Jou nal o Compu ing and In o ma ion Technology - CIT 21, 2013, 1, 47–56
doi:10.2498/ci .1002109
47
Digi al Measu emen o Myelo ib osis
Associa ed Pla ele De i ed G ow h
Fac o Recep o
ββ
(PDGFR
ββ
)
Exp ession in Bone Ma ow Biopsies
Szil ia Szeghalmy1, Judi Bedeko ics2,G
´
abo M´
ehes2and A ila Fazekas1
1Depa men o Compu e G aphics and Image P ocessing, Uni e si y o Deb ecen, Hunga y
2Depa men o Pa hology, Uni e si y o Deb ecen, Hunga y
In daily ou ine he e iculin sil e s aining is used on
bone ma ow biopsy samples as a gold s anda d o he
cha ac e iza ion o myelo ib osis, howe e his me hod
does no p o ide in o ma ion abou he p e ib o ic s age.
Recen ly a speci ic immunohis ochemical me hod was
in oduced which may o e come hese weaknesses o
e iculin s aining. Ac i a ed ib oblas s esponsible
o s omal p oli e a ion a e highligh ed by inc eased
PDGFR
β
exp ession, which can be p esen ed by im-
munohis ochemis y in bone ma ow samples. Using
his s aining he p e- ib o ic s age can become de ec able
and we ha e in o ma ion abou he disease ac i i y.
Du ing de elopmen o new s aining me hod i is impo -
an o p o e i s eliabili y and usabili y. In his pape we
in oduce a digi al image p ocessing me hod o measu e
pa anchymal damage in digi alized his ological slides
ha can aid co ec in e p e a ion o he s aining.
Keywo ds: myelo ib osis, PDGFR, image p ocessing
1. In oduc ion
Myelo ib osis (MF)is a linge ing disease which
eplaces no mal cells in he bone ma ow o i-
b o ic issue con aining a iable deg ees o e i-
culin and, in he ad anced phase, collagen i-
b es. The se e i y o he disease is usually cha -
ac e ized by he deg ee o ibe con en high-
ligh ed by he e iculin sil e s aining (G¨
om¨
o i’s
s aining)in bone ma ow biopsy samples.
Recen ly, di e en g ading sys ems ha e been
in oduced o assess bone ma ow ib osis, mos
o hem de i ed om he Baue meis e sco -
ing sys em [2]. A ew yea s ago, Thiele e
al. eached a consensus de ining a g ading sys-
em [17]. This scheme consis s o quali a i e
and quan i a i e analysis o bone ma ow ib o-
sis and dis inguishes ou inc easing ca ego ies,
anging om MF0 o MF3. This wide-sp ead
g ading sys em allows making mo e accu a e
p ognosis [6], desc ibes p ecisely he ibe con-
en o he bone ma ow, bu does no p o ide
in o ma ion abou ibe p oducing cells which
a e called ib oblas s.
La ely we iden i ied PDGFR
β
as a no el bio-
ma ke o ac i ed ib oblas s. PDGFR
β
con-
ols s omal p oli e a ions and u ned ou o be
selec i e o ib oblas s in he bone ma ow. The
amoun o PDGFR
β
is s ongly dependen on
hei numbe and ac i i y and in his con ex i
could be applied as a measu e o he ib o ic p o-
cess. A new sco ing sys em based on PDGFR
β
exp ession was also in oduced. This sys em e-
sul ed in excellen ag eemen wi h he classical
sil e s aining me hod, al hough di e en as-
pec s o he same p ocess a e analysed (Table 1,
Figu e 1). As a special ea u e, he new me hod
also co e s he ea lies , p e ib o ic phase o he
disease due o he p esen a ion o all ib ob-
las s po en ially pa icipa ing in e iculin ibe
syn hesis. The me hod is simple and equi es
only a s anda d immunohis ochemical s aining
o PDGFR
β
by one o he comme cially a ail-
able an ibodies (an i-PDGFR
β
clone ab-32570,
Abcam was used in his s udy).
48 Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)...
PDGFR
β
immunoposi i i y can be assessed in
a semi-quan i a i e way du ing isual inspec-
ion in he mic oscope. Howe e , a much be -
e and ep oducible assay would be equi ed
o he compa ison o di e en clinical sam-
ples o samples om di e en ime poin s om
he same pa ien . In o de o pe o m measu e-
men s on PDGFR
β
ela ed immunoposi i i y
in bone ma ow samples, digi al image analysis
was used and special algo i hms we e applied o
de ec he pa enchyma (use ul egion)and he
posi i e componen .
G ade Desc ip ion
0No signi ican amoun s o posi i e
ib oblas s.
1
Isola ed posi i e ib oblas s and/o
hei b anches clea ly ecognizable in
he in e cellula space.
2
Many posi i e ib oblas s wi h long
p ocesses a e p esen accompanied
by in e sec ions o ming a loose ne wo k.
3
Masses o posi i e ib oblas s and
p ocesses a e p esen accompanied
by equen in e sec ions and bands
o ming a dense ne wo k.
Table 1. G ading sys em based on PDGFR
β
exp ession
o ac i a ed ib oblas s.
Figu e 1. G ading sys em based on PDGFR
β
exp ession o ac i a ed ib oblas s.
1.1. Rela ed wo ks
Al hough manual me hods a e he gold s an-
da ds o analysis o immunohis ological s ain-
ing, e o s a e being made o apply au oma ic
digi al analysis as well. Mos o he solu ions
a e semi-au oma ic: manual segmen a ion is
pe o med by any image p ocessing ool (Im-
ageJ, Pho oShop, Tmaj, e c.), and only he mea-
su emen is au oma ic [5].
The e a e some ully au oma ic cha ac e isa-
ion algo i hms o simila s aining me hod o
ou s, o example he au ho s in [8]analyse lung
b eas cance in mice. Bu he bone ma ow e-
qui es speci ic analysing me hod because:
•No only he issue o be examined can be
ound in he sample.
•The disease eplaces no mal cells o ib o ic
issue.
•Fib o ic issue can de o m no mal cells.
Ou samples can include some kinds o haema-
opoie ic bone ma ow cells, a issue, bone a-
becules, blood clo s (Figu e 2), muscula issue,
connec i e issue (Figu e 9)and noises.
Figu e 2. The e a e bone ma ow nuclei (a), posi i e
ib oblas (b), plasma o bone ma ow cells (c) a (d)
and de o med nuclei (e)inside he black cu e. ( )is a
pa o bones, (g)is a blood clo (ou side he cu e)and
(h)shows a s aining ailu e pa .
The classi ica ion o di e en , some imes o e -
lapping and c inkled issues is no an easy ask.
In [12] he au ho s no e, he adi ional au-
oma ic classi ica ion algo i hms, such as K-
means and Fuzzy-C-means do no gi e co ec
esul . They inally de eloped a semi-au oma ic
analyse using Gene alised Reg ession Neu al
Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)... 49
Ne wo ks. The use can se he applied ea u es
(Fou ie ans o m o egion o common GLCM
ex u e ea u es)and some o he pa ame e s.
They did no men ion any s aining p ocedu e,
and he e is no in o ma ion whe he pa holog-
ical cases we e in es iga ed o no . Anyway,
ibe s do no appea as a sepa a e class.
In [7]au ho s p oposed a me hod o segmen
bone abeculae om bone ma ow biopsy dye
wi h haema ology and eosyn. Fi s , hey use
Wa e shed T ans o m and compu e ea u es o
each segmen . A e ha , hey classi y he ea-
u es o ou g oups by K-means. Using bone
egions as ma ke s, hey pe o m he Wa e shed
T ans o m again o ind he exac bo de s o he
bone abeculae.
I is e y impo an o us ha he p ocedu e
can dis inguish be ween he posi i e ibe s and
no mal cells, hus we de eloped an own me hod
aking ad an age o he p ope ies o immuno-
his ochemical s aining.
1.2. Ma e ials and me hods
Fo esea ch 41 bone ma ow biopsy samples
we e a ailable. Each sample was p ocessed by
immunohis ochemical s aining. In immunehis-
ochemis y e e y cell nucleus is displayed by
haema oxylin coun e -s aining which gi es he
nuclei a blue colo . Tha speci ic p o ein which
is in he ocus o in e es is bounded by a spe-
ci ic an ibody, in his case his was a p ima y
an ibody agains PDGFR
β
subuni . Then he
p ima y an ibody is bounded by a seconda y
an ibody which is linked o an enzyme. I we
add a subs a e o he sample, a chemical in-
e ac ion will p oceed be ween he enzyme and
i s subs a e which is accompanied by a b own
p ecipi a ion o ma ion. In o he wo ds, unde
he mic oscope he b own colo indica es he
localiza ion o he analysed p o ein wi hin he
issue. The sco e o PDGFR
β
was assigned o
s ained samples manually.
The samples we e ob ained a 40x magni ica-
ion using Ligh mic oscopy (Leica DM2500
mic oscope, DFC 420 came a and Leica Appli-
ca ion Sui e V3 so wa e, We zla , Ge many).
The Mi ax SlideAC SDK was used o ead he
slides. The size o he whole slides was 256216
×78336 pixels, ha is why we pe o med he
analysis wi h 1 : 4 magni ica ion, ile by ile.
Each ile size was 512 ×512 pixels.
2. Digi al Analysis
In his sec ion we in oduce ou own me hod.
Figu e 3 p esen s he main s eps.
In his pape , he wo ds nucleus and cy oplasm
will deno e he nucleus and cy oplasm which
belong o he haema opoie ic pa enchyma ( he
bone ma ow cells, excep cells o bone, a ,
and connec i e issue).Theuse ul egion is he
nuclei and he cy oplasms a ound hem. Figu e
4 shows wo ideal pa s o samples. No mal
nuclei a e pu ple, posi i e ib oblas s a e da k
b own. The backg ound is clea whi e, he e a e
no inapp op ia e hings, only nuclei, cy oplasm
and whi e, ound a cells.
Figu e 3. Main s eps o he algo i hm.
50 Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)...
Figu e 4. Ideal pa s o samples. The use ul egion is
inside he black lines.
On he o he hand, ideal case is a cu iosi y. The
examina ion p ocedu e implies many mis akes,
since each s ep can cause se e al e o s om he
biopsy o he digi aliza ion. Tha is why, man-
ual selec ion o some app op ia e egion o he
digi al analysis is a s anda d. In ou opinion,
his s ep should be omi ed because he samples
can be inhomogeneous, so he selec ion may
in luence he esul .
2.1. C ea ing mask o posi i e componen
The a io o he posi i e ib oblas s and he use-
ul egion is one o he mos impo an in o ma-
ion o measu e he s a e o disease. As s a ed
abo e, he posi i e ib oblas is da k b own, so
i s , we need o de ine he b own mask,
Bm=1,i H≤60 o 340 ≤H,
0,o he wise,(1)
whe e Hdeno es hue channel o HSV ile.
I is mo e di icul o ind h eshold le el, which
can sepa a e ib oblas om he plasma o o he
cells and he backg ound. I can also be im-
po an in o ma ion whe he he posi i e ib ob-
las s a e o ming a dense ne wo k, and i he
h eshold is high, he ne wo k will b eak up.
And when he h eshold is low, he de ec o
equen ly inds alse ne wo ks because he dis-
colou ed backg ound can connec sepa a ed pa s.
Using sa u a ion channel seems o be sui able
o pa i ion, al hough he shapes o his og am
a e a ious. To de ec in e es ing pa s he
B=1,i Bm=1andS> ,
0,o he wise,
o mula is used, whe e is he le el o h eshold
compu ed on he
Sb=S,i Bm=1,
0,o he wise,
whe e Sdeno es sa u a ion and Bmis de ined by
(1).
We es ed some well-known au oma ic h esh-
old algo i hms o ind sui able h eshold alue,
such as O su (1997), h ee-le el O su [14],Yen
[18], Ki le [11]. We obse ed, ha each es ed
me hod is sensi i e o he discolou ed back-
g ound and plasma, excep he Ramesh [15]
and A o a [1]wi h adequa e pa ame e s ha
pe o med well also in he p oblema ic cases.
The la e p ocedu e wi h he pa ame e s, which
wo ked well in noisy pic u es, was oo s ic
in cases o cleane pic u es, hus we chose he
Ramesh me hod. The esul o an ideal case
can be seen in Figu e 5b and Figu e 5c – 5e
gi es example ha discolou ed plasma can mis-
lead o he h eshold me hods. (Only he esul s
using Ramesh and he wo O su h esholds a e
p esen ed he e.)
Figu e 5. De ec ed ib oblas s using a ious h eshold
me hods. (a)The o iginal image, (b)Ramesh me hod
on (a),(c)O iginal image wi h discolou ed plasma, (d)
O su me hod on (c),(e)Highe h eshold o h ee-le el
O su me hod on (c),( )Ramesh me hod on (c).
2.2. C ea ing mask o use ul egion
To de ec use ul egion, he s a ing-poin is he
ideal case, when almos all he b own and pu -
ple pixels belong o i . A e ha , we elimina e
he unwan ed objec s. Finally, we ex end he
mask o nuclei inside he emaining pa .
Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)... 51
2.2.1. De ec ing use ul candida e pixels
Since he plasma belongs o he use ul egion
he c i e ia o highligh use ul candida e poin s
a e mo e pe missi e han when he aim is o
de ec posi i e ib oblas s. The mask o use ul
candida e om b own pa s is de e mined by Bc
and o ha om pu ple pa s is de e mined by
Pc.
Bc=1,i Bm=1andS>20,
0,o he wise,(2)
Pc=1,i 160 ≤Hand H≤320
and S>20,
0,o he wise,
(3)
whe e His he hue, Sis he sa u a ion channel
o he HSV ile. The h eshold o he sa u a ion
ensu es ha he whi e a issue is no included
in he mask.
2.2.2. De ec ing lobe-like objec s
In an ideal case he Bcand Pc o m oge he
he use ul egion, bu he pixels o o he is-
sue also can sa is y he o mula Pco Bc.The
non-use ul pa s in pu ple egion, excep he
olds, a e commonly smoo he han he use ul
segmen s. Whe eas he middle o nucleus is
b igh colou ed su ounded by pale cy oplasm
o b own ibe s, he lobe-like objec s a e ho-
mogeneous. Figu e 6 displays ypical kinds o
non-use ul egions.
Figu e 6. Typical kinds o unin e es ing pa s.
To de ec lobes, he homogeneous egion was
highligh ed om he pu ple egion by
L0=1,i 160 ≤Hand H≤320
and Eg+Es<40,
0,o he wise,
o mula, whe e Eg,Esa e he edge images com-
pu ed by Sobel ope a o wi h 5 ×5maskon he
g ay-le el espec i ely sa u a ion image. The
o mula causes he segmen s con aining g oup
o nucleus o all apa , while he lobe-like ob-
jec s emain whole. (Figu e 7b)
A e connec ed componen labelling [4]was
used on L0 o de ec each blob, we copied all
he blobs la ge han he gi en limi o L.(In
ou case he limi was se o 500.)Since he edge
o blobs does no belong o he L, we dila e L
by he ollowing mask:
M=
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎣
0011100
0111110
1111111
1111111
1111111
0111110
0011100
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎦
(4)
In he nex s ep, we used hole illing on L o
elimina e la ge hole in he blobs. Figu e 7c
p esen s he esul image. Finally, we emo e L
om Pc.Le Pdeno e he new image.
Figu e 7. S eps o elimina ing he lobes. (a)O iginal
image, (b)Raw lobe mask (L0),(c)De ec ed lobe (L),
(d)Pu ple nuclei (P).
2.2.3. Ex ending mask o nuclei
Remembe ha Bccon ains only posi i e ib ob-
las s, bu we also need he plasma be ween hem.
Since he in e es ing plasma egions a e e y
simila o lo s o non-use ul pa s, he egion
g ow algo i hms canno be used. Ins ead o
hem, we ill he small holes be ween he nu-
clei inside he candida e egion, bu hose pixels
which we e emo ed om Pc, canno become
use ul pixels.
U0=Po B,
U=(U0•M)and (Po Bc),

52 Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)...
whe e •is he mo phological close ope a o and
Mis he Fo mula 4. Figu e 8a p esen s esul
o his s ep on Figu e 7a. Figu e 8c is he de-
ec ed ib oblas and he no mal cells (U0)in
Figu e 8a. We ex end his egion applying M,
which can elimina e he small holes be ween
he ibe s and nuclei (Figu e 8d). In good case
he ex ended pa is plasma, bu backg ound o
lobe-like objec may be added o he esul as
well. So we keep only pixels alling in he
use ul candida e egion (wi hou lobes)(Figu e
8e). This is ensu ed by he “and” ope a o be-
ween he wo pa s o he o mula. We can
obse e in his example ha (Po B)almos
pe ec ly ma ches he use ul egion. I ’s a bi
misleading, because his pa can con ain also
he discolou ed backg ound, i he s ain has no
been pe ec .
Figu e 8. (a)De ec ed use ul egion on 7.a. (b)-( )The
de ailed example o compu ing U:(b)an o iginal
image, (c)posi i e ib oblas s and no mal cells (U0),
(d)U0dila ed by M,(e)use ul candida e egion wi hou
lobes (Po Bc),( ) he esul .
2.3. Sa e y il e
Howe e , mos o he noise was emo ed by
he p e ious s eps, some o unin e es ing pa s
could emain in he U. I can happen ha he
use ul egion in he slide is much mo e smalle
han he o he ( o eg ound) egions. Acco d-
ingly, we ha e o deal wi h his p oblem, else
lo s o small e o s could in luence he esul s.
One eason o he e o is i ano he so o is-
sue ge s in o he sample (See Figu e 9), o olds
emain in he de ec ed use ul egion.
Figu e 9. De ec ed bone ma ow cells in muscle issue
(a)and in connec i e issue (b).
When he mask o use ul egion was c ea ed,
mos o hese e o s we e il e ed ou . The
g ea e he di e ence be ween he Pcand P,
he mo e likely ha he emaining pa s o pu -
ple one a e bugs. Based on ou expe ience he
alse emaining pa s a e less han 15% o he
emo ed pa . The le side o Fo mula 6 es i-
ma es he a io o he emaining lobes a ea and
he nuclei a ea.
We also ha e o check whe he he ile con ains
enough da a (Fo mula 5), because we will in-
oduce ea u es o cha ac e ize he samples in
he nex sec ion, and some o hem may gi e
alse esul i use ul a ea can be ha dly ound in
he ile.
a(U)>s
10 (5)
0.15 ·(a(Pc)−a(P))
a(B)+a(P)<e,(6)
whe e a(X) he a ea o Xand sis he size o he
ile, and eis he gi en limi o e o .
When ea u es o he whole slide a e compu ed,
only iles sa is ying (5)and (6)a e used. On
a e age, 48% o he iles (con ains da a)sa is-
ies he condi ion. In many cases almos whole
sample is bone, a and connec i e issue. This
is consis en wi h he expe ience o o he e-
sea che s [12].
We ha e p ocessed 25870 Mi ax iles, o which
6819 con ained any da a (o he s only con ained
backg ound). Table 2 p esen s he de ailed e-
sul .
Rejec ed by All Con ains inapp op ia e obj.
Fo mula 5. 2488 1681
Fo mula 6. 884 883
Table 2. Numbe o iles il e ed ou by Fo mula 5 and
Fo mula 6 wi h e=0.01.
Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)... 53
To ind ou how e icien ly we il e ou he i -
ele an pa s, we implemen ed [7]and es ed i
on ou samples (Figu e 10). I is ob ious ha
he compa ison canno be comple e because he
s aining me hods a e di e en . Mo eo e , he
goal o he whole p ocedu e in [7]is o ind
bones, while we emo e lobe-like objec s (in-
cluding he bone)in a ious s eps.
Conside ing only he bone de ec ion, i is a y-
ing which algo i hm is mo e e icien . Gonza-
lez’s algo i hm uses ix numbe o clus e s, hus
i he e is no bone in he pic u e, his algo i hm
ine i ably ails. In gene al, he olds can also
cause an e o because K-means o en classi y
he pixels o he simple and he double laye
o bones o di e en classes (Figu e 10b).Ou
lobes de ec o is also no pe ec . I he lobes a e
discolou ed o b own a bi , ou algo i hm does
no ind hem co ec ly (Figu e 10g), bu o he
s eps can elimina e hese pa s om he sample
(Figu e 10h)and he sa e y- il e can h ow ou
he ile.
3. Cha ac e isa ion o Samples
A e c ea ing masks, some pa ame e s we e
compu ed o cha ac e ise he samples. The a-
io o posi i e ib oblas s a ea and use ul e-
gion a ea is commonly used in e iculin sil e
analysing [17]. Based on ou esul s, his a io is
no always app op ia e o dis inguish he s age
o samples, because he a ea o hin loose ne -
wo k (PDGFR
β
2)can be less han he a ea o
isola ed posi i e ib oblas (PDGFR
β
1). Thus,
o he pa ame e s a e also calcula ed. Table 3 in-
cludes calcula ed pa ame e s.
Pa ame e s Desc ip ion
PFib A A ea o posi i e ib oblas s egions.
F bCoun Numbe o he posi i e ib oblas blobs
(p b).
SumPe m Sum o he pe ime e o p b.
WPe m
Weigh ed sum o he pe ime e o p b.
Weigh is he numbe o c oss and end
poin s o he skele ons o p b.
Top50A Sum o he a ea o he 50 bigges p b
on whole slide.
Top50P Sum o he pe ime e o he 50 mos
leng h pe ime e o p b.
Top50S Sum o he leng h o skele on he 50
mos leng h pe ime e o p b.
Table 3. Compu ed ea u es. The PFib A,F bCoun ,
SumPe m and WPe m we e no malised by he a ea o he
use ul egion.
Figu e 10. (a),(e)O iginal images wi h anno a ion. (The bone abecula is inside he black lines.).
(b)Resul o Gonzalez me hod K-means s ep. (c)Resul o Gonzalez me hod on (a).
(d)De ec ed lobe by ou me hod on (a).( )Resul s o Gonzalez me hod on (e),(g)
De ec ed lobes (L)by ou me hod on (e).(h)De ec ed pu ple nuclei (P)on (e).
54 Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)...
Ou o he 41 scanned samples 11 a e PDGFR
β
0, 9 a e PDGFR
β
1, 10 a e PDGFR
β
2, 11
a e PDGFR
β
3 by manual g ading. Ou ob-
se a ion is ha , mos pa ame e s a e di ec ly
p opo ional o he manual g ades, bu he e a e
some excep ions. Fo example, he F bCoun is
o en la ge in PDGFR
β
2 hanPDGFR
β
3,
hence when he p bs connec oge he , he num-
be o p bs is dec eased. SumPe m also can be
almos equal in bo h cases, because when he
p bs close o each o he connec , he pe ime e
o he new objec can be less han he sum o he
pe ime e s o he o iginal p bs. The weigh in
he WPe m can compensa e ha phenomenon.
Figu e 11 p esen s he WPe m pa ame e s. The
“Top” pa ame e s a e mainly in oduced o de-
ec eme gence o posi i e ib oblas ne wo ks.
Figu e 11. Rela ionship o WPe m and PDGFR
β
g ade.
A p esen , only he F bCoun is no mal dis-
ibu ed based on Shapi o-Wilk Tes wi h 0.05
le el, hus we chose he K uskal-Wallis es o
compa e means o g oups. The H0( he median
o a iable a e he same ac oss ca ego ies o
PDGFR
β
)was ejec ed o each a iable. The
MANOVA analysis con i med his obse a ion.
Figu e 12 p esen s he CVA sca e plo [9].O
cou se, his esul mus be ea ed wi h cau ion,
because o he sample size.
Figu e 12. CVA sca e plo : g een, blue, pu ple and ed
deno e he a 0, 1, 2, 3 g ades in his o de .
3.1. Reliabili y
We de ined some o he ea u es ha can indi-
ca e he quali y o sample o eliabili y o e-
sul s (Table 4). Some cases con ain only a ew
usable iles, which is ele an in o ma ion abou
he sample, since his alue is be ween 50 and
200 in gene al. A e using he sa e y il e ,
mos samples a e homogeneous by PFib A.Oc-
casionally, he posi i e ib oblas is s onge in
a ew iles, ha can be a isen om de ec ion
e o , bu i also can be ue. The s ain e o is
mo e ypical, and we canno il e ou his ype
o e o well because he w ong segmen can be
easily mis aken o heal hy one. The ex u e is
he same, only he e o pa is smoo he a bi .
Figu e 13 is a scenic example because PDGFR
β
g ade o his sample is 3.
Figu e 13. Thumbnail o whole slide. Only he middle
o he sample was s ained well.
We c ea e he his og am PFib A wi h 0.1 bin.
I e e y hing goes well, he his og am is uni-
modal. O he wise one o he e o s (men ioned
abo e)occu s. We de e mine he O su h esh-
old, and compu e how many iles a e smalle
(NumLow)o la ge (NumUsable-NumLow) han
he h eshold. Ou expe ience shows ha i
he e a e only ew ou lie s, hese iles con ain
big essel o he issue is discolou ed nea he
bone. I he e a e almos equal iles in bo h
sides o h eshold, he sample was p obably
badly s ained. In his case he p og am will
no compu e he esul au oma ically. Fi s , an
expe decides which g oup o iles is sui able
o calcula ing he ea u es.
Digi al Measu emen o Myelo ib osis Associa ed Pla ele De i ed G ow h Fac o Recep o
β
(PDGFR
β
)... 55
Pa ame e s Desc ip ion
NumUsable Numbe o iles sa is y (5)and (6).
NumFil e ed Numbe o iles sa is y (5)bu do
no s a is y (6).
Range Range o his og am.
His oType Is his og am unimodal o no .
NumLow Numbe o iles wi h lowe PFib A
han h eshold.
Table 4. Reliabili y ea u es.
3.2. Conclusion and u u e plans
Immunohis ochemichal assessmen o PDGFR
β
on bone ma ow biopsy samples is a ypical
example o semi-quan i a i e sco ing sys ems
which a e equen ly used in pa hological p ac-
ice. Biological samples usually show he e o-
genei y, he ea u es o s age can be mixed, so
i can be di icul o ind he g ade which is he
bes o desc ibe whole sample.
We ha e been de eloping an image p ocessing
me hod ha enables us o analyse he bone ma -
ow samples p epa ed by immunohis ochemi-
cal s aining and make a possibili y o objec-
i e compa ison o he samples. A p esen he
me hod wo ks well on ideal samples, and can
elimina e om he samples he ino dina e ob-
jec s such as bone abecules, connec i e issue,
muscle issue ai ly well. Ou me hod can also
indica e ce ain s aining ailu e.
The incipien esul shows he compu ed pa am-
e e s enable a clea dis inc ion o ea ly (MF0-1)
om ad anced (MF2-3)myelo ib osis and p o-
ide a good basis o u u e classi ica ion o he
disease. Howe e , because o he small sample
size, u he in es iga ions a e necessa y.
The esea ch is going on. Now, we ha e been
dealing wi h some na u al phenomena caused
by he s aining me hod. Fo example posi i i y
may appea besides o bones edge in heal hy
samples as well, and he inside o blood essels
a e discolou ed ypically. In he immedia e u-
u e, u he samples will be a ailable, ha will
gi e us he oppo uni y o es ou analysing
me hod on mo e a ied samples and selec he
bes ea u es o de e mine he s a e o disease.
Re e ences
[1]S. ARORA,J.ACHARYA,A.VERMA,PRASANTA K.
PANIGRAHI, Mul ile el h esholding o image seg-
men a ion h ough a as s a is ical ecu si e algo-
i hm. Pa e n Recogni ion Le e s,29(2)(2008),
119–125.
[2]D. E. BAUERMEISTER, Quan i a ion o bone ma ow
e iculin – a no mal ange. Am. J. Clin. Pa hol.,
56(1)(1971), 24–31.
[3]J. BEDEKOVICS,SZ.SZEGHALMY,A.FAZEKAS,G.
M´
EHES, Pla ele de i ed g ow h ac o ecep o – an
ea ly ma ke o myelo ib osis. In P oc. o Mode n
sej anali ikai m´
odsze ek: a VII. Magya Sej anal-
i ikai Kon e encia, 31 May–2 Sep embe ,(2012)
Budapes , Hunga y, 160., ISBN 978-963-9070-89-
9(in Hunga ian)
[4]F. CHANG,C-J.CHEN,C-J.LU, A Linea -Time
Componen -Labeling Algo i hm Using Con ou
T acing Technique. Compu e Vision and Image
Unde s anding,93(2)(2004), 206–220.
[5]T. C. CORNISH,A.M.DEMARZO, Tissue Mic oa -
ays in Cance Resea ch. Mode n Molecula Biology
Applied Bioin o ma ics and Bios a is ics in Cance
Resea ch,(2010), pp. 157–184
[6]U. GIANELLI ET AL., The Eu opean consensus on
g ading o bone ma ow ib osis allows a be e
p ognos ica ion o pa ien s wi h p ima y myelo i-
b osis. Mode n Pa hology,25(2012), 1193–1202
[7]M. A. GONZALEZ,T.R.CUADRADO,V.L.BAL-
LARIN, Compa ing ma ke de ini ion algo i hms
o Wa e shed segmen a ion in mic oscopy images.
Jou nal o Compu e Science & Technology,8(3)
(2008), 151–157.
[8]T. GOUDAS,I.MAGLOGIANNIS,Ad anced Cance
Cell Cha ac e iza ion and Quan i ica ion o Mi-
c oscopy Images.LNAI(7297), Sp inge -Ve lag,
Be lin, Heidelbe g, (2012), 315-322, ISBN: 978-3-
642-30447-7
[9]O. HAMMER,D.A.T.HARPER,P.D.RYAN ,PAST:
Paleon ological s a is ics so wa e package o edu-
ca ion and da a analysis. Palaeon ologia Elec on-
ica,4(1)(2001).
[10]R. HOFFMAN,D.RONDELLI, Biology and ea -
men o p ima y myelo ib osis. Ame ican Socie y o
Hema ology, 2007(1)(2007), 346–354.
[11]J. KITTLER,J.ILLINGWORTH, Minimum e o h esh-
olding. Pa e n Recogni ion,19(1986), 41–47.
[12]G. MESCHINO,E.G.MOLNER,L.I.PASSONI,Semi-
au oma ed Segmen a ion o Bone Ma ow Biopsies
Images Based on Tex u e Fea u es and Gene alized
Reg ession Neu al Ne wo ks. In P oc o CACIC,
4-6 Oc obe 2004, Uni e sidad Nacional de La
Ma anza, ISBN: 987-9495-58-6
[13]N. OTSU, A h eshold selec ion me hod om g ay-
le el his og ams. IEEE T ans. Sys., Man. Cybe .,
9(1)(1979), 62–66.