J Pa hol In o m Edi o -in-Chie :
Anil V. Pa wani , Li on Pan anowi z,
Pi sbu gh, PA, USA Pi sbu gh, PA, USA J Pa hol In o m Edi o -in-Chie :
Anil V. Pa wani , Li on Pan anowi z,
Pi sbu gh, PA, USA Pi sbu gh, PA, USA
OPEN ACCESS
HTML o ma
Fo en i e Edi o ial Boa d isi : www.jpa hin o ma ics.o g/edi o ialboa d.asp Fo en i e Edi o ial Boa d isi : www.jpa hin o ma ics.o g/edi o ialboa d.asp
Resea ch A icle
Fea u e‑based analysis o mouse p os a ic in aepi helial neoplasia
in his ological issue sec ions
Pekka Ruusu uo i1,2, Mi a Valkonen1, Ma i Nyk e 1, Tapio Visako pi1,3, Leena La onen1,3
1Ins i u e o Biosciences and Medical Technology ‑ BioMediTech, Uni e si y o Tampe e, Tampe e, 2Tampe e Uni e si y o Technology, Po i, 3Fimlab Labo a o ies, Tampe e
Uni e si y Hospi al, Tampe e, Finland
E‑mail: *D . Leena La onen ‑ [email p o ec ed]
*Co esponding au ho
Recei ed: 03 Decembe 15 Accep ed: 20 Decembe 15 Published: 29 Janua y 2016
Abs ac
This pape desc ibes wo k p esen ed a he No dic Symposium on Digi al Pa hology
2015, in Linköping, Sweden. P os a ic in aepi helial neoplasia (PIN) ep esen s
p emalignan issue in ol ing epi helial g ow h con ined in he lumen o p os a ic acini.
In he a emp s o unde s and oncogenesis in he human p os a e, ea ly neoplas ic
changes can be modeled in he mouse wi h gene ic manipula ion o ce ain umo
supp esso genes o oncogenes. As wi h many ea ly pa hological changes, he PIN
lesions in he mouse p os a e a e mac oscopically small, bu mic oscopically spanning
a eas o en la ge han single high magni ica ion ocus ields in mic oscopy. This poses
a challenge o u ilize ull po en ial o he da a acqui ed in his ological specimens. We
use whole p os a es ixed in molecula ixa i e PAXgene™, embedded in pa a in,
sec ioned h ough and s ained wi h H&E. To isualize and analyze he mic oscopic
in o ma ion spanning whole mouse PIN (mPIN) lesions, we u ilize au oma ed whole
slide scanning and s acked sec ions h ough he issue. The egion o in e es s is masked,
and he masked a eas a e p ocessed using a cascade o au oma ed image analysis s eps.
The images a e no malized in colo space, a e which exclusion o sec e ion a eas
and ea u e ex ac ion is pe o med. Machine lea ning is u ilized o build a model o
ea ly PIN lesions o de e mining he p obabili y o his ological changes based on he
calcula ed ea u es. We pe o med a ea u e‑based analysis o mPIN lesions. Fi s , a
quan i a i e ep esen a ion o o e 100 ea u es was buil , including se e al ea u es
ep esen ing pa hological changes in PIN, especially desc ibing he spa ial g ow h
pa e n o lesions in he p os a e issue. Fu he mo e, we buil a classi ica ion model,
which is able o align PIN lesions co esponding o g ading by isual inspec ion o mo e
ad anced and mild lesions. The classi ie allowed bo h de e mining he p obabili y o
ea ly his ological changes o unca ego ized issue samples and in e p e a ion o he
model pa ame e s. He e, we de elop quan i a i e
image analysis pipeline o desc ibe mo phological
changes in his ological images. E en sub le changes
in mPIN lesion cha ac e is ics can be desc ibed wi h
ea u e analysis and machine lea ning. Cons uc ing
and using mul idimensional ea u e da a o ep esen
his ological changes enables iche analysis and
in e p e a ion o ea ly pa hological lesions.
Key wo ds: His opa hological image analysis,
machine lea ning, p os a ic in aepi helial neoplasia
Access his a icle online
Websi e:
www.jpa hin o ma ics.o g
DOI: 10.4103/2153-3539.175378
Quick Response Code:
© 2016 Jou nal o Pa hology In o ma ics | Published by Wol e s Kluwe ‑ Medknow
This a icle may be ci ed as: Ruusu uo i P, Valkonen M, Nyk e M, Visako pi T,
La onen L. Fea u e‑based analysis o mouse p os a ic in aepi helial neoplasia in
his ological issue sec ions. J Pa hol In o m 2016;7:5.
A ailable FREE in open access om: h p://www.jpa hin o ma ics.o g/ ex .
asp?2016/7/1/5/175378
This is an open access a icle dis ibu ed unde he e ms o he C ea i e Commons
A ibu ion-NonComme cial-Sha eAlike 3.0 License, which allows o he s o emix,
weak, and build upon he wo k non-comme cially, as long as he au ho is c edi ed
and he new c ea ions a e licensed unde he iden ical e ms.
Fo ep in s con ac : ep in [email protected]
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]
J Pa hol In o m 2016, 1:5 h p://www.jpa hin o ma ics.o g/con en /7/1/5
INTRODUCTION
This pape desc ibes wo k p esen ed a he No dic
Symposium on Digi al Pa hology 2015, in Linköping,
Sweden.
Cance is a majo heal h ca e challenge and one o he
mos s udied g oup o diseases wo ldwide.[1] In many
cance s, e.g., wi h p os a e cance , di e ences be ween
ad anced cance and no mal issue a e ela i ely well
known, ye he p ocess o cance de elopmen is s ill
inadequa ely unde s ood.[2] To gain u he knowledge
on he biological s eps in ol ed in oncogenesis, mo e
unde s anding is needed o ea ly changes leading o he
de elopmen o cance .
P os a ic in aepi helial neoplasia (PIN) ep esen s a ype
o p emalignan issue wi h neoplas ic epi helial g ow h
con ined in he lumen o p os a ic acini.[3] To s udy
oncogenesis in he p os a e, ea ly neoplas ic changes
in PIN a e o en modeled in he mouse by gene ic
manipula ion o ce ain umo supp esso genes o
oncogenes.[4] E.g., i is well known ha he e ozygous
dele ion o umo supp esso P en induces PIN o ma ion
in he mouse p os a e, in addi ion o hype plasia and
umo s in se e al o he issues.[5] He e, we analyze he
his ology o such mouse PIN (mPIN) lesions o med in
he p os a es o 10–11 mon hs old P en ± mice.
Ea ly pa hological changes in issue end o be small
and/o sub le as a as his ology is conce ned. The
neoplas ic mPIN lesions used as a model he e ep esen
a g oup o samples wi h such ea ly pa hological changes
wi h ela i ely na ow, bu de ec able, his ological scope o
al e a ions, and hus ep esen a challenge o his ological
quan i a ion. No p e ious s udies o ou knowledge exis
p o iding compu a ional models o assessmen o mPIN
his ology. The mPIN lesions a e mac oscopically so small
ha hey a e unde ec able by eye om he whole o gan
p epa a ion. Howe e , mic oscopically, he mPIN lesions
can span a eas la ge ha can be isualized in se e al high
magni ica ion ocus ields in mic oscopy – bo h wi h an
ac ual mic oscope o wi h digi al pa hology applica ions
wi h s anda d compu e sc eens. The gene ally accep ed
way o analyze such his ological da a is o snapsho
single o se e al ields wi hin he egion o in e es (ROI)
o pe o m he analyses. In his way, only a ac ion
o he ROI is u ilized o en lea ing mos o he da a
unanalyzed. In addi ion, subjec i e e o s a e in oduced
h ough selec ion bias. The e is a endency o selec a eas
no immedia ely adjacen o edges o he ROI, esul ing
in quan i a ion ep esen ing mo e he middle a ea o he
lesion and lea ing possible di e ences nea he edges
igno ed.
Mo e comp ehensi e ways o u ilize he ull po en ial o
whole-slide scans and sequen ial sec ions a e equi ed.
Visual inspec ion o only one o ew slides and pa s o he
ROIs pe lesion lea es a signi ican amoun o in o ma ion
unobse ed. Fu he mo e, o cap u e he po en ial spa ial
di e ences in he issue, i is o impo ance o u ilize
in o ma ion in h ee-dimensional (3D) o unde s and,
e.g., cance g ow h pa e ns be e . By employing he ull
ROI a eas and including 3D in o ma ion, we can help
basic scien i ic app oaches o secu e mo e comp ehensi e
da a. E en ually, using he enhanced me hods o analysis,
he goal is o ensu e be e clinical ools o he u u e.
Image analysis and machine lea ning p o ide powe ul
ools o p ocessing mic oscope images and a e
inc easingly applied in digi al pa hology.[6-8] Image analysis
can be used o segmen ing issue a eas and o ex ac ing
mul idimensional ea u e da a om ull esolu ion
whole slide images. Machine lea ning can be used o
compu a ionally lea ning a desc ip i e o disc imina i e
model om he da a ei he o segmen a ion o sample
classi ica ion. The a ionale behind using quan i a i e
app oach o e he common use o ew, simple eadou s,
such as size and olume, is ha when making decisions
o his ology, also expe s use se e al cha ac e is ics o
he in ensi y, spa ial dis ibu ion, and mo phology o he
issue componen s. By quan i ying such ea u es, and
using hem o lea ning-based analysis, in o ma ion abou
he p ope ies common o issue samples wi h his ological
changes can be acqui ed.
Ou app oach is o use a pipeline o au oma ed image
analysis me hods o he exclusion o unwan ed egions
om u he analysis and o ex ac ing a la ge se o
nume ical desc ip o s o he s udied issue a eas. Using
his nume ical ep esen a ion o he ROIs, i is possible
o isualize a ious p ope ies and o obse e simila i ies
and di e ences wi hin he popula ion o samples. The
ea u es, along wi h he anno a ion ob ained om ROIs
segmen ed by an expe , a e also used o aining a
classi ie model. Machine lea ning enables inco po a ing
expe knowledge in quan i a i e analysis in a
sophis ica ed manne and is applicable in a small sample
se ing. He e, he classi ie is used o de e mining
he p obabili y o his ological changes in a gi en issue
sample. The con ibu ion o indi idual ea u es o
classi ica ion p o ides in o ma ion abou di e ences in
he his ology o mPIN lesions.
METHODS
Tissue Ma e ial
We s udied p os a es o 10–11-mon hs-old male
FVB/N mice he e ozygous o umo supp esso P en
(n = 12; wi h 6 mPIN lesions/mouse on a e age).
Hal o he mice exp essed ARR2PB-miR32 ansgene
(La onen L, Visako pi T, unpublished). E hical app o al
o animal expe imen a ion has been admi ed by he
Regional S a e Adminis a i e Agency o Sou he n
Finland (ESAVI/6271/04.10.03/2011).
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]
J Pa hol In o m 2016, 1:5 h p://www.jpa hin o ma ics.o g/con en /7/1/5
P os a e issues we e ixed in PAXgene™ molecula ixa i e
(P eAnaly iX GmbH, Homb ech ikon, Swi ze land)
acco ding o manu ac u e ’s ecommenda ions and
embedded in pa a in. The issue blocks we e sec ioned
h ough, and H&E s aining was pe o med o h ee 5 µm
sec ions e e y 50 µm apa o s udy p os a e his ology
h oughou he p os a e. The slides we e scanned wi h a
Zeiss Axioskop 40 mic oscope (Ca l Zeiss Mic oImaging,
NY, USA) wi h ×20 objec i e and a cha ge-coupled
de ice colo came a (QICAM Fas ; QImaging, Canada)
and a mo o ized specimen s age (Mä zhäuse We zla
GmbH, Ge many). The au oma ed image acquisi ion was
con olled by he Su eyo imaging sys em (Objec i e
Imaging, UK). Uncomp essed bi map ou pu was
con e ed by JVSdicom Comp esso applica ion o
JPEG2000 WSI o ma .[9]
His ological Assessmen
Mouse p os a e his ology was assessed by an expe .
These mice de elop mPIN wi hou any e idence o
in asion h ough he basemen memb ane (Th ee,
unpublished). S ages I–IV o mPIN we e g aded acco ding
o he guidelines in oduced by Pa k e al.[10] The mPIN
lesions which included a eas wi h clea nuclea a ypia and
ep esen ed G ades II–IV, we e included in he s udy. The
p esence o p ominen nucleoli and inc eased a ypia in
nuclei we e ega ded as de e minan s in G ades III–IV.
Howe e , i mus be no ed ha as he analysis akes in o
accoun he whole o he lesions in 3D, mos o he
lesions include a eas ep esen ing se e al mPIN g ades
wi hin hem. Thus, i is no sui ed o he aims o he
s udy o ca ego ize lesions acco ding o he g ades. Ra he ,
he e ms “mild” and “ad anced” pheno ypes a e used o
desc ibe he mos a iable na u e o he al e a ions seen
wi hin indi idual lesions. “Mild” e e s o he lesion wi h
majo i y o S age II–III al e a ions, as “ad anced” e e s
o he lesion wi h majo i y o S age III–IV al e a ions.
Image P ocessing
In o al, he e we e 72 mPIN lesions, wi h 387 ROIs,
included in he analysis (1–25 ROIs/lesion; mean
alue 5.4). A mul i esolu ion app oach was used o
segmen a ion o ROIs. The bes quali y sec ion om
h ee adjacen H&E-s ained sec ions was selec ed om
each whole slide scan. The segmen a ion was done using
a eehand selec ion ool in ImageJ, (Na ional Ins i u es
o Heal h, Be hesda, MD, USA),[11] o a low- esolu ion
e sion o he o iginal image ob ained using he educ ion
le els in wa ele decomposi ion. The esul ing bina y
mask was hen esized o ma ch he o iginal image size
and used o ex ac ing he ROI om he ull esolu ion
o iginal H&E image o u he p ocessing. The image
p ocessing pipeline, including ea u e ex ac ion and
machine lea ning, was de eloped using Ma lab (The
Ma hWo ks, Inc., Na ick, MA, USA).
The colo s o he ex ac ed lesion a eas we e equalized
using con as -limi ed adap i e his og am equaliza ion.
This educes he colo a ia ion be ween di e en
lesion sec ions, and he colo s h ough he whole
s ack o images become mo e consis en . Exclusion o
sec e ion- illed and emp y a eas was pe o med o ob ain
e ec i e issue a ea wi hin ROI. These ex ac ion phases
a e based on he sub ac ion o di e en colo channels,
and he segmen a ion o hese a eas is done using basic
h esholding me hod minimizing he in aclass in ensi y
a iance.
Fea u e Based Analysis
In o ma ion o ROIs in each sec ion le el was included
in he calcula ion o ea u es o each lesion. In his
p oo -o -p inciple s udy, we ex ac ed ea u es desc ibing
each lesion a ea om one colo channel ( ed). Fo lesions
spanning o e mo e han one sec ion, he ea u e alues
a e a e aged o e he sec ions. Fea u es a e lis ed in
Table 1. Mo phological ea u es e e o he quan i a i e
analysis o o m. Mul iple ea u es desc ibing he shape
and size, including ea u es such as a ea, majo axis
leng h, and mino axis leng h and pe ime e we e
ex ac ed. Be o e subsequen analysis, he ea u es we e
scaled o ze o mean and uni a iance.
Tex u es a e complex isual pa e ns, p ope ies o
which can be quan i ied wi h ea u es desc ibing he
equency, egula i y, oughness, linea i y, o smoo hness
o he s udied a ea. The ex ac ed ex u e based ea u es
included, e.g., mean in ensi y alue, con as , co ela ion,
and ene gy, calcula ed om g ay le el co-occu ence
ma ix (GLCM). Tex u al ea u es we e also ex ac ed
using local bina y pa e n (LBP).[12,13] P ope ies o
he lesions we e also ex ac ed using his og am o
Table 1: Fea u e ca ego ies
Fea u e ype Desc ip ion Numbe o ea u es
Mo phological ea u e Fea u es desc ibing he olume and shape o he lesion 18
Unwan ed egions The amoun o sec e ion and “emp y a ea” inside he lesion 2
In ensi y ea u es Tex u e ea u es desc ibe he spa ial a angemen o in ensi y alues in an
image egion. These ea u es can be s uc u al o s a is ical
11
Local bina y pa e n Desc ibe he appea ance o an image in a small neighbo hood a ound a pixel 10
His og am o o ien ed g adien s Coun s occu ences i g adien o ien a ions in localized image egions 81
Maximally s able ex emal egions A me hod o blob de ec ion om an image 5
To al 127
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]
J Pa hol In o m 2016, 1:5 h p://www.jpa hin o ma ics.o g/con en /7/1/5
o ien ed g adien s (HOG) desc ip o [14,15] as well as wi h
maximally s able ex emal egions (MSER).[16] MSER
implemen a ion by VLFea [17] was used in his wo k.
Machine Lea ning o De ec ing His ological
Changes
Machine lea ning was applied o de ec ion o
his ological changes in issue as ollows. We used he
ea u e ep esen a ions o lesions o building a model
which de e mines he p obabili y o a gi en lesion o
belong o he g oup wi h mo e ad anced his ological
changes when compa ed o he g oup wi h milde
changes. The g ouping o lesions o he aining samples
wi h ad anced and mild his ological changes was done
by an expe o a subse o lesions (see abo e; mild,
n = 11; ad anced, n = 9). A logis ic eg ession classi ie
wi h he lasso (leas absolu e selec ion and sh inkage
ope a ion) egula iza ion[18,19] was cons uc ed using he
aining da a. The ou pu om he classi ie is he class
condi ional p obabili y o he ad anced g oup, i.e., issues
wi h mo e p ominen his ological changes.
RESULTS
We sc eened mPIN lesions in whole mouse p os a es
by sec ioning h ough he issue wi h 5 µm sec ions
[Figu e 1]. To u ilize ull da a and omi he subjec i e
e o s made by human-di ec ed ield selec ion, we
pe o med ou analysis o lesion a eas in a s ack in z.
H&E-s ained his ological images e e y 50 µm apa
we e p ocessed and used in he analysis. All mPIN
lesions we e masked as ROIs, and unwan ed emp y and
sec e ion- illed egions we e selec ed ou wi hin he ROI
a eas [Figu e 2]. Resul ing issue a eas wi hin ROIs we e
subjec ed o ea u e analysis.
We ex ac ed in o al 127 ea u es assessing shape,
ex u e and spa ial a angemen o he lesions [Table 1].
The ea u es included mo phological, ex u al, LBP,
scale-in a ian ea u e ans o m, HOG, and MSER
ea u es. All sec ions h ough each lesion we e p ocessed
and combined in o a ea u e ec o , which p o ides
a nume ical ep esen a ion o each lesion. P incipal
componen analysis o he lesions acco ding o he
ea u es is shown in Figu e 3a. The analysis shows he
con ibu ion o each ea u e o he wo i s p incipal
componen s, and he ea u es a e shown as lines s a ing
om he o igo. The ed do s ep esen mPIN lesions
p ojec ed in o he p incipal componen space. The alues
o each lesion a e calcula ed ac oss all ROIs o ha
pa icula lesion and a e aged pe lesion. As se e al ex u e
ea u es (mainly HOG ea u es) show sepa a ion oughly
in bo h di ec ions along he y-axis, emphasis on size- and
shape- ela ed ea u es s. composi ion and ex u e
(e.g., solidi y, Eule numbe , ex en , homogenei y, con ex
a ea, a ea, axis leng hs, bounding box, and equi alen
diame e ) is isible oughly along he x-axis [Figu e 3a].
As an example o sepa a ion o lesions by composi ion
and shape desc ip o s, a sca e plo o a e age solidi y
e sus sum o he a ea o lesions is shown in Figu e 3b.
Lesions a e sepa a ed by hese desc ip o s o wha can be
in e p e ed as di e en g ow h pa e ns and/o s ages o
mPIN ad ancemen , as is shown o ep esen a i e lesions
wi h example sec ion masks [Figu e 3c].
We applied machine lea ning o de ec ion o he
ela i ely small his ological changes wi hin he g oup o
mPIN lesions. We used ce ain lesions as aining samples
o pheno ypically mild (n = 11) and sligh ly ad anced
(n = 9) lesions and used he ea u e ep esen a ions
o lesions o building a model which de e mines
Figu e 1: P ocessing o mouse p os a e ma e ial o ea u e‑based analysis. Whole mouse p os a es we e sec ioned h ough wi h 5 µm
sec ions. Sec ions we e H&E s ained, and he whole slide scanned o ob ain high‑ esolu ion images. H&E‑s ained his ological image e e y
50 µm apa was p ocessed and used o ma k mouse p os a ic in aepi helial neoplasia lesions as a egion o in e es s and subjec ed o
image p ocessing. All egion o in e es s in a pa icula lesion ob ained om a s ack o his ological images in z‑di ec ion we e included
and subjec ed o ea u e analysis
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]
J Pa hol In o m 2016, 1:5 h p://www.jpa hin o ma ics.o g/con en /7/1/5
he p obabili y o a gi en lesion o belong o ei he
g oup [Figu e 4a]. The plo ed p obabili ies a e a e ages
om 1000 epe i ions o lea e-one-ou expe imen s,
whe e a andomly picked lesion was ou o he aining
samples in Figu e 4a, posi i e aining samples a e shown
in ed and nega i e in g een. The hold-ou expe imen
is a simula ion o model s abili y in a small-sample
se up. In addi ion o he le -ou sample, he majo i y
o all a ailable lesions did no ha e an expe -de ined
class, and hus, we e no used o aining (blue ma ke s
in Figu e 4a). The model success ully dis inguished
he emaining ad anced lesions o he p ope g oup
wi h posi i e class condi ional p obabili ies >0.5 and
le he milde lesions below he h eshold, examples
o lesion ROIs a e shown in Figu e 4b. The models
ob ained du ing he 1000 epe i ions composed o ou
ea u es ep esen ing aspec s o lesion size combined o
shape (majo axis leng h, eccen ici y, equi alen diame e ,
and pe ime e ) and se e al ex u e ea u es (LBPs, HOGs,
and MSER2) [Figu e 4c]. The his og am bins show he
numbe o imes each ea u e was selec ed in he model.
CONCLUSIONS
S udies o many diseases would bene i o enhanced
quan i a i e assessmen o ea ly pa hological changes.
E.g., by unde s anding he al e a ions occu ing ea ly in
cance de elopmen , we could lea n impo an aspec s o
neoplasia o ma ion, which could help us ind be e ways
o p e en , diagnose, and ea cance . Ea ly pa hological
changes leading o cance o ma ion a e p ac ically
impossible o s udy in human samples, as ea ly lesions go
unde ec ed and human samples lack he possibili y o
ime se ies expe imen a ion. Thus, as wi h many o he
diseases, cance is o en modeled in he mouse. Mice
p o ide he possibili y o ollow neoplas ic de elopmen
and g ow h in ime, as he gene ically homogenous
backg ound o labo a o y mice in combina ion wi h
ce ain gene ic manipula ions p o ides ela i ely
homogenous sample ma e ial o e he specimens wi h a
ela i ely cons an de elopmen a e.
Ea ly pa hological changes in issue end o be
mac oscopically small and his ologically sub le. Howe e ,
hey can span a eas la ge han single high magni ica ion
ocus ield in mic oscopy. To u ilize he ull po en ial o
he da a acqui ed in his ological specimens wi h mode n
whole slide scanning sys ems, en i e lesions should be
used in quan i a i e assessmen s. This challenges he
adi ional way o analyzing ew single ocus ields wi h
mic oscopes o snapsho s om i ual slide iewe s o
ob ain quan i a i e da a. In his s udy, we ackled his
p oblem by using sec ions wi h ce ain in e als (50 µm;
e e y 10 h o 5 µm sec ions) h ough he whole mouse
p os a e issue and analyzing all ROIs o each lesion.
Figu e 2: Image p ocessing s eps. An example o a PIN lesion a ea in a H&E‑s ained image (a) and i s selec ion as a egion o in e es (b).
Exclusion o sec e ion‑ illed (c) and emp y (d) a eas is pe o med o ob ain e ec i e issue a ea wi hin egion o in e es (e) excluded
a eas shown in u quoise
d
c
b
a e
Figu e 3: Fea u e analysis o mouse p os a ic in aepi helial
neoplasia lesions. (a) Sepa a ion o indi idual lesions ( ed do s,
n = 72) acco ding o p incipal componen analysis and he
ela i e weigh s o ea u es (blue lines). The alues o each lesion
a e calcula ed ac oss all egion o in e es s o ha pa icula
lesion. (b) Sca e plo o mouse p os a ic in aepi helial neoplasia
lesions wi h a e age solidi y and sum o he a ea ac oss lesion
(i.e., he sum o a eas o all lesion egion o in e es s in all sec ions
in z). (c) Examples o he ype o lesion egion o in e es masks
ep esen ing he mo phological g oups sepa a ed by lesion solidi y
and size (e.g., lesion a ea o olume). Colo s in (b) indica e he
same lesions o which an indi idual egion o in e es mask is shown
in (c). Region masks in (c) a e in he same scale
c
b
a
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]

J Pa hol In o m 2016, 1:5 h p://www.jpa hin o ma ics.o g/con en /7/1/5
The mPIN lesions, depending on hei size and g ow h
di ec ion ela i e o cu o ien a ion, spanned om 1 o
25 images.
As expec ed, he size- and shape- ela ed mo phological
ea u es we e good desc ip o s o he mPIN lesion
ad ancemen . Size and isual complexi y o g ow h
pa e n oughly co ela e o g ow h a e and po ency o
neoplas ic lesions, and cha ac e iza ion o such lesions
by eye is pa ly dependen on hese cha ac e s. As
hese pa ame e s co espond well o wha is obse ed
by isual inspec ion o he lesion his ology, i may no ,
a i s , seem like hei quan i a ion could b ing no el
in o ma ion compa ed o human-inspec ed pa hological
sc eening. Howe e , human eye has di icul ies keeping
ack o g ow h pa e ns o e se e al sec ions, especially
i he g ow h cu ls and in e wines wi h o he s uc u es
in he issue and/o se e al lesions a e included in he
same issue. Thus, compu e -aided quan i ica ion enables
inspec ion o lesion g ow h pa e ns in an addi ional
dimension compa ed o wo-dimensional (2D) isual
inspec ion.
Pa ame e s ela ed o lesion size a e sensi i e o issue
cu o ien a ion. E.g., in ou case, mPIN in p os a e issue
g ow wi hin he p os a ic acini, c ea ing mo e o less
cylinde -shaped lesions. I needs o be aken in o accoun
in which o ien a ion acini a e cu in ela ion o how many
sec ions he e a e. As in many o he pa ame e s, mean o
ROIs (ac oss sec ions) desc ibes well he gene al esul
o each ea u e, in a ea- ela ed ea u es he en i y o he
lesion needs o be aken in o accoun using, e.g., sums o
all ROIs o a pa icula lesion o ela ion o olume ic
es ima ions.
Quali ies such as shape, ex u e, and spa ial a angemen
o lesions can expand ou unde s anding o umo biology
and e en ually could ha e he po en ial o imp o e
he accu acy o clinical p ognosis. His opa hologic
classi ica ion based on isual inspec ion o an expe
is subjec i e e en wi h expe ienced pa hologis s, a
challenge o, e.g. Gleason sco ing o p os a e cance .[20]
Fea u es such as ex u e migh be e en impossible o
accu a ely cha ac e ize by isually inspec ing he images.
Compu e -aided quan i a i e analysis o his ological
issue images is an e ec i e ool o de ec speci ic and
small al e a ions oo sub le o small o be sepa a ed by
eye. Fo example, he e, he his ological changes desc ibed
in ea u es such as solidi y o con ex a ea may be spo ed
by eye when p ominen enough, howe e , compu e -aided
quan i a i e analysis is a mo e accu a e in posi ioning
a pa icula sample wi hin he ange o a ia ion wi hin
each pa ame e . Fu he mo e, he mo e complex and
especially up o pixel- ideli y ea u es, such as LBP and
HOG ex u e ea u es, compose o in o ma ion no
necessa ily de ec ed by human eye, and hus can p o ide
an addi ional le el o esolu ion o his ological sub yping.
Machine lea ning and s a is ical pa e n ecogni ion
ypically equi e la ge sample sizes in o de o p o ide
Figu e 4: Mouse p os a ic in aepi helial neoplasia lesion classi ie model based on ea u es ob ained by machine lea ning. (a) Lesions
so ed acco ding o posi i e class condi ional p obabili ies ob ained du ing he 1000 epe i ions o he classi ie design by andom
hold‑ou o a aining sample. T aining samples a e ma ked wi h colo s ( ed, ad anced pheno ype; g een, mild pheno ype). A p obabili y
h eshold o 0.5 is ma ked wi h a dashed line. (b) Examples o lesion egion o in e es s ep esen ing he wo pheno ypes in he classi ica ion
model. Examples o bo h aining samples and classi ied samples a e shown. I mus be no ed ha he lesions a e no in he same
scale. (c) The his og am bins showing he numbe o imes he ea u es ha e been selec ed in he classi ie model
c
b
a
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]
J Pa hol In o m 2016, 1:5 h p://www.jpa hin o ma ics.o g/con en /7/1/5
eliable models o p edic ion pu poses. Howe e ,
machine lea ning may also be seen as a way o using
he expe labeled aining o lea ning mo e abou he
a ailable da a e en when he numbe o samples is
low, as is o en he case wi h he cos ly mouse model
s udies. He e, we used a classi ie o de e mining he
likelihood o his ological changes, enabling anking o
unca ego ized samples. Fu he mo e, machine lea ning
is po en ially use ul in p o iding in o ma ion abou he
impo ance o indi idual ea u es in g ouping o samples.
Fo he pu pose o in e p e ing he da a using he
classi ie model pa ame e s, we chose egula ized logis ic
eg ession classi ie . The mo i a ion o his selec ion
is ha logis ic eg ession, whe e model complexi y is
penalized wi h egula iza ion, end o p oduce spa se
models,[18] which can be used o p o iding a selec ion o
an in o ma i e subse o ea u es in a ious applica ions,
e.g., in he analysis o g ow h and mo phology.[21] He e,
he ea u es selec ed by he classi ie a e po en ially
in e es ing pa ame e s ela ed o mPIN his ology.
The ROI size-dependen pa ame e s selec ed du ing he
epea ed aining ounds, namely he sum o con ex
a eas, eccen ici y, equi alen diame e , and pe ime e ,
can all be conside ed nume ical desc ip o s o egula i y
e sus i egula i y in lesion shape. Ea ly neoplas ic
lesions a e small and egula in shape. As hey g ow,
he space limi a ions o he issue en i onmen esul
in pene a ion o neighbo ing a eas by g ow h p essu e
squeezing he su ounding issue o , e.g., g ow h along
lumen o an acinus. The mo e p onounced he g ow h,
he s onge he p essu e agains he su ounding
issue, esul ing in inc easing acini diame e s and he
in oduc ion o i egula i y o o me ly nea -sphe ical
o -cylind ical objec s isible as ound o ellip ical shapes
in 2D. Fu he p og ess o malignancy induces in asion
o su ounding issue. This gene alized iew o e olu ion
in neoplas ic g ow h pa e ns comply wi h he idea ha
inc easing i egula i y is a decen measu e o his ological
s age o he lesion.
The ex u e ea u es selec ed o he classi ie a e
pa icula ly in e es ing and po en ially p o ide no el
indica o s o his ological changes in ea ly neoplas ic
issue. Con as , calcula ed om he GLCM, desc ibes
he di e ence o in ensi ies be ween neighbo ing pixels.
HOGs a e edge and g adien based desc ip o s. The
appea ance and shape o an objec can be cha ac e ized
wi h he dis ibu ion o local g adien s and edge
di ec ions wi hou knowing he posi ions o he g adien s
and edges. The p ocessed image is di ided in o smalle
windows, and he dis ibu ion o local g adien s and edge
di ec ions is calcula ed o each window.[15] MSER is an
image elemen de ec o me hod, which gene a es ea u es
ha a e in a ian o a ine ans o ma ions.[16] The
me hod is use ul in baseline ma ching and in all, e y
obus and as ea u e de ec o . LBP me hod in ex u e
analysis is compu a ionally simple and obus in e ms
o g ayscale a ia ions. These bo h a e e y impo an
ac o s when conside ing his ological image analysis, as
he g ayscale a ies in hese issue sec ion images due o
lack o s anda d s aining and scanning sys ems.
In summa y, we epo he de elopmen o quan i a i e
image analysis pipeline o desc ibe mo phological changes
in his ological images using mPIN in mouse p os a e
issue as a model o ea ly neoplas ic changes in he
p os a e. Ou app oach is o in oduce quan i a i e igo
h ough image analysis and machine lea ning ega dless
o he sample size. To he bes o ou knowledge, such
ho ough compu a ional pipeline o quan i a i e
analysis o mouse PIN lesions using machine lea ning
has no been p esen ed elsewhe e in he li e a u e.
Impo an ly, we show how hese compu a ional me hods
a e powe ul e en in small sample se ings, ypical in
mouse model s udies. Rep esen ing his ological changes
by cons uc ing and using mul idimensional ea u e da a
can signi ican ly con ibu e o in e p e a ion and esea ch
o ea ly pa hological lesions. The me hods desc ibed he e,
used in combina ion wi h po en ially au oma ed digi al
pa hology applica ions, could p o ide imp o ed and
as e pipelines o g ade ea ly neoplas ic lesions and o
sc een o po en ially mo e malignan lesions om la ge
se s o da a. These me hods could bo h enhance esea ch
o umo models in, e.g., gene ically manipula ed model
o ganism issues, as well as aid in u u e a emp s o
de elop be e ools o digi al pa hology.
Acknowledgmen s
Funding o his wo k has been ob ained om he
Finnish Funding Agency o Inno a ion, p ojec “ h ee
dimensional his ogenomic modeling o whole p os a e,”
Academy o Finland (g an no. 279270), he Cance
Socie y o Finland, he Sig id Juselius Founda ion,
and he Medical Resea ch Fund o Tampe e Uni e si y
Hospi al. We hank Ma ika Vähä-Jaakkola and Ka ja
Liljes öm o skill ul echnical assis ance.
Financial Suppo and Sponso ship
Nil.
Con lic s o In e es
The e a e no con lic s o in e es .
REFERENCES
1. S ewa BW, Wild CP, edi o s. IACR Wo ld Cance Repo 2014. Lyon,
F ance: WHO P ess; 2014.
2. Tomlins SA, Rubin MA, Chinnaiyan AM. In eg a i e biology o p os a e
cance p og ession. Annu Re Pa hol 2006;1:243‑71.
3. Me imen JL, E ans AJ, S igley JR. P eneoplasia in he p os a e gland wi h
emphasis on high g ade p os a ic in aepi helial neoplasia. Pa hology
2013;45:251‑63.
4. Shappell SB, Thomas GV, Robe s RL, He be R, I mann MM, Rubin MA,
e al. P os a e pa hology o gene ically enginee ed mice: De ini ions and
classi ica ion. The consensus epo om he Ba Ha bo mee ing o
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]
J Pa hol In o m 2016, 1:5 h p://www.jpa hin o ma ics.o g/con en /7/1/5
he Mouse Models o Human Cance Conso ium P os a e Pa hology
Commi ee. Cance Res 2004;64:2270‑305.
5. Di C is o ano A, Pesce B, Co don‑Ca do C, Pandol i PP. P en is essen ial
o emb yonic de elopmen and umou supp ession. Na Gene
1998;19:348‑55.
6. Ghazna i F, E ans A, Madabhushi A, Feldman M. Digi al imaging in pa hology:
Whole‑slide imaging and beyond. Annu Re Pa hol 2013;8:331‑59.
7. Hamil on PW, Bankhead P, Wang Y, Hu chinson R, Kie an D, McA DG,
e al. Digi al pa hology and image analysis in issue bioma ke esea ch.
Me hods 2014;70:59‑73.
8. Ve a M, Pluim JP, an Dies PJ, Vie ge e MA. B eas cance his opa hology
image analysis: A e iew. IEEE T ans Biomed Eng 2014;61:1400‑11.
9. Tuominen VJ, Isola J. Linking whole‑slide mic oscope images wi h DICOM
by using JPEG2000 in e ac i e p o ocol. J Digi Imaging 2010;23:454‑62.
10. Pa k JH, Walls JE, Gal ez JJ, Kim M, Aba e‑Shen C, Shen MM, e al. P os a ic
in aepi helial neoplasia in gene ically enginee ed mice. Am J Pa hol
2002;161:727‑35.
11. Schneide CA, Rasband WS, Elicei i KW. NIH image o imagej: 25 yea s o
image analysis. Na Me hods 2012;9:671‑5.
12. Ojala T, Pie ikäinen M, Mäenpää T. Mul i esolu ion g ay‑scale and o a ion
in a ian ex u e classi ica ion wi h local bina y pa e ns. IEEE T ans
Pa e n Anal Mach In ell 2002;24:971‑87.
13. Pie ikäinen M, Ojala T, Xu Z. Ro a ion‑in a ian ex u e classi ica ion using
ea u e dis ibu ions. Pa e n Recogni 2000;33:43‑52.
14. Ludwig O, Delgado D, Goncal es V, Nunes U. T ainable Classi ie ‑ usion
Schemes: An Applica ion o Pedes ian De ec ion. Vol. 1. In:
12 h In e na ional IEEE Con e ence on In elligen T anspo a ion Sys ems.
2009. p. 432‑7.
15. Dalal N, T iggs B. His og ams o o ien ed g adien s o human de ec ion.
In: IEEE Compu e Socie y Con e ence on Compu e Vision and Pa e n
Recogni ion Vol. 1. 2005. p. 886‑93.
16. Ma as J, Chum O, U ban M, Pajdla T. Robus wide‑baseline s e eo om
maximally s able ex emal egions. Image Vis Compu Pu ing 2004;22:761‑7.
17. Vedaldi A, Fulke son B. VLFea : An Open and Po able Lib a y o Compu e
Vision Algo i hms; 2008. A ailable om: h p://www. l ea .o g/. [Las
accessed on 2016 Jan 16].
18. Tibshi ani R. Reg ession sh inkage and selec ion ia he lasso. J R S a Soc
1996;58:267‑88.
19. F iedman J, Has ie T, Tibshi ani R. Regula iza ion pa hs o gene alized
linea models ia coo dina e descen . J S a So w 2010;33:1‑22.
20. Iczkowski KA, Lucia MS. Cu en pe spec i es on Gleason g ading o
p os a e cance . Cu U ol Rep 2011;12:216‑22.
21. Ruusu uo i P, Lin J, Sco AC, Tan Z, So sa S, Kallio A, e al. Quan i a i e
analysis o colony mo phology in yeas . Bio echniques 2014;56:18‑27.
[Downloaded ee om h p://www.jpa hin o ma ics.o g on Monday, Augus 15, 2016, IP: 153.1.49.22]