Automated Skin Lesion Segmentation using K-

means Clustering from Digital Dermoscopic Images

Ashi Agarwal, Ashish Issac, Malay Kishore Dutta Kamil Riha, Vaclac Uher
Department of Electronics and Communication Department of Telecommunications
Engineering, Amity University, Noida, India Brno University of Technology,
[email protected], [email protected], Technicka 12, 616 00, Brno, Czech Republic
[email protected] rihak@[email protected], [email protected]

Abstract—Melanoma can prove fatal if not diagnosed at early As mentioned, it is a topic of great concern and has always
stage. The accuracy of identification of skin cancer from gained attention of researchers to accurately detect, segment
dermoscopic images is directly proportional to the accuracy of and classify the skin lesion. And hence a lot of work has been
the skin lesion segmentation. This work proposes a skin lesion done on the topic. Sforza et al [1] proposes an adaptive
segmentation method using clustering technique. The use of
thresholding method for segmentation of the skin lesion.
smoothing filter and area thresholding is competent enough to
sufficiently reject the noisy pixels from the finally segmented Soumya et al [2] proposes a thresholding based technique for
image. The results of skin lesion segmentation obtained from the segmentation which is followed by active contour
proposed algorithm has been compared with the annotated implementation for estimation of boundaries of the lesion.
images. The results have been expressed in the form of Kasmi et al [3], proposed a method for skin lesion
overlapping score and correlation coefficient. The maximum segmentation using gabor filter and geodesic active contour.
values of overlapping score and correlation coefficient obtained Moussa et al [4] proposed an edge detection techniques for skin
from the algorithm are 96.75% and 97.66% respectively. The lesion segmentation from the dermoscopic images. Arabi et al
results are convincing and suggests that the proposed work can [5] has also proposed a method for melanoma detection. An
be used for some real time application.
image enhancement technique followed by low pass filtering is
proposed for selection of ROI. Sowmya et al [6] has proposed
Keywords—Dermoscopic Image; Skin Lesion; K-means Clustering;
use of different fuzzy grouping procedures like FCM, PCM
Intensity Threshold; Mathematical Morphology.
Algorithm and HCM Algorithm for segmentation of skin
I. INTRODUCTION lesion. Pirnog et al [7] proposed another approach of histogram
thresholding for point by point division of moles in view of
Melanoma is all the more regularly, yet not generally, an histogram examination of the immersion shading segment.
ailment of the skin. It begins in melanocytes – the cells that
Santy et al [8] proposed the use of measurable locale
convey the shade melanin that tone the skin, hair and eyes.
combining, iterative stochastic district consolidating, versatile
Melanocytes similarly outline moles, where melanoma
generally develops. Having moles can be a peril for melanoma, thresholding, shading improvement and iterative division,
yet it's basic to remember that most moles don't get the chance multilevel thresholding for segmentation. Sujitha et al [9]
to be melanoma. Inquire about suggests that around 90% of proposed a blend of Active contour and Watershed
melanoma manifestations can simply be associated with Transformation calculations for division handle. Beuren et al
prologue to brilliant (UV) rays from typical or recreated [10] proposed a method for ROI segmentation by applying
headsprings. As melanoma can be developed from any filtering operations to dermoscopic images. This is achieved
melanocyte all over the body including the ones that are never with the help of morphological tools which are applied using a
displayed to the sun, UV light can't be exclusively in charge of lexicographic order onto HSI color space followed by
analysis. Current research focuses to a blend of family history, thresholding for ROI segmentation. Although a lot of work has
hereditary qualities and ecological variables that are likewise to been done in the field of skin lesion segmentation, still there is
fault. Melanoma is most commonly seen on the skin unlike the a need for some real time method which is able to segment the
other cancers which make it simpler to detect in early stages. In lesion correctly so that the segmented mole can be analyzed
the event that left undetected, notwithstanding, melanoma can further for classification as benign or malignant.
spread to inaccessible locales or far off organs. And in that
condition which is generally referred to stage IV, it becomes The dominant contribution of the proposed work lies in the
extremely difficult to cure. When extended to later stages, it utilization of color as a parameter in distinguishing the lesion
tend to effect the liver, lungs, bones and cerebrum. from other objects in the images. Since, a visual difference in
the color of the lesion and the background pixels can be
This work was supported in part by the Czech Science Foundation (GACR) observed, this property has been used as a basis for
under Grant 17-19638S and by the European Research Council within the segmentation of lesion from the images. A new grayscale
National Sustainability Programme of the European Commission under Grant
LO1401.
image is generated by strategically combining the channels
which is able to sufficiently distinguish the objects of different
colors. The clustering approach has been used for extraction of of the image. The final lesion is selected as an object with its
lesion and has proved beneficial in correct identification of centroid lying in a specific range from the center of the image.
lesion pixels. However, some noisy pixels similar to lesion
pixels are also segmented. It has been observed that the noisy A. Pre-processing of Dermoscopic Images
pixels are smaller in size and can be easily rejected using area
The input image is a dermoscopic RGB image. The
based thresholding.
segmentation of the lesion directly from the RGB image on
The remaining paper is structured as follows: Section II the basis of intensity becomes difficult. Also, it has been
discusses various image processing based methods involved for observed from the images that the lesions have visual
segmentation of optic disc. Section III discusses the results difference in color from the background pixels. A colored
obtained from the experimentation. Section IV discusses the image is combination of gray values from individual R, G and
conclusions. B channels. The color analysis can be achieved using
individual channels of the RGB images, but it would be
II.PROPOSED METHODOLOGY computationally inefficient to work on 3 channels. So, keeping
Melanoma is a kind of skin disease that rises when the shade the computational efficiency in mind, the input RGB image is
delivering cells (melanocytes) change and get the opportunity converted to L*a*b color space to clearly observe the color
to be particularly cancer-causing. Melanoma is just a single differences between the mole and the lesion. As the a* and b*
kind of skin growth, and relatively less basic than basal cell channel stores the color data, both of them are added together
and squamous cell skin disease. In any case, melanoma is an and then the lightness channel is subtracted from the sum to
especially dangerous kind of skin ailment since it will probably obtain a new gray scale image with clear difference between
spread. the lesion and the skin.

Input Dermoscopic Image

Convert image into L*a*b

color space

a channel b channel L channel

Fig. 1 Dermoscopic Images (a) Normal Lesion (b) Melanoma Affected
Lesion

Fig. 1 differentiates the mole which is melanoma effected

from a non-melanoma mole. A non-melanoma mole is shown
in Fig. 1(a) and has a regular boundary, single color and
symmetry. Fig. 1(b) shows a melanoma affected lesion and Average Filtering
has an irregular boundary, more than one color and is
asymmetrical.
Contrast Adjustment
The proposed methodology can be majorly divided into three
main sections:
K-means Clustering
 Pre-processing of Dermoscopic images
 Segmentation of the skin lesion
Thresholding
 Post-processing to finally segment the skin lesion
The complete methodology is graphically explained in Fig. 2.
The RGB image is first converted into the L*a*b color space. Reject noise pixels with area
A strategic combination is applied to the individual channels thresholding
of the L*a*b color space to get a new grayscale image in
which the mole and skin pixels are more differentiable. The
Selecting the final lesion
image hence obtained is subjected to smoothening for the
rejection of small intensity changes followed by the contrast
enhancement to improve the visual difference. The K-means
clustering is implemented to segment out the mole from the
Segmented Lesion
preprocessed image. Post segmentation, the rejection of false Fig. 2 Block Diagram of Proposed Method
positives is achieved on the basis of distance from the center
Aiming to achieve good results, the resulting grayscale image B. Segmentation of the skin lesion
is subjected to some pre processing steps. The averaging filter
The preprocessed image obtained from the previous section is
[13] is applied on the image in order to reduce the small
subjected to a clustering technique for lesion segmentation. The
intensity changes due to skin scales and false positives due to
clustering technique used is K-means clustering which is one
hair present on the skin. The smoothened image is further
of the most common technique used for color based
subjected to contrast enhancement. Fig. 3(a) shows the input
segmentation. Each pixel present in the image is initially
RGB dermoscopic image and Fig. 3(b) shows the image in
assigned to a cluster such that cluster sum of squares is
L*a*b color space. Fig.3(c) shows the grayscale image. Fig.
minimized and finally cluster pixels are updated on the basis of
3(d) is the preprocessed grayscale image which is obtained
the new mean obtained for each cluster.
averaging filter and Fig. 3(e) shows the pre processed image
after contrast enhancement. The assignment of a pixel to a particular cluster is done on the
basis of the following equation:
(𝑡) (𝑡) 2 (𝑡) 2
𝑆𝑖 = {𝑥𝑝 ∶ ‖𝑥𝑝 − 𝑚𝑖 ‖ ≤ ‖𝑥𝑝 − 𝑚𝑗 ‖ ∀ 𝑗, 1 ≤ 𝑗 ≤ 𝑘}
(11)
(t)
Where, S = cluster,
Fig. 3 Preprocessing of Dermoscopic Images: (a) Input RGB Image (b) L*a*b xp = observation,
Color space transformed image (c) Grayscale Image generated by strategic mi(t) and mj(t) = initial means
combination of channels (d) Average filter smoothed image (e) Contrast
Enhanced Image The clusters are updated by determining the new mean using
the following equation:
The complete process is summarized in the form of an
algorithm as follows: (𝑡+1) 1
𝑚𝑖 = (𝑡) ∑𝑥 (𝑡) 𝑥𝑗 (12)
|𝑆𝑖 | 𝑗 ∈ 𝑆𝑖
Algorithm 1: Pre-processing of dermoscopic image
Where, S(t) = cluster,
Step 1: Load the input image in RGB format.
xj = observation
Step 2: Convert the input image from RGB color space to mi(t+1) = updated mean
L*a*b color space. The transformation from RGB
to L*a*b color space is not a direct conversion. In the proposed work, the number of clusters used are two,
First the image in RGB format is converted to XYZ since, the segmentation is needed to differentiate into lesion
format and then to L*a*b format. and non-lesion pixels. The output of K-means clustering
algorithm results in an image which has pixels divided into
RGB to XYZ conversion takes place as follows: different clusters. The clustered image is subjected to an
X = 0.4124 * Ir + 0.3576 * Ig + 0.1805 * Ib (1) intensity based thresholding to result in a binary image which
Y = 0.2126 * Ir + 0.7152 * Ig + 0.0722 * Ib (2) consists of lesion and some noisy pixels. This is demonstrated
Z = 0.0193 * Ir + 0.1192 * Ig + 0.9505 * Ib (3) in Fig. 4 below. Fig. 4(a) shows the preprocessed images which
was obtained as a result of the grayscale operation performed on
XYZ to L*a*b conversion takes place as follows: the color image obtained after K-means clustering process. Fig.
L = (116 * Yr) – 16 (4) 4(b) shows the binary image which has been obtained after
A = 500 * (Xr – Yr) (5) applying intensity based threshold to K-means clustering
B = 200 * (Yr – Zr) (6) applied color image.

Where,
Xr = X / 95.047 (7)
Yr = Y / 100 (8)
Zr = Z / 108.883 (9)

Step 3: Apply strategic combination of mathematical

operations to individual channels of the L*a*b Fig. 4 Segmentation of the lesion: (a) Preprocessed Image obtained as a result
color format to form a new grayscale image. of the grayscale operation performed on the color image obtained after K-
Ig = a + b – L (10) means clustering process (b) Binary Image obtained after threshold the
grayscale image shown in (a).

Step 4: Apply averaging filter to remove the effects of hair

C. Post-processing to select the skin lesion
and skin scales during segmentation.
Step 5: Improve the contrast of the smoothed image to As discussed earlier, the output of the clustering algorithm
clearly distinguish the boundaries of the lesion. comprises of some false positive pixels. These noisy pixels
might result because of hair development on the skin and non- pixels which have been segmented as lesion pixels using
uniform illumination. It has been observed that the proposed method and S2 represents the lesion pixels in
objectsbelonging to hair are very small in size as compared to annotated images, then the overlapping score can be expressed
the mole. Hence, an area based thresholding is applied to mathematically as follows:
remove such objects. The hole filing operation is applied to the
Overlapping Score = (S1 ∩ S2) / (S1 U S2) (13)
resulting image in order to improve the segmentation accuracy
and ensure that the centroid of the image lies on the image only Where, ∩ = intersection
which will be calculated to select the mole from all segmented U = union
objects.
In order to remove the remaining false positives, the centroid of
all the objects is found out. It has been observed that the lesion
in all the images under experimentation is located in the center
of the image. So, a distance based approach has been used to
select the final lesion from the binary image. The distance
between the centroid of each object and image center is
calculated. If the calculated distance is under a specific value,
then the respective object is considered as mole and otherwise
rejected as false positive. The specific distance value used in
the proposed method is 15 pixels provided that the image size
is [500,500]. Fig. 5(a) show the binarized image obtained after
hole filling operation. Fig 5(b) demonstrates each object
marked with their centroid. Fig. 5(c) demonstrates the final
contour obtained after final selection based on the distance.

Fig.5 Post-processing: (a) Binary Image after hole filling (b) Each object with Fig. 6Experimental Results: (a) Input Image (b) Segmented Lesion (c)
its centroid marked on it (c) Final segmented lesion obtained after removal of Ground Truth
distant objects
A correlation is established between the binary image of the
III. RESULTS segmented lesion and ground truth using the following
The proposed method has been tested on 50 images from the equation:
publicly available database of Dermatology Information ∑𝑥 ∑𝑦(𝐺− 𝜇𝐺)(𝐿− 𝜇𝐿 )
System [11] and DermQuest [12]. This database consists of 𝐶= (14)
√(∑𝑥 ∑𝑦(𝐺− 𝜇𝐺)2 )(∑𝑥 ∑𝑦(𝐿− 𝜇𝐿)2 )
dermoscopic images classified for melanoma for both dermis
and dermquest. An annotated version of ground truth for Where, C = correlation coefficient
lesions is provided with the database and has been used in the G represents the ground truth image
proposed work to determine the accuracy of the proposed L represents the segmented lesion image
algorithm. A correlation and overlapping score between the µG= mean of ground truth image
ground truth and segmented results is calculated. The proposed µL= mean of segmented lesion image
method is competent enough to strategically segment the lesion x = number of rows
from the images. y = number of columns
Fig. 6 shows the segmentation results for some samples from
the database. Fig. 6(a) shows the input images which have Table 1 demonstrates the computation time, overlapping score
been used to test the proposed method. Fig. 6(b) shows the and correlation for all the images under test. The computation
segmented lesions using the proposed method and Fig. 6(c) time for each image demonstrates the time involved in lesion
shows the ground truth marked by some experts. segmentation. The comparison between ground truth and
segmented results is reported in terms of overlapping score and
The overlapping score has been obtained as the ratio of the correlation. An average overlapping score value of 89.87% has
intersecting pixels to the pixels combined in both, ground truth been obtained using the proposed method with a maximum
and segmented, binary images. If S1 represents the number of value of 96.75%. Similarly, a maximum correlation achieved
between the segmented and annotated lesion is 97.66% with an
Correlation Coefficient Vs Overlapping Score
average value of 94.13% for 50 samples.
1.0
TABLE 1. COMPUTATION TIME, OVERLAPPING SCORE AND
CORRELATION FOR THE SEGMENTED LESION WITH THE GROUND
0.8
TRUTH USING PROPOSED METHOD
0.6
S No Time Elapsed Correlation Overlapping score
(seconds) 0.4
Sample01 4.3041 0.9696 0.9494
Sample02 4.4189 0.9295 0.8845 0.2
Sample03 4.3998 0.9669 0.9453
0.0
Sample04 4.2534 0.9191 0.8923 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Sample05 4.0117 0.8913 0.8247
Fig.7 Graphical representation of correlation coefficient vs overlapping score
Sample06 4.0784 0.9766 0.9675
Sample07 3.9205 0.9666 0.9481
Sample08 3.9117 0.9699 0.9466
Fig. 7 shows a graph plotted with correlation coefficient on one
Sample09 3.9815 0.9710 0.9509
axis and overlapping score on the other axis for all the samples.
Sample10 4.1974 0.9525 0.9237
The values corresponding to each sample is demonstrated in
Sample11 4.3928 0.9345 0.8838 blue color. The graph demonstrates that the markings for most
Sample12 4.3143 0.9660 0.9444 of the samples lies in the top-right grid of the graph indicating
Sample13 3.9410 0.9570 0.9248 that a high value for both parameters has been achieved using
Sample14 3.8746 0.9721 0.9500 the proposed method. A high value of these parameters is an
Sample15 4.4635 0.9609 0.9309 indication of high segmentation accuracy of skin lesions and
Sample16 4.2642 0.9546 0.9429 that most of the pixels from the ground truth are retained
Sample17 4.3402 0.9500 0.9087 during the segmentation process using the proposed method.
Sample18 4.3646 0.9528 0.9161 The proposed algorithm was developed and tested using
Sample19 4.4757 0.9582 0.9265 MATLAB R2013b technical tool on a computing machine with
Sample20 4.4841 0.9062 0.8247 clock speed of 2GHz, 4GB RAM and 64 bit operating system.
Sample21 4.3687 0.9526 0.9126 An average computation time of 4.21 seconds is elapsed in the
Sample22 4.5217 0.9341 0.8794 segmentation of lesion from the images using the proposed
Sample23 4.3095 0.9359 0.8814 method. The computation time using the proposed method can
Sample24 4.2825 0.9635 0.9325
be considered a significant contribution and beneficial in
Sample25 4.3859 0.9738 0.9607
development of some real-time system for determination of
Sample26 4.3677 0.9519 0.9238
skin cancers using image processing techniques.
Sample27 4.3696 0.9618 0.9394
Sample28 4.4438 0.9657 0.9392 IV. CONCLUSION
Sample29 4.4192 0.9494 0.9261
Sample30 4.2329 0.9287 0.8826
An important task in identification of skin growth utilizing
Sample31 4.4177 0.9394 0.9025
picture preparing procedures is the exact segmentation of the
Sample32 4.2931 0.9564 0.9175
skin sore. A completely automated method has been proposed
Sample33 4.2494 0.9196 0.8531
which is able to segment the skin lesion from the images. A
Sample34 4.2520 0.9277 0.8817
strategic combination of the individual channels of the L*a*b
Sample35 4.2506 0.9498 0.9085
color space is advantageous in highlighting the differences
Sample36 4.2926 0.8659 0.7624 between the skin lesion and background pixels. The grayscale
Sample37 4.2872 0.9107 0.8393 image has high contrast between the lesion and non-lesion
Sample38 4.1450 0.9573 0.9263 pixels. The noise produced by the hair and skin scales are
Sample39 4.1120 0.9652 0.9378 reduced by applying the smoothing filter. A high value of
Sample40 3.1861 0.9481 0.9030 correlation and overlapping score indicates that the proposed
Sample41 4.2968 0.9029 0.8206 method can be used for some real time application in
Sample42 4.2644 0.9187 0.8571 segmentation of skin lesion from digital images.
Sample43 4.3895 0.9177 0.8469
REFERENCES
Sample44 4.2811 0.9313 0.8746
Sample45 4.1584 0.9412 0.8929 [1] G. Sforza, G. Castellano, S. A. ArikaLeAnder et al., "Using adaptive
thresholding and skewness correction to detect gray areas in melanoma in
Sample46 4.2252 0.8662 0.7550 situ images", IEEE Transactions on Instrumentation and Measurement,
Sample47 3.9168 0.9356 0.8898 vol. 61, no. 7, pp. 1839-1847, 2012.
Sample48 3.8770 0.9330 0.8876 [2] R. S. Soumya, S. Neethu, T. S. Niju, A. Renjini and R. P. Aneesh,
Sample49 3.9058 0.9348 0.8837 "Advanced earlier melanoma detection algorithm using colour
correlogram," 2016 International Conference on Communication Systems
Sample50 3.8886 0.9020 0.8295 and Networks (ComNet), Thiruvananthapuram, 2016, pp. 190-194.
 [3] R. Kasmi and K. Mokrani, "Classification of malignant melanoma and
benign skin lesions: implementation of automatic ABCD rule," in IET
Image Processing, vol. 10, no. 6, pp. 448-455, 6 2016.
[4] R. Moussa, F. Gerges, C. Salem, R. Akiki, O. Falou and D. Azar,
"Computer-aided detection of Melanoma using geometric features," 2016
3rd Middle East Conference on Biomedical Engineering (MECBME),
Beirut, 2016, pp. 125-128.
[5] P. M. Arabi, G. Joshi, S. Nigudgi and N. V. Deepa, "Computer vision for
diagnosis of malignant melanoma by pixel intensity matrix
parameters," 2016 10th International Conference on Intelligent Systems
and Control (ISCO), Coimbatore, 2016, pp. 1-4.
[6] R. Sowmya Devi, L. Padma Suresh and K. L. Shunmuganathan,
"Intelligent fussy system based dermoscopic image segmentation for
melanoma detection," International Conference on Sustainable Energy
and Intelligent Systems (SEISCON 2011), Chennai, 2011, pp. 739-743.
[7] I. Pirnog, R. O. Preda, C. Oprea and C. Paleologu, "Automatic lesion
segmentation for melanoma diagnostics in macroscopic images," 2015
23rd European Signal Processing Conference (EUSIPCO), Nice, 2015,
pp. 659-663.
[8] A. Santy and R. Joseph, "Segmentation methods for computer aided
melanoma detection," 2015 Global Conference on Communication
Technologies (GCCT), Thuckalay, 2015, pp. 490-493.
[9] S. Sujitha, M. L. Priya, J. Premaladha and K. S. Ravichandran, "A
combined segmentation approach for melanoma skin cancer
diagnosis," 2015 IEEE Seventh National Conference on Computing,
Communication and Information Systems (NCCCIS), Coimbatore, 2015,
pp. 11-16.
[10] A. T. Beuren, R. J. G. Pinheiro and J. Facon, "Color approach of
melanoma lesion segmentation," 2012 19th International Conference on
Systems, Signals and Image Processing (IWSSIP), Vienna, 2012, pp. 284-
287.
[11] Dermatology Information System, http://www.dermis.net, 2016.
Accessed August 2016.
[12] DermQuest, http://www.dermquest.com, 2016. Accessed August 2016.
[13] https://in.mathworks.com/help/images/ref/fspecial.html