Embed Size (px)
Citation preview
Ut
Qa
1b
c
d
1
a
A
R
R
2
A
K
S
M
D
A
B
R
E
1
Sammai
R
0d
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
jo ur n al hom ep age : www.int l .e lsev ierhea l th .com/ journa ls /cmpb
nsupervised skin lesions border detection viawo-dimensional image analysis
aisar Abbasa,b,∗, Irene Fondónc, Muhammad Rashidd
Department of Computer Science and Technology, Huazhong University of Science and Technology,037 Luoyu Road, Wuhan 430074, ChinaCenter for Biomedical Imaging and Bioinformatics, Key Laboratory of Image Processing, 1037 Luoyu Road, Wuhan 430074, ChinaDepartment of Signal Theory and Communications, School of Engineering Path of Discovery, s/n C. P., 41092 Seville, SpainDepartment of Life Science and Technology, Biochemistry and Molecular Biology, Huazhong University of Science and Technology,037 Luoyu Road, Wuhan 430074, China
r t i c l e i n f o
rticle history:
eceived 9 November 2009
eceived in revised form
5 June 2010
ccepted 28 June 2010
eywords:
kin cancer
elanoma
ermoscopy
rtifacts removal
order detection
a b s t r a c t
The skin cancer was analyzed by dermoscopy helpful for dermatologists. The classifi-
cation of melanoma and carcinoma such as basal cell, squamous cell, and merkel cell
carcinomas tumors can be increased the sensitivity and specificity. The detection of an
automated border is an important step for the correctness of subsequent phases in the
computerized melanoma recognition systems. The artifacts such as, dermoscopy-gel, spec-
ular reflection and outline (skin lines, blood vessels, and hair or ruler markings) were also
contained in the dermoscopic images. In this paper, we present an unsupervised approach
for multiple lesion segmentation, modification of Region-based Active Contours (RACs)
as well as artifact diminution steps. Iterative thresholding is applied to initialize level
set automatically; the stability of curves is enforced by maximum smoothing constraints
on Courant–Friedreichs–Lewy (CFL) function. The work has been tested on dermoscopic
database of 320 images. The border detection error is quantified by five distinct statistical
egion-based active contour
xemplar-based image inpainting
metrics and manually used to determine the borders from a dermatologist as the ground
truth. The segmentation results were compared with other state-of-the-art methods along
with the evaluation criteria. The unsupervised border detection system increased the true
detection rate (TDR) is 4.31% and reduced the false positive rate (FPR) of 5.28%.
Canada [3].
. Introduction
kin cancer is the most common cancer in U.S. [1]. Therere two major types of skin cancer [2], namely malignantelanoma and non-melanoma (basal cell, squamous cell, and
erkel cell carcinomas, etc.). Melanoma is more dangerousnd can be fatal if untreated. These tumors develop changesn the skin texture and color, but they can be cured in more
∗ Corresponding author at: Department of Computer Science and Technooad, Wuhan 430074, China. Tel.: +86 27 13437168375; fax: +86 27 87559
E-mail address: [email protected] (Q. Abbas).169-2607/$ – see front matter © 2010 Elsevier Ireland Ltd. All rights resoi:10.1016/j.cmpb.2010.06.016
© 2010 Elsevier Ireland Ltd. All rights reserved.
than 90% of cases, if they are detected and treated in theearly stages. Some cutaneous tumors remain an enigma. Agood example, one even better than melanoma, is perhaps thedeadliest of all skin cancers “Merkel cell carcinomas” whichhave tripled in the incidence from 1986 to 2001 in the U.S. and
logy Huazhong University of Science and Technology, 1037 Luoyu091.
It is well known that early finding and treatment of skincancer can reduce the mortality and morbidity of patients.Digital dermoscopy [4] is widely considered as one of the
erved.
e2 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
Fig. 1 – Problems with border detection within both clinical and dermoscopic view for various types of lesions whichillustrate: (a) benign melanocytic naevi, (b) scapula malignant melanomas, (c) non-pigmented basal cell carcinoma, (d)
ted t
merkel cell carcinomas, (e) keratosis tumors, (f) non-pigmenmost cost-effective means to identify and classify skin can-cer. In order to differentiate benign lesions from malignantones, expert dermatologists advocate the use of computerizedimage analysis schemes. An automatic dermoscopy imageanalysis system [5–9] has usually three stages: (1) boundarydetection, (2) features extraction and selection, and (3) lesionrecognition. The border detection step is one of the mostimportant, since it affects the precision of the subsequentsteps. Unsupervised border detection is a difficult task dueto the great variety of lesion shapes, sizes, and colors alongwith diverse skin types and textures as shown in (Fig. 1). Inaddition to this, artifacts such as dermoscopy-gel (Fig. 1(a)),specular reflection (Fig. 1(b)) and outline (markers), skin lines,hair (Fig. 1(f))) might decrease the accuracy of border detectionboth in clinical (Fig. 1(b)) and dermoscopy (Fig. 1(f)) images.
1.1. Background
For designing an unsupervised segmentation method, theremust be single system, which consists of lesions border detec-tion and artifact reduction step. The borders finding of lesionsand artifact removal systems are very challenging and vitalconsiderations in fast and precise skin border description.In recent years, many studies have been focused on unsu-pervised and supervised segmentation of lesions in digitaldermoscopy. However, none of the research efforts have beenfully described to remove all artifacts.
As a consequence of advances in skin imaging technology,Celebi et al. [10] recommended an unsupervised approach todetect border in dermoscopy images by statistical region merg-ing algorithm. Also, Celebi et al. [11] discussed the difficultyand subjectivity of human understanding with computerized
analysis in terms of denoising and border detection tech-niques. Moreover, Celebi et al. [12] suggested a modified Jsegalgorithm to segment a tumor from its surrounding. A multi-direction gradient vector flow (GVF) snake-based scheme isumor.
proposed by Tang [13]. He put forward an AD filter by usingadaptive threshold selection and a new gradient computa-tion way, which is robust to noise and can effectively removethe hair. Further Yuan et al. [14] suggested a novel multi-modal technique based on a region fusion and narrow bandenergy graph partitioning to accurate lesion segmentation. In[15], Schmid proposed a method to segmentation tumor byfuzzy c-means (FCM) clustering and mathematical morphol-ogy. Erkol et al. [16] used gradient vector flow (GVF) snakes tofind the border of skin lesions. The region-growing methodproposed by Iyatomi et al. [17,18] to yield the results close tothose obtained by dermatologists. Gómez et al. [19] designedan unsupervised algorithm called the Independent HistogramPursuit (IHP), for segmentation of dermatological lesions. Next,Zhou et al. [20] developed a new mean shift based fuzzy c-means that requires less computational time than previoustechniques, while providing good segmentation results. Sil-veira et al. [21] evaluated six different segmentation systemsin favor of three skin tumor types. These algorithms are: (1)adaptive thresholding (AT), (2) gradient vector flow (GVF), (3)adaptive snake (AS), (4) level set method of Chan et al. (C-LS),(5) expectation–maximization level set (EM-LS), and (6) fuzzy-based split-and-merge algorithm (FBSM).
2. Aim and our approach
Although the significant research efforts have gone into devel-oping computerized algorithms to reduce artifacts, detectlesion boundary, and categorization remains a technical chal-lenge. Because segmentation step is highly dependent uponnoise likes illumination, dermoscopic-gel or air bubbles and
outlines such as line markers, skin and hair, etc. It has beenalso observed that there are three major problems present inprevious studies in particular. First problem is that none of thestudy efforts have been fully devoted to develop an efficient
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15 e3
Fig. 2 – Flow diagram of the proposed unsupervised skin lesion border detection.
rftccw
ssoFreabataebiwhmb
t(r
outine for every type of artifact attenuation. Second, veryew studies dealt with the automatic segmentation of mul-iple lesions. Third problem is that in some cases for basal cellarcinoma (BCC), squamous cell carcinoma (SCC), and merkelell carcinoma. (MCC) tumors, the skin lesion borders are notell clear, which reduces the accuracy of border detection.
In adding up to this, Celebi et al. [11] suggested that imagemoothing effectively removes the skin lines and blood ves-els, bubbles can be removed by a morphological top-hatperator followed by a radial search procedure introduced byleming et al. [22], and both Wighton et al. [23], Zhou et al. [24]ecommended hair removing algorithms. Although, Wightont al. developed hair disocclusion using inpainting with thenisotropic diffusion scheme that performed on average 32.7%etter than the linear interpolation of DullRazor [25] hairlgorithm. Lately, Zhou et al. have developed the more sophis-icated feature-preserving algorithms to remove artifacts suchs dermoscopic-gel and hair with the help of curve fitting andxemplar-based image inpainting [26], which can be extendedy Criminisi et al. [27]. Similarly, unsupervised PDE-basedmage inpainting algorithm for the hair-occluded informationas also described by Xie et al. [28]. However, these techniquesave several drawbacks including high computational require-ents and the fact that the required number of iterations can
e unbounded.
For automatic border detection, our approach consists ofhree main steps: (1) image rescaling, (2) preprocessing, and3) tumors border detection. In the first step, all images wereescaled to a standard size. In the second step, numerous fil-
ters have been used to reduce each type of artifacts with thehelp of homomorphic filtering, weighted median filtering withlocal cost function, and an efficient exemplar-based inpaint-ing scheme. In the third step, a region-based active contourmodel is applied to segment the lesion boundaries. The flowdiagram of an unsupervised border detection method is shownin (Fig. 2).
3. Materials and methods
3.1. Image dataset
A clinical online database of 320-color dermoscopic and clini-cal view lesion images were obtained from varies sources butmost images came from the Department of Dermatology [29],Health Waikato New Zealand. In addition, most of these skincancer images were captured from Nikon 995 with the digitalacquisition system. The images have been stored in the RGB-color format with dimensions vary from 640 × 480 to 640 × 405or 620 × 300 to 340 × 460. The database was subdivided intosix categories consisting of total 320 images: (1) 50 benignmelanocytic lesions, (2) 70 malignant melanoma lesions, and(3) 40 basal cell carcinoma lesions, with, (4) 40 merkel cell car-
cinoma lesions, (5) 90 seborrhoeic keratosis lesions, and (6) 30non-melanocytic lesions. We have left 60 out of 380 images,which having non-lesions objects to show significance of ourproposed method for tumor delineation.
s i n
e4 c o m p u t e r m e t h o d s a n d p r o g r a m3.2. Image rescaling
In this subsection, it is an initial and necessary step for pre-processing and segmentation results comparisons with othermethods. In this step, all images were rescaled to a standardsize of 450 × 360 pixels.
3.3. Preprocessing
In this step, we make it possible to accurate detect the lesionborders. It has been observed that a dermoscopic image oftencontains artifacts. Most recently, many studies suggested theirworks but none of them have discussed about all cases ofartifacts.
For this rationale, we have developed an effective prepro-cessing method which can be divided into following threesteps namely, (1) specular reflection reduction, (2) air bubblesor dermoscopic-gel reduction, and (3) hair and lines reduction.
3.3.1. Specular reflection reductionA dermoscope such as DermLite has a magnifying lens of10×, and 32 light-emitting diodes (LED) of polarized light withthe output of approximately 1800 foot-candles. When usingsuch imaging devices, non-uniform illumination can result inunsatisfactory border detection results. To address this issue,the homomorphic, FFT and high pass filter used to compen-sate for uneven illumination or specular reflection variationsin order to obtain the high contrast lesion images. Homomor-phic filtering [30] is a general technique for nonlinear imageenhancement and correction. It simultaneously normalizesthe brightness across an image and increases contrast.
Let a skin image s(x,y) can be decomposed as the productof illumination I(x,y) and reflectance R(x,y).
s(x, y) = I(x, y).R(x, y) (1)
If the illumination and reflectance can be acted upon sepa-rately, the illumination problem will be solved. Consequently,the log transforms is used to Eq. (1) as:
log[s(x, y)] = log[I(x, y)] + log[R(x, y)]
i.e.
S(u, v) = I(u, v) + R(u, v) (2)
Subsequently, the Fourier transforms is applied to Eq. (2)and filtering is performed in the frequency domain. Then, 2-DFourier transforms is used. The coordinate variables becomeu and v. As H(u,v) is the homomorphic filter function appliedto the illumination and reflectance. After taking the inverseFourier transform and the exponent transforms, an enhancedskin image s(x,y) is obtained. H(u,v) used in this paper has thefollowing form:
H(u, v) = (HH − HL) ×(
1 − exp(
−Bw ×(
D0
D(u, v)
)))+ HL (3)
b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
and
Bw =(
1 +(((
u
2
)2+(v
2
)2)0.5
)˛)−1
(4)
where Bw is the Butterworth filter which can be calculated byEq. (4), D(u, v) is the distance between point (u,v) and point oforigin in a frequency domain; D0 is the normalized transition,and ̨ is the constant parameter having the value in the rangeof ̨ ≥ 1.4 and ̨ ≥ 2.5. Similarly, if the parameters HL and HH
are chosen to be HL < 1and HH > 1, then the filter H(u,v) willdecrease the contribution of the low frequency (illumination)and amplify the contribution of the mid and high frequencies(reflectance).
We can compute the inverse Fourier transforms as:
s(x, y) = I(x, y) + R(x, y) (5)
To reverse the logarithm used at the beginning of the pro-cess, compute an exponential function by:
I(x, y) = exp[I(x, y)] + exp[R(x, y)] (6)
In order to achieve simultaneous dynamic range compres-sion and contrast enhancement, the Butterworth equationsallow such as setting. We used the following Butterworthwith highpass filter, which has been calculated by Eq. (4).It should be noted that after this, the RGB skin image isconverted to grayscale. Then in order to get illuminationcorrection without this conversion effect, used again RGBconversion by previous storing of the color map HSV infor-mation with 0.5 gamma adjustment as shown in (Fig. 3). Inthis figure, we have displayed that the light variation resultsby setting different value of ̨ ≥ 1.5 and ̨ ≥ 1.6 with a con-stant value of HL = 0.5 and HH = 1.4. The fixed values of HL
and HH constants parameters have been chosen from exper-iments by averaging of all threshold values in that dataset.
3.3.2. Air bubbles or dermoscopic-gel reductionDermoscopy images often contain air bubbles or dermoscopic-gel, and it is vital to recognize them to avoid them frominterfering with border detection. These artifacts were foundto possess two individual characteristics: speckles (Fig. 3(b))are nearby in their interiors, and a strong, bright edge (Fig. 4(a))is here at their boundaries.
In order to remove dermoscopic-gels, we utilized adaptiveand recursive weighted median filter developed by Sweet [31].This type of median filter has an edge persevering capability.Although, we used this filter to reduce both speckle noise (seeFig. 3(b)) and air bubbles (see Fig. 4(a)) with radius 3 × 3 pix-els, white level pixel weights for brightest pixel ranges from234 to 245 pixels, and by filter type recursive and adaptive.Although, this practice may create blur effect in some degreewhen choose higher weights and radius size. For this purpose,
we should define criteria to replace those pixels which arebrighter than others. Therefore, we just added one step tomake criteria for pixel replacement strategy based on max-imum and minimum gradients values in blue RGB channel
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15 e5
Fig. 3 – An example of preprocessing step to reduce specular reflection. (a) Original images, (b) after illumination correctionw
annr
FOb
ith ̨ = 1.5, and (c) by ̨ = 1.6, HL = 0.5, HH = 1.4.
s shown in Fig. 4. We selected the blue channel because theoise part in this channel is typically clearer than other chan-els of RGB color space. This image gradient information iselevant to design the local cost function. The cost function
ig. 4 – An example of preprocessing step to reduce dermoscopicriginal ground truth image, (b) maximum gradient magnitude iubble reduction step.
C (i,j) from pixel i to a neighboring pixel j is defined as:
nCn(i, j) = (max G(N) − (g(i) − g(j))1/2)(max G(N) − min G(N) + K)
(7)
-gel or air bubbles by using weighted median filtering. (a)n the red channel, and (c) blue channel output of the air
s i n
y) +∣∣
wise
e6 c o m p u t e r m e t h o d s a n d p r o g r a m
where Cn(i,j) denotes a control cost function of n pixel map,max G(N) and min G(N) are maximum and minimum gradientfunctions of an overall image g, and k denotes intensity value,which is calculated from a histogram of the absolute valuesof the gradient all over the image. Then this brightest pixel’sreplacement criterion function will be defined by using Eq. (8)as:
Cn(i, j) ={
1 if g(x + i)or g(y + j) = [max G(N), min G(N)]
0 otherwise
}(8)
As shown in (Fig. 4), the maximum magnitude of thebrightest pixels in the red and blue channel becomes likespeckle noise. Afterwards these brightest pixels in thisrange were selected to signify air bubbles and removed byneighbourhood pixels. Therefore, when brightest pixels orwhite-level noise is inside the tumor region, then it may dis-turb the tumor structure to some degree due to our pixelreplacement strategy. However, we observed that such aside-effect has negligible influence on border detection accu-racy.
3.3.3. Hair and lines reductionThe detection and removal of hair pixels, blood vessels, andskin lines with ruler markings is an important early step ina dermoscopy image analysis system. As it was pointed outbefore, many skin lesion artifact removal algorithms havebeen proposed in the literature [22–25] just to remove hair pix-els in dermoscopy images. But, these techniques often leavebehind undesirable blurring; disturb the texture of the tumor;and result in color bleeding. Due to these problems, it is verydifficult to use the color diffuse image for further skin tumordifferentiation. In contrast, a new artifact removal algorithmthat focuses on accurate detection of curvilinear artifacts andpays special attention to lesion structure during the removalstage has been introduced by Zhou et al. [20]. They confirmedthat lines can be accurately detected by using explicit curve
modeling [32] and then removed by exemplar-based inpaint-ing. This approach effectively removes artifacts such as rulermarkings and hair, but it has high computational require-ments.
To address these issues, we developed a novel methodfor the removal of hair, blood vessels, and ruler markingsusing line detection and exemplar-based inpainting [27]. Weobserved some properties of these lines’ structures, and thesestructures can be accurately identified using the line detectionscheme. These properties can be defined as thickness, magni-tude, and length other than just direction. The lesion features,on the other hand, can be distinguished because they do nothave these properties. Instead of repeatedly using the local
Ci(x, y) ={Ri(x + d,
0 other
neighborhood average, we utilize an exemplar-based inpaint-ing technique, which actively searches for image regions thathave the similar characteristics to the patch of pixels we aregoing to replace.
b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
In order to detect lines, we propose a line detection proce-dure based on the 2-D derivatives of Gaussian (DOG) [33]. Forcompleteness purposes, here we describe this filter briefly. Theresponse filter for thin and thick line detection is calculatedas:
Ri(x, y) = gi(x, y) ∗ i(x, y) + g�(x, y) ∗ i(x, y) (9)
where i(x,y) is the input image, * is a convolution operator,and gi(x,y) is the Gaussian derivative for thin smooth linescalculated by Eq. (10) along with g�(x,y) is used for thick linesdetection by Eq. (11). Then the 2-D derivative of Gaussian (DOG)for smooth thin lines detection filter [33] is defined as:
gi(x, y) = −x√2�.�3
i
e−x2/2�2
i , for |x| ≤ 3�i,∣∣y∣∣ ≤ Li/2 (10)
This derivative of Gaussian (DOG) efficiently detects linesin all directions. Where �i is the standard deviation of theGaussian function at a scale i, and Li is the span of thefilter in y direction at that scale. Parameter Li is used tosmooth a line along its tangent direction. Then the rotation ofgi(x,y) with angle � is applied by using g�(x′,y′) = gi(x,y), wherex′ = x cos � + y sin � and y′ = y cos � + x sin �. There are also somelines such as blood vessels, which are significantly thickerthan normal ruler markings or hair lines. In order to achievethis characteristic, a Gaussian function with weight towardsthe centre of the image is defined as:
g�(x, y) = cos(�x′
2˛
)cos(�y′
˛
)e−(x2+y2)/2(0.5˛)2 (11)
where ̨ is the fixed parameter equal to 0.5 weighted value.This parameter can be used for reducing noise while stillretaining all the blood vessel information. The appearance ofblood vessels was darker than the neighbouring pixels, so weresolve this by inverting the convolved image. We repeat thisprocess for values of � between 0 and �. The following criterionwas then used to eliminate superfluous curves
Ri(x + d, y)∣∣ , if Ri(x − d, y) > 0 and Ri(x + d, y) < 0
(12)
where d is used to attain the degree around the “zerocrossing”.Since, the advantage of using d parameter for “zerocrossing”is that it can effectively separate smooth thin and thick lines.Moreover, the line direction is determined as the directionwith maximal centreline filter. The line width can be approx-imated by measuring the distance between a local maximumand local minimum along the perpendicular direction of theline. After we got the center, the direction and the width oflines can be interpolated through the information calculatedby thresholding to this filter. By calculating the properties,unwanted curves can be eliminated, which were part of tumorstructure. The line segments extracted after the line point fill-ing with green RGB color is shown in Fig. 5(b).
After detecting the lines, we adopted the exemplar-basedinpainting and patch ordering mechanism. The detail of thisalgorithm has been displayed in [27]. For an optimization solu-tion, we have just added target region lines to be replaced
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15 e7
Fig. 5 – A close-up view for edge and lines detection of image from the original image which has been displayed in Fig. 6. (a)F n an
atpwrmrbiitdttdbsr
oitfhmsat
4
Fusasaeatsi
alse-positive line detections, (b) true-positive lines’ detectio
nd control the iterations. In this implementation, the con-our of the target region is modeled as a dense list of imageoint locations through green lines as shown in Fig. 5(b). Theidth of lines vary depends upon the lines detection algo-
ithm. These lines have been selected automatically by thisethod. Finally, pixels are classified as belonging to the target
egion ˝, the source region � or the remainder of the imagey assigning contrast values to their alpha component. Themage alpha channel is, therefore, updated (locally) at eachteration (t ≤ 144) of the filling algorithm. The value of itera-ion (t) is determined through experiments on the availableatasets. However, in case of much hair surrounded by theumor, we should increase this value. Moreover, if there wereoo many lines close to each other then it is very difficult toetect the accurate lines and also replacement of these linesy exemplar-based inpainting method. In some cases, if hairurrounded completely by the tumor surface, then the pixeleplacement method may not give satisfactory results.
The presented artifact reduction algorithm has been testedn a data set of 147 images. About half of the images have vis-
ble artifacts like this. This algorithm automatically detectshese visible artifacts and removes them. Examples of arti-act reduction results are shown in Fig. 6. Notice that in otherair detection studies, the associated pixels with hairs or rulerarkings are picked up; a few non-artifact pixels are falsely
elected. It should be noted that, in the presence of a largemount of hair, this procedure may damage the structure ofhe tumor.
. Tumors border detection
ollowing artifact reduction, we determine lesion borderssing a modified region-based active contours (RACs) [34]cheme. The segmentation of tumor is also critical due to
fuzzy or irregular border of lesions. In order to segmentingle or multiple tumors from dermoscopic images, therere several studies have been developed, which can be cat-gorized as thresholding, region growing, clustering-based
long with active contour or snake based. In the active con-our framework, they used mostly edge-based or region-basedegmentation techniques. These methods demonstrated sat-sfactory results in case of a single lesion, clear boundariesd filling with green intensity value through proposed work.
and no extra artifacts. In order to avoid these problems, thereare two algorithms have been developed multi-direction gra-dient vector flow (DGVF) and region fusion and narrow bandenergy graph partitioning (NBGP). These methods can effec-tively segment tumors in case of smooth edges, strong hair,and bubbles. However, these techniques have several draw-backs including difficulty with handling multiple lesions andproper choice of parameter values, and requirement of level-set re-initialization.
To address this issue, in this paper, we developed morelocalized region-based active contours (RACs) model withconstraints to automatically segment skin lesions. The imple-mentation of this algorithm is available in [35]. The basicconcept of this model is based on an active contour with-out edges named as Chan–Vese (CV) segmentation technique.The CV form of this snake curve is an energy minimizationmethod, which has been widely used in several computervision or medical applications. We used this model [34] to seg-ment multiple regions simultaneously and its region-basedapproach as compared to edge-based methods. The advan-tages of using this method for skin lesion segmentation is thatit has capabilities to separate heterogeneous objects, insensi-tive to noise, and automatic convergence along with controlof overlapping contours into some extent. In this framework,the localization was presented in improved form, which cansegment multiple tumors with “local energies” when “globalenergies” fail. However, this type of implementation of RACsmethod is also good for initial curve placement, but requiresmore time when contour is placed far away from the tumors.Therefore, good initialization is often essential to increase theconvergence speed. For this purpose, we have defined level setinitialization method by using the iterative histogram thresh-olding. By following this approach, we explain the RACs modelfor multiple tumors segmentation based on CV model in thefollowing sections.
4.1. Automatic level set initialization
In this subsection, a lesion region position is determined andsegmented with the initial detection of a lesion shape withina skin image by performing iterative thresholding methodon luminance values. The monochrome luminance image
e8 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
Fig. 6 – An example of preprocessing step for Hair and outlines such as blood vessels or ruler markers diminution. (a, c, e)nts r
Original ground truth input images whereas (b, d, f) represeis calculated by combining the RGB values according to theNTSC 1953 coefficients. Then iterative thresholding is per-formed by calculating the mini–max intensity histograms onthis luminance image. This type of thresholding is calledglobal thresholding because it used histogram method. Themethod performs thresholding in an iterative way, which hasbeen developed by Ridler and Calvard in 1978. In this simplethresholding method, first the algorithm computes the meanintensity values of an image. Next, by using these thresholdvalues, it computes mean values, which are above and belowto these mean intensity values. Finally, this algorithm applied
iterative method to find out threshold value. This techniqueis visually demonstrated in Fig. 7. In this figure, first we con-vert RGB color image into the luminance image. By calculatingthis type of histogram thresholding, we will get the segmentedFig. 7 – Level set initialization through adaptive thresholding whsteps, (b,f) convert into monochrome luminance image in greyscthresholding, and (d,h) initialize contour.
esults after hair and lines artifacts reduction.
binary image as shown in Fig. 7(c) and (g). In this image, thetumors represent 0 intensity value, while a background is rep-resented by 1 value. Therefore, we can initialize a level setcurve closer towards the tumors as displayed in Fig. 7(d) and(h). Moreover, in order to get lesions smooth boundaries, weutilized the localized RACs in a variational level set framework.
4.2. Lesions segmentation model
We then applied a localized region-based active-contourmodel to lesion boundaries segmentation. This multiple
border detection model is based on the implemented ofChan–Vese (CV) segmentation technique. In this model, wecan integrate varies energies models such as uniform mod-eling energy, the means separation energy, and the histogramere (a,e) are original input RGB images after preprocessingale, (c,g) initial tumors boundaries determine by
i n
selc“
MitzC
rttlwvsetmdg
E
wofcbemxirmb
E
Bw
aitC
c o m p u t e r m e t h o d s a n d p r o g r a m s
eparation energy. However, we used the uniform modelingnergy technique due to its easy implementation. For thisocal energy, each point is considered individually. In order toalculate local energy, the neighborhoods are split into localinterior” and “exterior” by the evolving curves.
Now, we briefly describe this active contour model. Theatlab code of this model has been downloaded from [35] and
ntegrated it into our application by the level set initializa-ion method, and modified curve stability enforcement. Let theero level set of a signed distance function is denoted by �, i.e.
={�(x) = 0
}. Similarly, the interior and exterior of C with
espect to smoothed Heaviside function was also defined inhe same way as [34]. Then the generic force function F allowshe curve to move freely. And B(x,y) function is used to denoteocal mask regions and will be return 1 when the point y is
ithin a ball of radius r centered at x and 0 otherwise. Thealue of radius is very important consideration in this lesionegmentation model. The reason is that by choosing r alsonforces the smoothness inside and outside the contour. Forhis reason, we used 2 pixels value of r parameter. We deter-
ined this value, by experiments. Then by using B(x,y) theefinition of energy function as define in [34] in terms of aeneric force function as:
(�) =∫˝x
ı �(x)
∫˝y
B(x, y).F(I(y), �(y))dy dx (13)
here E(�) represents energy function for the level set, whichnly used points near the contour. By taking ı�(x) Dirac deltaunction outside the force integral represents that we can getapture much broader range of objects. The points selectedy this function still allow contours to merge and split. Forvery point selected by Dirac function ı�(x), we can use B(x,y)ask to ensure that F operates only on local regions about
. Furthermore, in order to keep the curve smooth, a regular-zation term � weighted factor has been also added [37]. Thisegularization term is kept constant throughout the experi-
ents as 0.4. If this value is higher than this, then image wille smoother.
(�) =∫˝x
ı(�(x))
∫˝y
B(x, y).F(I(y), �(y))dy dx
+ �
∫˝x
ı(�(x))∥∥∇�(x)
∥∥dx (14)
y taking the first variation of this energy with respect to �,e obtain the evolution curves calculated by Eq. (15).
∂�
∂t(x) = ı(�(x))
∫˝y
B(x, y).∇�(y)F(I(y), �(y))dy
+ �ı(�(x))div
(∇�(x)∥∥∇�(x)
∥∥)
(15)
From Eq. (15), the only restriction on F is that its first vari-
tion with respect to � can be computed. This means that allmplemented of region-based energies can be integrated intohis framework. However, for easy implementation, we usedV energy model.b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15 e9
4.3. Energy minimization
In this subsection, we introduce the energy function and theway of optimization this energy to numerous tumors withoutoverlapping the contours. We used this energy function as aconstant energy model called CV energy in the same way asdenoted in [34]. Hence, this energy function operates as a con-stant intensity model named uniform modelling energy. Forcurvature flow of this energy, the global region-based energythat uses this local intensity model is given by:
∂�
∂t(x) = ı(�(x))
∫˝y
B(x, y).ı�(y).(
(I(y) − Ux)2
− (I(y) − Vx)2) dy + �ı(�(x))div
(∇�(x)∥∥∇�(x)
∥∥)
(16)
where u and v are the mean of interior and exterior points,which can be used to force the curve. In this localized versionof curvature flow, every point on the curve has moved withbest approximated by local mean ux and vx. These local meanshave been calculated in the same way as presented in [34]. Byintroducing the energies in terms of a signed distance func-tion implement flows in a level set framework as proposed by[36,37]. In order to improve efficiency [38], we only computevalues of in a narrow band around the zero level set. Further-more, in order to segment multiple lesions, we are using themethod of Brox and Weickert [39], which presents an effectivesolution for multiple region segmentation based on the ideaof N computing regions. For this purpose, we allow multiplecontours to compete with each lesion objects on a single inter-face and split until energy does not decrease for zero level set,i.e. (x) = 0 ⇔ �(x) = 0. In this case, the scheme for n level setfunctions �n(x) is given by:
∂�i∂t
(x) = ı�i(x)
⎛⎜⎜⎜⎜⎜⎝ maxı�i(x) > 0
j#i
(ai − rj
)+ minı�i(x) > 0
j#i
(ri − aj
)⎞⎟⎟⎟⎟⎟⎠
(17)
where ai represents forces for level set function to movemultiple contours advance and other parameter ri retreatforce to split and merge contours by preventing overlaps.Then, these parameters have been calculated as:(18)ai =∫˝yB(x, y).ı�i(y).(I(y) − vx)
2dy + �2div
(∇�i(x)
‖∇�i(x)‖)
and the retreat
forces (ri) are given by
ri =∫˝y
B(x, y).ı�i(y).(I(y) − ux)2dy + �
2div
(∇�i(x)∥∥∇�i(x)
∥∥)
(19)
We can therefore, enforce the stability by using theCourant–Friedreichs–Lewy (CFL) condition [40], that moves the
curve very fast in case of multiple tumors. In CFL condition,this also asserts that the numerical waves should propagateat least as fast as the physical waves within a time frame. Thiscondition is a coarse measure of stability and convergence of
s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
Table 1 – Parameters values used for experiments andcomparisons.
Methods IT TS (s) r (pixels) � ˛
DGVF 280 0.2 1.5 0.5GPAC 320 0.2 1.5 0.5RACs 175 0.2 2 0.4 0.5
IT: maximum iterations, TS (s): times step in seconds, RACs: region-based active contours, DGVF: directional gradient vector flow, GPAC:
e10 c o m p u t e r m e t h o d s a n d p r o g r a m
curves. However, this stability can be achieved by using a muchmore restrictive CFL condition, although this is computation-ally costly. Various methods to enforce curve stability such astemporal discretization, Runge–Kutta methods, etc. have beenproposed in the literature. Although these methods may be tootime consuming to enforce this condition for multiple objects.In our case, we modify this CFL condition in the RACs modelby the maximum smoothing constraint on the level set curves,which is defined as:
�n(x) = �n(x) + D(�) × ∂�
∂t(x) (20)
and derivation of level set function is denoted by D(�) andcalculated as
D(�) = Sm(max
(∂�∂t (x)
)+ ε) (21)
where Sm is a speed distance function for CFL condition andthe maximum flow of curvature depends upon the cost ofintroducing this parameter (Sm). By choosing an appropriateSm parameter value is very important to increase the speedof computations for faster curvature flow. We presented amethod (Sm) by using the mean values of interior and exteriorpoints in each curvature vectors which is given by:
Sm =∣∣∣∣255 − �n(x) + ux
vx
∣∣∣∣ (22)
Therefore, the speed distance function and maximumsmoothing constraints improves the CFL function. Conse-quently, this will reduce the iterations to converge thecontours towards the actual boundaries. By updating signeddistance function �n(x), we can evolve the curves, which movein a smooth manner. Moreover, we used the fast marchingscheme to re-initialize the contours [41].
5. Evaluation results
This algorithm was applied to 320 images with type of pig-mented and non-pigmented lesions. These were 24-bit RGBcolor images with dimensions 640 × 480 pixels. Manual trac-ing of the tumors was also performed by an experiencedexpert serving as a gold standard to evaluate the performanceof the proposed multi-phase region-based active contours(RACs). The algorithm was preliminarily implemented in Mat-lab 7.6.0.324® (The Mathworks, Natick, MA).
The proposed procedures took in total 2475 ms to seg-ment (including 175 numbers of iterations) and 3328 ms forartifact reduction (including 144 numbers of iterations forexemplar-based inpainting) for all tumors present in an imageof 640 × 480 (rescaled to 450 × 360) dimensions on average.On average, it took region-based active contour 119.27 iter-ations to reach a stable border for dermoscopic images. Allcomputations were performed on a 2.3 GHz dual-core 64-bitAMD processor with 4 GB DDR2 RAM, running Windows 2003
server edition. As discussed before, quantitative evaluationswere also performed on segmentation by using four statisticalmetrics. These border detection results were compared withthe reference images (GT).graph based active contours, r (pixels): radius in pixels, �: regular-ization parameter, ˛: elasticity of curve.
For experimental step, the parameters used by eachscheme were almost identical in all the cases and conductedtests. As pointed out before, we enhanced each image usingthe proposed preprocessing method. The Homomorphic fil-tering, weighted median filter with control function and anexemplar-Based image inpainting for lines’ reduction filterswere applied to reduce the artifact’s effects. In order to find outthe significance of our segmentation method, we comparedthe border detection results with existing state-of-the-arttechniques. And, we should also set up an environment forfixing parameters and threshold values.
5.1. Experimental setup
In order to find out the efficiency of our method, we usedthe Matlab implementation of these algorithms GVF [42],graph-partitioning active contours (GPAC) [43], and color seg-mentation (Jseg) [44] algorithms. However, in case of multipleGVF snake, we have just added the directional gradient infor-mation with a varying direction as done by Tang [13]. We calledthis snake model also DGVF. Now, we will discuss the initialparameters setting of each technique.
The overall experimental parameters have been displayedin Table 1. Although, mostly parameters for deformable mod-els are the elasticity, rigidity, viscosity, and regularizationparameters were ̨ = 0.5, ̌ = 0.1, = 0.1 and � = 1.5, respectively.For DGVF and GPAC snakes, we used the parameter valuesgiven in Table 1. For initialization point of view of thesesnakes, we used the iterative thresholding method on lumi-nance channel. In case of Jseg algorithm, we used whole RGBimage to segment the tumors based on color and texture infor-mation. Finally, we extracted the ROI boundaries of tumors,which were detected by Jseg, DGVF, and GPAC algorithms forperformance evaluation.
5.2. Statistical analysis
Consequently, we employed four dissimilar metrics to quan-tify the boundary differences by recent systematic study [21]as well as Mean border error (MBE). The Hammoude dis-tance (HM), the true detection rate (TDR) and the false positiverate (FPR) are an area based metrics; in addition, the Haus-dorff distance (HD) which measures the distance between theboundaries in pixels. We denote (AS) for automatic segmenta-tion and (GT) denote ground truth segmentation obtained by
the expert. And, we also calculated MBE, which shows the pix-els of AS disagree with the GT. These metrics are calculated asfollows:
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15 e11
Table 2 – Comparison results of the segmentation with proposed region-based active contour for poly skin lesionsthrough 6 miscellaneous cases.
HM (%) TDR (%) FPR (%) HD (pixels) MBE (%)
Methods case-1 50 Benign melanocytic lesionsDGVF 12.74 78.28 12.01 47.81 08.91JSeg 15.67 79.20 22.81 52.06 19.03GPAC 13.45 83.10 14.82 32.67 10.03Proposed RACs 10.25 92.18 4.26 28.95 06.00
Methods case-2 70 Malignant melanoma lesionsDGVF 17.50 83.00 10.11 59.61 14.91JSeg 19.67 79.70 21.82 78.21 18.16GPAC 14.67 85.70 16.82 49.45 11.51Proposed RACs 11.37 93.18 03.78 35.39 02.04
Methods case-3 40 Basal cell carcinoma lesionsDGVF 14.68 76.24 24.48 78.61 16.76JSeg 16.38 79.34 31.78 44.72 21.47GPAC 13.10 83.69 11.69 42.34 18.49Proposed RACs 13.21 94.10 05.48 31.28 08.89
Methods case-4 40 Merkel cell carcinoma lesionsDGVF 21.34 81.28 23.53 47.50 13.41JSeg 29.67 67.50 25.90 62.20 19.18GPAC 19.61 77.43 19.60 52.36 16.26Proposed RACs 14.25 88.17 06.13 24.95 07.02
Methods case-5 90 Seborrhoeic keratosis lesionsDGVF 31.34 78.28 14.57 37.66 09.25JSeg 35.00 71.20 24.12 42.31 15.52NBGP 25.67 76.00 19.22 52.45 13.70Proposed RACs 13.00 79.97 04.03 21.75 02.01
Methods case-6 30 Non-melanocytic lesionsDGVF 22.11 81.24 11.27 68.17 09.62JSeg 24.64 64.30 14.29 82.28 14.15GPAC 17.28 79.49 13.69 44.10 08.35Proposed RACs 14.68 90.18 10.10 36.27 04.48
Total average for proposed RAC 92.17 05.62 04.58
Results of the segmentation results calculated for the 320 images. The region-based active contour shows the best performance for segmentationof skin lesions where (HM: Hammoude distance, TDR: true detection rate, FPR: false positive rate, HD: Hausdorff distance, MBE: mean border
and
wnb
H
Tb
T
Ft
) = m
T, AS) = max{
max d(gti, AS), max d(ASj, GT)} (26)
error), lesions as: DGVF: directional gradient vector flow, JSeg: textureregion-based active contours.
HM metric makes a pixel by pixel comparison in such aay; pixels classified as lesion by the dermatologist that wereot classified as such by the AS and pixels classified as lesiony the automatic segmentation as:
ammoude distance (HM) = #(AS ∪ GT) − #(AS ∩ GT)#(AS ∪ GT)
(23)
DR metric measures the rate of pixels classified as lesion byoth the automatic and the medical expert segmentation as:
rue detection rate (TDR) = #(AS ∩ GT)#(GT)
(24)
Hausdorff distance (HD)
d(gti, AS
i.e. HD(G
PR metric measures the rate of pixels classified as lesion byhe automatic segmentation that were not classified as lesion
color segmentation, GPAC: graph partitioning active contours, RACs:
by the medical expert:
False positive rate (FPR) = #(AS ∩ GT)#(GT)
(25)
HD metric finds the largest distance between the bound-ary points. This metric, like the HM, does not distinguishbetween distance from inside or from outside of the ref-erence curve. Where GT =
{gt1, gt2, gt3, ..., gtm
}and SR =
{AS1, AS2, AS3, ..., ASn} denote the set of points belonging tocontour GT and SR contour, respectively. The distance from gti
to its closest point (DCP) in SR is given by:
inj
∥∥ASj − gti∥∥
iLikewise, mean border error in percentage is calculated usingEq. (27). Where ⊕ is an exclusive-OR operation that gives the
e12 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
bor
Fig. 8 – Unsupervisedpixels for which the AS and GT disagree as:
Mean border error (MBE) = area(AS) ⊕ area(GT)area(GT)
× 100 (27)
As a result, these five statistical measures are adequateto quantify border detection errors. Table 2 shows the borderdetection performance for each method in every case of (1)benign melanocytic, (2) malignant melanoma, and (3) basalcell carcinoma, in addition to, (4) merkel cell carcinoma, (5)seborrhoeic keratosis, and (6) non-melanocytic lesions. The
best result according to on average mean border error is the ASwith a score of 4.58%. The best true detection rate is achievedby the AS of RACs technique is 92.17%. The average best FPRis obtained by the proposed method that is (5.62%) which isFig. 9 – Border detection results of primary 3-cm merkel cell carcwoman with (a) represents original MCC tumor, (b) the outline m(GT), segmentation results by use of (c) likely proposed RACs, (d)
der detection results.
reduced from 11.32%. The HD behaves similarly to the HM.The TDR is especially important since it measures the fractionof lesion pixels, which are detected in all images.
As presented in (Fig. 8), the border detection results havebeen highlighted proposed by our method but for clarity ofcomparisons and explanation of steps; consider this case ofmerkel cell carcinoma. We consider an experimental exampleof one of the results, which have been displayed in (Fig. 9). Inthis experiment, we explored the difficulty in finding accurateimage edges where the transition between the lesion and thehealthy skin was very smooth. To demonstrate the results for
clinical images, we selected this 3 cm merkel cell carcinoma(MCC) on the dorsal left forearm of a 55-year old woman.The border detection difficulty is encountered even fromthe point of view of detectors found in the literature one, can
inoma (MCC) on the dorsal left forearm of a 55-year-oldanually marked by a dermatologist illustrate ground truthGPAC, (e) DGVF (f) JSeg algorithms.
c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15 e13
Fig. 10 – For Border detection a challenging task in case of: (a) skin-color lesion of acral volar, (b) non-pigmented basal cellc tagio
ewibblps
6
Iuitmmcwp
blgitswIomllottiplabfioa
arcinoma, (c) malignant melanoma, and (d) molluscum con
xamine that many of these methods do not give good resultshen applied to images with this same degree of complex-
ty. This is an apparent when we compared the result obtainy the proposed method (Fig. 9 (c)) with the result obtainedy other three systems. The work proposed here only exactesion edges, while other three methods (DGVF, JSeg and GPAC)rocess many undesirable edges, which may be part of thekin.
. Discussion and Conclusion
n this paper, we have presented an algorithm designed fornsupervised border detection for numerous tumors, which
s also very useful for the elimination of artifacts. We haveested to this method not only melanomas but also in pig-
ented and non-pigmented (basal cell, squamous cell, anderkel cell carcinomas) lesions. The algorithm is robust, effi-
ient and intuitive. This type of border detection techniqueith artifact reduction steps has not been proposed in theast.
This segmentation algorithm delivers good results for theig majority of the tested images; there are two groups of
esions that are not almost correctly border detected. Oneroup is when there are areas belonging to the lesion thats lighter than the skin and is situated on the boundaries ofhe lesion. In these cases, the light area will be regarded askin sees (Fig. 10) for more details. The other group is imagesith more than one skin lesion are much closed to each other.
n that situation, our border detection can accurately trackuter borders but this cannot be precisely separate lesionsuch clearer. If the lesions are well separated, then the prob-
em can be solved by letting the physician give the number ofesions as input. When the lesions lie close, the result is anver-segmentation. It may be possible in these cases to dividehe image pixels into three classes, based on the location ofhe skin mode, the first and the second minimum. Moreover,mages in the training data set sample have been taken fromeople with fair (white) skin. As this is the group, which is most
ikely to develop melanoma. Even if, there is no guarantee that person with darker skin never will develop melanoma. Theorder detection algorithm has to be tested on images taken
rom a person with darker skin, to verify that the algorithm isndependent of skin-color. If it shows that the method worksnly on images from fair skinned people, this issue has to bedjusted.sum tumors.
From the evaluation of the 320 test images, the problemis that the border detection often fails when the lesion con-sists of the area diffused with skin or not well separated byskin. The hair-removing algorithm performs the way it should;removing the hair that otherwise would affect the borderdetection. Testing the border detection algorithm on a samplewith more malignant melanomas needs to be done to assurethe algorithm performs well in all cases. The evaluation ofthe performance of the border deduction is not very limited.But, at least one more expert dermatologist should evaluatethe borders, and more lesions from the test sample shouldbe included in the evaluation. Moreover, from Fig. 10(c), it isdifficult to differentiate between skin lesion and non-lesionobjects from this model, mostly in clinical images. In additionto this, to segment multiple tumors from single initializationof contour is also hard as shown in Fig. 10(d). Accordingly,this segmentation modal can be enhanced by shape priorconstraint, automatic contours initialization and selection ofconstant parameters.
These techniques evaluated unsupervised border detec-tion of skin lesions in both clinical and dermoscopic view.This method was applied to a set of 320 dermoscopic imagesfrom the database including 50 benign melanocytic, 70 malig-nant melanoma, and 40 basal cell carcinoma, with 40 merkelcell carcinoma, 90 keratosis, and 30 non-melanocytic lesions.These images were segmented manually by an expert derma-tologist. The output of the automatic border detection wascompared with the manually segmented images using fivemetrics (Hammoude distance, TDR, FPR, and Hausdorff dis-tance with MBE) to other three schemes (JSeg, GPAC, and DGVF).The TDR was considered the most relevant metric from a clin-ical point of view. In addition, this algorithm is robust in thisdata set since the number of average mean border error is 4.58.The best results were obtained by the presented algorithmwith achievements. From this figure, the results obtained incase of mostly multiple tumors present in the selected dataset.
There are several potential improvements, which havebeen developed in this study. These novel ideas related to skincancer analysis for dermatologists includes: (1) a new and effi-cient method for analyzing skin images with multiple lesions(2) fastest convergence in case of poly lesions with the level set
initialization through iterative thresholding, and modify CFLcondition by maximum smoothing constraints (3) to reducethe artifact noises and recommended a solution for betterreduction of camera flash, bubbles and lines such as skin lines
s i n
r
e14 c o m p u t e r m e t h o d s a n d p r o g r a m
or ruler markings, hair, and blood vessels (4) in case of BCC andMCC for which early and correct diagnosis is also important.They are as well included in this proposed method.
In addition, we have performed a separate evaluation forthe border detection of varies types of lesions. This kind ofcomparisons has been also addressed [21] to some extent; theyonly considered melanomas and basal cell carcinoma lesions.These studies displayed increased TDR for the melanomacases and also an increased FPR in the case of basal cell carci-noma. Thus, it is interpreted as the low contrast between thelesion and the skin which characterizes this type of lesion.We found that both preprocessing and segmentation work isrobust and useful for any type of lesion border detection in acomputer-aided diagnosis system to assist the clinical diag-nosis of dermatologists.
Conflict of interest statement
All authors in this paper have no potential conflict of interest.
Acknowledgements
This study was supported by the Chinese Scholarship Coun-cil (CSC) (grant no. 2008GXZ143). We would like to thankShawn Lankton for providing us the implementation of localregion-based active contours to start the basic segmentationalgorithm. We would also like to thank Assoc. Prof. AmandaOakley for providing us patient‘s skin cancer database fromthe New Zealand Dermatological Society.
e f e r e n c e s
[1] Skin cancer facts and figures, 2009. Available from:http://www.skincancer.org/Skin-Cancer-Facts.
[2] I. Papamichail, N. Nikolaidis, C. Nikolaidis, I. Glava, K.Lentzas, Marmagkiolis, et al., Merkel cell carcinoma of theupper extremity: case report and an update, J. Surg. Oncol. 6(32) (2008).
[3] A. Schwartz, W.C. Lambert, The merkel cell carcinoma: a50-year retrospect, J. Surg. Oncol. 89 (1) (2005) 1–4.
[4] G. Argenziano, H.P. Soyer, S. Chimenti, R. Talamini, R.Corona, F. Sera, et al., Dermoscopy of pigmented skinlesions: results of a consensus meeting via the Internet, J.Am. Acad. Dermatol. 48 (5) (2003) 679–693.
[5] M.E. Celebi, H.A. Kingravi, B. Uddin, et al., A methodologicalapproach to the classification of dermoscopy images,Comput. Med. Imaging Graph. 31 (2007) 362–373.
[6] P. Rubegni, M. Burroni, G. Cevenini, et al., Digitaldermoscopy. Analysis and artificial neural network for thedifferentiation of clinically atypical pigmented skin lesions:a retrospective study, J. Invest. Dermatol. 119 (2002) 471–474.
[7] H. Blum, U. Luedtke, Ellwanger, et al., Digital image analysisfor diagnosis of cutaneous melanoma. Development of ahighly effective computer algorithm based on analysis of 837melanocytic lesions, Brit. J. Dermatol. 151 (2004) 1029–1038.
[8] M. Burroni, P. Sbano, G. Cevenini, et al., Dysplastic naevusvs. in situ melanoma: digital dermoscopy analysis, Brit. J.
Dermatol. 152 (2005) 679–684.[9] S.V. Patwardhan, A.P. Dhawan, P.A. Relue, Classification ofmelanoma using tree structured wavelet transforms,Comput. Methods Programs Biol. 72 (2003) 223–239.
b i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15
[10] M.E. Celebi, H.A. Kingravi, H. Iyatomi, et al., Border detectionin dermoscopy images using statistical region merging, SkinRes. Technol. 14 (2008) 347–353.
[11] M.E. Celebi, H. Iyatomi, G. Schaefer, W.V. Stoecker, Lesionborder detection in dermoscopy images, Comput. Med.Imaging Graph. 33 (2009) 148–153.
[12] M.E. Celebi, Y.A. Aslandogan, W.V. Stoecker, et al.,Unsupervised border detection in dermoscopy images, SkinRes. Technol. 13 (2007) 454–462.
[13] J. Tang, A multi-direction GVF snake for the segmentation ofskin cancer images, Pattern Recognit. 42 (6) (2009) 1172–1179.
[14] X. Yuan, N. Situ, G. Zouridakis, A narrow band graphpartitioning method for skin lesion segmentation, PatternRecognit 42 (6) (2009) 1017–1028.
[15] P. Schmid, Segmentation of digitized dermatoscopic imagesby two-dimensional color clustering, IEEE Trans. Med.Imaging 18 (1999) 164–171.
[16] B. Erkol, R.H. Moss, R.H. Stanley, et al., Automatic lesionboundary detection in dermoscopy images using gradientvector flow snakes, Skin Res. Technol. 11 (2005) 17–26.
[17] H. Iyatomi, H. Oka, M. Saito, et al., Quantitative assessmentof tumor extraction from dermoscopy images andevaluation of computer-based extraction methods forautomatic melanoma diagnostic system, Melanoma Res. 16(2006) 183–190.
[18] H. Iyatomi, H. Oka, M.E. Celebi, et al., An improvedinternet-based melanoma screening system withdermatologist-like tumor area extraction algorithm,Comput. Med. Imaging Graph. 32 (2008) 566–579.
[19] D.D. Gómez, C. Butakoff, B.K. Ersboll, W.V. Stoecker,Independent histogram pursuit for segmentation of skinlesions, IEEE Trans. Biomed. Eng. 55 (2008) 157–161.
[20] H. Zhou, G. Schaefer, A. Sadka, M.E. Celebi, Anisotropicmean shift based fuzzy C-means segmentation ofdermoscopy images, IEEE J. Sel. Top. Signal. 3 (2009) 26–34.
[21] M. Silveira, J.C. Nascimento, J.S. Marques, comparison ofsegmentation methods for melanoma diagnosis indermoscopy images, in: Proceedings of IEEE J. Sel. Top.Signal. 3 (February) (2009) 35–45.
[22] M.G. Fleming, C. Steger, J. Zhang, Techniques for a structuralanalysis of dermatoscopic imagery, Comput. Med. ImagingGraph. 22 (5) (1998) 375–389.
[23] P. Wighton, T.K. Lee, M.S. Atkinsa, Dermoscopic hairdisocclusion using inpainting, in: Proceedings of the SPIEMedical Imaging 2008 Conference, 6914, pp.691427–691427-8.
[24] H. Zhou, M. Chen, R. Gass, J.M. Rehg, L. Ferris, J. Ho, et al.,Feature-preserving artifact removal from dermoscopyimages, in: Proceedings of the SPIE Medical Imaging 2008Conference, 6914, 69141B–69141B-9.
[25] T.K. Lee, V. Ng, R. Gallagher, A. Coldman, D. McLean,Dullrazor: a software approach to hair removal from images,J. Comput. Biol. Med. 27 (6) (1997) 533–543.
[26] A. Criminisi, P. Perez, and K. Toyama Object removal byexemplar-based inpainting, in: Proceedings of Comput. Vis.Pattern Rec., Madison, WI, June, 2003, pp. 721–728.
[27] A. Criminisi, P. Perez, K. Toyama, Region filling and objectremoval by exemplar-based image inpainting, IEEE Trans.Image Process. 13 (9) (2004).
[28] F.-Y. Xie, S.-Y. Qin, Z.-G. Jiang, R.-S. Meng, PDE-basedunsupervised repair of hair-occluded information indermoscopy images of melanoma, Comput. Med. ImagingGraph. 33 (2009) 275–282.
[29] Dermatologic Image Database, University of Auckland, NewZealand. Available from:
http://dermnetnz.org/doctors/dermoscopy-course/(accessed on 15.06.2009).
i n b
c o m p u t e r m e t h o d s a n d p r o g r a m s[30] H.G. Adelmann, ‘Butterworth equations for homomorphicfiltering of images, J. Comput. Biol. Med. 2 (28) (1998) 169–181.
[31] M.R. Sweet, adaptive and recursive median filtering.Available from:http://www.easysw.com/∼mike/gimp/despeckle.html(accessed on 22.07.2009).
[32] C. Steger, Unbiased detector of curvilinear structures, IEEETrans. Pattern Anal. 20 (2) (1998) 113–125.
[33] Q.L.L. Zhang, J. You, D. Zhang, P. Bhattacharya, Dark linedetection with line width extraction, in: Proceedings of theIEEE International Conference on Image Processing (2008)621–624.
[34] S. Lankton, A. Tannenbaum, Localizing region-based activecontours, IEEE Trans. Image Process. 17 (11) (2008) 2029–2039.
[35] Active Contour Segmentation matlab implementation.Available from:http://www.mathworks.com/matlabcentral/fileexchange/19567-active-contour-segmentation (accessed on
26.01.2009).[36] J.A. Yezzi, A. Tsai, A. Willsky, A fully global approach toimage segmentation via coupled curve evolution equations,J. Vis. Commun. Image Represent. 13 (1) (2002) 195–216.
i o m e d i c i n e 1 0 4 ( 2 0 1 1 ) e1–e15 e15
[37] S. Osher, R. Fedkiw, Level Set Methods and Dynamic ImplicitSurfaces, Cambridge Univ. Press, New York, 2003.
[38] J.A. Sethian, Level Set Methods and Fast Marching Methods,second ed., Springer, New York, 1999.
[39] T. Brox, J. Weickert, Level set based image segmentationwith multiple regions, Pattern Recognit. (2004)415–423.
[40] R. Courant, E. Issacson, E. Rees, On the solution of nonlinearhyperbolic differential equations by finite differences,Commun. Pure Appl. Math. 5 (1952) 243–255.
[41] M. Sussman, P. Smerka, S. Osher, A Level set approach forcomputing solutions to incompressible two-phase flow, J.Comput. Phys. 119 (1994) 146–159.
[42] Gradient Vector Flow (GVF). Available from:http://www.iacl.ece.jhu.edu/static/gvf/ (accessed on3.09.2009).
[43] Graph Partitioning Active Contours (GPAC). Available from:http://vision.ece.ucsb.edu/∼lbertelli/soft GPAC.html
(accessed on 3.09.2009).[44] Color and texture segmentation (JSEG). Available from:http://vision.ece.ucsb.edu/segmentation/jseg/ (accessed on3.09.2009).
