Indonesian Journal of
Electrical
Engineer
ing and
Computer Science
V
o
l. 10
, No
. 3, Jun
e
20
18
, pp
. 10
30
~
1
035
ISSN: 2502-4752,
DOI: 10.115
91/ijeecs
.v10.i
3.pp1030-1035
1
030
Jo
urn
a
l
h
o
me
pa
ge
: http://iaescore.c
om/jo
urnals/index.php/ijeecs
Fusion of Random Projection, Multi-resolution Features and
Distan
ce Wei
g
hted K Nearest Nei
g
hbor f
o
r Mass
es
Det
e
ction
in Mammographic Images
Viet Dun
g
N
g
uyen
1
,
Minh Dong
Le
2
1
Department of Biomedical
Eng
i
neering
,
Hano
i Un
ivers
i
t
y
of
S
c
ienc
e
and T
echn
o
log
y
, Vi
etn
a
m
2
Department of Computer
Scien
ce,
C
honnam National University, South Kor
e
a
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
Ja
n 11, 2018
Rev
i
sed
Mar
19
, 20
18
Accepted
Mar 30, 2018
Breast can
cer
is
the top cancer
in women bot
h in the develo
ped and the
develop
i
ng world. For
early detection of the dise
ase, ma
mmogra
phy is stil
l
the most effective method b
e
side ultrasoun
d
and magnetic
resonance
imaging.
Computer A
i
ded D
e
tection systems h
a
ve been
deve
l
oped t
o
ai
d
radiolog
ists in
diag
nosing
b
r
east cancer. D
i
ffe
rent meth
ods w
e
re
proposed t
o
ov
ercome the mai
n
draw
back of
producin
g large
number of
False Po
sitiv
e
s.
In this paper, we presen
ted a novel method for masses
detection
in mammogra
m
s. To descri
be masses, multi-resolution featur
es
were ut
ili
zed
. In
featur
e ex
tra
c
tio
n s
t
ep,
we
cal
cul
a
ted
m
u
lti-res
o
lu
tion Blo
c
k
Differenc
e Inve
rs
e P
r
obabili
t
y
featur
es
and
m
u
lti-res
o
lut
i
o
n
s
t
atis
ti
ca
l
featur
es. Once the descriptor
s we
re extracted, we deplo
y
ed random
projection
and distance wei
ght
ed
K Neares
t Ne
igh
bor to c
l
as
s
i
f
y
th
e de
tec
t
ed
ma
sse
s.
The
re
sult is quite
sa
nguine
with
sensit
i
v
it
y,
false
positi
ve redu
ctio
n
and time for
carr
y
ing
out the algo
rithm
K
eyw
ords
:
Distance weighted K nea
r
est
nei
g
hb
o
r
Ma
mm
o
g
r
aphy
Mass detection
Mu
lti-reso
l
u
tion
featu
r
es
R
a
nd
om
pro
j
ec
t
i
o
n
Copyright ©
201
8 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
V
i
et
Du
ng
Ngu
y
en
,
Depa
rt
m
e
nt
of
El
ect
roni
c
Tec
h
n
o
l
o
gy
a
n
d
B
i
om
edi
cal
Engi
neeri
n
g
,
Hano
i
Un
i
v
ersi
ty o
f
Scien
ce an
d Tech
no
log
y
,
No
1
.
Dai Co
Viet Str.
,
Han
o
i
, Vietnam
.
Em
a
il: d
u
n
g
.
ng
u
y
en
v
i
et1@hu
st.ed
u
.vn
1.
INTRODUCTION
B
r
east
cancer
is the most c
o
mmon cancer in women
wo
rldwide, with nearly 1.7 mill
ion new
cases diagnosed
in 201
2 [
1
]. Ab
no
rmal tissue sc
reen
ing usin
g X-r
a
y mammogr
aphy is curre
ntly the
most effective method o
f
early detection o
f
the di
sease [2-3]. T
h
e introduction
of di
gital mammography
gave the o
p
p
o
rtunity
of
in
creasing the num
ber
of c
o
mmercial Computer
Aide
d Detection (
C
AD
)
systems, which has significantly e
nhanced the radiolog
ists’ ability
to detect and diagnose cancer and
take immediat
e precautions
for its earliest
prevention [4]. One
problem with
C
A
D
syste
m
s
is
due
to a
large number of false positive
(FP) marks when hi
gh sensitivity
is required [5]. Too many fals
e
positives may confuse the radiologist of the most
common types
of
cancer among wome
n all over the
world is breast cancer. Grea
t effort
has be
en devoted in
recent years
to
the development
of CAD which
pro
pose a lot
of features to
red
u
ce
false
positives
[6]. However, many feat
ures are not ke
y features of
masses and they make high di
mensions f
o
r classificatio
n.
In this
pape
r
,
we introduce novel method
using mome
nt
and basic characteris
tic of the masses.
B
l
ock Difference Inverse Probability (
B
DIP
)
and basic f
eatures
are c
a
lcul
at
ed in different
multi-
resolutions. O
n
ce the features are
extracted, random p
r
ojection [7] a
nd k nearest neighb
or (k N
N
) [8]
with
distance
weighting are used to classif
y
the suspici
ous areas into real mass or no
rmal
parenchyma.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
d
onesi
a
n
J
E
l
ec En
g &
C
o
m
p
Sci
ISS
N
:
2
5
0
2
-
47
52
Fu
sion
o
f
Rand
om Pro
j
ection
,
Mu
lti-Reso
l
u
tio
n
Fea
t
u
r
es a
n
d
Distan
ce…
(Viet
Dun
g
Ng
u
y
en
)
1
031
2.
PROP
OSE
D
METHO
D
2.
1.
D
a
t
a
b
a
se
In t
h
i
s
st
u
d
y
,
we use m
a
m
m
og
ram
dat
a
base M
i
ni
- M
I
AS [9]
t
o
t
e
st
t
h
e
m
e
t
hod p
r
ese
n
t
e
d. M
I
A
S
i
s
t
h
e p
ubl
i
c
dat
a
base o
f
M
a
m
m
og
ra
phi
c Im
age Anal
y
s
i
s
S
o
c
i
et
y
- an or
ga
ni
zat
i
on o
f
U
n
i
t
e
d Ki
ng
d
o
m
research
gr
o
ups
. T
h
i
s
d
a
t
a
base i
n
cl
u
d
e
s 3
22 m
a
m
m
og
ram
s
from
161
pat
i
e
nt
s.
Fi
l
m
s
t
a
ken f
r
om
t
h
e U
n
i
t
e
d
Ki
ng
d
o
m
National B
r
eas
t Screeni
ng
Program
have been di
gitized
to 50-m
i
cron pixel edge and
prese
n
ted
each pixe
l
w
ith
an
8
-
b
it w
o
r
d
. Ev
er
y imag
e in
d
a
tabase alw
a
ys h
a
s ex
tr
a in
for
m
atio
n
or
gr
ound
tr
u
t
h
as show
n
in
Fi
gu
re
1
fr
om
the ra
di
ol
o
g
i
s
t
s
abo
u
t
c
h
aract
e
r
i
s
t
i
c
of
bac
k
gr
ou
n
d
t
i
ssue
,
t
y
pe
of a
b
no
rm
ali
t
y
present
,
sev
e
ri
t
y
of a
b
no
rm
al
ity
, t
h
e co
or
di
n
a
t
e
s of ce
nt
er
and a
p
pr
o
x
i
m
at
e radi
us (i
n pi
xel
s
)
of a
ci
rcl
e
encl
osi
ng t
h
e
abn
o
r
m
a
li
t
y
.
M
i
ni
-M
IA
S da
t
a
base i
s
a red
u
ced t
y
pe
of
th
e orig
in
al MIAS d
a
tab
a
se (d
ig
itized
at 5
0
-m
icro
n
pi
xel
e
dge
)
ha
s bee
n
red
u
ce
d t
o
2
0
0
-m
i
c
ron
pi
xel
ed
ge
an
d cl
i
p
ped/
p
a
dde
d s
o
e
v
e
r
y
im
age has s
i
ze of
10
2
4
x 10
2
4
pi
xel
s
.
Fi
gu
re
1.
R
e
d l
i
ne s
h
o
w
s
g
r
o
u
n
d
t
r
ut
h i
n
M
I
NI
-M
I
A
S
dat
a
base
2.2. Prepr
o
ces
sing
The aim of the step is to re
move
unnecessa
ry in
fo
rm
at
i
on i
n
m
a
m
m
ogram
s such as l
a
bel
,
pect
ora
l
m
u
scle o
r
o
t
h
e
r no
ise. To
separate th
e breast reg
i
on
fro
m
i
m
ag
e lab
e
l, we j
u
st th
res
h
old
the im
age and
keep
t
h
e bi
ggest
t
h
r
e
sh
ol
d
regi
on
.
The
pect
o
r
al
m
u
scl
e
in a ma
mm
ographi
c im
age appea
r
s as a
predominant
den
s
i
t
y
regi
o
n
.
It
can
af
fect
negat
i
vel
y
t
h
e resul
t
of
det
ect
i
on m
e
t
hod
[1
0]
. F
o
r t
h
i
s
reas
on
, t
h
e
regi
o
n
represe
n
ting t
h
e pectoral m
u
s
c
le shou
ld
be e
l
iminated. In t
h
e m
a
mmogra
m
,
there are al
so s
o
m
e
s
m
all
bri
ght
sp
o
t
s
w
h
ich
h
a
v
e
gray lev
e
l ap
pro
x
i
m
a
te
th
at o
f
circu
m
scrib
e
d
m
a
ss. Med
i
an
filterin
g
w
ith
a w
i
nd
ow
of 3x3
is app
lied
for eli
m
in
atin
g
t
h
ese spo
t
s as illu
strated
i
n
Fi
g
u
re
2
.
Fi
gu
re
2.
O
r
i
g
i
n
al
(l
eft
)
a
n
d
p
r
ep
roce
ssed
(
r
i
ght
) m
a
m
m
ogram
s
2.
3. M
a
ss dete
cti
o
n
In
t
h
i
s
st
a
g
e,
s
u
spi
c
i
ous
re
gi
o
n
s a
r
e e
x
t
r
act
e
d
fr
om
t
h
e pre
p
r
o
cesse
d m
a
m
m
ogram
. Th
e ra
di
ol
o
g
i
s
t
s
sh
ou
l
d
fo
cu
s t
h
eir atten
tio
n
t
o
th
ese
ext
r
acted re
gions. T
h
e steps of this
pr
oce
d
u
r
e are f
u
l
l
y
descri
be
d
i
n
[1
1]
.
Sho
w
n
i
n
Detected
ROIs are
mask
ed
are m
a
sk
ed
as t
r
u
e
positiv
e ROIs (TP-ROIs) or
false po
sitiv
e R
O
Is (FP-
ROIs)
as
illu
st
rated
in
Figu
re 3
b
a
sed
o
n
th
e p
r
ov
id
ed
g
r
o
und
tru
t
h.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
502
-47
52
I
ndo
n
e
sian
J Elec Eng
& Com
p
Sci, V
o
l. 10
,
No
.
3
,
Jun
e
2
018
:
10
30
–
1
035
1
032
Fi
gu
re
3.
Det
e
ct
ed R
O
Is
(g
re
en)
an
d
gr
o
u
n
d
t
r
ut
h (
r
e
d
)
2.
4.
Fe
ature
e
x
tr
acti
on
In
h
u
m
a
n vi
si
o
n
, e
d
ges a
nd
v
a
l
l
e
y
s
[12]
i
n
a
n
i
m
age are ve
ry
im
port
a
nt
fe
at
ures,
especi
a
l
l
y
val
l
e
y
s
are fun
d
a
m
e
n
t
al in
th
e v
i
sion
p
e
rcep
tion
of an
o
b
j
ect
shap
e [13
-
14
].
B
l
o
c
k
Differen
ce
In
v
e
rse Pro
b
ab
ility
(BDIP) is th
e
tex
t
u
r
e feat
u
r
e wh
ich
m
easu
r
es th
e
v
a
ria
tio
n
in in
ten
s
ities of an
im
ag
e b
l
o
c
k
.
It effect
iv
ely
extracts edges
and
valleys. The large
r
the variations of in
t
e
nsi
t
y
or t
h
e si
ze of t
h
e bl
ock
,
t
h
e hi
g
h
er t
h
e
val
u
e
of
B
D
I
P
[1
5]
.
B
D
IP
o
f
a
bl
oc
k
of
si
ze
Wx
W
i
s
de
fi
ne
d as:
2
(,
)
(,
)
(,
)
1
ma
x
(
,
)
(
,
)
ma
x
(
,
)
ij
B
ij
B
ij
B
Ii
j
I
i
j
W
BD
I
P
Ii
j
where I(i,j) denotes the inten
s
ity of
a pixel
(i,j) in the blo
c
k B.
As th
e d
e
tected
ROI is no
t in
size o
f
WxW
so
we
sub
titu
te th
e ter
m
“W
2”in
ab
ov
e equ
a
tio
n
b
y
size
o
r
nu
m
b
er of
p
i
x
e
ls i
n
th
e
ROI t
o
calcu
late th
e BDIP
featu
r
e at
first
reso
lu
tion
,
wh
i
c
h
th
en
is
ju
st
si
m
p
l
y
called BDIP.
Other B
D
IP feat
ures
at di
ffe
rent reso
l
u
tio
n are calcu
lated
as fo
llo
w:
a.
Divide
each si
de
of the m
i
nim
a
l
rectangular that c
o
ntains
the RO
I by
2, 3...n
to get 4, 9... n
2
bl
oc
ks.
b.
For
eac
h
bl
oc
k
usi
n
g a
b
ove
eq
uat
i
o
n
t
o
c
a
l
c
ul
at
e B
D
I
P
feat
u
r
es
w
h
i
c
h a
r
e cal
l
e
d B
D
IP
2
x
2
a
n
d
BD
I
P
3x
3... BD
I
P
nx
n.
c.
Expectation a
n
d va
riation
of BDIPs a
r
e
use
d
as
BDIP
features
for each RoI. They a
r
e
BD
I
P
2x
2
m
ean
, BDI
P
2
x2var
,
B
D
IP3x3mean
, BDI
P
3
x3v
ar,...BDI
Pn
xn
m
ean
, B
D
I
P
nx
nv
ar
respectively.
On the
ot
her
hand,
we
compute basics fe
at
ures of each ROI:
a.
Mean: the
av
er
a
g
e
grey
level
b.
Var: the standard devi
ation o
f
grey
level
c.
Max: the high
est grey
level
d.
Min: the lowest grey
level
Ho
we
ver
hi
g
h
or l
o
w i
n
t
e
ns
i
t
y
val
u
es i
s
n
o
t
abs
o
l
u
t
e
, i
n
put
i
m
ages oft
e
n ha
ve
di
f
f
e
r
ent
bri
ght
ness
.
W
e
pr
o
pose
t
w
o e
x
t
r
a feat
u
r
es
f
o
r
ens
u
ri
n
g
t
h
e
pe
rsuasi
ve
of
o
u
r
al
go
ri
t
h
m
a.
Ratio_1: Mean/Max
b.
Ratio_2: Max/
Max_I
where Max_
I is the highest
gray level of the whole image.
Mu
lti-reso
l
u
tion
b
a
sic feat
u
r
es are calcu
lated
in th
e
sam
e
man
n
e
r as m
u
lti-reso
l
u
tio
n B
D
IP feat
u
r
e.
2.
5.
R
a
n
d
om
Projec
ti
o
n
In
m
a
th
e
m
atic
s and
statistics, rando
m
p
r
oj
ectio
n
is a tech
n
i
q
u
e
u
s
ed
t
o
red
u
c
e th
e
d
i
m
e
n
s
ion
a
lity o
f
a set
of
poi
nt
s whi
c
h l
i
e
i
n
Eucl
i
d
ea
n spac
e. R
a
n
dom
pro
j
ect
i
on m
e
t
h
o
d
s
are p
o
w
er
ful
m
e
t
hods
kn
o
w
n
f
o
r
t
h
ei
r sim
p
l
i
c
i
t
y
and l
e
ss erro
neo
u
s o
u
t
p
ut
com
p
ared wi
t
h
ot
he
r
m
e
t
hods
.
Accor
d
i
n
g t
o
expe
ri
m
e
nt
al
r
e
sul
t
s
,
random
projec
tion pres
erve distances
we
l
l
,
but
em
pi
ri
cal
r
e
sul
t
s
are
s
p
ar
se [
15]
.
I
n
ran
dom
p
r
o
j
ect
i
o
n, t
h
e
ori
g
i
n
al
D-
di
m
e
nsi
o
nal
dat
a
i
s
p
r
o
j
ect
ed
t
o
a
L-
di
m
e
nsi
ona
l
(L <<
D
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
d
onesi
a
n
J
E
l
ec En
g &
C
o
m
p
Sci
ISS
N
:
2
5
0
2
-
47
52
Fu
sion
o
f
Rand
om Pro
j
ection
,
Mu
lti-Reso
l
u
tio
n
Fea
t
u
r
es a
n
d
Distan
ce…
(Viet
Dun
g
Ng
u
y
en
)
1
033
LxN
L
x
D
DxN
XR
X
where X
LxD
, X
DxN
den
o
te output and inp
u
t matrix and R
LxD
is a rand
om
projection
matrix.
The ra
nd
om
mat
r
i
x
R
can be
gene
rat
e
d usi
ng a Ga
ussi
a
n
di
st
ri
but
i
o
n.
Achl
i
o
pt
as [
1
5
]
has sho
w
n
t
h
at
t
h
e Gau
ssi
an di
st
ri
b
u
t
i
o
n can be repl
ace
d by
a
m
u
ch
si
m
p
l
e
r
di
st
ri
but
i
on suc
h
as:
,
1
w
i
t
h
probabi
l
i
t
y
1/
6
3
0
w
i
th pr
ob
a
b
il
ity
2/3
1
w
i
t
h
probabi
l
i
t
y
1/
6
ij
R
2.
6.
K Neares
t Nei
g
hbor
Let T = {(x
i
, y
i
): i=1:N} denote the traini
ng set w
h
ere
x
i
is the training
vecto
r
in m
-dim
e
nsional
feature
space a
n
d y
i
is th
e
corr
espon
d
i
n
g
class lab
e
l.
Giv
e
n un
kno
w x’
, class y’
is assigned
b
y
two
steps
a.
First, a set of
k labelled target neighbours
for the x’
is id
entified and sorted in ascending o
r
der
in term of Euclidean distanc
e
to x’
.
b.
Second, the class label y’
is predicted by major voting
of it nearest n
e
ighbo
urs.
A wei
ghte
d
voting sc
hem
e
for kNN,
whic
h
is calle
d
distance-wei
ghte
d
k nearest neighbor (wkNN)
rule is
p
r
op
os
ed i
n
[
1
6]
.
In
wk
N
N
, t
h
e cl
o
s
er
neig
h
b
o
r
s
are
weig
hted
m
ore heavily
t
h
an
the
fa
rthe
r
o
n
es,
usin
g the dista
n
ce-
weig
hted
f
unctio
n. T
h
e
n
the classifi
cation
result o
f
th
e que
ry
is
m
a
de by
the m
a
jority
weighted
voting a
neighbor
with sm
aller distance is we
ig
hted
m
ore hea
v
ily
than one
with greater di
stance:
the nearest n
e
igh
b
or
gets weight o
f
1
,
the f
urt
hest ne
ig
hb
or a wei
ght o
f
0 and the
oth
e
r weig
hts are
scaled
linearly to the inte
rval in
between.
3.
RESULTS
The
num
ber
of detected R
O
I i
s
10
00
[11]. For each R
O
I, B
D
IP and
basic
feature
s are cal
culated at n
level. The m
a
xim
a
l value o
f
n is the m
i
nim
a
l radius
o
f
a circle enclos
ing the a
b
no
r
m
ality
pro
vide
d in the
M
i
ni-M
IA
S
da
tabase. T
o
tally
we
ha
ve 2
4
0
0
featu
r
es.
Dif
f
e
r
ent
values of
K are
tested a
n
d
value
of K
whic
h
gives highest sensitivity
is
selected.
Figure 4 shows the
perform
a
nce with
di
fferent
K val
u
e. The
selected
value of K
is 21 with
sensitivity of
90 %.
Figu
re
4.
O
r
igi
n
al (left
)
a
n
d
p
r
ep
roce
ssed
(
r
ight
) m
a
m
m
ogram
s
Table 1 gives
com
p
arison
s o
f
ou
r
m
e
thod
t
o
dif
f
ere
n
t a
p
pr
oac
h
es.
It is
o
bvi
o
u
s th
at
ou
r m
e
thod
provides hi
gher sensitivity at
lower num
b
er of false
posi
tives per im
ag
e.
On the other hand, we
also
compare the performance in terms of
sensitivity, false positive per im
age, time of random projection and
time of runni
ng between di
fferent sizes of rand
om pr
ojection matrix. The results
are given in Table 2.
The result sho
w
s rand
om p
r
ojection he
lp
to reduce tim
e of running. This
tool should be effecti
v
e with
big data and a
lot of features
but in sma
ll d
a
ta
it
can influ
e
nce to other
performance.
Table 1.
C
o
m
p
arison
to ot
her
approaches
Approach
Sensitivity (%
)
False Positives per Im
age
Density
slicing,
textur
e flow field
analysis
81
2.
2
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN:
2
502
-47
52
I
ndo
n
e
sian
J Elec Eng
& Com
p
Sci,
Vo
l. 10
,
No
.
3
,
Jun
e
2
018
:
10
30
–
1
035
1
034
Multi-level thresh
old seg
m
entation
80
2.3
K m
e
an
cluster
i
ng
85
1
Multi-resolution f
eatures, distance
weighted k near
est neighbor
90
1.
04
Table 2. Per
f
o
r
m
a
nce
with dif
f
ere
n
t
size of r
a
nd
om
pr
ojecti
o
n
m
a
trix
Size of
matrix
Sensitivity
False po
sitive per
image
Time of rando
m
pro
j
ection per im
age
(s
)
Running time per
image (s)
2000x2400 89
1.
1
2.
1
19
1500x2400 87
1.
2
1.
9
17
1000x2400 85
1.
4
1.
6
16
F
u
ll 90
1.
04
24
4.
CO
NCL
USI
O
NS
This stud
y
pr
opo
ses a
new meth
od
t
o
de
tect mas
s
es in
mammographic image based
on
combination
of m
u
lti-resolution features and dist
ance weighted K
nearest neighbor alg
o
rithm
.
The
highest sensitivity is observed with
small fals
e positive per image. Compar
isons with other re
lat
e
d
wor
k
s p
r
o
v
e that our meth
od is effective and has p
o
t
ential to be further in
vestigated. When
using
random
projection, this tool will be
effective with
big
data.
In the future, we will evalu
a
te the
method o
n
larger set of mammog
r
ams and use different features.
REFERE
NC
ES
[1]
Ghoncheh M
, et al.
, “
I
nc
idenc
e
and Mortal
it
y a
nd Epidem
iolog
y
of Bre
a
st Can
cer in
the W
o
rl
d,”
Asian Pac J
Cancer
Pr
e
v
, vo
l. 17(S3)
,
pp
. 43
-46, 2016
.
[2]
H.
D.
Cheng,
et
al
.,
“
A
pproaches
for au
tom
a
ted
d
e
te
ction
and
cl
as
s
i
fica
tion of
m
a
s
s
e
s
in
ma
mmogr
a
m
s,
”
Pat
t
ern
Recogn
ition
,
vo
l. 39(4)
, pp
. 646-
668, 2006
[3]
Elah
e Chagh
a
ri
, et
al.
, “A Novel Approach for Tu
mor De
tection in
Mammography
Images,”
Indonesian Journal o
f
E
l
ec
t
r
i
c
al
E
n
gi
ne
e
r
i
n
g
and
Computer Science
, v
o
l. 12(8)
, pp
. 62
11-6226, 2014
.
[4]
V. M. R
a
o,
et al.
, “How Widely
I
s
Computer-Aided Detection Us
ed in
Screening
and Di
agnostic Mammography?,”
Journal of the American College of
Radiolog
y,
v
o
l. 7(10)
, pp
. 80
2-805, 2010
.
[5]
P.
T
a
y
l
or,
et al.
, “Impact of
computer-aided
d
e
tection
prompts
on the sensitivity
and specificity
of scr
eenin
g
mammograph
y
,”
He
alth Te
c
hnol
Asse
ss.
vol. 9(6), pp. 1-58, 2005.
[6]
F
a
riha Nos
h
een,
et al.
,
“False p
o
sitive and
false negative reduction in
digital ma
mmograms usin
g binary rotatio
n
invariant and noise toler
ant texture descriptor,”
2017 International Conferen
ce
on Co
mmunication Techno
logies
(ComTech), 201
7
[7]
P
.
P
unithavath
i
,
et al.,
“Rando
m Projection-based Cancelable Templa
te Gen
e
ration for Sp
a
r
se
ly
Distributed
Biom
etric P
a
tter
n
s
,
”
Indonesian
Journal of Electrical
Engin
eerin
g and Computer Scien
c
e
, vol. 7(
3), pp. 877-886
,
2017
[8]
Zhigao Zheng
,
e
t
a
l
.
,
“
Tim
e
-Wei
ghted Uncer
tain
Nearest Neighbo
r Collabor
ativ
e
Filtering Algor
it
hm
,
” Indonesian
Journal of Electrical
Engi
neerin
g and Computer
Scien
c
e
, vol. 12(
8), pp
. 6393-640
2, 2014
[9]
http:
//pe
i
pa.essex.ac.uk/info/mi
a
s.html
[10]
N. R. Mudigond
a,
et a
l
.
, “Detection of br
east m
a
sses in mammograms b
y
density
slicing and
texture flow-f
ield
anal
ys
is
,
”
I
EEE Transactions
on Medica
l
Imagin
g
, vol. 20(12)
, p
p
. 1215-1227
[11]
V
.
D
.
N
guyen,
et a
l
.
,
"
D
e
t
e
c
t
i
on of
t
u
m
o
r i
n
ma
m
m
o
g
r
a
p
h
i
c
i
m
a
g
e
s
by
hi
e
r
arc
h
y
of
bl
oc
k
'
s f
e
at
ure
s
,"
19th
Internationa
l Conference on D
i
gital S
i
gnal P
r
ocessing, DS
P
2
0
1
4
[12]
Y. D. Chun,
et a
t
.
, “Image r
e
tr
ieval using BDIP and BVLC moments,”
IEEE Transactions on Cir
c
uits and System
s
for Vid
e
o Techn
o
logy
, vol. 13(9)
, pp
. 951-957
[13]
D. E. Pearson and J. A. Robinson,
“Visual co
mmunication at very low data
rates,”
P
r
oceedi
ngs
of the IEE
E
,
vol.73(4), pp. 79
5-812.
[14]
Y, J
.
Ry
oo
, N. C. Kim
,
“
V
alle
y operator for ext
r
act
ing s
k
etch fe
atures
: DIP
,
”
El
ectr
oni
cs
L
e
tter
s
, vol. 24(8), pp
.
461-463.
[15]
T. D
.
Ngu
y
en
,
et a
l
.
, “Surface Extr
action Using
SVM-Based Tex
t
ure Classifica
tion for
3D
Fetal Ultr
asound
Imaging,” 1
st
In
ternational Conference on Communica
tions and
Electronics (
I
CCE2006), 2006
.
[16]
D. Achlioptas,
“Database-fri
en
dly
r
a
ndom projections,”
20
th
ACM SIGMOD-
SIGACT-SIGART symposium o
n
Principles of
database systems
PODS2001
[17]
S. A. Dudani, “The Distance-
Wei
ghted k-Nearest-Neighbor Rule,”
IEEE Transactions on Systems, Man, and
Cybernetics
, vol. 6(4), pp. 325-32
7
Evaluation Warning : The document was created with Spire.PDF for Python.
In
d
onesi
a
n
J
E
l
ec En
g &
C
o
m
p
Sci
ISS
N
:
2
5
0
2
-
47
52
Fu
sion
o
f
Rand
om Pro
jection
,
Mu
lti-Reso
l
u
tio
n
Fea
t
u
r
es a
n
d
Distan
ce…
(Viet
Dun
g
Ng
u
y
en
)
1
035
BIOGRAP
HI
ES OF
AUTH
ORS
Viet Dung Ng
uy
en
rec
e
iv
ed Doctorat
e degr
ee from
Hanoi Univers
i
t
y
of
S
c
ience
and
Techno
log
y
, Hanoi, Vietnam, in Electron
i
c En
gi
neer
ing in
20
16. Dr. Ngu
y
en
is curr
ently
working as Sen
i
or Lectu
r
er
, Vice Head of
th
e Department o
f
Electronic Technolog
y
and
Biomedical Engineering of Scho
ol of Electroni
cs
and Telecommu
nications, Hano
i University
of
Science and Technolog
y
,
Hanoi, Vi
etnam which
he joined in 20
00. His main research in
ter
e
sts
includ
e biosign
a
l and
m
e
dic
a
l
im
age
ana
l
y
s
is;
m
e
dica
l instrum
e
nt
ation
.
M
i
nh Dong Le
rece
ived h
i
s
Eng
i
neer
Degre
e
an
d M
a
s
t
er Degr
ee
of Engin
eer
ing
in Biom
edic
al
Engineering at
Hanoi Universi
ty
of Science and
Techno
log
y
, Hanoi, Vietnam in
2014 and 2016
res
p
ect
ivel
y.
He
is
current
l
y
w
o
rking as
a r
e
s
earch
er a
t
Depa
rtm
e
nt of Com
puter S
c
i
e
nc
e,
Chonnam National University
,
South
Korea. His research in
ter
e
sts are in sign
al processing
,
biomedical engineering
,
mach
in
e learning
& pa
ttern recognition.
Evaluation Warning : The document was created with Spire.PDF for Python.