Volume 11 Supplement 6
Proceedings of the Seventh Annual MCBIOS Conference. Bioinformatics: Systems, Biology, Informatics and Computation
Lesion detection in demoscopy images with novel densitybased and active contour approaches
 Mutlu Mete^{1}Email author and
 Nikolay Metodiev Sirakov^{1, 2}
DOI: 10.1186/1471210511S6S23
© Mete and Sirakov; licensee BioMed Central Ltd. 2010
Published: 7 October 2010
Abstract
Background
Dermoscopy is one of the major imaging modalities used in the diagnosis of melanoma and other pigmented skin lesions. Automated assessment tools for dermoscopy images have become an important field of research mainly because of inter and intraobserver variations in human interpretation. One of the most important steps in dermoscopy image analysis is the detection of lesion borders, since many other features, such as asymmetry, border irregularity, and abrupt border cutoff, rely on the boundary of the lesion.
Results
To automate the process of delineating the lesions, we employed Active Contour Model (ACM) and boundarydriven densitybased clustering (BDDBSCAN) algorithms on 50 dermoscopy images, which also have ground truths to be used for quantitative comparison. We have observed that ACM and BDDBSCAN have the same border error of 6.6% on all images. To address noisy images, BDDBSCAN can perform better delineation than ACM. However, when used with optimum parameters, ACM outperforms BDDBSCAN, since ACM has a higher recall ratio.
Conclusion
We successfully proposed two new frameworks to delineate suspicious lesions with i) an ACM integrated approach with sharpening and ii) a fast boundarydriven densitybased clustering technique. ACM shrinks a curve toward the boundary of the lesion. To guide the evolution, the model employs the exact solution [27] of a specific form of the Geometric Heat Partial Differential Equation [28]. To make ACM advance through noisy images, an improvement of the model’s boundary condition is under consideration. BDDBSCAN improves regular densitybased algorithm to select query points intelligently.
Introduction
Melanoma is the fifth most common malignancy in the United States. Malignant melanoma, the most deadly form of skin cancer, is one of the most rapidly increasing cancers in the world. An estimated amount of 8,441 deaths out of 68,720 cases were recorded in the United States in 2009 [1]. Early diagnosis is particularly important, since melanoma can be cured with a simple excision if detected early.
Dermoscopy, a noninvasive skin imaging technique, has become one of the most important instruments in the diagnosis of melanoma and other pigmented skin lesions. It involves optical magnification of the regionofinterest, which makes subsurface structures more easily visible compared to what can be seen via the nakedeye [2]. This, in turn, improves screening characteristics and provides greater differentiation between difficult lesions such as pigmented Spitz nevi and small, clinically equivocal lesions [3]. However, it has also been demonstrated that dermoscopy may actually lower the diagnostic accuracy in the hands of inexperienced dermatologists [4]. Therefore, new frameworks for the understanding of computerized images are needed to minimize the diagnostic errors that result from the difficulty and subjectivity of visual interpretation [5][6].
For melanoma investigation, delineation of the regionofinterest is the first and most important step in the computerized analysis of skin lesion images for many reasons. First of all, the border structure provides important information for accurate diagnosis. Asymmetry, border irregularity, and abrupt border cutoff are just a few of the clinical features calculated based on the border lesion. Furthermore, the extraction of other important clinical indicators such as atypical pigment networks, globules, and bluewhite areas (irregular, structureless areas of confluent blue pigmentation with an overlying white groundglass film) critically depends on the border detection [7][8].
In the literature, many algorithms were proposed regarding border detection in dermoscopy images. These include the PCT/median cut algorithm [9], adaptive thresholding in the first image plane of the PCT [10], thresholding in the blue image plane[11], kmeans clustering [12], splitandmerge [9][13], a segmentation technique based on a Markov random field (MRF) image model [14], and a nonlinear diffusion technique [12]. Furthermore Schmid et al. [15] proposed an algorithm based on color clustering. In their study, a twodimensional histogram is calculated first from the first two principal components of CIE L*u*v* color space. The histogram is then smoothed, and initial cluster centers are obtained from the peaks using a perceptron classifier. In the final step, the image of the lesion is segmented using a modified version of the fuzzy Cmeans clustering algorithm. Gao et al. [12] created two methods: one based on stabilized inverse diffusion equations, a form of nonlinear diffusion, and another one based on Markov random fields in which the model parameters are estimated using the mean field theory.
The active contour approach was developed in late eighties by the work of Kass, Witkin, Terzopoulos [16], Osher and Sethian [17] and quickly became very popular, providing excellent results in almost all areas of its application. One such area is biomedical image analysis with a number of strong methods and algorithms. Without underestimating the contributions of the other works, Acton and Ray wrote one important book that analyzes the advantages and limitations of the active contour methodology [18].
From the angle of cluster boundaries, Lee and Castro [19] introduced a new algorithm of polygonization based on the boundaries of resulting point clusters. Recently, Nosovskiy et al. [20] used an adaptive function approach to find the boundary of a cluster in order to infer accurate boundaries between close neighboring clusters. These two works principally focus on the boundaries of finalized data groups (clusters), which is not the case for our present work.
Border error, precision and recall measures for images in the dataset.
ACM  BDDBSCAN  

Img. ID  Border Error  Precision  Recall  Border Error  Precision  Recall 
1  13.3%  0.76  0.94  8.2%  0.98  0.79 
2  6.7%  0.92  0.91  8.0%  0.93  0.86 
3  3.8%  0.86  0.96  4.9%  0.89  0.85 
4  8.0%  0.93  0.84  6.2%  1.00  0.82 
5  3.9%  0.92  0.95  4.6%  1.00  0.83 
6  11.0%  0.75  0.99  3.9%  0.96  0.91 
7  22.2%  0.95  0.27  3.2%  1.00  0.87 
8  6.2%  0.80  0.90  3.4%  1.00  0.82 
9  3.8%  0.87  0.98  2.2%  1.00  0.91 
10  2.8%  0.79  0.99  0.9%  1.00  0.91 
11  5.7%  0.99  0.66  6.5%  1.00  0.61 
12  10.0%  1.00  0.79  14.8%  1.00  0.70 
13  3.1%  1.00  0.82  5.9%  1.00  0.67 
14  4.6%  1.00  0.84  6.8%  1.00  0.76 
15  7.6%  1.00  0.60  6.0%  1.00  0.67 
16  4.4%  1.00  0.80  6.4%  1.00  0.71 
17  4.5%  0.99  0.89  8.8%  1.00  0.78 
18  5.3%  0.97  0.91  12.6%  1.00  0.73 
19  5.5%  0.99  0.85  8.6%  1.00  0.76 
20  5.0%  1.00  0.82  5.7%  1.00  0.79 
21  8.2%  1.00  0.76  9.0%  1.00  0.74 
22  4.5%  0.99  0.80  8.0%  1.00  0.65 
23  5.3%  0.99  0.87  10.6%  1.00  0.75 
24  6.6%  0.99  0.85  11.3%  1.00  0.74 
25  7.0%  1.00  0.85  10.8%  1.00  0.77 
26  2.3%  0.98  0.96  4.2%  1.00  0.88 
27  3.2%  0.99  0.89  4.0%  1.00  0.85 
28  7.9%  1.00  0.71  8.0%  1.00  0.71 
29  4.7%  0.99  0.84  8.0%  1.00  0.73 
30  1.5%  0.97  0.96  3.2%  1.00  0.85 
31  5.3%  1.00  0.81  7.3%  1.00  0.74 
32  2.5%  0.98  0.90  3.6%  1.00  0.84 
33  11.4%  0.68  1.00  3.0%  1.00  0.87 
34  5.6%  1.00  0.83  10.9%  1.00  0.68 
35  3.9%  0.99  0.88  7.4%  1.00  0.78 
36  3.6%  0.93  0.98  4.2%  1.00  0.88 
37  14.6%  1.00  0.61  14.9%  1.00  0.60 
38  5.3%  1.00  0.87  9.4%  1.00  0.77 
39  3.2%  0.91  0.92  4.6%  1.00  0.75 
40  1.2%  0.98  0.94  2.9%  1.00  0.81 
41  8.8%  0.79  0.76  6.5%  0.90  0.74 
42  5.1%  1.00  0.77  5.8%  1.00  0.74 
43  4.5%  0.99  0.82  5.6%  1.00  0.77 
44  1.7%  0.97  0.92  2.9%  1.00  0.82 
45  7.4%  0.99  0.74  8.3%  0.89  0.79 
46  4.6%  0.95  0.92  6.3%  0.98  0.83 
47  30.8%  0.39  1.00  3.2%  1.00  0.79 
48  0.7%  0.99  0.95  2.4%  1.00  0.79 
49  16.9%  0.53  0.96  4.6%  1.00  0.74 
50  5.0%  0.98  0.84  8.8%  1.00  0.71 
Min  0.7%  0.39  0.27  0.9%  0.89  0.60 
Max  30.8%  1.00  1.00  14.9%  1.00  0.91 
Mean (µ)  6.6%  0.93  0.85  6.6%  0.99  0.78 
Stdev (σ)  0.05%  0.12  0.13  0.03%  0.03  0.08 
Results and discussion
Dataset
As we mentioned above, the two methods are tested on a set of 50 dermoscopy images obtained from the Edra Interactive Atlas of Dermoscopy [2]. These are 24bit RGB color images with dimensions ranging from 577 × 397 pixels to 1921 × 1285 pixels. The benign lesions include nevocellular nevi and dysplastic nevi.
Preprocessing for BDDBSCAN
In the next step, a grayscale image is segmented with the intermeans algorithm, developed by Ridler and Calvard [23]. Actually, this is an iterative technique for choosing a threshold. Similar to many other segmentation algorithms, it takes the image histogram and initially assigns a threshold T. Compiling a histogram is an efficient and simple mean of representing an image as a onedimensional array based on quantity of color samples. A histogram is the graphical representation of a resultant array using a bar chart or different visual modalities. Throughout this study, the histogram is denoted by y_{0}, y_{1}, y_{2}, …, y_{n}, where y_{i} is the number of pixels having grayvalue i. The maximum value for the subscript i = {0, 1, …, n} is 255 for all images used in the present study. Thus the threshold T = {0, 1, …, n} is used to split the histogram into two groups. All pixel values less than or equal to T are assigned to one group, and those that are greater than T are assigned to the second group.
Intermeans segmentation
Preprocessing for ACM
Results
In addition to border error, we also reported precision (positive predictive value) and recall (sensitivity) for each experimental image in Table 1. Precision and recall are defined as respectively. In addition, we can express border error with these widely used definitions,
Note that border error and accuracy measurements run over number of pixels in particular regions. Analogously, Area(.) function returns the number of positive pixels in a thresholded image.
Both methods presented in this study give the same average border error of 6.6% in all 50 images, which is a very promising result in terms of a computeraided framework. This means that, on average, given a dermoscopy image, either algorithm will identify 93.4% of the targeted skin lesions correctly. However, ACM’s standard deviation is slightly greater than that of BDDBSCAN (0.05 > 0.03), indicating that BDDBSCAN is more resistant to underlying variation in images. This is also confirmed when minimum and maximum values are compared. Although ACM’s minimum border error is less than that of BDDBSCAN (0.7% < 0.9 %), this small difference is not seen when comparing maximum border errors (30.8% >> 14.8%). Statistically higher standard deviation and a few extreme values of boundary errors, such as 30.8% and 22.2%, cause ACM to have the same mean on results. However, a pairswise comparison shows that ACM performs better than BDDBSCAN in 38 out of 50 cases. In Table 1, bold numbers in the boundary error column indicate better results in the corresponding images.
We did not perform the ttest for this comparison because the major assumption of the ttest is not met within the results of ACM. The ttest requires normally distributed variables to compare. To check the normality of boundary errors, we used the Lilliefors test [25], which shows that the boundary errors of ACM is not normally distributed (pvalue:0.001, critical value: 0.1245). Although BDDBSCAN passed the normality test (pvalue:0.3906, critical value:0.1245), the ttest would be inappropriate in this study.
Discussion
Skin cancer is the one of most common malignancies in the United States and should be treated accurately by means other than manual delineation. The frameworks presented in this study play a key role in alleviating inter and intravariability in medical assessments.
By its nature, ACM is more sensitive to noisy images. As seen in Figure 6(b), three spikes are caused by noisy pixels. BDDBSCAN can also be negatively affected by noise, but, unlike ACM, the erroneous region is bounded locally. BDDBSCAN usually finds a marginally shrunk version of the lesion, (Figure 6(a), 6(b), 6(c)) having several precision values of 1.00, as seen in many cases in Table 1. In a few cases, it also marked outer regions (FP) of skin lesions. The average ratio of precision implies that the regions found by BDDBSCAN are usually smaller than those defined by ACM because of high means of precision (0.99 > 0.93).This observation is also examined in each of the three samples of Figure 6. In each of them, the green boundary (generated by BDDBSCAN) is narrower than the blue, comparatively. On the other hand, recall rates on two groups suggest that ACM is more successful in finding more pixels of targeted skin lesion. BDDBSCAN outperforms ACM only in these three images.
Apparently, the region between red (drawn by the dermatologist) and green boundaries seems to be a major problem for BDDBSCAN. Alternatively, it means that the manual border tolerates background errors (FP) in order to increase recall. Therefore, a dermatologist’s manual selection might not be accurate in delineating the exact region of lesion at a very fine level. Based on this assumption, the transition regions found between red and green boundaries would be expected consequences in this framework. This observation opens the door to the problem of intraobserver variability that needs more attention from researchers. It is suggested that the users of these frameworks look for various parameters that can be optimized for the underlying data. The mask obtained with Eq. 1 and Eq. 2 significantly affects the image in the first step. The small changes in parameters t and s are less effective inputs regarding the general design of ACM. Therefore, we will focus on newer techniques that lessen the negative impact of preprocessing in ACM. Similar to that of ACM, the preprocessing step of BDDBSCAN can change delineation significantly. Other than a nonparametric histogrambased thresholding mechanism, a new set of binarization methods can be investigated. Parameters R and MP of BDDBSCAN are less questionable in the context of this study, since they are good for each image once an agreement is reached.
In this study, we introduce and compare two frameworks based on the novel ACM and BDDBSCAN in order to automatically detect skin lesions in dermoscopy images. A large number of active contour and level set algorithms are available in the field. A good survey is given in [18]. We have used the one presented in [27] because it has a larger capture rate, fewer arithmetic operations, better accuracy, but worse performance with noise when compared to the others. Similarly, BDDBSCAN presents an innovative solution for fast calculation of lesion boundaries. Thanks to boundary definition of the cluster, it eliminates a huge number of region queries. The proper preprocessing steps are explained within each of the frameworks. Both of these algorithms have the same average boundary errors, 6.6%.
Methodology
Theoretical derivation of active counter
The present sections develop the theoretical fundament of the novel algorithm, described in the previous part of the paper, designed to automatically determine the boundary of lesions. As discussed in the Introduction, the new algorithm combines two approaches: the first one uses the socalled high boost filtering idea presented in [26]; the second one applies a version of the new ACM recently reported in [27].
The filtering approach is designed to facilitate the convergence of the active contour model, which could catch unnecessary objects or noise at the time of evolution. To avoid doing this, the image is processed in a way to eliminate such objects through sharpening and increasing the homogeneity of its background. One useful approach is to apply the highboost approach to develop a new and useful unsharpening mask [26].
The above idea is applied in the present study, and a new formulation of the framework is given hereafter. Assume the image is presented with the function f(x,y), where (x,y) gives the coordinate of a pixel whereas f(x,y) shows the gray level of this pixel. For simplicity’s sake, we consider a single color channel in this presentation. In the case of a colored image, the work is extended to the three used channels.
is the sum of the directional derivatives in the directions of the vectors (±1, ±1). The last term of Eq. 1 represents the Laplacian of the image.
Now varying A, B, and C in Eq. 1, we could generate a mask with different entries. Each mask manipulates the image in a different manner. For example, if the sum of the mask’s entries is zero, the image will appear with a dark background, and the objects will have tiny white boundaries. Increasing the sum of the mask’s entries will sharpen the image, lighten the background, and make it more homogenous by erasing the small details. One potential disadvantage from a mask with a high sum of entries is that some light zones of the lesion may disappear as well. Using this knowledge, it is not difficult for a user to determine the right values of A, B, and C with respect to the given set of images.
where , nc denotes the number of columns of the image, whereas nr denotes the number of rows. The boundary conditions states that:
if for t>0.001 and s=s* then (5)
otherwise (6).
Eq. 5 states that, if the inequality is satisfied, the new value of the active contour at the point s* in time is set to be the same as the value at time t. If the inequality is not satisfied, the new value of the parametric function r, at time is calculated and used further by ACM.
The complexity of the calculation of this approach is in the order of where The accuracy of boundary approximation is in the order of , where Δt is a value given by the user and could be selected in a way to minimize the error of boundary delineation.
Boundary driven density based algorithm
Clustering, a major problem in the scope of unsupervised learning, deals with recognizing meaningful groups that include similar items. With the increase of digital data all around the world, more powerful tools are required to exploit piles of socalled useless datasets. Even though there is now substantial body of research on clustering, the constraints (e.g., efficiency and effectiveness) of current approaches require more practical algorithms. BDDBSCAN demonstrates how the efficiency of prominent densitybased clustering algorithm DBSCAN [29] is improved for skin lesion detection.
The boundarydriven densitybased algorithm is an intelligent technique that can be applied to any thresholded image to find the most represented objects(s) in the current scene. The rationale behind BDDBSCAN is to evaluate pixels regarding their likelihood of expanding the boundary of current cluster. Since a significant part of computational time of DBSCAN is spent for the region queries, BDDBSCAN focuses on this problem so that the improved version chooses data objects more intelligently for region queries. Also, being a novel idea in the literature of densitybased clustering, this approach introduces the notion of a cluster boundary, which is exploited in the selection of influential points –the term point corresponds to an individual data object in a dataset, such as a pixel in an image. The improvement reported in [21] saves a huge number of queries (20%  39% in virtual slides) when compared to DBSCAN, in which one neighborhood query fired for each of points in dataset semirandomly.
The idea of BDDBSCAN in 2D relies on the cluster's boundary, which is a new concept introduced in [21]. Having these borders, we can identify those points that are likely to change the current shape of the cluster’s boundary. Note that the area of a cluster always expands out and never shrinks. In cases where queries cannot affect the cluster's area, the current region query is considered as unnecessary and omitted to fasten the regular DBSCAN. This preverification is very helpful in keeping the running time of the whole algorithm low. The intuition behind determining the border of a cluster is derived from the border of a primitive cluster. When boundaries of region queries are united, the outer boundary of this process gives the boundary of the current cluster. Simply, line segments connecting inner and outer boundaries (not only outer boundaries because of the donut problem) are exploited to indicate the border for a cluster. The concept of a convex hull is used so as to i) construct the initial boundary of a cluster and ii) expand the current cluster.
Initialization of cluster boundary
Selecting leading points
BDDBSCAN fires only those queries that are likely to expand the boundaries of the cluster in an effort to increase the efficiency of the DBSCAN. To select leading points, it uses boundaries of the polygons that delineate the cluster body. In contrast with DBSCAN, BDDBSCAN does not keep track of the status of the points, such as core. The cluster body can be enlarged only via points that are qualified for R neighborhood query. In other words, if a point is close enough to a cluster boundary, we fire an R neighborhood query around it; otherwise, the query will be omitted, i.e., some queries will not alter the shape of a cluster. Hence, we maintain a set of points that are likely to change the boundaries of a cluster at a given time.
Expansion of clusters
The innovative algorithm BDDBSCAN principally behaves similar to DBSCAN and tries to enlarge an existing cluster using unprocessed cluster points. However, we need to inspect the next query in terms of not only newly added points, but also its effect on boundaries. Once the first convex hull is formed around an initial point, it becomes the initial boundary of current cluster. Afterwards, each convex hull around a query point is combined with the main body of the cluster. Principally, this operation corresponds to the union of two polygons. Note that the convex hull is also a special case in the domain of the simple polygon. The notion of expanding a cluster is given in the following definition.
Definition 6. Let C be a cluster of points bounded by polygon(s) P_{1}, P_{2} … P_{ i } (i > 0), and let P_{1} be outer polygon for C, i.e., P_{2} … P_{ i } (i > 1) are to show the holes in P_{1}. Let T be a newly found convex hull to be merged to main body of a cluster C. Just after merging with convex hull T, cluster C is formulized by C = (P_{1} ∪ T)−(P_{2} ∪ P_{3} … ∪P_{ i }).
Definition 6 carefully considers the donut problem and uses P_{2} ∪ P_{3} … ∪P_{i} to exclude these regions from C, shown primarily with P_{1}.
Future study
In the future, we will elaborate on ACM to make it more robust against noisy images. Regarding the densitybased method, we will focus more on intravariability and postassessment during the performance analysis of the intelligent systems. Additionally, the performance of BDDBSCAN will be evaluated over different polygonunion algorithms.
Abbreviations
 ACM:

Active Contour model
 BDDBSCAN:

Boundary driven density based clustering algorithm with noise
 HSL:

Hue, Saturation, Lightness
 CIE L*u*v:

International Commission on Illumination (CIE) color model (1976)
 PCT:

Principal component transform
 R:

Algorithm parameter, Radius of search region in BDDBSCAN
 MP:

Algorithm parameter, Number of minimum points in BDDBSCAN
 A:

Parameter in Eq. 1
 B:

Parameter in Eq. 1
 C:

Parameter in Eq. 1
 t:

Parameter of time step
 FP:

False Positive
 FN:

False Negative
 TP:

True Positive
 FN:

False Negative
 ACHM:

Active Convex Hull Model
 p^{A}:

An arbitrary pixel in image
 S^{o}:

Set of all pixel found in region query
 C:

Cluster of points
 P_{n}:

Polygon
 T:

Convex hull.
Declarations
Acknowledgements
Authors thank Mr. K. Ushkala for implementation of ACM in Java.
This article has been published as part of BMC Bioinformatics Volume 11 Supplement 6, 2010: Proceedings of the Seventh Annual MCBIOS Conference. Bioinformatics: Systems, Biology, Informatics and Computation. The full contents of the supplement are available online at http://www.biomedcentral.com/14712105/11?issue=S6.
Authors’ Affiliations
References
 Jemal A, Siegel R, Ward E, Hao Y, Xu J, Murray T, Thun MJ: Cancer statistics. CA: A Cancer Journal for Clinicians 2008, 59: 225–249. 10.3322/caac.20006
 Argenziano G, Soyer HP, De Giorgi V: Dermoscopy: A Tutorial. Milan: Edra Medical Publishing & New Media; 2002.
 Steiner K, Binder M, Schemper M, Wolff K, Pehamberger H: Statistical evaluation of epiluminescence dermoscopy criteria for melanocytic pigmented lesions. Journal of the American Academy of Dermatology 1993, 29: 581–588. 10.1016/01909622(93)70225IView ArticlePubMed
 Binder M, Schwarz M, Winkler A, Steiner A, Kaider A, Wolff K, Pehamberger H: Epiluminescence microscopy: a useful tool for the diagnosis of pigmented skin lesions for formally trained dermatologists. Archives of Dermatology 1995, 131: 286–291. 10.1001/archderm.131.3.286View ArticlePubMed
 Fleming MG, Steger C, Zhang J, Gao J, Cognetta AB, Pollak I, Dyer CR: Techniques for a structural analysis of dermatoscopic imagery. Comput Med Imaging Graph 1998, 22: 375–389. 10.1016/S08956111(98)000482View ArticlePubMed
 Celebi EM, Aslandogan AY, Stoecker WV, Iyatomi H, Oka H, Chen X: Unsupervised Border Detection in Dermoscopy Images. Skin Res Technol 2007, 13: 454–462. 10.1111/j.16000846.2007.00251.xPubMed CentralView Article
 Celebi ME, Iyatomi H, Schaefer G, William StoeckerV: Lesion Border Detection in Dermoscopy Images. Computerized Medical Imaging and Graphics 2009, 33(2):148–153. 10.1016/j.compmedimag.2008.11.002PubMed CentralView ArticlePubMed
 M. Emre Celebi, Hitoshi Iyatomi, Stoecker WV, Moss RH, Rabinovitz HS, Argenziano G, Soyer HP: Automatic Detection of BlueWhite Veil and Related Structures in Dermoscopy Images. Computerized Medical Imaging and Graphics 2008, 32: 670–677. 10.1016/j.compmedimag.2008.08.003View Article
 Hance GA, Umbaugh SE, Moss RH, Stoecker WV: Unsupervised color image segmentation with application to skin tumor borders. IEEE Engineering in Medicine and Biology 1996, 15: 104–111. 10.1109/51.482850View Article
 Carlotto MJ: Histogram Analysis Using a ScaleSpace Approach. IEEE Transactions on Pattern Analysis 1987, 1: 121–129. 10.1109/TPAMI.1987.4767877View Article
 GutkowiczKrusin D, Elbaum M, Szwaykowski P, Kopf AW: Can early malignant melanoma be differentiated from atypical melanocytic nevus by in vivo techniques? Skin Research and Technology 1997, 3: 15–22. 10.1111/j.16000846.1997.tb00154.xView Article
 Gao J, Zhang J, Fleming MG, Pollak I, Cognetta AB: Segmentation of dermatoscopic images by stabilized inverse diffusion equations. Proceedings of IEEE Int. Conf. on Image Process 1998, 823–827.
 Dhawan AP, Sicsu A: Segmentation of images of skin lesions using color and texture information of surface pigmentation. Comput Med Imaging Graph 1992, 16: 163–177. 10.1016/08956111(92)90071GView ArticlePubMed
 MeasYedid V, Tilie S, OlivoMarin J: Color Image Segmentation Based on Markov Random Field Clustering for Histological Image Analysis. Proceedings of 16th International Conference on Pattern Recognition 2002, 796–799.
 Schmid P: Segmentation of digitized dermoscopic images by twodimensional color clustering:. IEEE Trans Med Imaging 1999, 18: 164–171. 10.1109/42.759124View ArticlePubMed
 Kass M, Witkin A, Terzopoulos D: Snakes: Active Contour Models. Inter. J. Computer Vision 1987, 1: 211–221.
 Osher S, Sethian JA: Fronts Propagating with Curvature Dependent Speed: Algorithms Based on HamiltonJacobi Formulations. J Comp. Physics 1988, 79: 12–49. 10.1016/00219991(88)900022View Article
 Acton ST, Ray N: Biomedical Image Analysis: Tracking. Arlington, Virginia: Morgan and Claypool Publishers; 2006.
 Lee I, EstivillCastro V: Polygonization of point clusters through cluster boundary extraction for geographical data mining. In Proceedings of the 10th International Symposium on Spatial Data Handling. Ottawa, Canada; 2002:27–40.View Article
 Nosovskiy GV, Liu D, Sourina O: Automatic clustering and boundary detection algorithm based on adaptive influence function. Pattern Recognition 2008, 41: 2757–2776. 10.1016/j.patcog.2008.01.021View Article
 Mete M: Delineation of Malignant Areas in Histological Images of Head and Neck Cancer. University of Arkansas at Little Rock, Applied Science Department; 2008. PhD thesis.
 Levkowitz H, Herman GT: A Generalized Lightness, Hue, and Saturation Color Model. Graphical Models and Image Processing 1993, 55: 271–285. 10.1006/gmip.1993.1019View Article
 Ridler T, Calvard S: Picture thresholding using an iterative selection method. IEEE Trans. Systems Man Cybernet 1978, 8: 630–632. 10.1109/TSMC.1978.4310039View Article
 Celebi EM, Kingravi HA, Iyatomi H, Aslandogan AY, Stoecker WV, Moss RH, Malters JM, Grichnik JM, Marghoob AA, Rabinovitz HS, Menzies SW: Border Detection in Dermoscopy Images Using Statistical Region Merging. Skin Res Technol 2008, 14: 347–353. 10.1111/j.16000846.2008.00301.xPubMed CentralView ArticlePubMed
 Lilliefors H: On the Kolmogorov–Smirnov test for normality with mean and variance unknown,. Journal of the American Statistical Association 1967, 62: 399–402. 10.2307/2283970View Article
 Gonzalez RC, Wood RE: Digital Image Processing, 3rd Edition. Upper Saddle River, NJ: Pearson Prentice Hall; 2008.
 Sirakov NM, Ushkala K: An Integral Active Contour Model for Convex Hull and Boundary Extraction,. Edited by: G. Bebis. Springer; 2009:1031–1040. LNCS 5876
 Sirakov NM: A New Active Convex Hull Model For Image Regions. Journal of Mathematical Imaging and Vision 2006, 26: 309–325. 10.1007/s1085100690046View Article
 Ester M, Kriegel HP, Sander J, Xu X: A densitybased algorithm for discovering clusters in large spatial databases with noise. Proceedings of 2nd International Conference on Knowledge Discovery and Data Mining 1996, 226–231.
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.