the highest mean vessel segmentation accuracyin less iterations. We observe that, the highest mean ACCl

occurs for α = 1, k = 1.8, with the mean iteration numbertl = 4. Thus, the threshold function φ2(t) for Algorithm 1was estimated as φ2(t) = 205 + (t− 1)1.8.

2) Stopping criterion design: In Fig. 1, we observe that thebest vessel estimate occurs at some iteration between t = 0to t = 10. Also, Algorithm 1 requires a stopping criterion tostop the iterative vessel addition process at iteration numbertf where a segmentation vasculature with highest accuracyexists. This stopping criterion can be determined by analyzingthe quality of segmented vessel estimates in every iteration.

The initial vessel segment V0 contains the major portions of

(a) (b)

Fig. 2. Estimation of threshold function φ2(t) for region-growing. (a) Thehighest mean vessel segmentation accuracy (ACCl) versus the thresholdfunction parameters [α, k]. (b) The mean iteration number (tl) correspondingto highest ACCl versus threshold function parameters [α, k] on the DRIVETrain set of images.

the blood vessels and hence it is free from false edge pixels.In each iteration, new pixels are added to the existing bloodvessels. While actual vessel pixels are added in each iteration,pixels corresponding to false edges also get introduced intothe segmented vessel structure. Using the manually markedvessels V and the iterative vessel estimate Vt in each iterationt, the criteria for any pixel (i, j) to be a true positive (tpt),true negative (tnt), false positive (fpt) and false negative (fnt)are defined in (2)-(5). The total number of pixels that are truepositives (TPt), true negatives (TNt), false positives (FPt)and false negatives (FNt) in each image are then computedusing (6)-(9).

For the binary image Vt, in iteration t we define: (2)∀(i, j), i ∈ [1, 2...n1], j ∈ [1, 2....n2],

tpt(i, j) = 1, if, [V (i, j) = 1, Vt(i, j) = 1].

tnt(i, j) = 1, if, [V (i, j) = 0, Vt(i, j) = 0]. (3)fpt(i, j) = 1, if, [V (i, j) = 0, Vt(i, j) = 1]. (4)fnt(i, j) = 1, if, [V (i, j) = 1, Vt(i, j) = 0]. (5)

n1∑i=1

n2∑j=1

tpt(i, j) = TPt,

n1∑i=1

n2∑j=1

tnt(i, j) = TNt, (6)

n1∑i=1

n2∑j=1

fpt(i, j) = FPt,

n1∑i=1

n2∑j=1

fnt(i, j) = FNt. (7)

The sum of all pixels in binary image Vt representsthe total number of TPt and FPt pixels in Vt.

⇒n1∑i=1

n2∑j=1

Vt(i, j) = TPt + FPt. (8)

Also, n1 · n2 = TPt + TNt + FPt + FNt. (9)

For a fundus image with [n1 · n2] pixels, the metricsidentifying the quality of segmented vessel (Qt), the changein vessel estimate Ct and error in the vessel estimate Et, aredefined in (10) - (13). In (12), we observe that Et has twonon-zero components, one due to the FNt pixels and theother due to FPt pixels. The vessel segmentation accuracyACCt increases as Et decreases as described by (13). Thus,the iteration tf corresponding to minimum Etf will ensure asegmented vessel estimate with highest accuracy.



TBME-01423-2014 5

Define :Qt =

∑n1

i=1

∑n2

j=1[tpt(i, j)]∑n1

i=1

∑n2

j=1[fpt(i, j)]=TPt

FPt. (10)

Ct =1

n1 · n2

n1∑i=1

n2∑j=1

Vt(i, j)− V0(i, j) . (11)

Et =1

n1 · n2

n1∑i=1

n2∑j=1

|V (i, j)− Vt(i, j)|. (12)

⇒ Et =

∑n1

i=1

∑n2

j=1[fpt(i, j) + fnt(i, j)]

TPt + TNt + FPt + FNt= 1−ACCt. (13)

Intuitively, Ct can be modeled as an odd-degree polynomial,and Et can be modeled as an even-degree polynomial. Thegoal of segmentation is to stop at the iteration with least errorto maximize the vessel segmentation accuracy. However, Et

requires knowledge of the manual annotated image and is notknown. Therefore, Et cannot be used as a parameter in thestopping criteria. However, an analysis of the trends in Ct andEt based on known images demonstrates a correlation betweenthese parameters. This allows us to derive a stopping criterionusing Ct. Later in the section we propose to stop the iterationswhen the first three derivatives of Ct are non-negative. Thisis illustrated in Fig. 3.

Fig. 3. The stopping criterion for the iterative algorithm. If the number ofiteration ‘t’ is less than 3, or if the sign of the C1

t , C2t , C

3t are not all non-

negative, then iterations are continued.

For trend analysis of the curves Ct and Et, a theoreticalexample is shown in Fig. 4, where curves Cx and Ex are oddand even functions of the same repeated root, respectively.Ex is a function of 4th degree repeated root, and hence weanalyze the trends of the first three derivatives of Ex and Cx

computed using backward difference rule as ‘x’ varies from[1 to 20] in intervals of 1. We observe that although Ex andCx are both functions of repeated roots at x = 10.5, the firstthree derivatives of both Ex and Cx become non-negative, i.e.,[E1

x, E2x, E

3x] ≥ 0, [C1

x, C2x, C

3x] ≥ 0 for [x ≥ 12]. The region

from x = 9 to x = 12 has very small variations in Cx and Ex.Also, in Fig. 4, three significant regions exist correspondingto high, medium and low quality of segmented vessel pixels(Q). The region with high Q corresponds to the early iterationswhen the number of actual vessel pixels identified are more

than the noisy false edge pixels (TPt >> FPt). The regionwith low Q corresponds to iterations when the number offalse edge pixels being added exceed the actual vessel pixels(TPt << FPt). The third significant region is when Q ismedium and change in iterative vessel estimates is significantlysmall. The curve trends in regions when Q is high and Q islow are analyzed in (14)-(16). The curve Ct always has anon-decreasing trend (15).

Ct+1 − Ct =1

n1 · n2

n1∑i=1

n2∑j=1

Vt+1(i, j)− Vt(i, j) , (14)

Since new vessel pixels are added in each iteration,n1∑i=1

n2∑j=1

Vt+1(i, j) ≥n1∑i=1

n2∑j=1

Vt(i, j)⇒ Ct+1 − Ct ≥ 0. (15)

⇒ Ct always has non− decreasing trend.

Now, Et+1 − Et =1

n1 · n2

n1∑i=1

n2∑j=1

[|V (i, j)− Vt+1(i, j)|

−|V (i, j)− Vt(i, j)|],

⇒ Using (12), Et+1 − Et =1

n1 · n2

n1∑i=1

n2∑j=1

[fpt+1(i, j)

+fnt+1(i, j)− fpt(i, j)− fnt(i, j)]. (16)

However, for Et, we consider two separate cases to analyzethe trends of Et in high Q and low Q regions. In case 1with high Q, TPt >> FPt. Here, the number of false vesselpixels in Vt (FPt) are very small compared to the number oftrue vessel pixels (TPt) added in consecutive iterations (17).Also, for each image, the sum of vessel pixels in the manuallymarked image remains constant, i.e., TPt +FNt = constant(18). Hence in the high Q region, Et has a non-increasingtrend (19). In case 2 with low Q, TPt << FPt. Here, thenumber of false vessel pixels in Vt (FPt) are significantlylarger than the true vessel pixels (TPt) (20). We observe thatEt has a non-decreasing trend in this region (21).

Case 1: High Q, [TPt >> FPt].

Iteratively, TPt+1 ≥ TPt, FPt+1 ≈ FPt. (17)

⇒ TPt + FNt =

n1∑i=1

n2∑j=1

V (i, j) = constant. (18)

Using, (17), (18), FNt+1 ≤ FNt.

From (16), ⇒ Et+1 − Et ≤ 0.

⇒ Et has non− increasing trend. (19)Case 2: Low Q, [TPt << FPt].

Iteratively, FPt+1 ≥ FPt, TPt+1 ≈ TPt (20)⇒ FNt+1 ≈ FNt.

From (16), ⇒ Et+1 − Et ≥ 0.

⇒ Et has non− decreasing trend. (21)

From the experimental standpoint, we observe that for allthe test images the best fit polynomial for Ct is of degree 3and the best fit polynomial for Et is of degree 4. Thus, the



TBME-01423-2014 6

Fig. 4. Theoretical example for curves corresponding to the iterative change in blood vessels Cx and iterative error incurred Ex. The repeated root for curvesEx, Cx occurs in the region with medium Q. As iterations proceed beyond the repeated root, the first three derivatives of Ex and Cx become non-negative.

first, second and third derivatives must exist for both curvesCt and Et. Let these derivatives be denoted by Cm

t and Emt ,

for m = [1, 2, 3], respectively. These derivatives are computediteratively using the backward difference rule as shown in (22).

If, C0t = Ct, E

0t = Et, then, (22)

Emt = Em−1

t − Em−1t−1 ,m = [1, 2, 3].

Cmt = Cm−1

t − Cm−1t−1 ,m = [1, 2, 3].

Further, from Fig. 4, we observe that the iterative changein Ct and Et are significantly small in the medium Q regionthan the high and low Q regions. This is intuitive since inthe medium Q region, fine vessel branches are detected andadded to the existing vessel estimate. However, as soon asthe noisy false edge regions, that are significantly larger inarea when compared to the fine vessel branches, start gettingidentified and added to the vessel estimates, the change in Ct

and Et become considerably large, and this is when the low Qregion begins. Based on this observation, a stopping criterionis designed for Algorithm 1 using the following property.

Property 1. Assume Ct, Et exist for t = [1, 2...10..], and,they can be approximated as Ct ≈ Λ0(t − t∗)3 + Λ1, andEt ≈ Ω0(t − t∗)4 + Ω1, where Λ0,Λ1,Ω0,Ω1 are constants.If the region with medium Q exists for at least 3 consecutiveiterative steps (i.e., the change in the iterative vessel estimatesare very small for at least 3 iterations), then the iteration tfat which the first instance of [C1

tf≥ 0, C2

tf≥ 0, C3

tf≥ 0]

occurs will result in the best segmented vessel estimate.

Proof. Let the region medium Q exist for at least 3 consecutiveiterations starting at iteration t∗ such that beyond the (t∗ +3)th iteration the quality of pixel addition deteriorates, i.e.,

the number false positive pixels added as vessels increases.Then,

n1∑i=1

n2∑j=1

Vt∗(i, j) ≈n1∑i

n2∑j=1

Vt∗+1(i, j), (23)

n1∑i=1

n2∑j=1

Vt∗+1(i, j) ≈n1∑i=1

n2∑j=1

Vt∗+2(i, j),

n1∑i=1

n2∑j=1

Vt∗+3(i, j) ≥n1∑i=1

n2∑j=1

Vt∗+2(i, j).

While transitioning from medium Q to low Q region,both Ct and Et have non-decreasing trends.

⇒ Ct∗ ≈ Ct∗+1 ≈ Ct∗+2, Ct∗+3 ≥ Ct∗+2.

⇒ Et∗ ≈ Et∗+1 ≈ Et∗+2, Et∗+3 ≥ Et∗+2. (24)

Using (24), and (22), we have,

⇒ C1t∗+1 = 0, C1

t∗+2 = 0, C1t∗+3 ≥ 0, (25)

⇒ E1t∗+1 = 0, E1

t∗+2 = 0, E1t∗+3 ≥ 0.

⇒ C2t∗+2 = 0, C2

t∗+3 ≥ 0, (26)⇒ E2

t∗+2 = 0, E2t∗+3 ≥ 0.

⇒ C3t∗+3 ≥ 0, E3

t∗+3 ≥ 0. (27)

Combining (24)-(27) at t∗ + 3, we have,

[C1t∗+3 ≥ 0, C2

t∗+3 ≥ 0, C3t∗+3 ≥ 0], (28)

[E1t∗+3 ≥ 0, E2

t∗+3 ≥ 0, E3t∗+3 ≥ 0].

For iterations t > t∗ + 3, the number of false edge pixels thatare identified and added becomes higher than the number of



TBME-01423-2014 7

actual vessel pixels, which in turn reduces the accuracy of thesegmented vessel estimate. Thus, for Algorithm 1, tf = t∗ +3is the iteration at which the best segmented vasculature withhighest accuracy can be estimated.

Even though our assumption that Ct and Et have repeatedroot at same iteration number (with constant terms removed)does not hold, the proposed stopping criterion still workswell. In general, Ct behaves like an odd-degree polynomial.Thus, it may behave like a 5th-degree polynomial instead of3rd. This may require stopping at the iteration where first 5derivatives are non-negative. However, the proposed stoppingcriterion still leads to high accuracy due to two reasons. First,Ct is approximately constant in the entire medium Q region.Constant Ct over few iterations means fewer new vesselsare added. Second, the medium Q region often spans a fewiterations. Thus, stopping at any iteration in this region leadsto similar accuracy. This is illustrated using sample imageschosen from the datasets: DRIVE Test and STARE in Fig. 5.In Fig. 5 (a), the three consecutive derivatives of Ct and Et

become non-negative at the same iteration number; however,in Fig. 5 (b), Ct and Et do not represent polynomials for thesame repeated root, and hence, the first three derivatives of Ct

become non-negative in iteration t = 4, while the derivativesof Et become non-negative at t = 6. However, since iterationst = [4, 5, 6] lie in the medium Q region, where the change initerative vessel estimates is very small, the accuracy of thesegmented vessel in iteration t = 4 is similar to the accuracyof the vessel estimate at t = 6.

Based on (28) from Property 1, the stopping criterion forAlgorithm 1 is defined as the iteration when the first threederivatives of curve Ct become non-negative. It is important tonote that this iterative process continues for at least 3 iterationsto obtain the first three derivatives for Ct.

C. Limiting Conditions

The proposed vessel segmentation algorithm is limited inits performance of estimating fundus vasculature with highaccuracy by one primary and two secondary constraints. Theprimary constraint is the major vessel estimate (V0) at thebeginning of the iterative process. The two secondary con-straints govern the rate at which new vessels are iterativelyadded using functions φ1(t) and φ2(t).

First, we analyze the error propagation and limiting con-ditions posed by the major vessel estimate. The proposediterative segmentation process is based on the inherent as-sumption that V0 has no false positives. However, if a largenumber of false positive pixels are introduced in major vessels,then the error in segmented vasculature propagates throughthe iterations, resulting in low segmentation specificity andaccuracy of the final vessel estimate. The estimation of themajor vessels is dependent on the thresholding pixel intensityvalue p ∈ [0, 1]’ for extracting the major vessels from image T ,i.e., V0 = [T > p, area > 200]. If ‘p’ is set to a very low pixelintensity, then most of the regions in the image are estimatedas the major vessels, thereby introducing more false positivesthan true positive pixels in V0 as shown in (29). In suchsituations, the trend of curve Et in (16) becomes consistently

non-decreasing in (30). This implies that the iterative processwill only decrease the accuracy of the final segmented vesselestimate when compared to the major vessel estimate (sinceEt = 1 − ACCt). However, if ‘p’ is set to a very high pixelintensity, then very few regions are estimated as major vessels,and almost no false positives are introduced in the estimateV0. As the iterative process begins, the first few iterations addmajor vessel pixels followed by fine vessel pixels added insubsequent iterations. Hence, the number of iterations beforethe stopping criterion is met increases. Thus, a low value of‘p’ introduces large errors in the segmented vessel estimatewhile a large value of ‘p’ increases the time complexityof the process. Further analysis shows that for all the testdata sets, p ∈ [0.4 − 0.7] ensures maximum accuracy in thefinal segmented vasculature without significant increase in thenumber of iterations.

If FP0 >> TP0 >> FN0, (29)

Et =1

n1 · n2[FPt+1 − FPt + FNt+1 − FNt].

Since FPt+1 − FPt >> FNt+1 − FNt, (30)Et+1 − Et ≥ 0,Et has non− decreasing trend.

Next, we analyze the limiting conditions due to the rate ofnew vessel addition using functions φ1(t), φ2(t) to analyzethe Receiver Operating Characteristic Curves (ROC). While alow rate of vessel addition increases the number of iterations,a high rate of vessel addition introduces large number of falsepositives before the stopping criterion is met. The performancemetrics that are used to evaluate the final segmentation perfor-mance are: pixel-based sensitivity (SEN), specificity (SPEC),accuracy (ACC) of vessel segmentation, time taken to achievethe segmented vasculature and area under the ROC curves(AUC).

For this analysis, the rate at which new vessels are addediteratively is varied by changing φ1(t), φ2(t) as (31)-(32), andapplying the stopping criterion for that particular choice ofthreshold functions.

If, k = [1, 1.1, 1.2, .....3], (31)φ1(t) = 0.905− 0.005[10k(t− 1)− 9t],

φ2(t) = 205 + (t− 1)k. (32)

The ROC curves for all the three data sets with respectto ground-truth O1 are shown in Fig. 6 (a). In Fig. 6 (b), thevariations in mean ACC by varying the thresholds as given in(31)-(32) are shown. Here, we observe that the images from theDRIVE and STARE data sets maintain significantly high ACCas k varies in the range [1 ≤ k ≤ 2.2]. For k > 2.2, the rateof change in vessel estimates becomes significantly high, suchthat Ct does not maintain a medium Q region for 3 consecutiveiterations or more. This causes the iterative process to pick upa lot of false positive pixels before stopping, thus reducingthe segmentation ACC. The images from the CHASE DB1



TBME-01423-2014 8

(a) (b)

Fig. 5. The vessel estimate curves Ct and Et for a particular image from (a) DRIVE Test, and (b) STARE, respectively. The stopping iterations are (a)tf = 5, (b) tf = 4. In (a) the three consecutive derivatives for Et and Ct become non-negative at the same iteration step. In (b), the iteration for threeconsecutive non-negative derivatives corresponding to Ct and Et are different.

(a) (b)

Fig. 6. (a) ROC curves for blood vessel segmentation on DRIVE test, STARE and CHASE DB1 data sets with respect to ground-truth O1. (b) Variation in meansegmentation accuracy by varying the thresholds. Highest ACC is achieved for the DRIVE Test, STARE and CHASE DB1 data sets with k = [1.4, 1.6, 1.6],respectively.

data set are 30o FOV images centered at the papilla. For theseimages, if [1 ≤ k ≤ 1.4], rate of new vessel addition is verysmall, resulting in low segmentation SEN and ACC. Fork > 2.4, the rate of change in vessel estimates becomes veryhigh causing the medium Q region to be limited to less than 3consecutive iterations, thereby increasing the number of falseedge pixels in the final segmented vessel estimate. Thus, forthe test images, rate of vessel addition obtained by k ∈ [1.4−2.2] in (31)-(32) results in high segmentation ACC.

III. EXPERIMENTAL EVALUATION AND RESULTS

The performance of the proposed vessel segmentation al-gorithm is evaluated using the segmented vasculature and the

manually marked ground-truth O1, O2 and O3. To assess theoverall performance of vessel segmentation by the proposediterative method, we performed the following three sets ofexperimental evaluations. In the first experiment, the segmen-tation performance metrics were compared to the existingmethods. In the second experiment, abnormal images withbright and red lesions from the STARE Abnormal data setwere analyzed for segmentation performance. In the thirdexperiment, segmentation performance of the peripapillaryvessel pixels were analyzed.

A. Vessel Segmentation PerformanceThe inter-observer variability in the ground-truths is as-

sessed by analyzing the segmentation performance metrics



TBME-01423-2014 9

of the proposed vessel segmentation algorithm with respectto the ground-truths O1, O2 and O3 are given in Table II.Here, we observe that the STARE data set has high inter-observer variability in the segmentation performance metricsdue to the wide variety of pathological images in it. Next, theperformance and computational complexity of the proposedblood vessel segmentation algorithm with respect to ground-truth O1 in comparison with existing vessel segmentationmethods is shown in Table III. Most existing methods analyzethe automated segmentation performance with ground-truthO1. The implementation time complexity of the method in[30] is taken from [21] and the method in [29] is implementedon our Matlab system for comparative analysis.

From Table III, we observe that the proposed iterativevessel segmentation approach outperforms all the existingmethods except for the supervised approach by Ricci. et. al.[16] and Roychowdhury et. al. [13] on the DRIVE Test dataset, and the unsupervised Lam et. al. [27] method on theSTARE data set. In the method by Ricci et. al. [16], supportvector machine (SVM) classifiers were separately trained using20,000 manually segmented randomly chosen pixels from theDRIVE and STARE data sets. This method is computationallymore complex due to the use of the SVM classifier, and theclassifier may need retraining for new data sets. The supervisedmethod in Roychowdhury et. al. [13] uses GMM classifierstrained on the DRIVE train data set and hence it is moretrained to segment vessel pixels in the DRIVE Test data setthan the STARE data set. Also, from Table III, we observethat the AUC of the supervised methods is comparable to thatof the proposed method. This is indicative of the robustness ofthe stopping criterion design with varying threshold functions[φ1(t), φ2(t)].

B. Abnormal Image Analysis

Existing works in [36], [16], [26], [27], [22] have shownthat robustness of a vessel segmentation algorithm dependson the segmentation performance on abnormal retinal imageswith red lesions such as micro-aneurysms or hemorrhagesand bright lesions such as exudates or cotton-wool spots.Vessel segmentation from two images with significant redand bright lesions from the STARE Abnormal data set havebeen analyzed by a number of existing methods [26], [27],[15] and [17]. In Table IV, we compare the segmentationperformance of all known methods with the proposed approachon the 10 images from the STARE Abnormal data set withrespect to ground-truth O1. We observe that the proposedmethod outperforms all other methods except the perceptivetransform based method in Lam et. al. [27]. However, thecomputational complexity of the proposed approach is lowerthan most existing methods. From Table III, we observethat although the unsupervised method in Nguyen et. al.[21] has the lowest computational complexity on the STAREdata set (segmentation time of 2.5s per image), it suffersfrom false positive detection around the optic disc regionand over detection in pathological images, there by havinglower segmentation performance metrics when compared tothe proposed method.

TABLE IVSEGMENTATION PERFORMANCE ON THE STARE ABNORMAL DATA SET

WITH RESPECT TO GROUND-TRUTH O1 . SEGMENTATION TIME ISRECORDED PER IMAGE.

Method ACC AUC TimeHoover et. al. [20] 0.9211 0.7590 5 minsJiang et. al. [34] 0.9352 0.9343 8-36 s

Mendonca et. al. [22] 0.9426 - 3 minsSoares et. al. [15] 0.9425 0.9571 3 mins

Vermeer et. al. [25] 0.9287 0.9187 -Marin et. al. [17] 0.9510 - 90 sLam and Yan [26] 0.9474 0.9392 8 mins

Lam et. al. [27] 0.9556 0.9707 13 minsRoychowdhury et. al. [13] 0.9453 0.9596 8.36 s

Proposed 0.9535 0.9638 3.87 s

C. Peripapillary vessel analysis

In retinal fundus images, the blood vessels in and aroundthe optic disc are referred to as the peripapillary blood vessels.Many retinal abnormalities such as proliferative DR (PDR),glaucoma, central retinal vein occlusion, cilio-retinal arteryocclusion can lead to changes in the blood vessel structuremainly in the peripapillary region. For instance, neovascular-ization at the disc (NVD) caused due to PDR is evident ifnew blood vessels are visible within 1-optic disc diameter (1-DD) centered at the papilla [11]. Thus, the performance ofthe proposed vessel segmentation system for extracting theperipapillary blood vessels in normal and abnormal images isanalyzed here. For this purpose, the images from the DRIVE,STARE and CHASE DB1 data set were manually annotatedfor the optic disc boundary and optic nerve head in each image,and then, a mask was created centered at the optic nervehead with 1-DD radius to extract the peripapillary region. Thefirst 19 images from the STARE vessel segmentation data setcontain the peripapillary region and hence peripapillary vesseldetection was performed on these 19 images only.

The segmented vessel images of the DRIVE data set usingthe Marin et. al. method [17]1, Soares et. al. method [15]2, Jiang et. al. method [34] and Staal et. al. method [31] 3

are downloaded from the respective websites for comparison.The segmented vessel images from the STARE data set arecompared to the segmentation results produced by the Marinet. al. method [17], Soares et. al. method [15] and Hoover et.al. method [20] 4.

We analyze the performance of peripapillary vessel segmen-tation for the three data sets with respect to the three sets ofground-truths O1, O2 and O3 in Table V. We know that fundusimages with smaller FOV, such as images from STARE andCHASE DB1 data sets, capture a more detailed view of the fineblood vessels when compared to the images with larger FOV(images from DRIVE data set). From Table V, we observe thatthe proposed vessel segmentation method has significantly bet-ter peripapillary vessel segmentation performance with respect

1Available at http://www.uhu.es/retinopathy/eng/bd.php2Available at http://sourceforge.net/apps/mediawiki/retinal/index.php

?title=Segmentation results3Available at http://www.isi.uu.nl/Research/Databases/DRIVE/browser.php4Available at http://www.parl.clemson.edu/ ahoover/stare/probing/index.html



TBME-01423-2014 10

TABLE IIPERFORMANCE OF THE PROPOSED METHOD ON THE TEST DATA SETS. SEGMENTATION TIME IS RECORDED PER IMAGE. MEAN PERFORMANCE METRICS

AND THEIR STANDARD DEVIATION IS GIVEN IN ().

Data Ground-Truth AUC ACC SEN SPEC Time(s)DRIVE Test O1 0.9672 0.9494 (0.005) 0.7395 (0.062) 0.9782 (0.0073)

O2 0.9640 0.9507 (0.006) 0.7601 (0.0581) 0.9772 (0.01)O3 0.9613 0.9442 (0.006) 0.7305 (0.0525) 0.9787 (0.0076) 2.45 (0.3196)

STARE O1 0.9673 0.9560 (0.0095) 0.7317 (0.053) 0.9842 (0.0069)O2 0.9601 0.9384 (0.0131) 0.7498 (0.06) 0.9801 (0.075)O3 0.9617 0.9332 (0.0125) 0.7552 (0.052) 0.9719 (0.0076) 3.9587 (0.2732)

STARE Normal O1 0.9706 0.9586 (0.0061) 0.7571 (0.0088) 0.9877 (0.0035)O2 0.9621 0.9382 (0.0131) 0.7310 (0.0559) 0.9844 (0.0067)O3 0.9642 0.9368 (0.0073) 0.7568 (0.05) 0.9819 (0.009) 4.0387 (0.2865)

STARE Abnormal O1 0.9638 0.9535 (0.017) 0.7062 (0.100) 0.9808 (0.026)O2 0.9577 0.9387 (0.008) 0.7586 (0.05) 0.9795 (0.008)O3 0.9596 0.9320 (0.02) 0.7161 (0.073) 0.9705 (0.015) 3.8786 (0.2476)

CHASE DB1 O1 0.9623 0.9467 (0.0076) 0.7615 (0.0516) 0.9575 (0.003)O2 0.9600 0.9454 (0.008) 0.7514 (0.0442) 0.9622 (0.006)O3 0.9620 0.9455 (0.007) 0.7653 (0.05) 0.9579 (0.004) 7.9138 (0.6774)

TABLE IIICOMPARATIVE PERFORMANCE OF PROPOSED MODEL WITH EXISTING WORKS ON THE DRIVE TEST AND STARE DATA SETS WITH RESPECT TO

GROUND-TRUTH O1 . SEGMENTATION TIME IS RECORDED PER IMAGE.

Test Data: DRIVE Test STAREMethod ACC SPEC SEN AUC Time ACC SPEC SEN AUC Time SystemSupervised MethodsNiemeijer et. al. [14] 0.942 0.969 0.689 0.93 - - - - - - -Staal et. al. [31] 0.944 0.977 0.719 0.952 15 mins 0.952 0.981 0.697 0.961 15 mins 1.0 GHz, 1GB RAMSoares et. al. [15] 0.946 0.978 0.733 0.961 ∼3 mins 0.948 0.975 0.72 0.967 ∼3 mins 2.17 GHz, 1GB RAMRicci et. al. [16] 0.959 0.972 0.775 0.963 - 0.965 0.939 0.903 0.968 - -Marin et. al. [17] 0.945 0.98 0.706 0.958 ∼90 s 0.952 0.982 0.694 0.977 ∼90 s 2.13 GHz, 2GB RAMFraz et. al. [18] 0.948 0.981 0.74 0.974 ∼100 s 0.953 0.976 0.755 0.976 ∼100 s 2.27GHz, 4GB RAMRoychowdhury et. al. [13] 0.952 0.983 0.725 0.962 3.11 s 0.951 0.973 0.772 0.969 6.7 s 2.6GHz,2GB RAMUnsupervised MethodsHoover et. al. [20] - - - - - 0.9275 0.81 0.65 0.75 5 mins Sun SPARCstation 20Jiang et. al. [34] 0.891 0.90 0.83 0.932 8-36 s 0.901 0.90 0.857 0.929 8-36 s 600MHz PCMendonca et. al. [22] 0.945 0.976 0.734 - 2.5 mins 0.944 0.973 0.699 - 3 mins 3.2 GHz, 980MB RAMLam and Yan [26] - - - - - 0.947 - - 0.939 8 mins 1.83 GHz,2GB RAMAl-Diri et. al.[35] - 0.955 0.728 - 11 mins - 0.968 0.752 - - 1.2GHzLam et. al. [27] 0.947 - - 0.961 13 mins 0.957 - - 0.974 13 mins 1.83 GHz 2GB RAMBudai et. al.(2010)[28] 0.949 0.968 0.759 - 11 s 0.938 0.975 0.651 - 16 s 2.0 GHz, 2GB SDRAMBudai et.al.(2013)[29] 0.957 0.987 0.644 - ∼5 s 0.938 0.982 0.58 - ∼6 s 2.3GHz,4GB RAMPerez et. al. [30] 0.925 0.967 0.644 - ∼2 mins 0.926 0.944 0.769 - ∼2 mins Parallel ClusterMiri et.al. [24] 0.943 0.976 0.715 - ∼50 s - - - - - 3GHz, 1GB RAMNguyen et. al. [21] 0.941 - - - 2.5 s 0.932 - - - 2.5 s 2.4GHz, 2GB RAMProposed 0.949 0.978 0.739 0.967 2.45 s 0.956 0.984 0.732 0.967 3.95 s 2.6GHz, 2GB RAM

to the three sets of ground-truths in the images with smallerFOV than on the images with larger FOV.

IV. DISCUSSION AND CONCLUSIONS

In this paper, we have proposed an unsupervised iterativeblood vessel segmentation algorithm using fundus imagesand tested it on three public data sets of DRIVE, STAREand CHASE DB1. This algorithm iteratively extracts vesselpixels from a morphological vessel enhanced image from thenegative green plane from each fundus image. The algorithminitiates with an initial estimate of the major portions of theblood vessels. Next, iteratively new vessel pixels are addedto the existing vessel estimate by adaptive global threshold-ing until a stopping criterion is met. The proposed iterativesegmentation algorithm is computationally simple, general

and faster than all supervised and unsupervised algorithmsdeveloped so far.

The proposed algorithm depends on three limiting con-straints that include the initial major vessel estimate and twothreshold functions that determine the rate at which vessel esti-mates change iteratively. The ROC curves that are constructedby varying these threshold functions demonstrate that thestopping criterion is robust to change in the adaptive thresholdvalues since an average AUC of 0.96 is achieved across thethree data sets for three sets of manual annotations. Also,the proposed algorithm relies on the first 3 derivatives of thechange in iterative vessel estimates to identify a novel stoppingcriterion that terminates the iterative vessel addition processwhile ensuring high accuracy of the segmented vasculature.Future efforts may be directed towards the analysis of thefirst 5 orders of derivatives for the change in iterative vessel



TBME-01423-2014 11

TABLE VPERIPAPILLARY VESSEL ANALYSIS. ACC AND THE STANDARD

DEVIATION IS GIVEN IN ().

Method O1 O2 O3

DRIVE TestSupervised

Marin et. al. [17] 0.914(0.02) 0.919(0.02) 0.901(0.02)Soares et. al. [15] 0.912(0.01) 0.917(0.01) 0.899(0.02)Staal et. al. [31] 0.910 (0.02) 0.915(0.02) 0.898(0.02)

Roychowdhury et. al. [13] 0.915(0.02) 0.926(0.02) 0.904(0.02)Unsupervised

Jiang et. al. [34] 0.863(0.03) 0.866(0.02) 0.846(0.03)Proposed 0.920(0.02) 0.923(0.02) 0.903(0.02)

STARESupervised

Marin et. al. [17] 0.805(0.04) 0.719(0.06) 0.663(0.06)Soares et. al. [15] 0.804(0.05) 0.724(0.05) 0.671(0.05)


Hoover et. al. [20] 0.768(0.03) 0.718(0.04) 0.705(0.04)Proposed 0.818(0.03) 0.774(0.06) 0.747(0.06)

CHASE DB1Supervised


Proposed 0.904(0.03) 0.905(0.04) 0.889(0.03)

estimates to develop a stopping criterion for high resolutionfundus images.

The proposed vessel segmentation algorithm is very effec-tive in abnormal retinal images with large red or bright lesionsin close vicinity of the main vasculature, since the algorithmwill stop before the iteration in which the large lesion re-gions are identified and added to the existing vessel estimate.However, this algorithm has one limitation. In retinal imageswith small microaneurysms close to the vasculature, the mi-croaneurysms will be included as a part of the vasculaturedue to the regiongrow operation. Future work will be directedtowards combining the proposed method with decision makingapproaches to further enhance the segmentation performanceon abnormal retinal images with such small lesions close tothe vasculature.

The most significant features of the proposed segmentationalgorithm is that it provides segmented vessels with highaccuracy for peripapillary blood vessel extraction. It achievesgreater than 90% peripapillary vessel segmentation accuracyfor the DRIVE and CHASE DB1 data set and greater than77% accuracy for the STARE data set with respect to bothhuman observers. Thus, the proposed segmentation algorithmcan be useful for automated algorithms that detect the density,tortuosity or width of these peripapillary vessels for severityanalysis of pathologies such as vein occlusions and glaucoma.

Unlike some methods that are fine tuned to handle abnormalretinal images in [27], the proposed vessel segmentation algo-rithm is generalizable, and robust against false edges, unevenillumination and pigmentation. The high speed, computationalsimplicity, and data independence of the proposed methodmakes it ideal for automated vessel pathology analysis forfuture research problems that analyze changes in the retinalvessel width [5] [7] [8]. Future work will be directed towards

combining the proposed algorithm with DR screening systemssuch as [9] and to detect neovascularization in the peripapillaryregion. Redesigning the proposed method for vessel segmen-tation on wide field images with greater than 200o FOV is atopic that requires further study.

REFERENCES

[1] C. M. Wilson, K. D. Cocker, M. J. Moseley, C. Paterson, S. T. Clay,W. E. Schulenburg, M. D. Mills, A. L. Ells, K. H. Parker, G. E. Quinn,A. R. Fielder, and J. Ng, “Computerized analysis of retinal vessel widthand tortuosity in premature infants,” Investigative Ophthalmology andVisual Science, vol. 49, no. 8, pp. 3577–3585, 2008.

[2] A. Karperien, H. F. Jelinek, J. J. Leandro, J. V. Soares, R. M. Cesar Jr,and A. Luckie, “Automated detection of proliferative retinopathy inclinical practice,” Clinical ophthalmology (Auckland, NZ), vol. 2, no. 1,p. 109, 2008.

[3] CDC, “Diabetic retinopathy. Atlanta, GA: National for chronic diseaseprevention and health promotion.” March 23,2011. [Online]. Available:http://www.cdc.gov/visionhealth/pdf/factsheet.pdf

[4] M. D. Abramoff, M. K. Garvin, and M. Sonka, “Retinal imaging andimage analysis,” IEEE Transactions on Medical Imaging, vol. 3, pp.169–208, January 2010.

[5] H. M. Pakter, S. C. Fuchs, M. K. Maestri, L. B. Moreira, L. M. Dei Ri-cardi, V. F. Pamplona, M. M. Oliveira, and F. D. Fuchs, “Computer-assisted methods to evaluate retinal vascular caliber: What are theymeasuring?” Investigative Ophthalmology and Visual Science, vol. 52,no. 2, pp. 810–815, 2011.

[6] K. Kotliar, E. Nagel, W. Vilser, S.-F. Seidova, and I. Lanzl, “Mi-crostructural alterations of retinal arterial blood column along the vesselaxis in systemic hypertension,” Investigative Ophthalmology and VisualScience, vol. 51, no. 4, pp. 2165–2172, 2010.

[7] A. Kochkorov, K. Gugleta, C. Zawinka, R. Katamay, J. Flammer, andS. Orgul, “Short-term retinal vessel diameter variability in relation tothe history of cold extremities,” Investigative Ophthalmology and VisualScience, vol. 47, no. 9, pp. 4026–4033, 2006.

[8] E. Nagel, W. Vilser, and I. Lanzl, “Age, blood pressure, and vesseldiameter as factors influencing the arterial retinal flicker response,”Investigative Ophthalmology and Visual Science, vol. 45, no. 5, pp.1486–1492, 2004.

[9] S. Roychowdhury, D. Koozekanani, and K. Parhi, “Dream: Diabeticretinopathy analysis using machine learning,” IEEE Journal of Biomed-ical and Health Informatics, vol. 18, no. 5, pp. 1717–1728, Sept 2014.

[10] T. Teng, M. Lefley, and D. Claremont, “Progress towards automateddiabetic ocular screening: A review of image analysis and intelligentsystems for diabetic retinopathy,” Medical and Biological Engineeringand Computing, vol. 40, pp. 2–13, 2002.

[11] K. Goatman, A. Fleming, S. Philip, G. Williams, J. Olson, and P. Sharp,“Detection of new vessels on the optic disc using retinal photographs,”IEEE Transactions on Medical Imaging, vol. 30, no. 4, pp. 972–979,2011.

[12] M. Fraz, A. Remagnino, P.and Hoppe, B. Uyyanonvara, A. Rudnicka,C. Owen, and S. Barman, “Blood vessel segmentation methodolo-gies in retinal images-a survey,” Computer Methods and Programs inBiomedicine, vol. 108, pp. 407–433, 2012.

[13] S. Roychowdhury, D. Koozekanani, and K. Parhi, “Blood vessel seg-mentation of fundus images by major vessel extraction and sub-imageclassification,” IEEE Journal of Biomedical and Health Informatics, p.doi:10.1109/JBHI.2014.2335617, 2014.

[14] M. Niemeijer, J. Staal, B. van Ginneken, M. Loog, and M. D. Abramoff,“Comparative study of retinal vessel segmentation methods on a newpublicly available database,” Proc. SPIE 5370, Medical Imaging, vol.5370, pp. 648–656, 2004.

[15] J. Soares, J. Leandro, R. Cesar, H. Jelinek, and M. Cree, “Retinal vesselsegmentation using the 2-D Gabor wavelet and supervised classifica-tion,” IEEE Transactions on Medical Imaging, vol. 25, no. 9, pp. 1214–1222, 2006.

[16] E. Ricci and R. Perfetti, “Retinal blood vessel segmentation usingline operators and support vector classification,” IEEE Transactions onMedical Imaging, vol. 26, no. 10, pp. 1357–1365, 2007.

[17] D. Marin, A. Aquino, M. Gegundez-Arias, and J. Bravo, “A new super-vised method for blood vessel segmentation in retinal images by usinggray-level and moment invariants-based features,” IEEE Transactions onMedical Imaging, vol. 30, no. 1, pp. 146–158, 2011.



TBME-01423-2014 12

[18] M. Fraz, P. Remagnino, A. Hoppe, B. Uyyanonvara, A. Rudnicka,C. Owen, and S. Barman, “An ensemble classification-based approachapplied to retinal blood vessel segmentation,” IEEE Transactions onBiomedical Engineering, vol. 59, no. 9, pp. 2538–2548, 2012.

[19] C. Lupascu, D. Tegolo, and E. Trucco, “Fabc: Retinal vessel segmenta-tion using adaboost,” IEEE Transactions on Information Technology inBiomedicine, vol. 14, no. 5, pp. 1267–1274, Sept 2010.

[20] A. Hoover, V. Kouznetsova, and M. Goldbaum, “Locating blood vesselsin retinal images by piecewise threshold probing of a matched filterresponse,” IEEE Transactions on Medical Imaging, vol. 19, pp. 203–210, 2000.

[21] U. T. V. Nguyen, A. Bhuiyan, L. A. F. Park, and K. Ramamohanarao,“An effective retinal blood vessel segmentation method using multi-scaleline detection,” Pattern Recogn., vol. 46, no. 3, pp. 703–715, Mar. 2013.

[22] A. Mendonca and A. Campilho, “Segmentation of retinal blood vesselsby combining the detection of centerlines and morphological reconstruc-tion,” IEEE Transactions on Medical Imaging, vol. 25, no. 9, pp. 1200–1213, 2006.

[23] F. Zana and J.-C. Klein, “Segmentation of vessel-like patterns usingmathematical morphology and curvature evaluation,” IEEE Transactionson Image Processing, vol. 10, no. 7, pp. 1010–1019, Jul 2001.

[24] M. Miri and A. Mahloojifar, “Retinal image analysis using curvelettransform and multistructure elements morphology by reconstruction,”IEEE Transactions on Biomedical Engineering, vol. 58, no. 5, pp. 1183–1192, 2011.

[25] K. A. Vermeer, F. M. Vos, H. G. Lemij, and A. M. Vossepoel, “A modelbased method for retinal blood vessel detection,” Computers in Biologyand Medicine, vol. 34, no. 3, pp. 209–219, 2004.

[26] B. Lam and H. Yan, “A novel vessel segmentation algorithm forpathological retina images based on the divergence of vector fields,”IEEE Transactions on Medical Imaging, vol. 27, no. 2, pp. 237–246,2008.

[27] B. Lam, Y. Gao, and A.-C. Liew, “General retinal vessel segmentationusing regularization-based multiconcavity modeling,” IEEE Transactionson Medical Imaging, vol. 29, no. 7, pp. 1369–1381, 2010.

[28] A. Budai, G. Michelson, and J. Hornegger, “Multiscale blood vesselsegmentation in retinal fundus images,” Proc. Bildverarbeitung fr dieMed., pp. 261–265, March 2010.

[29] A. Budai, R. Bock, A. Maier, J. Hornegger, and G. Michelson, “Ro-bust vessel segmentation in fundus images,” International Journal ofBiomedical Imaging, no. 154860, p. doi:10.1155/2013/154860, 2013.

[30] M. Palomera-Perez, M. Martinez-Perez, H. Benitez-Perez, and J. Ortega-Arjona, “Parallel multiscale feature extraction and region growing:Application in retinal blood vessel detection,” IEEE Transactions onInformation Technology in Biomedicine, vol. 14, no. 2, pp. 500–506,March 2010.

[31] J. Staal, M. Abramoff, M. Niemeijer, M. Viergever, and B. van Gin-neken, “Ridge-based vessel segmentation in color images of the retina,”IEEE Transactions on Medical Imaging, vol. 23, no. 4, pp. 501–509,2004.

[32] K. U. Research, “Chase db1,” January 2011. [Online]. Available:http://blogs.kingston.ac.uk/retinal/chasedb1/.

[33] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 2nd ed.Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc.,1992.

[34] X. Jiang and D. Mojon, “Adaptive local thresholding by verification-based multithreshold probing with application to vessel detection inretinal images,” IEEE Transactions on Pattern Analysis and MachineIntelligence, vol. 25, no. 1, pp. 131–137, 2003.

[35] B. Al-Diri, A. Hunter, and D. Steel, “An active contour model for seg-menting and measuring retinal vessels,” IEEE Transactions on MedicalImaging, vol. 28, no. 9, pp. 1488–1497, 2009.

[36] R. Perfetti, E. Ricci, D. Casali, and G. Costantini, “Cellular neural net-works with virtual template expansion for retinal vessel segmentation,”IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 54,no. 2, pp. 141–145, 2007.

S ohini Roychowdhury received her M.S. fromKansas State University in 2010 and PhD fromUniversity of Minnesota in 2014 in Electrical andComputer Engineering. She is currently an assis-tant professor in the Department of Electrical En-gineering at University of Washington, Bothell. Herresearch interests include medical image process-ing, signal processing, pattern recognition, machinelearning and artificial intelligence.

D ara Koozekanani received a PhD from the OhioState University in Biomedical Engineering in 2001.His research dissertation involved application ofcomputer vision techniques to the analysis of opticalcoherence tomography images. He subsequently re-ceived an MD degree from The Ohio State Univer-sity in 2003, completed ophthalmology residency atthe University of Wisconsin in 2007, and completeda surgical retinal fellowship at the Medical Collegeof Wisconsin in 2009.

He is currently an assistant professor of ophthal-mology on the clinical faculty at the University of Minnesota. He sees patientswith a variety of surgical and medical retinal diseases. His research interestsare the application of ophthalmic imaging technologies and automated analysisof those images.

K eshab K. Parhi (S’85-M’88-SM’91-F’96) receivedthe B.Tech. degree from the Indian Institute of Tech-nology, Kharagpur, in 1982, the M.S.E.E. degreefrom the University of Pennsylvania, Philadelphia,in 1984, and the Ph.D. degree from the Universityof California, Berkeley, in 1988. He has been withthe University of Minnesota, Minneapolis, since1988, where he is currently Distinguished McKnightUniversity Professor and Edgar F. Johnson Profes-sor in the Department of Electrical and ComputerEngineering. He has published over 550 papers, has

authored the textbook VLSI Digital Signal Processing Systems (Wiley, 1999)and coedited the reference book Digital Signal Processing for MultimediaSystems (Marcel Dekker, 1999). His research addresses VLSI architecturedesign and implementation of signal processing, communications and biomed-ical systems, error control coders and cryptography architectures, high-speedtransceivers, secure computing and molecular computing. He is also currentlyworking on intelligent classification of biomedical signals and images, forapplications such as seizure prediction and detection, schizophrenia clas-sification, biomarkers for mental disorder, brain connectivity, and diabeticretinopathy screening.

Dr. Parhi is the recipient of numerous awards including the 2013Distinguished Alumnus Award from IIT, Kharagpur, India, 2013 Gradu-ate/Professional Teaching Award from the University of Minnesota, 2012Charles A. Desoer Technical Achievement award from the IEEE Circuits andSystems Society, the 2004 F. E. Terman award from the American Societyof Engineering Education, the 2003 IEEE Kiyo Tomiyasu Technical FieldAward, the 2001 IEEE W. R. G. Baker prize paper award, and a GoldenJubilee medal from the IEEE Circuits and Systems Society in 2000. He hasserved on the editorial boards of the IEEE TRANSACTIONS ON CIRCUITSAND SYSTEMS PART I and PART II, VLSI Systems, Signal Processing,Signal Processing Letters,and Signal Processing Magazine, and served asthe Editor-in-Chief of the IEEE TRANSACTIONS ON CIRCUITS ANDSYSTEMSART I (2004-2005 term), and currently serves on the EditorialBoard of the Springer Journal of Signal Processing Systems. He has servedas technical program cochair of the 1995 IEEE VLSI Signal Processingworkshop and the 1996 ASAP conference, and as the general chair of the2002 IEEE Workshop on Signal Processing Systems. He was a distinguishedlecturer for the IEEE Circuits and Systems society during 1996-1998. Heserved as an elected member of the Board of Governors of the IEEE Circuitsand Systems society from 2005 to 2007.

tbme-01423-2014 1 iterative vessel segmentation of … ieee papers/matlab(image... · iterative...

Documents