independent component analysis for assessing therapeutic response in vitiligo skin disorder
TRANSCRIPT
Independent component analysis for assessing therapeuticresponse in vitiligo skin disorder
M. H. AHMAD FADZIL{, S. NORASHIKIN{, H. H. SURAIYA{ and H. NUGROHO*{
{Electrical and Electronic Engineering Department, Universiti Teknologi PETRONAS, Tronoh,Bandar Sri Iskander, Perak Darul Ridzuan, Malaysia
{Dermatology Department, Hospital Kuala Lumpur, Jalan Pahang,Wilayah Persekutuan Kuala Lumpur, Malaysia
This paper describes an image analysis technique that objectively measures skin
repigmentation for the assessment of therapeutic response in vitiligo treatments. Skin
pigment disorders due to the abnormality of melanin production, such as vitiligo, cause
irregular pale patches of skin. The therapeutic response to treatment is repigmentation of
the skin. However the repigmentation process is very slow and is only observable after a
few months of treatment. Currently, there is no objective method to assess the therapeutic
response of skin pigment disorder treatment, particularly for vitiligo treatment. In this
work, we apply principal component analysis followed by independent component
analysis to represent digital skin images in terms of melanin and haemoglobin
composition respectively. Vitiligo skin areas are identified as skin areas that lack melanin
(non-melanin areas). Results obtained using the technique have been verified by
dermatologists. Based on 20 patients, the proposed technique effectively monitored the
progression of repigmentation over a shorter time period of six weeks and can thus be
used to evaluate treatment efficacy objectively and more effectively.
Keywords: Therapeutic response; Vitiligo; Principal component analysis; Independent
component analysis
1. Introduction
Skin pigment disorders cause the skin to appear lighter or
darker than normal skin. The disorders occur due to
abnormal melanin production [1–4]. Melanin is the colour
pigment found in hair, eye and skin. Albinism, melasma
and vitiligo are examples of skin pigment disorders.
Vitiligo is an idiopathic disease that causes destruction of
melanocytes (the pigment producing cells). Due to the lack
of melanin, vitiligo lesions are paler than normal skin, as
shown in figure 1. The prevalence of vitiligo varies from
0.1% to 2% in various global populations without any sex,
racial, skin types or socioeconomic predilection [2,4].
Currently, there is no objective method to assess the
therapeutic response to vitiligo treatment. A global
assessment of therapeutic response is made by physicians
using a scoring system based on an ordinal scale, e.g. the
physician’s global assessment (PGA). The scale is based on
the degree of repigmentation within lesions over time and
an example is shown in table 1. PGA is largely dependent
on the human eye and visual judgment to produce the
scorings. Therefore it is a highly subjective method of
assessment as inter-rate and intra-rate variation is inevi-
table. Moreover, it can be months before repigmentation
within the lesions can be observed. Treatment efficacy thus
cannot be determined quickly.
*Corresponding author. Email: [email protected]
Journal of Medical Engineering & Technology, Vol. 33, No. 2, February 2009, 101–109
Journal of Medical Engineering & TechnologyISSN 0309-1902 print/ISSN 1464-522X online ª 2009 Informa Healthcare USA, Inc.
http://www.informaworld.com/journalsDOI: 10.1080/03091900802454459
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
The use of digital image techniques may help in an objec-
tive assessment of repigmentation and in follow-up studies
of skin pigment disorder progression to test the efficacy of
therapeutic procedures [5]. This paper describes the develop-
ment of image processing techniques for a computer-based
image analysis intended for the assessment of therapeutic
response in a skin pigment disorder such as vitiligo.
2. Approach
2.1. Interaction between light and skin
Skin is made up of multiple layers of epithelial tissues which
protect muscles and organs [1]. It has three primary layers: the
epidermis, the dermis and the hypodermis (subcutaneous
tissue). Melanin is a colour pigment found in epidermis layer
of skin. The blood vessels are all located in the dermis [1].
Anderson and Parrish found that when an incident light
approaches the skin surface, a small portion of the light will
be reflected because of the difference in the index of reaction
between the air and skin surface [6]. From their experiment,
the surface reflectance was found to be typically 4–7%
over the entire spectrum of 250–3000 nm [6,7]. Figure 2
illustrates the interaction between skin and incoming light.
The light reflections of skin can be defined by several
components, as illustrated in figure 1. At incident angles
close to normal, +5% of the incident light coming in
contact with skin is directly reflected at the surface. This is
mainly due to the change in refractive index between air
(nD=1.0) and skin (nD=1.5).
The primary chromophores in skin are melanin and
haemoglobin [7], as shown in figure 3. Most of the incident
light (nearly 95%) penetrates into skin and follows
a complex path until it exits back out of the skin or gets
attenuated by skin chromophores [7].
2.2. Flowchart of the algorithm
The spatial distribution of melanin and haemoglobin in
digital images of skin can be separated by employing
independent component analysis of a skin colour image [9].
In our work, principal component analysis (PCA) is used to
reduce the dimensions from three (RGB) to two principal
components: principal component 1 (PC 1) and principal
component 2 (PC 2). Next, independent component
analysis (ICA) is used to align principal component axes
to represent pure density vectors of melanin and haemo-
globin [10]. As a result, we will have skin images that
represent skin areas due to melanin and haemoglobin. The
vitiligo areas in the melanin images are then segmented
using a thresholding method based on the Euclidean
distance. The overall flowchart is shown in figure 4.
2.2.1. Principal component analysis. Principal component
analysis is used as a dimensional reduction tool. To find the
principal components of an image, the mean value is
initially subtracted from each of image spectral bands
(RGB) to obtain zero mean variable data.
R ¼ R0 � mR0; G ¼ G0 � mG0
; B ¼ B0 � mB0; ð1Þ
Figure 1. Vitiligo lesion.
Table 1. Physician’s global assessment scale.
Repigmentation Scale
0–25% Mild
26–50% Moderate
51–75% Excellent
76–100% Complete
Figure 2. Schematic representation of the major optical
pathways in human skin [8].
Figure 3. Distribution model of melanin and haemoglobin
in skin.
102 M. H. Ahmad Fadzil et al.
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
where R0, G0 and B0 denote the image spectral band before
subtraction and mR0, mG0
and mB0denote the mean value of
image spectral bands.
The covariance matrix of the three dimensional RGB
dataset is then calculated. The covariance matrix of an
image in the RGB dataset is defined as follows:
COV ¼CRR CGR CBR
CRG CGG CBG
CRB CGB CBB
24
35; ð2Þ
where
CXX ¼1
N
XNi¼1
Xi � mið Þ2 ð3Þ
CXY ¼ CYX¼1
N
XNi¼2
XiYi
!� mXmY ð4Þ
mX ¼1
N
XNi¼1
Xi; ð5Þ
where X;Y 2 R;G;Bf g and Nand m denote number of
pixels in the image and means value, respectively.
The eigenvectors are then determined from the covar-
iance matrix by solving the following equation:
COV ¼ glgT; ð6Þwhere l is a diagonal matrix representing eigenvalues of
covariance matrix, COV and g is a matrix of eigenvectors of
covariance matrix, COV, arranged as a columns.
The eigenvectors are used as linear transform of original
(R, G, B) values. It is reported that the resulting vectors
have uncorrelated components, or in other words, the
primary axis of the data has been aligned where the
variance is maximal [11,12]. The vectors in new space
[X1 X2 X3]T are obtained by
X1
X2
X3
24
35 ¼ g11 g12 g13
g21 g22 g23g31 g32 g33
24
35 R
GB
24
35 ð7Þ
whereg11 g12 g13g21 g22 g23g31 g32 g33
24
35 are the eigenvectors of the covar-
iance matrix.
It has been shown that the RGB skin images can be
adequately represented by using two principal components
with an accuracy of 99.3% [11,12].
2.2.2. Independent component analysis. Skin colour dis-
tribution can be expressed as functions of pure density
vector of melanin and haemoglobin [13]. Let S1(x,y)
and S2(x,y) represent the quantities of the two colour
pigments on image coordinate (x,y), which are independent
variables. The colour vectors of the two pigments per unit
quantity are denoted a1 and a2, respectively. It is also
assumed that the compound colour vector v(x.y) can be
calculated using linear combination of the colour vectors as
follows:
vðx; yÞ ¼ a1s1ðx; yÞ þ a2s2ðx; yÞ: ð8ÞThis equation can be written as:
vðx; yÞ ¼ Asðx; yÞ ð9Þ
A ¼ a1½ a2� ð10Þ
Equation (9) is known as independent component analysis
(ICA), where A is a mixing matrix and s contains the
independent components.
The goal of ICA is to find a linear transformation W of
the dependent sensor signals v that makes the output
independent as possible [9,14–16] (equation (9)). ICA is
illustrated in figure 5. In addition, it is also necessary to
make vector, v, zero mean and unit variance in order to get
stronger independence conditions [14]. A preliminary study
using historical patient data has been reported earlier [17].
u ¼Wv ð11Þ
Figure 4. Flowchart of the algorithm.
Assessing therapeutic response in vitiligo skin disorder 103
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
Matrix W is estimated using an ICA method developed by
Hyvarinen, and known as FastICA [9,15,16]. The outputs
resulting from the ICA process are images that represent
skin areas due to melanin and haemoglobin.
2.3. Thresholding based on the Euclidean distance
Thresholding is an image segmentation technique that
segments foregrounds/objects and their background in a
digital image. In the algorithm, thresholding based on
Euclidean distance [18] is performed on the melanin of a
skin colour image to separate vitiligo skin lesion from
healthy skin.
In the technique, two samples of the pigment melanin,
each representing the vitiligo skin lesion and healthy skin
segments are obtained. The Euclidean distance is used to
map each pixel to its closest segments.
2.4. Determination of repigmentation area
Vitiligo areas are defined as skin areas that lack melanin.
From skin images that represent skin areas due to melanin,
we can determine the area of vitiligo lesions and thus,
progression of repigmentation.
The difference in the vitiligo areas between skin images
before and after a treatment period will be expressed as
a percentage of repigmentation in each vitiligo lesion.
This percentage will represent the repigmentation pro-
gression of a particular body region. Dermatologists
choose the vitiligo surface areas whilst the details of
camera position and locations of the lesions are recorded
by clinicians. The information will be used to ensure
similar condition in subsequence camera shots during
treatment process. Errors due to geometrical changes are
reduced.
The measurement is used as an objective method to
determine the progression of repigmentation during treat-
ment, and the efficacy of vitiligo treatment.
3. Performance measurement
The performance of the algorithm is measured by employ-
ing the algorithm on reference model images.
3.1. Reference model
Reference model images are simulated images that repre-
sent healthy skin and vitiligo skin images. The images are
modelled based on the distribution of colour combinations
in the three spectral bands, namely red, green and blue
(RGB). The distribution models are developed using sam-
ples of skin colour taken from historical data of patients.
These samples are chosen by dermatologists to obtain valid
Figure 5. Independent component analysis.
Figure 6. Simulated (a) healthy skin and (b) vitiligo lesion
images.
Figure 7. (a) Reference model A; (b) reference model B;
(c) reference model C; and (d) reference model D.
104 M. H. Ahmad Fadzil et al.
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
reference model. The distribution of each spectral value is
then modelled using Gaussian distribution. Equation (12)
shows the Gaussian distribution function.
fðx; m; sÞ ¼ 1
sffiffiffiffiffiffi2pp e�
x�mð Þ2
2s2 ; ð12Þ
where s is the standard deviation and m is the mean value.
The Gaussian distribution is chosen based on previous
studies of skin modelling by Caetano and Barone [10], Zhu
et al. [19] and Chang et al. [20]. They reported that the skin
colour distribution can be modelled using Gaussian
distributions.
A healthy skin model image as shown in figure 6(a) is
produced using a distribution model of healthy skin. The
vitiligo lesion model image, as shown in figure 6(b) is
produced using the distribution model of vitiligo lesions.
Using the healthy skin and vitiligo lesion model images,
four reference images, i.e. reference images A, B, C and D
(see figure 7) are constructed and used in investigating the
capability of the developed algorithm. Each reference
image is 2006 200 pixels in size. The size of the reference
image is constructed based on the size of vitiligo lesion
found on the data.
Reference image A models a skin image having a vitiligo
lesion with an area of 406 50 pixels (see figure 7(a)). Refer-
ence image B is created like reference image A (figure 7(b)),
Figure 8. Reference model A, (a) melanin and (b)
haemoglobin.
Figure 9. Reference model B, (a) melanin and (b)
haemoglobin.
Figure 10. Reference model C, (a) melanin and (b)
haemoglobin.
Figure 11. Reference model D, (a) melanin, (b) haemoglo-
bin and (c) 1-by-1 pixel repigmentation.
Figure 12. Patient C (a) RGB image taken at baseline,
(b) RGB image taken six weeks later.
Assessing therapeutic response in vitiligo skin disorder 105
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
but in its vitiligo lesion area three areas which represent
repigmented skin areas are added. The size of each
repigmentation area is 56 5 pixels. Reference images C
and D (figures 7(c), (d)) are created like reference image B.
In reference image C the size of the repigmentation areas is
36 3 pixels, whilst in reference image D the size is reduced
to 16 1 pixel.
3.2. Testing the algorithm
By applying the developed system to all the reference ima-
ges, it has been able to discern vitiligo lesions, healthy skin
and skin repigmentation areas, as depicted in figures 8–11.
Under ideal conditions, the system can even detect skin
repigmentation areas whose sizes are only 1-by-1 pixel.
4. Results and analysis
4.1. Sample data
The vitiligo digital image analysis system is applied to the
skin images provided by Hospital Kuala Lumpur, Malay-
sia. A group of 20 patients with different clinical features of
vitiligo (generalized vitiligo, acrofacial vitiligo, segmental
vitiligo, focal vitiligo and universal vitiligo) are randomly
chosen by a physician. The images of the vitiligo lesions are
then captured using Nikon digital camera SLR D100.
Figure 12 shows samples from patient C. The images refer
to the same skin area and are taken within a 6-week interval.
It is observed that some repigmentation has occurred.
Figures 13(a)–(f) show the same skin area with lesion
(digitally enlarged) of patient C being processed by the digital
image analysis technique. The processed images now repres-
ent skin areas due to the melanin component (figures 13(b)
and (e)) and the haemoglobin component (figures 13(c)
and (f)). As discussed in x1, vitiligo lesions occur due to the
abnormality of melanin production. From the melanin
component image, we can determine melanin and non-
melanin (vitiligo) areas. By comparing previous and current
data the progression of repigmentation can be assessed.
Figure 13. Processed images of patient C: (a) RGB image at
baseline, (b) melanin at baseline, (c) haemoglobin at
baseline, (d) RGB image, six weeks later, (e) melanin, six
weeks later, and (f) haemoglobin, six weeks later.
Figure 14. Patient C (a) segmented melanin image at
baseline, (b) segmented melanin image, six weeks later.
106 M. H. Ahmad Fadzil et al.
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
Segmented melanin images are shown in figure 14. To
ensure the accuracy of the system, the actual size of the
pixel in each image is calibrated using a reference marker.
For example, in figure 8(a), the marker (which has an area
of 113 mm2) is represented by 45 972 pixels. Thus, in
figure 12(a), 1 pixel represents 2.4586 1072 mm2. This
calibration process is performed in every image before we
measure the vitiligo lesion area in the image.
4.2. Analysis
Table 2 shows the vitiligo areas measured by the proposed
image analysis algorithm based on two visits. Table 3
Table 2. Vitiligo area measurements.
Vitiligo areas (mm2)
Body area 1st visit 2nd visit (6-week interval)
A Head 170.35 161.1
B Head 2447.58 2227.4
C Head 66.94 54.3
Lower limb 282.3 249.9
Trunk 577.7 511.6
Upper limb 3632.5 3595
D Head 3102.596 3049.7
E Trunk 4021.67 3796.23
Upper limbs 5666.8 5236.2
Head 164.92 85.4
F Trunk 2463.4 2142.37
G Head 363.5 318.6
H Feet 155.38 151.52
Head 389.85 347.1
Upper limbs 319.79 289.9
Lower limbs 1010.24 1000.88
Trunk 35.7 34.8
I Hands 83.169 82.06
Face 486.62 459.2
J Feet 95.8 94.5
Hand 61.8 57.6
Face 272.7 223.12
Upper limb 744.6 739.5
Lower limb 813.47 809
K Face 91.9 79.52
L Lower limb 89.6 84.68
Upper limb 47.09 37.8
M Head 81.61 78.7
Trunk 1882.7 1805.18
N Feet 196.306 171.69
Hand 175.9 164.05
Head 98.2 77.29
Lower limb 308.79 306.08
Upper limb 302.11 265.4
O Feet 141.61 134.7
Trunk 420.2 334.1
Upper limb 336.4 312.27
P Head 41.95 38.52
Q Feet 87.9 82.9
Hand 73.72 65.53
Head 58.4 54.44
Lower limb 128.77 113.34
Trunk 591.36 580.4
R Feet 677.7 611.29
Hand 223.6 213.78
Neck 69.74 69
Lower limb 1695.14 1420
Upper limb 457.1 448.8
S Head 156.06 142.88
T Head 1148.1 938.46
Table 3. Comparison between PGA and the proposedalgorithm.
Body areas PGA Proposed algorithm
A Head Mild 5% (Mild)
B Head Mild 9% (Mild)
C Head Mild 19% (Mild)
Lower limb Mild 11% (Mild)
Trunk Mild 11% (Mild)
Upper limb Mild 1% (Mild)
D Head Mild 2% (Mild)
E Trunk Mild 6% (Mild)
Upper limbs Mild 8% (Mild)
Head Moderate 48% (Moderate)
F Trunk Mild 13% (Mild)
G Head Mild 12% (Mild)
H Feet Mild 2% (Mild)
Head Mild 11% (Mild)
Upper limbs Mild 9% (Mild)
Lower limbs Mild 1% (Mild)
Trunk Mild 3% (Mild)
I Hands Mild 1% (Mild)
Head Mild 6% (Mild)
J Feet Mild 1% (Mild)
Hand Mild 7% (Mild)
Head Mild 18% (Mild)
Upper limb Mild 1% (Mild)
Lower limb Mild 1% (Mild)
K Face Mild 13% (Mild)
L Lower limb Mild 5% (Mild)
Upper limb Mild 20% (Mild)
M Head Mild 4% (Mild)
Trunk Mild 4% (Mild)
N Feet Mild 13% (Mild)
Hand Mild 7% (Mild)
Head Mild 21% (Mild)
Lower limb Mild 1% (Mild)
Upper limb Mild 12% (Mild)
O Feet Mild 5% (Mild)
Trunk Mild 20% (Mild)
Upper limb Mild 7% (Mild)
P Head Mild 8% (Mild)
Q Feet Mild 6% (Mild)
Hand Mild 11% (Mild)
Head Mild 7% (Mild)
Lower limb Mild 12% (Mild)
Trunk Mild 2% (Mild)
R Feet Mild 10% (Mild)
Hand Mild 4% (Mild)
Neck Mild 1% (Mild)
Lower limb Mild 16% (Mild)
Upper limb Mild 2% (Mild)
S Head Mild 8% (Mild)
T Head Mild 18% (Mild)
Assessing therapeutic response in vitiligo skin disorder 107
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
compares repigmentation assessment by physician’s global
assessment and the developed algorithm. Two dermatolo-
gists estimated the physician’s global assessment scores by
comparing the digital images of vitiligo lesions before and
after six weeks of treatment.
As recorded in table 2, after six weeks of treatment, it can
be seen that vitiligo areas are somewhat reduced indicating
positive response to the treatment. The changes in vitiligo
areas allow us to determine repigmentation progression.
This information can be used in assisting physician to
evaluate the efficacy of the treatment in a shorter time
period. In comparison to physician’s global assessment
scores (table 3), the percentages obtained using the
proposed method are found to be within the physician’s
global assessment range. It is shown that the proposed
method is consistent with the physicians. Based on the 20
patients, the proposed method is very accurate as it did not
produce any errors.
5. Conclusion
Skin pigment disorders cause irregular pale patches of the
skin. The disorders occur due to the abnormality of
melanin production. The use of digital image analysis in
this research enables the objective assessment of repigmen-
tation progression and evaluation of efficacy of therapeutic
procedures. A global assessment of therapeutic response is
made by the physicians using a scoring system based on an
ordinal scale, of which physician’s global assessment (PGA)
is one example. It is shown that in order to determine
repigmentation progression objectively, the conversion of
RGB images to melanin related skin images is essential.
This is achieved by using PCA to dimensionally reduce 3D
RGB images into 2D data consisting of the principal
components. By using ICA, we are able to align the 2D
data into melanin and haemoglobin related data, resulting
in a skin image that effectively represents melanin. The
digital image analysis method is able determine vitiligo
lesion (non-melanin area) and repigmentation progression
on a shorter time frame.
Results from 20 patients show that the proposed
algorithm has the ability to objectively measure the
therapeutic response of vitiligo treatment. It is found
that all of the percentages obtained using the proposed
method are comparable to the physician’s global assess-
ment range as verified by dermatologists (two dermatol-
ogists scored by comparing the lesion). Thus, the
proposed method is highly accurate as it is consistent in
comparison to physician’s global assessment performed by
dermatologists.
Acknowledgement
This research is a collaborative work between Universiti
Teknologi PETRONAS and the Dermatology Department,
Hospital Kuala Lumpur. The authors would like to
acknowledge Ahmad Tarmidzi Zakaria from the Derma-
tology Department, Hospital Kuala Lumpur for his
assistance in data collection.
Declaration of interest: The authors report no conflicts of
interest. The authors alone are responsible for the content
and writing of the paper.
References
[1] Hann, S.K. and Nordlund, J.J., 2000, Vitiligo. A monograph of the
basic and clinical science (Oxford: Blackwell Science).
[2] Kovacs, S., 1998, Continuing medical education: vitiligo. Journal of
the American Academy of Dermatology, 38, 647–666.
[3] Lerner, A., 1959, Vitiligo. Journal of Investigative Dermatology, 32,
285–310.
[4] Roberts, N. and Lesage, M., 2003, Vitiligo: causes and treatment. The
Pharmaceutical Journal, 270, 440–442.
[5] Maglogiannis, I., 2003, Automated segmentation and registration of
dermatological images. Journal of Mathematical Modelling and
Algorithms, 2, 277–294.
[6] Anderson, R.R. and Parrish, J.A., 1981, The optics of human skin.
Journal of Investigative Dermatology, 77, 13–19.
[7] Stamatas, G.N., Zmudzka, B.Z., Kollias, N. and Beer, J.Z., 2004,
Review: innovative technology, non-invasive measurements of skin
pigmentation in situ. Pigment Cell Res, 17, 618–626.
[8] Storring, M., Andersen, H.J. and Granum, E., 2000, Skin colour
detection under changing lighting conditions. Journal of Autonomous
System.
[9] Hyvarinen, A. and Oja, E., 2000, Independent component analysis:
algorithm and applications. Neural Networks, 13, 411–430.
[10] Caetano, T.S. and Barone, D., 2001, A probabilistic model
for the human skin colour. In Proceeding of the 11th Inter-
national Conferene on Image Analysis and Processing, 26–28
September 2001, Palermo, Italy, 279–283. IEEE Computer
Society.
[11] Hance, G., Umbaugh, S., Moss, R. and Stoecker, W., 1996, Unsuper-
vised colour image segmentation with application to skin tumor
borders. IEEE Engineering in Medicine Biology Magazine, 15, 104–
111.
[12] Umbaugh, S., Moss, R. and Stoecker, W., 1993, Automatic colour
segmentation algorithms. IEEE Engineering in Medicine and Biology
Magazine, 12, 75–82.
[13] Tsumura, N., Haneishi, H. and Miyake, Y., 1999, Independent
component analysis of skin colour image. Journal of the Optical
Society of America A, 16, 2169–2176.
[14] Karhunen, J., Oja, E., Wang, L., Vigario, R. and Joutensalo, J., 1997,
A class of neural networks for independent component analysis. IEEE
Transactions on Neural Networks, 8, 486–504.
[15] Hyvarinen, A., 1999, Fast and robust fixed–point algorithms for
independent component analysis. IEEE Transactions on Neural
Networks, 10, 626–634.
[16] Hvyarinen, A., 1998, New approximations of differential
entropy for independent component analysis and projection
pursuit. Advances in Neural Information Processing System, 10, 273–
279.
[17] Nugroho, H., Ahmad Fadzil, M.H., Yap, V.V., Norashikin, S. and
Suraiya, H.H., 2007, Determination of skin repigmentation progres-
sion. Conf Proc. IEEE Eng Med Biol Soc, 3442–3445.
[18] Heckbert, P., 1982, Colour image quantization for frame buffer
display. In Proceedings of ACM SIGGRAPH. Computer Graphics,
16(3), 297–307.
108 M. H. Ahmad Fadzil et al.
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.
[19] Zhu, E., Yang, J. and Waibel, A., 2000, Segmenting hands of arbitrary
color. In Proceedings of the Fourth IEEE International Conference on
Automatic Face and Gesture Recognition, 28–30 March 2000,
Grenoble, France, 446–453.
[20] Chang, F., Ma, Z. and Tian, W., 2007, A region-based skin color
detection algorithm. In Proceedings of the 11th Pacific-Asia Conference
on Knowledge Discovery and Data Mining, 22–25 May 2007, Nanjing,
China, 417–424. Lecture Notes in Computer Science.
Assessing therapeutic response in vitiligo skin disorder 109
J M
ed E
ng T
echn
ol D
ownl
oade
d fr
om in
form
ahea
lthca
re.c
om b
y U
nive
rsity
of
Con
nect
icut
on
08/2
9/13
For
pers
onal
use
onl
y.