The Median Split Algorithm for Detection of Critical Melanoma
Color Features
Kaushik V. S. N. Ghantasala
1
, Raeed H. Chowdhury
2
, Uday Guntupalli
1
, Jason Hagerty
1,2
,
Randy H. Moss
1
, Ryan K. Rader
2
and WilliamV.Stoecker
2
1
Missouri University of Science And Technology, G20 Emerson Electrical Co. Hall, Rolla, MO 65409, U.S.A.
2
Stoecker & Associates, 10101 Stoltz Drive, Rolla, MO 65401, U.S.A.
Keywords: Median Split, Melanoma, Image Analysis, Color Processing, Dermoscopy.
Abstract: Detection of melanoma remains an empirical clinical science. New tools for automatic discrimination of
melanoma from benign lesions in digitized dermoscopy images may allow an improvement in early
detection of melanoma. This research implements a fast version of the median split algorithm in an open
source format and applied to four-color splitting of the lesion area to capture the architectural disorder
apparent in melanoma colors. Our version of the median split algorithm splits colors along the color axis
with maximum Range. For a set of 888 dermoscopy images, the best model for discrimination produces an
area under the receiver operating characteristic curve of 0.821. Logistic regression analysis of 242
parameter variables obtained from 888 images shows that the most important features in the final model,
measured by Wald Chi-square significance, are the lengths of two peripheral inter-color boundaries and one
measure of boundary overlay by different colors. The median split algorithm is fast, requiring less than one
second per image and only a four-color splitting, but it captures sufficient critical information regarding
color disorder, with peripheral inter-color boundaries showing the highest significance for melanoma
discrimination.
1 INTRODUCTION
Early detection of melanoma may be aided by
analytic methods applied to dermoscopy images of
melanoma, which offer the possibility of detecting
potential melanomas before they are sufficiently
advanced to affect life expectancy. Color methods
splitting the entire lesion were investigated by
Andreassi et al. Eccentricity of color components
and presence of color islands were important in
discriminating melanomas from benign lesions
(Andreassi et al., 1999). Colors were also used to
discriminate melanomas from benign lesions based
on three-dimensional color probability histograms
that were measured by both crisp methods (Faziloglu
et al., 2003) and fuzzy logic methods (Khan et al.,
2009). In this paper, we describe a technique termed
the median split technique (Umbaugh, 2011), which
we use to capture the architectural disorder of early
in situ melanoma. This technique has the advantages
of speed, simplification of lesion architecture yet
retention of critical features, and high discriminatory
power for melanoma.
2 MEDIAN SPLIT ALGORITHM
The median split algorithm has been previously
applied to entire images using CVIPtools
(http://cviptools.ece.siue.edu). This algorithm is
based on the Heckbert color compression algorithm
(Heckbert, 1982). In this research, we apply the
median split algorithm to a specific region of interest
(ROI)—the lesion only. The motivation is to
quantize the ROI so that fewer colors are used in
order to describe the ROI. The simplified image
allows exact quantization of color values, color
areas, and inter-color boundaries. The color space
segmentation is performed by splitting the pixel
histogram of a color segment. Each iteration splits
this color segment into two segments with equal
pixel populations. The segment with the highest
range in any color axis is chosen for the subsequent
split. Within the chosen segment, the split is
performed along the color axis with the highest
range. The division occurs at the median pixel m on
the chosen axis. Formally, the chosen color axis and
chosen color bin satisfies:
492
Ghantasala K., Chowdhury R., Guntupalli U., Hagerty J., Moss R., Rader R. and Stoecker W..
The Median Split Algorithm for Detection of Critical Melanoma Color Features.
DOI: 10.5220/0004304904920495
In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 492-495
ISBN: 978-989-8565-47-1
Copyright
c
 2013 SCITEPRESS (Science and Technology Publications, Lda.)

where {
}
conditio
n
histogra
m
new bin
s
for the
R
list is u
p
range al
split. T
h
number
colors s
u
(Umbau
g
process
decision
space a
s
lesion o
n
region r
e
2.1
M
Figu
r
Figure 2
:
regions.
}
denotes the
n
within {}
m
of
p
ixels
w
s
replace the
R
OI. After a
p
dated, and
a
l
ong any col
o
h
e algorithm
of colors is
r
u
fficed
t
o rep
r
g
h, 2011). F
for a melan
o
variable in r
e
s
performed
n
ly. Pixels in
e
maining in t
h
M
edian Spl
r
e 1: Original
m
:
Median spli
t
number of pi
x
; and hist(
b
w
ithin the ch
o
old bin in th
e
given split,
t
a
gain the seg
m
or
axis is ch
o
continues u
n
r
eached. Um
b
r
esen
t
the col
o
igures 1-3 s
h
o
ma image u
s
e
d, green and
on the RO
I
a bin are ma
p
h
e final image
it Exampl
e
m
elanoma: der
m
t
, 3 splits (R
G
x
els satisfyin
g
b
in) denotes
o
sen bin. The
e
list of colo
r
t
he segment
(
m
en
t
with hi
g
o
sen for the
n
til the spec
i
b
augh found
or
s within a l
e
h
ow the spli
t
s
ing range a
s
blue (RGB)
c
I
—the melan
p
ped to the la
r
.
e
m
oscopy image
.
G
B) create 4
c
g
the
the
two
r
ins
(
bin)
g
hest
next
i
fied
four
e
sion
t
ting
s
the
c
olor
n
oma
r
gest
.
c
olor
Fig
u
lesi
o
2.
2
for
en
d
W
e
eac
thr
e
co
m
2.
3
Ra
n
var
i
var
i
do
m
dis
t
Pre
sig
n
ski
n
Ac
c
de
c
Th
i
im
a
ca
p
u
re 3: Media
n
o
n.
2
Media
n
no_of_splits
(
for no_of
_
-> find
c
inner loo
p
-> find a
b
-> arrang
e
order of c
h
-> split b
as possibl
-> remap
in bin are
d
of outer loo
p
e
set no_of_s
p
h
bin: red,
g
e
e channels
m
m
ponent axes.
3
Media
n
n
ge is chos
e
i
able. Altern
a
i
ance and
i
m
ain, contac
t
t
ributed LED
-
p
rocessing
n
ificant artif
a
n
and bu
b
c
ordingly, w
e
c
ision variabl
e
i
s has
p
otenti
a
a
ge, such as
p
tured well by
split, 3 split
s
Split Pseu
(
outer loop b
e
_
current_bins
c
hannel with
p
b
in with the l
a
e
pixels in bi
n
h
annel with l
a
i
n into nearl
y
e after splitti
n
colors in bi
n
set to mean o
p
lits = 3. Thre
e
g
reen and blu
m
ay be used,
Split Deci
s
e
n as the
m
a
te decision
v
i
nte
r
-quartile
dermoscop
y
l
ighting, ther
e
a
llows elimi
n
a
cts, e.g. hair
s
b
bles from
e
retained t
h
e
: range with
i
a
l to capture
s
small netwo
r
variance or i
n
s
create 4 col
u
docode
e
gins)
(inner loop)
largest rang
a
rgest range
n
with larges
t
a
rgest range
y
equal bins
(
n
g)
n
s to image (
t
o
f bin )
e
channels ar
e
u
e, but in pra
c
for example,
sion Varia
b
m
edian split
v
ariables ma
y
distances.
y
images wi
t
e
is little rand
o
i
nation of
n
r
s
p
rotruding
the gel
h
e CVIPtool
i
n a given co
l
small variati
o
r
ks that wou
l
n
te
r
-quartile
d
o
rs in the
e end of
t
range in
(
as equal
t
he pixels
e
used for
c
tice, any
principal
b
le
decision
y
include
For this
t
h evenly
o
m noise.
n
early all
from the
interface.
s
default
l
o
r
plane.
o
ns in the
l
d not be
d
istance.
TheMedianSplitAlgorithmforDetectionofCriticalMelanomaColorFeatures
493

3 EXPERIMENTS
3.1 Instrumentation and Images
The contact dermatoscope used in this study is the
3Gen DermLite Fluid attachment (3Gen LLC, San
Juan Capistrano, CA). This device uses bright white
LED lights, 10X magnification and a gel interface.
3.2 Images
A set of 195 melanoma and melanoma in situ images
were obtained in the study SBIR R44 CA-101639-
02A2 of the National Institutes of Health (NIH). A
similar benign set of 693 images were obtained for
the same study. Lesion borders (ROIs) were
manually drawn using second-order b-splines.
3.3 Median Split Features Obtained
for Each Lesion
1. Ring Value: Fraction of pixels in the border of
the lesion that overlap with the pixels in the color.
2. Total Lesion Area
3. Ratio of Each Color to Peripheral Ring
4. Average R, G, B values in Each Color Region
5. Number of Blobs before filling for each Color
region: This gives the number of blobs in each
region without filling holes or filtering the small
blobs. These blobs are the connected contours
present in that segment.
6. Number of Blobs after filling for each Color:
Each color blob with area less than 26 pixels is
eliminated and interior holes in the remaining blob
are filled.
7. Area of Each Color
8. Area of largest Blob: Area occupied by largest
blob in each color before and after filling the holes.
9. Internal/External Perimeter of each Color
10. Perimeter of largest Blob: Total internal and
external perimeter of largest blob for each color
11. Normalized Perimeter: Found by dividing
perimeter by square root of lesion area.
12. Perimeter Intensity Drop: Average RGB color
difference between the pixels that are on the border
of each color and those one pixel outside the border.
13. Normalized Intensity Drop: Perimeter intensity
drop divided by square root of total lesion area.
14. Centroid of each Color: The centroid of the
lesion and of each segment by Matlab region props.
15. Euclidean Distance: Distance between the
centroid of each color and the centroid of the lesion.
16. Normalized Distance of each Color: The ratio
of Euclidean distance and square root of color area.
17. Absolute, Background and Relative
luminance for each Color: Absolute luminance is
the measure of brightness of the desired segment:
Luminance = 0.30R + 0.59G + 0.11B.
18. Average background skin R, G, B Values:
Average Red Green Blue values of the skin color
that surrounds the lesion part in the original image.
19. Relative RGB values for each Color
20. Average Euclidean distance of each Color
21. Red Chromaticity
4 RESULTS
4.1 Performance
The median split algorithm in C running on a
Core™ 2 duo processor requires < 1 second/image.
4.2 Most Significant SAS Features
Logistic regression analysis using Statistical
Analysis Software (SAS) was used to analyze the
results. The significant features described in §3.3 are
shown by Chi-sq and p-value statistics (Table 1).
Table 1: Significant parameters (Section 3.3).
Feature Chi-Sq P value
Perimeter of largest
blob before filling (§
3.3.10)
84.31 <.0001
Ring value (§ 3.3.1) 28.30 <.0001
Normalized perimeter of
largest blob before filling
(§ 3.3.11)
19.72 <.0001
4.3 Receiver Operating Characteristic
(Roc) Curves
Accuracy of the algorithm was tested using the
Receiver Operating Characteristic (ROC) curve. The
ROC curve plots sensitivity versus one minus
specificity. We report here the area under the curve
(AUC) for five experiments. The five ROC curves
(Figure 4) depict results for five cases, adding
different parameters from perimeters measured on
all four color segments.
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
494

Figure 4:
p
erimete
r
with sep
a
AUC = 0
.
5 D
I
5.1
T
a
A
A new
d
clinicia
n
“A” in
t
CASH
Symmet
r
differen
t
asymme
t
split i
m
characte
r
number
o
make t
h
color
b
o
u
are hig
h
regressi
o
appears
in mela
n
5.2
M
a
The ch
o
alternat
e
algorith
m
includin
g
ROC curves,
r
cases. The b
e
a
rate external
a
.
82.
I
SCUSSI
O
T
he Media
n
a
nd the Go
a
A
rchitectu
r
d
ermoscopy c
l
n
s emphasiz
e
t
he CASH al
g
is an acro
n
r
y and Ho
m
t
colors, h
a
t
ry, and are
m
ages show
n
r
istics. The i
m
of colors (fo
u
h
e measurem
e
u
ndaries. The
h
in import
a
o
n analysis. T
h
to capture t
h
n
oma.
M
edian Spl
a
nd Chann
e
o
ice of the
R
e
color spaces
m
could be e
x
g
L*a*b*, H
S
sensitivity vs.
e
st case for no
r
a
nd internal c
o
O
N
n
Split Al
go
a
l of Capt
u
r
al Disorde
r
l
assification
a
e
s architectu
r
g
orith
m
(Hen
n
n
ym for Co
l
m
ogeneity.
M
a
phazard ar
c
inhomogene
o
n
capture al
l
m
age is quan
t
u
r colors suffi
c
e
nt of the le
n
se inter-color
a
nce as mea
s
h
e median sp
l
h
e haphazard
it Channel
e
l Pre-pro
c
R
GB color
s
is arbitrary.
x
tended to o
t
S
V, and CIE-
L
1-specificity,
f
r
malized peri
m
o
lor perimeter
s
o
rithm
u
rin
g
r
a
lgorithm use
d
r
al disorde
r
—
n
ing et al, 2
0
o
r
, Architec
t
M
elanomas
h
c
hitecture,
c
o
us. The me
l
four of t
h
t
ified into a s
m
c
e) enabling
u
n
gth of the i
n
b
oundary le
n
s
ured by lo
g
l
it color algor
i
color distrib
u
s
c
essin
g
s
pace rather
The median
t
her color sp
a
L
UV color sp
a
f
or 5
m
eters
s
has
d
by
—
the
0
07).
t
ure,
h
ave
c
olo
r
e
dian
hese
m
all
u
s to
n
te
r
-
n
gths
g
istic
i
thm
u
tion
than
split
a
ces,
a
ces.
Th
e
co
m
6
Th
e
me
l
Se
p
me
a
as
0
me
l
the
r
(A
n
arb
i
dis
o
fro
m
R
E
An
d
Faz
i
He
c
He
n
Kh
a
U
m
e
method cou
m
ponent trans
f
CONCL
U
e
median sp
l
l
anoma an
d
p
aration of m
e
a
sured by are
a
0
.82. This me
t
l
anomas, the
r
efore is not
d
n
dreassi et al
i
trary selecti
o
o
rder. The
m
m
inte
r
-color
p
E
FEREN
C
d
reassi, L., P
e
Cevenini, G.,
B
P. (1999).
D
differentiation
new quantitati
v
135(12), 1459
-
i
loglu Y, Sta
n
McLean RP.
melanoma di
s
R
esearch and
T
c
kbert, P. (198
2
b
uffer displa
y
297.
n
ning, J.S., Du
s
Rabinovitz, H
.
The CASH
homogeneity)
the American
A
a
n, A., Gupta,
K
R.H., Argenz
i
Fuzzy logic t
e
in dermosco
p
Imaging and
G
b
augh, S.E. (
2
A
nalysis: Hu
m
with CVIPtool
s
l
d also be ex
t
f
orm applied
t
U
SIONS
l
it algorithm
d
benign
d
e
lanomas fro
m
a
under the
R
t
hod was app
l
majority at
t
d
irectly comp
a
., 1999), (K
h
o
n of four
m
ost critical
f
p
erimeters.
C
ES
e
rotti, R., Ru
b
B
iagioli, M., T
D
igital dermo
s
of atypical ne
v
v
e semiology.
A
-
65.
n
ley RJ, Moss
(
2003). Colou
r
s
crimination i
n
T
echnology. 9(
2
2
). Color ima
g
. In Proceedi
n
s
za, S.W., Wan
g
.
S., Polsky, D.
(color, archit
e
algorithm for
A
cademy of De
r
K
., Stanley, R.
J
i
ano, G., …
chniques for b
p
y images.
G
raphics, 33(1)
,
2
011). Digital
m
an and Comp
s
(2
nd
ed.). Boc
a
t
ended to the
t
o these color
was applie
d
dermoscopy
m
the benign
l
R
OC curve w
a
l
ied to a diffi
c
the in situ s
t
a
rable to earli
h
an et al., 2
0
colors captu
r
features wer
e
b
egni, P., Bu
r
T
addeucci, P.,
…
s
copy analysi
s
v
i and early m
e
A
rchives of De
r
RH, Van St
o
r
histogram a
n
n
clinical im
a
2
), 147-156.
g
e quantization
n
gs of SIGG
R
g, S.Q., Marg
h
, & Kopf, A.
W
t
ecture, sym
m
dermoscopy.
J
r
matology, 56(
1
J
., Stoecker,
W
Cognetta, A.
B
b
lotch features
Computerized
,
50-57.
Image Proc
e
p
uter Vision A
p
a
Raton: CRC
P
principal
spaces.
d
to 888
images.
l
esions as
a
s as high
c
ult set of
t
age, and
er studies
0
09). The
r
es color
e
derived
r
roni, M.,
…
Barbini,
s
for the
e
lanoma: a
r
matology,
o
ecker W,
n
alysis for
a
ges. Skin
for frame
R
APH ’82,
oob, A.A.,
W
. (2007).
m
etry, and
J
ournal of
1
), 45-52.
W
.V., Moss,
B
. (2009).
evaluation
Medical
e
ssing and
p
plications
P
ress.
TheMedianSplitAlgorithmforDetectionofCriticalMelanomaColorFeatures
495