Bw31494497

Mrs.Mongra Sahu, Mr.Devesh Narayan / International Journal of Engineering Research and
Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 3, Issue 1, January -February 2013, pp.494-497
Computer image analysis of skin lesions
Mrs.Mongra Sahu, (M.Tech)1 Mr.Devesh Narayan, (M.Tech)2
Pursuing M.Tech
(Rungta College Of Engineering & Technology, Bhilai (C.G))1
Reader (C.S.E)
(Rungta College Of Engineering & Technology, Bhilai (C.G))2

Abstract
An automatic method for segmentation 1. Preprocessing
of images of skin cancer and other pigmented The first step in our image segmentation
lesions is presented. This method first reduces a method can be Considered a preprocessing
color image into an intensity image and operation that transforms a color image into an
approximately segments the image by intensity intensity image. This operation is motivated by
thresholding. Then, it refines the segmentation two observations:
using image edges. Double thresholding is used to 1. Skin lesions come in a variety of colors; therefore,
focus on an image area where a lesion boundary absolute colors are not very useful in segmenting
potentially exists. Image edges are then used to images. However, changes in color from a lesion
localize the boundary in that area. A closed to its background (it’s surrounding healthy skin)
elastic curve is fitted to the initial boundary, and are similarly observed in all images; therefore,
is locally shrunk or expanded to approximate changes in color can be used to effectively segment
edges in its neighborhood in the area of focus. images.
2. When segmenting a skin image, significant color
Key words: Early diagnosis, image analysis. variations may exist within a lesion or in the
background. Such variations should be suppressed
Introduction since our interest is in color changes from the
Skin Cancers are the most common form background to a lesion or from a lesion to the
of cancers in humans [1]. The American Cancer background.
Society estimates that more than 700 000 new skin Observation 1 suggests that we should use
cancers are diagnosed annually in the United States., changes in color rather than absolute colors to
Image segmentation is perhaps the most studied area segment images. Therefore, we transform pixel
in computer vision, with numerous methods reported colors that are vector quantities into intensities that
[2,3]. are scalars and represent color differences.
A segmentation method is usually designed Observation 2 states that, among the color changes,
taking into consideration the properties of a only those belonging to a lesion boundary are
particular class of images. In this paper, we develop important in Image segmentation, and color changes
a three-step segmentation method using the inside a lesion or in the background should be
properties of skin cancer images. The steps of our ignored.
method are as follows: We transform our images that are in RGB
1. Preprocessing: a color image is first transformed color coordinates into images that are in CIELAB or
into an intensity image in such a way that the CIE 1976 L*a*b*color coordinates [4]. CIELAB is
intensity at a pixel shows the color distance of that a color space standardized by the CIE (
pixel with the color of the background. The color of Commission Internationale de l’E´ clairage ) in
the background is taken to be the median color of 1976 to measure color differences. This is a uniform
pixels in small windows in the four corners of the color space defined in such a way that Euclidean
image. distance between two colors (defined as DE) is
2. Segmentation: a threshold value is determined proportional to their visual difference. Color in the
from the average intensity of high gradient pixels in CIELAB space can be described with less
the obtained intensity image. This threshold value is redundancy than in the RGB space.RGB color
used to find approximate lesion boundaries. coordinates can be transformed into L*a*b* color
3. Region Approaches: a region boundary is refined coordinates using the following formulae [4]:
1/3
using Edge information in the image. This involves L* =116 Y/Yn -16 Y/Yn 0.008856
>
initializing a closed elastic curve at the approximate
boundary , and shrinking and expanding it to fit L* =903.3Y/Yn 
 Y/Yn 0.008856
=<
to the edges in its neighborhood.
a* =500 
f X/Xn 
-f Y/Yn

b*=200 
f Y/Yn 
-f Z/Zn

494 | P a g e

Vol. 3, Issue 1, January -February 2013, pp.
where f (t) = t1/3 when t > 0.008856 and f(t) =7.787t background. Similarly, it reduces image gradients
+ 16/ belonging to a lesion. For intensities falling on
116 when t =<0.008856. Xn, Yn and Zn are the lesion boundaries, however, we see that gradients
coordinates of the CIELAB reference white, which are increased. Therefore, if we map image intensities
are usually chosen to be 0.9642, 1.0 and 0.8249, according to the function depicted in
respectively. If we require that the images be taken Fig. 1(a), we will increase gradients on lesion
such that lesions do not fall on image corners, we boundaries,
can then use colors in the four corners of an image to while decreasing gradients inside a lesion or in the
estimate the color of the background. We take small background. Mapping intensities in this manner
windows, typically 10 * 10 pixels in size, from the facilitates detection of lesion boundaries. As can be
four corners of an image and determine the median observed, most details within the lesion and some
L*, a* and b* of the pixels. We use this median details in the background have been suppressed,
color as an estimate to the color of the background. while variations from the lesion to the background
We use median color rather than average color and from the background to the lesion have been
because image averaging uses the hair colors as well enhanced. We segment the image of Fig. 2(a) to
as the skin colors to estimate the color of the isolate lesion boundaries. Note that the
background. Since the number of hair pixels is preprocessing operation not only reduces a color
usually much smaller than the number of skin pixels image into an intensity image, it enhances the
in an image, when the median color is used, the boundary of a lesion while suppressing details inside
color of a pixel belonging to the hair will not be used and outside a lesion.
and the color of a pixel belonging to the skin will be
used to estimate the color of the background. If the
intensities assigned to pixels are proportional to
color distances of the pixels to the color of the
background, we will obtain an image that has high
values in lesions and small values in the
background. An image generated in this manner
will, therefore, show lesions as bright spots.
Therefore, they are more likely to belong to
a lesion. Image gradients have been used in the past
to determine region boundaries [16]. However,
detection of lesion boundaries using pure image
gradients is a difficult task. We need to
preprocessing operation so that edges on lesion Fig. 2. (a) Transforming intensities , (b) Smoothing
boundaries are distinguished from edges inside a of (a) with a 2D Gaussian kernel of standard
lesion or in the background. deviation 2 pixels.
To implement observation 2, we will need a function
that provides the property shown in Fig. 1(a). For a 2. Segmentation
wide range of intensities in the background, this To reduce the effect of image noise and
function produces very similar intensities. Therefore, intensity variations due to skin’s repetitive texture
the function reduces image and hair, an image is first low-pass filtered before
gradients in the input corresponding to details in the being segmented. Fig. 2(b) shows the image of Fig.
2(a) after being smoothed with a 2D Gaussian kernel
of standard deviation 2 pixels. As can be observed,
although smoothing reduces details in the image, the
smoothed image still contains information about the
lesion, which is brighter than the background. The
objective in the initial segmentation is to determine
the approximate position and shape of a lesion, and
then the optimal lesion boundary exists. Since the
optimal threshold value at one boundary point may
differ from that at another boundary point, the
objective in double thresholding is to select a range
of threshold values that includes the optimal
Fig. 1. (a) A desirable function for mapping color threshold value at every boundary point. To
distances to image intensities. i and o show the input determine an initial threshold value automatically,
and output image intensities, respectively. we observe that gradients of pixels on lesion
boundaries are generally higher than gradients of
pixels inside or outside lesions. We will, therefore,
use the average intensity of the top p% highest

495 | P a g e

gradient pixels in the image to compute the threshold 3. Region Approaches
value. p is typically a small number, e.g. 5. Within this category the thresholding
Because noise and details from skin texture and operation is most often used [4]. The pixels of an
hair could also result in high gradients, this process image are grouped into regions using some
may detect details from noise, skin texture and hair. similarity criteria of some characteristic features
Such regions, however, are often small and can be such as intensity. The digital photography is made
removed. Threshold techniques can be categorized up of a number of pixels each of a known brightness
into two classes: global threshold and local (shade of grey). Two peaks can be seen representing
(adaptive) threshold. In the global threshold, a single the paler skin tones and the darker lesion.
threshold value is used in the whole image. In the Segmentation can be performed by choosing a value
local threshold, a threshold value is assigned to each of the threshold between the peaks [9].
pixel to determine whether it belongs to the
foreground or the background pixel using local Conclusions
information around the pixel. Because of the Image segmentation is the first step in
advantage of simple and easy implementation, the many image analysis problems. To analyze skin
global threshold has been a popular technique in lesions, it is necessary to accurately locate and
many years [6][7][8]. isolate the lesions. In this paper, an automatic
method for segmentation of skin cancer images was
2.1 Thresholds presented. The method starts with an initial
Threshold is one of the widely methods segmentation and uses edge information in the
used for image segmentation. It is useful in neighborhood of the initial segmentation to refine
discriminating foreground from the background. By the results. An elastic curve model is used to
selecting an adequate threshold value T, the gray represent the final segmentation. Although the
level image can be converted to binary image. The method is devised for segmentation of color images,
binary image should contain all of the essential early on in processing, a color image is transformed
information about the position and shape of the into an intensity image where the intensity at a pixel
objects of interest (foreground). The advantage of shows the color distance of that pixel to the
obtaining first a binary image. background. Intensities in the image obtained in this
manner are then transformed according to a function
2.2 Watershed Algorithm shown in Fig. 2 to suppress details in the background
The watershed segmentation has proven to and in a lesion while enhancing details across lesion
be a powerful and fast technique for both contour boundaries. Transformation of a color image into an
detection and region-based segmentation. In intensity image and mapping of image intensities to
principal, water-shed segmentation depends on enhance lesion boundaries are considered to be the
ridges to perform a proper segmentation, a property main contributions of this work.
which is often fulfilled in contour detection where
the boundaries of the objects are expressed as ridges. References
For region-based segmentation it is possible to [1] A.W. Kopf, T.G. Salopek, J. Slade, A.A.
convert the edges of the objects into ridges by Marghood, R.S. Bart, Techniques of
calculating an edge map of the image. Watershed is cutaneous examination for the detection of
normally implemented by region growing based on a skin cancer, Cancer Supplement 75 (2)
set of markers to avoid severe over-segmentation (1994) 684–690.
[10, 11,10]. Different watershed methods use [2] R.M. Haralick, L.G. Shapiro, Image
slightly different distance measures, but they all segmentation techniques, Computer Vision
share the property that the watershed lines appear as Graphics, and Image Processing 29 (1)
the points of equidistance between two adjacent (1985) 100–132.
minima. Meyer [9] use the topographical distance [3] R.K. Sahoo, S. Soltani, A.K.C. Wong, A
function for segmenting images using watershed survey of thresholding
segmentation, while Najman and Schmitt [8] present techniques,Computer Vision, Graphics, and
the watershed differences with classical edge Image Processing 41 (1988) 233–260.
detectors. Felkel et al. [10] use the shortest path cost [4] A.P. Dhawan, A. Sicsu, Segmentation of
between two nodes which is defined as the smallest images of skin lesions using color and
lexicographic cost of all paths between two points, texture information of surface
which reflects the flooding process when the water pigmentation, Computerized Medical
reaches a plateau. The success of watershed Imaging and Graphics 16 (3) (1992) 163–
segmentation relies on a situation where the de-sired 177.
boundaries are ridges. Unfortunately, the standard [5] M.M. Wick, A.J. Sober, T.B. Fitzpatrick,
watershed framework has a very limited flexibility Clinical characteristics of early cutaneous
on optimization parameters. As an example, there melanoma, Cancer 45 (1980) 2684–2686.
exists no possibility to smooth the boundaries.

496 | P a g e

[6] Alan Wee, Chung Liew , Hong Yan, “
Current Methods in the Automatic Tissue
Segmentation of 3D Magnetic Resonance
Brain Images,” Current Medical Imaging
Reviews, Vol.2, pp.1-13 ,2006.
[7] R. A. Lerski, K. Straughan, L. R. Schad,
D.Boyce, S. Bluml, and I. Zuna, “MR
ImageTexture Analysis- An approach To
Tissue Characterization”, Magnetic
Resonance Imaging, Vol. 11, pp. 873-
887,1993.
[8] H. Selvaraj1, S. Thamarai Selvi2, D.
Selvathi3, L. Gewali1, “ Brain MRI Slices
Classification Using Least Squares Support
Vector Machine ” IC-MED, Vol.1, No. 1,
Issue 1, Page 21 of 33 ,2007.
[9] P.N.Hall, E.claridge and J.D.morris Smith,
"Computer Screening for early detection of
melanoma - is there a future?", British
Journal of dermatologiy 1995; 132; 325-
338
[10] Marcel Prastawa a, Elizabeth Bullitt c, Sean
Ho a, Guido Gerig, “A Brain Tumor
Segmentation Framework Based on Outlier
Detection” Medical Image Analysis ,1-9
,2004.
[11] Jiayin Zhou, Vincent Chong,Tuan-Kay
Lim, Jing Huang, “MRI Tumor
Segmentation for Nasopharyngeal
Carcinoma Using Knowledge-based
Fuzzy Clustering” International Journal of
Information Technology Vol. 8, No. 2
,September 2002.
[12] A. Green, N. Martin, J. Pfitzner, M.
O’Rouke, N. Knight, Computer image
analysis in the diagnosis of melanoma, J.
American Academy of Dermatology 31 (6)
(1994) 958–964.

497 | P a g e

Bw31494497

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to Bw31494497 (20)

Bw31494497