1
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
www.na u e.com/scien i ic epo s
Analysis o spa ial he e ogenei y
in no mal epi helium and
p eneoplas ic al e a ions in mouse
p os a e umo models
Mi a Valkonen1,*, Pekka Ruusu uo i1,2,*, Kimmo Ka asalo1,3, Ma i Nyk e 1, Tapio Visako pi1,4
& Leena La onen1,4
Cance in ol es his ological changes in issue, which is o p ima y impo ance in pa hological diagnosis
and esea ch. Au oma ed his ological analysis equi es abili y o compu a ionally sepa a e pa hological
al e a ions om no mal issue wi h all i s a iables. On he o he hand, unde s anding connec ions
be ween gene ic al e a ions and his ological a ibu es equi es de elopmen o enhanced analysis
me hods sui able also o small sample sizes. He e, we se ou o de elop compu a ional me hods o
ea ly de ec ion and dis inc ion o p os a e cance - ela ed pa hological al e a ions. We use analysis o
ea u es om HE s ained his ological images o no mal mouse p os a e epi helium, dis inguishing he
desc ip o s o a iabili y be ween en al, la e al, and do sal lobes. In addi ion, we use wo common
p os a e cance models, Hi-Myc and P en+/− mice, o build a ea u e-based machine lea ning model
sepa a ing he ea ly pa hological lesions p o oked by hese gene ic al e a ions. This wo k o e s a
se o compu a ional me hods o sepa a ion o ea ly neoplas ic lesions in he p os a es o model
mice, and p o ides p oo -o -p inciple o linking speci ic umo geno ypes o quan i a i e his ological
cha ac e is ics. The esul s ob ained show ha sepa a ion be ween di e en spa ial loca ions wi hin
he o gan, as well as classi ica ion be ween his ologies linked o di e en gene ic backg ounds, can be
pe o med wi h e y high speci ici y and sensi i i y.
Tissue his ology is one o he main de e minan s in s udying and diagnosing many pa hologies, including cance .
Solid umo s change he s uc u e o he issue due o al e ed mo phologies and localiza ions o umo cells wi hin
he no mal issue. His opa hology is adi ionally a e y in ui i e me hod, whe e decisions a e mos o en based
on isual inspec ion. O en, howe e , abili y o objec i ely ecognize and quan i y pa hological changes in issue
his ology would be desi ed. Fu he mo e, gaining he decisi e pa hological in o ma ion om basic his ological
s ainings, mos o en hema oxylin and eosin (HE), would help o a oid using cos ly special s ainings. Se e al
cu en app oaches aim o de elop ools o help clinical pa hologis s o de e mine p esence o s a e o a pa icula
lesion om HE-s ained images, e.g. o s age cance o diagnos ic and p ognos ic pu poses1,2. Ye , a p essing need
o diagnose cance a ea lie s ages conce ns se e al cance ypes, e.g. p os a e cance and b eas cance ; when
umo s a e s ill small and changes in hem mo e benign, ea men op ions and p ognoses a e be e . Abili y o
ecognize and quan i y small and sub le changes in issue mo phology a e c ucial also o basic and p eclinical
esea ch aiming o iden i y ea ly changes p eceding and leading o malignan s ages o cance .
To be able o quan i y changes in issue mo phology, measu able de e minan s o he pa hology in ques ion
need o be de e mined. Key ques ions a e how o di e en ia e be ween no mal and pa hological issue, how o
measu e he s age o he pa hological change, and e en how o di e en ia e be ween se e al possible ypes o
change, e.g. sub ypes o cance . Fo example, accu a e sepa a ion o pa hologies om no mal issue his ology
equi es unde s anding and inclusion o all s a es and a iables o he no mal issue, whe he o igina ing om
1P os a e Cance Resea ch Cen e , Facul y o Medicine and Li e Sciences and BioMediTech, Uni e si y o Tampe e,
Tampe e, Finland. 2Tampe e Uni e si y o Technology, Po i, Finland. 3BioMediTech Ins i u e and Facul y o
Biomedical Sciences and Enginee ing, Tampe e Uni e si y o Technology, Tampe e, Finland. 4Fimlab Labo a o ies,
Tampe e Uni e si y Hospi al, Tampe e, Finland. *These au ho s con ibu ed equally o his wo k. Co espondence
and eques s o ma e ials should be add essed o L.L. (email: [email p o ec ed])
Recei ed: 28 Decembe 2016
accep ed: 13 Feb ua y 2017
Published: 20 Ma ch 2017
OPEN
www.na u e.com/scien i ic epo s/
2
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
cha ac e is ics o he issue i sel , o echnical a ia ion due o e.g. o ien a ion o he his ological sec ion cu el-
a i e o issue s uc u es.
In cance esea ch, ecen yea s o nex -gene a ion sequencing ha e e ealed he ex en o gene ic and gene
exp ession al e a ions in cance 3,4. Howe e , he pheno ypic e ec s o many gene ic al e a ions and hei com-
bina ions is s ill unde esea ch. The common goal is o be able o sub ype umo s be e o enhanced pa ien
s a i ica ion and ca e in he u u e. Combina ion o gene ic in o ma ion wi h his ology, howe e , equi es ha
he his ological in o ma ion can be ans o med in o a quan i a i e, objec i e o m. This can be achie ed h ough
digi al imaging and compu a ional analysis o he his ological cha ac e is ics. While compu a ional image in o -
ma ics can p o ide a ple ho a o quan i ied desc ip o s o a gi en image, he challenge in his ology is o so ou
he ele an cha ac e is ics which can be p esen ed in he o m o use ul ea u e ep esen a ions.
Fea u e-based analysis combined wi h supe ised lea ning has been a common app oach in decision suppo
sys ems and compu e aided diagnosis based on whole slide images1,5,6,. Such app oaches ha e been success ully
used o quan i a i ely desc ibing cha ac e is ics o p os a e his ology in neoplas ic lesions bo h o a mouse
model7 and o human issue8. Howe e , p e ious s udies ha e le oom o imp o ed ea u e enginee ing and
classi ica ion pe o mance.
We aim a imp o ing his ological ecogni ion and quan i ica ion o pa hological ea u es in p os a e cance ,
and sea ch o desc ip o s o di e en ia e ea ly pa hological lesions om no mal p os a e issue. Fu he mo e, he
goal is o sepa a e gene ically di e en ypes o ea ly neoplas ic changes om each o he . In he e, we use wo clas-
sical and popula gene ic p os a e cance mouse models, namely he e ozygous P en9 and Hi-Myc10, o pe o m
quan i a i e image analysis on ea ly neoplasms compa ed o no mal mouse p os a ic issue. Wi h a compu a ional
sepa a ion o hund eds o ea u es om he whole slide images o his ological issue sec ions and a andom o es
based machine lea ning app oach, we ind a combina ion o issue ea u es able o dis inguish be ween 1) no mal
spa ial he e ogenei y in he p os a e issue, 2) ea ly P en+ /− o Hi-Myc-induced neoplasms om no mal issue,
and 3) he wo ypes o neoplasms om each o he . Ou s udy se es as he i s s ep owa ds de eloping ools
o au oma ed analysis o ea ly neoplas ic changes in p os a e issue and hei linkage o di e en gene ic g oups.
Resul s
Spa ial a ia ion in epi helium o no mal p os a e. Fi s , we wan ed o assess no mal spa ial a ia ion
in he his ology o he p os a e. Mouse co e p os a e can be di ided in h ee lobula a eas: en al p os a e (VP),
la e al p os a e (LP) and do sal p os a e (DP), which su ound he u e h a (Fig.1A). All h ee lobes a e domi-
na ed by p os a e acini co e ed wi h an epi helial cell laye and a e su ounded by a basemen memb ane and
loose connec i e issue. Be ween he lobes, sub le di e ences exis in he o ganiza ion and di ec ion o he acina
ubes, somewha a ec ing he appea ance and lumen size o acini in his ological p epa a ions. The epi helium
is o speci ic ele ance due o being he issue componen whe e p os a e cance s o igina e om. The no mal
appea ance o he epi helium a ies be ween he di e en lobes (Fig.1A). Epi helium in he DP is columna , cy o-
plasm is ela i ely eosinophilic, and he nuclei a e cen ally o basally loca ed. The epi helium can be u ed. LP
epi helium has only spa se in oldings. The cells a e cuboidal o low columna , and he nuclei a e small, uni o m
and basally loca ed. VP has only ocal epi helial u ing. The cells a e cuboidal o columa , and he cy oplasm is
less eosinophilic. Nuclei a e small and basally loca ed (Fig.1A).
We manually selec ed 227 acini o ep esen a ia ion in p os a e epi helium in his ological sec ions, including
bo h he he e ogenei y in no mal appea ance o he issue, and he echnical a iance (e.g. acini cu in di e en
o ien a ions). Fo hese images, we pe o med p ep ocessing o, o example, co ec colo a ia ion ac oss he
issue samples, and o exclude a eas no including issue (e.g. emp y and sec e ion-con aing a eas inside he acini)
(Fig.1B). We applied colo decon olu ion o sepa a e hema oxylin and eosin s ains as sepa a e colo channels,
and pe o med nuclea segmen a ion. The esul ing image in o ma ion we used o compu e a compila ion o 241
ea u es (lis ed in Supplemen a yTable1). These ea u es included nume ous desc ip o s ela ed o issue ex u e
and local en i onmen s, as well as nume ic ep esen a ions o p ope ies, spa ial a angemen , and dis ibu ion o
nuclei. When hese 241 ea u es a e used o ep esen he samples in dimension- educing -Dis ibu ed S ochas ic
Neighbo Embedding ( -SNE)11 plo , he ex en o he ex ac able spa ial a ia ion be ween he no mal lobes o
he p os a e is shown (Fig.1C). While he VP sha es cha ac e is ics wi hin he ange o LP, he DP is mo e clea ly
dis inguished om he o he lobes.
Quan i a i e cha ac e is ics o mPIN lesions. To s udy dis inc ion o small pa hological changes om
no mal epi helium, we wan ed o compa e no mal issue o ea ly pa hological lesions. Mice he e ozygous o
umo supp esso P en o m mouse p os a ic in aepi helial neoplasia (mPIN) wi hin 8–12 mon hs9. In he e, we
used p os a e samples om P en+ /− mice o 10-11 mon hs, when ecognizable mPIN is e iden (Fig.2A).
We selec ed 199 a eas o mPIN, and pe o med image p ocessing and ea u e compu a ion as abo e. A -SNE
plo (Fig.2B) shows ha he PIN lesions o di e en p os a e lobes a e mixed a he han sepa a ed as lobe-speci ic
clus e s. This indica es ha , compa ed o no mal epi helium, he spa ial he e ogenei y is dec eased in mPIN
(compa e Fig.2B o Fig.1C). When compa ing he ela i e dis ibu ions o no mal epi helium and mPIN lesions,
PIN a eas a e clea ly sepa a e om no mal issue a eas by he compu ed compila ion o ea u es (Fig.2C). When
compa ing he di e en p os a ic lobes, i is e iden ha LP is u hes and DP closes o PIN lesions based on he
compu ed ea u e p o ile, co esponding o he issue cha ac e is ics obse ed by eye (Fig.2A).
To es whe he he compu ed ea u e cha ac e is ics can be used o eliably sepa a e mPIN lesions om no -
mal epi helium, we applied machine lea ning. We de eloped a andom o es based model and applied i in
lea e-one-ou c oss alida ion (LOOCV) o es ima e he p obabili y o a sample o belong o he g oup o PIN
lesions based on he ea u e da a. Figu e3A shows he classi ica ion con idence gi en by he machine lea ning
model o each sample o belong o he g oup o PIN. The accu acy o he es ima ions by he model was analysed
using ecei e ope a ing cha ac e is ic (ROC) cu e, om which he a ea unde cu e (AUC) measu e can be
www.na u e.com/scien i ic epo s/
3
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
used o quan i ying he sepa a ion be ween he no mal and p eneoplas ic issues (Fig.3B; AUC 0.988). P edic o
impo ances o 20 mos in luen ial ea u es in he model o dis inquish be ween he mo phology ypes a e shown
in Fig.3C. These ea u e impo ances we e gi en by a model ha was ained wi h all a ailable samples.
F om he LOOCV expe imen , used o alida e he obus ness o ou andom o es model, we collec ed
al oge he 426 models om which he a e age impo ances a e shown in Supplemen a yFigu e1. The a e aged
ea u e impo ance lis con ains a simila se o ea u es as ha gi en by he o iginal model ained wi h all a ail-
able samples (Fig.3C). These include se e al ypes o ex u e ea u es, such as LBPs and SIFT- ea u es. Nuclea
ea u es include se e al desc ip o s o nuclea size, densi y and neighbou hood (NhoodMaxDis , NhoodS dDis ,
HhoodSkewness, meanNucSize, meanNucDis InNucNB, NhoodMeanDis ). Ano he se o impo an ea u es
a e he ea u es desc ibing he ela i e posi ions and o ien a ions o he nuclei (NhoodNucAngleSkewABS,
NhoodNucAngleKu ABS, NhoodNucAngleVa , NhoodNucAngleS d0). These cap u e he dis inc i e p ope y
o no mal epi helial issue, whe e nuclei a e mos o en o ien ed as “beads in a ow” as opposed o sca e ed dis-
ibu ion in a umo (Supplemen alFig.2).
We u he es ed he abili y o he model o dis inguish mPIN om no mal epi helium in each lobe
(Supplemen a yFigu e3). Acco ding o he di e en incidence o mPIN in he p os a ic lobes in he P en+ / −
model mice, he numbe o mPIN samples in he analysis a ied be ween lobes being g ea es in LP and lowes in
VP (no mal epi helium and mPIN nVP = 37, nLP = 282, nDP = 107). PIN was dis inguished om no mal p os a e
mos accu a ely in LP (AUC 0.997), likely due o he clea pheno ypic di e ence in he his ology. mPIN in VP
and DP we e simila ly challenging, al hough wi h hese, ela i ely small sample se s, a e y high accu acy was s ill
eached (AUC 0.972).
Compu a ional dis inc ion be ween his ologies o di e en gene ic g oups. High exp ession o
oncogenic Myc in he mouse p os a e induces neoplas ic lesions isible al eady a 1 mon h o age10. These lesions
de elop la e on o adenoca cinoma, in con as o he mPIN in he P en+ / − he e ozygous mice which does
no de elop in o ca cinoma wi hou addi ional gene ic o ca cinogenic manipula ion9. We wan ed o compa e
hese wo ypes o ea ly neoplasms wi h gene ic di e ences, and o ind compu a ional ea u es sepa a ing hem
om each o he and om he no mal epi helium. As mos o he umo s in hese models o m in he LP, we
Figu e 1. Quan i a i e image analysis o no mal mouse p os a e his ology. (A) Gene al appea ance o
mouse p os a e in a his ological sec ion (la ge image) and a ia ion in he appea ance o no mal epi helium
wi hin h ee lobula a eas o mouse co e p os a e: en al p os a e (VP), la e al p os a e (LP) and do sal p os a e
(DP) (small images). U e h a (U) and he muscle laye (ml) su ounding i a e ma ked. Scale ba s: 1 mm
(la ge image) and 25 μm (small images). (B) O e iew o he image analysis wo k low ha includes masking
o ROI, co ec ing colo a ia ion using his og am ma ching, excluding a eas no including issue, sepa a ing
hema oxylin and eosin s ains, segmen ing cell nuclei, and ex ac ing quan i a i e ea u e da a. (C) A -SNE plo
p esen ing he spa ial a ia ion o quan i a i e ea u es in he epi helium be ween he no mal mouse p os a e
lobes (VP, LP, DP).
www.na u e.com/scien i ic epo s/
4
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
concen a ed on his issue a ea selec ing only LP samples om P en+ /− model mice (no mal epi heliu n = 137,
mPIN n = 145), and manually selec ed samples om LP o Hi-Myc mice (no mal epi helium n = 111, mPIN
n = 189). Examples o ep esen a i e his ologies on no mal and p eneoplas ic LP epi helium a e shown in Fig.4A.
As he wo p eneoplas ic change ypes occu a di e en ages, ou samples om he wo di e en models
ep esen ed p os a e epi helium om mice o e y di e en age (10–11 mon hs in P en+ /− mice compa ed o 1
mon h in Hi-Myc mice). Thus, o ensu e consis ency and compa abili y o he da a om hiMyc and P en+ /− is-
sues, we selec ed ea u es whose dis ibu ions did no show s a is ically signi ican di e ence ( h eshold α = 0.05)
acco ding o Kolmogo o -Smi no es in he wo con ol g oups o use in he u he analysis. Al oge he 59 ea-
u es ul illed he c i e ion (Supplemen a yTable2, Supplemen a yFigu e4). I is e iden al eady by he ea u e
Figu e 2. Quan i a i e cha ac e is ics o mouse PIN. (A) Examples o p os a ic acini con aining no mal
epi helium and PIN lesions om he h ee lobula a eas (VP, LP, DP). (B) Rep esen a ion o he quan i a i e
ea u es o PIN lesions in h ee mouse p os a e lobes (VP, LP, DP) in wo-dimensional ea u e space using
-SNE. (C) A -SNE isualisa ion e eals dec easing spa ial he e ogenei y in PIN lesions compa ed o no mal
epi helium in he mouse p os a e.
www.na u e.com/scien i ic epo s/
5
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
alues ha he h ee g oups o samples ha e hei own, dis inc signa u es (Fig.4B). Fu he mo e, hese g oups
a e clea ly dis inguished in ep esen a ion o he samples acco ding o he ea u e alues in -SNE (Fig.4B).
We de eloped a machine lea ning model based on he selec ed subse o ea u es o es ima e he p obabili y o
a sample o belong o any o he h ee g oups (no mal epi helium, P en he e ozygous mPIN, o Hi-Myc-induced
ea ly neoplasia). The model is success ul in p edic ing accu a ely he his ological classes o he samples, as shown
in a con usion ma ix o he p edic ions (Fig.5A) and a ROC-cu e o accu acy analysis (Fig.5B; AUC 0.997 o
no mal, 0.990 o P en+ /− , and 0.995 o Hi-Myc). Simila ly as in dis inguishing mPIN om no mal epi helium,
he p edic o impo ances o he mos aluable ea u es in his andom o es model (Fig.5C) include nuclea
densi y and angle- ela ed ea u es (NhoodNucAngleKu ABS; NhoodMeanDis , numbe O NucInNucNB). As
expec ed, howe e , he sepa a ion o wo mo phologically di e en neoplas ic his ologies b ings ea u es cap-
u ing mo e de ailed ex u al in o ma ion om he s ain in ensi ies and om nuclea densi y map o he op lis
o ea u es (e.g. Densi yLBP6, LBP4H). These emain in he lis o mos in luen ial ea u es e en when a e aging
he p edic o impo ances o 582 ainings o he andom o es model (Supplemen a yFigu e5), and p o ide
de ailed, quan i a i e in o ma ion o he di e ences be ween Hi-Myc and P en+ /− -p o oked ea ly neoplasms
(Fig.5D).
Discussion
Au oma ic ecogni ion and quan i a i e analysis o issue pa hologies equi es ways o compu a ionally iden i y
ep esen a i e his ological ea u es sepa a ing no mal and al e ed issue. In his wo k, we ex ac ed a se o 241
ea u es om images o p os a e issue, and used hem o analyze spa ial he e ogenei y wi hin no mal p os a e, as
well as o sepa a e di e en ypes o ea ly neoplas ic changes om no mal issue and om each o he . Ou esul s
show ha sepa a ion be ween di e en spa ial loca ions wi hin he o gan, as well as classi ica ion be ween his ol-
ogies p o oked by di e en gene ic lesions, can be pe o med wi h e y high speci ici y and sensi i i y.
We applied adi ional machine lea ning app oach based on ex ac ion o a la ge se o enginee ed ea u es
ollowed by a andom o es ensemble classi ie . Gi en he e y high accu acy ob ained o he ela i ely low
numbe o samples and small size o egions o in e es , his app oach appea s o be well jus i ied. Recen ly, neu al
ne wo k based deep lea ning app oaches12 ha e gained much a en ion in e o s o ecognize cance ous issue
om no mal issue, and in de eloping p e-sc eening ools o pa hologis s o indica e suspec a eas in issue13,14.
The se -back o hese app oaches so a is he di icul y o examining he ele ance and meaning o model p op-
e ies, i.e. ne wo k weigh s and ou pu laye alues used in he decision low in he con ex o he issue. In o de
o in e p e he model p ope ies used compu a ionally o a pa hologis o a esea ch biologis , meaning ul and
ecognizable ea u es o he issue a e p e e ed, especially i he in o ma ion o he issue ea u es need o be com-
bined wi h eadou s ob ained om o he measu emen modali ies. Me hods o a oid he in e p e abili y issues
in using deep lea ning a e likely o ollow, while combina ion app oaches o deep lea ning and adi ional, ea u e
based machine lea ning a e also being in es iga ed15,16.
In his s udy, we used wo popula p os a e cance mouse models17, o compa e he his ology o ea ly pa hol-
ogies o igina ing om di e en gene ic al e a ions. P en is a umo supp esso ha unc ions by inhibi ing Ak
pa hway, and i is o en dele ed in human p os a e cance 18. Mice he e ozygous o P en o m cance in many
o gans9. In he p os a e, hese mice de elop mPIN9, which is a ype o in si u lesion no epo ed o de elop o
in asi e ca cinoma wi hou addi ional gene ic o ca cinogenic manipula ion. Myc, on he o he hand, is a an-
sc ip ion ac o and a powe ul oncogene o e exp essed in many cance s, including p os a e cance 19. Hi-Myc
mice10 a e a popula model o s udy p os a e cance , as i is one o he ew gene ic mouse models o ming adeno-
ca cinoma, hus exhibi ing he cance ype mos common in human p os a e. Bo h ou models ep esen an ea ly
phase in he s ep-wise de elopmen o cance , and hus can be used in s udying he ea ly pa hological changes
in issue. We achie ed a success ul sepa a ion be ween he his ologies p o oked in hese wo models by hei
Figu e 3. Classi ica ion o mouse PIN using supe ised machine lea ning. (A) Classi ica ion con idence
gi en by he andom o es model o each P en+ /− sample o belong o he g oup o PIN using LOOCV.
(B) The sensi i i y and speci i y o he andom o es model classi ica ion shown in A p esen ed as a ROC cu e,
and he pe o mance measu ed by a ea unde he cu e (AUC). (C) Lis o 20 mos signi ican ea u es and hei
ela i e impo ances in he sepa a ion o mPIN om no mal epi helium gi en by a classi ica ion model ained
wi h all P en+ /− samples.

www.na u e.com/scien i ic epo s/
6
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
di e en gene ic al e a ions wi h e y high speci ici y and sensi i i y. This gi es p omise o u u e aims o au o-
ma ically link gene ic and quan i a i e his ological in o ma ion o mo e a ied gene ic popula ions and umo
sub yping.
The me hod we p esen ed he e is gene ic, and he applicabili y is no limi ed o mouse issue. Tumo ma e-
ial om human pa ien s, howe e , includes highe gene ic and pheno ypic a iabili y compa ed o gene ically
es ic ed mouse model ma e ial. Thus, well anno a ed, la ge enough da ase s a e equi ed o u he de elop and
alida e au oma ed image analysis pipelines o u u e use in clinical and esea ch applica ions in human cance .
Ano he cu en challenge is o au oma e he de ec ion s ep o egions o in e es . Recen ly, machine lea ning has
been applied in au oma ed ROI de ec ion in, e.g., me as asis de ec ion om human b eas cance samples13. Ou
app oach o ROI classi ica ion could also be applied o ROI de ec ion. This equi es quan i a i e p ocessing o
no jus he in e es a eas, bu all he neighbou ing issue ypes and s uc u es as well, including he basemen
memb ane, s omal cells and ibe s, ne es, ascula u e, smoo h muscle, and adipose issue. Acqui ing quan i a-
i e ep esen a ions o hese di e en issue componen s and ypes will bene i applica ions o digi al pa hology
also beyond cance esea ch.
In addi ion o umo geno ypes, da a ep esen ing umo pheno ypes a e desi ed o combine wi h quan i a i e
ep esen a ions o his ology p o ided by me hods such as he ones p esen ed he e. Recen ad ances in spa ial
ansc ip omics20,21 in eg a ed wi h compu a ional his ological analyses will undoub edly p o ide unde s anding
o spa ial a ia ion and e olu ion o umo s. Fu he , quan i a i e ep esen a ion o issue ea u es and he e oge-
nei y along wi h h ee-dimensional econs uc ions o whole o gans om se ial sec ions, such as he p os a e22,
Figu e 4. Quan i a i e cha ac e is ics o no mal epi helium, P en he e ozygous mPIN, and Hi-Myc-
induced p eneoplasia in mouse la e al p os a e. (A) Examples o ep esen a i e his ologies o no mal and
p eneoplas ic LP epi helium om P en+ /− and Hi-Myc mouse model p os a es. Scale ba s 25 μm. (B) Dis inc
ea u e alue pa e ns o P en he e ozygous PIN, Hi-Myc-induced ea ly neoplasia, and no mal epi helium
p esen ed in a hea map a e hie a chical clus e ing o no malized ea u e alues. (C) Th ee-dimensional -SNE
isualisa ion shows dis inc i e pa e ns o he h ee his ological popula ions.
www.na u e.com/scien i ic epo s/
7
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
will enable in ui i e isuliza ion and p o ide no el insigh in o he spa ial a ia ion wi hin issue, as well as umo
g ow h pa e ns, in a na u al con ex .
Me hods
E hical pe missions. All animal expe imen a ion and ca e p ocedu es we e ca ied ou in acco dance
wi h guidelines and egula ions o he na ional Animal Expe imen Boa d o Finland, and we e app o ed
by he boa d o labo a o y animal wo k o he S a e P o incial O ices o Sou h Finland (licence numbe s
ESAVI/6271/04.10.03/2011 and ESAVI/5147/04.10.07/2015).
Tissue samples. FVB/N mice ei he he e ozygous o P en (P en+ / −
9) o ansgenic o MYC oncogene
(Hi-Myc10) we e used. P os a es we e ixed in PAXgeneTM issue ixa i e acco ding o manu ac u e ’s p o ocol,
and embedded in pa a in. 5 μm issue sec ions we e cu , a ached o glass slides, and s ained wi h hema oxylin
and eosin.
Imaging and ROI sepa a ion. HE-s ained slides we e whole slide imaged wi h Zeiss Axioskop40 mic o-
scope (Ca l Zeiss Mic oImaging, NY, USA) wi h 20x objec i e and a CCD 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 acqui-
si 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 23. Snapsho images o Figu es
we e ob ained h ough JVSView i ual mic oscope (h p://j smic oscope.u a. i) and ImageJ so wa e (Na ional
Ins i u es o Heal h, Be hesda, MD, USA24). Regions o in e es we e manually ma ked using a eehand selec ion
ool in ImageJ. The esul ing bina y mask was used o ex ac ing he ROI om he ull esolu ion o iginal HE
image o u he p ocessing. In no mal issue, epi helial laye o p os a e acini was included, excluding o he is-
sue componen s in he o gan such as s oma, u e h a, essels and ne e bundles. Pa hological lesion masks each
included solely mPIN/neoplas ic epi helium. In he case o Hi-Myc samples, hese we e always wi hin a single
acinus each. In he case o P en+ /− lesions, whe e one mPIN umo could each se e al acina lumen wi hin a
ce ain sec ion, all a ec ed lumen we e included in a single mask.
P ep ocessing o images. The p ep ocessing o he images included his og am ma ching in o de o
emo e he colo a ia ion be ween samples, exclusion o unwan ed egions, colo decon olu ion o sepa a e
Figu e 5. Classi ica ion be ween no mal epi helium, P en he e ozygous mPIN, and Hi-Myc-induced
p eneoplasia in mouse la e al p os a e using supe ised machine lea ning. (A) Classi ica ion accu acies
o sepa a ing he h ee his ological popula ions gi en by he p oposed classi ica ion model as a con usion
ma ix. (B) ROC cu es ep esen ing he pe o mance o he classi ica ion model when dis inguishing P en
he e ozygous mPIN, Hi-Myc-induced ea ly neoplasia, and no mal epi helium. Each cu e p esen s he
classi ica ion accu acy o sepa a ing one g oup o he h ee om he wo o he o he g oups. Pe o mances
as measu ed by he a ea unde he cu e (AUC) a e shown. (C) Lis o 20 mos signi ican ea u es and hei
ela i e impo ances in he sepa a ion o P en he e ozygous mPIN, Hi-Myc-induced ea ly neoplasia, and
no mal epi helium by he andom o es model ained wi h all o he LP samples om P en+ /− , Hi-Myc,
and no mal epi helium. (D) Values o ou mos signi ican ea u es in C as boxplo s showing he di e ences
be ween P en he e ozygous mPIN, Hi-Myc-induced ea ly neoplasia, and no mal epi helium.
www.na u e.com/scien i ic epo s/
8
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
hema oxylin and eosin s ains, and nuclei segmen a ion. These s eps we e implemen ed o bounding box a eas
a ound each ROI.
His og am ma ching was pe o med o balance s aining a ia ion be ween sec ions. Fo a e e ence his og am,
a mean his og am was compu ed om a ep esen a i e se o samples consis ing o ROI images o p eneoplas ic
lesions and no mal p os a e epi helium om all h ee lobula a eas. A e his, his og am ma ching was pe -
o med by using a ans o m unc ion compu ed be ween he image’s his og am and he e e ence his og am.
To segmen he e ec i e issue a ea wi hin each ROI, a mask o sec e ion- illed egions and emp y a eas was
ob ained by sub ac ing di e en colo channels and pe o ming con as limi ed mapping o he in ensi y alues
simila ly as in Ruusu uo i e al.7. The inal bina y mask was ob ained by h esholding using O su’s me hod25. To
smoo hen he bina y segmen a ion mask, mo phological opening, closing, and illing we e pe o med.
A colo decon olu ion algo i hm26 was applied o con e he ed, g een, and blue channels o each image in o
hema oxylin s ain, eosin s ain, and backg ound. Hema oxylin s ains mainly he cell nuclei and he e o e, hema-
oxylin channel was u he p ocessed o segmen cell nuclei. Maximally s able ex emal egions (MSER)27 we e
ex ac ed om he g ayscale image o hema oxylin s ain. MSER is a me hod o blob de ec ion om an image. Se
o de ec ed egions we e selec ed based on he size co esponding o po en ial nuclei size. Addi ionally, egions
ha did no con ain high g ayscale in ensi y alues ela ed o high amoun o hema oxylin s ain, we e excluded.
To ge he map o high a e o hema oxylin, opha il e ing, maximum il e ing, Gaussian il e ing, and image
in ensi y adjus men was pe o med. Bina y mask o high hema oxylin a e was ob ained by h esholding using
O su’s me hod25. To ge he inal bina y mask o cell nuclei, MSER egions ha we e o e lapping wi h mask o
high a e o hema oxylin we e selec ed.
Fea u e ex ac ion. P ope ies o each ROI we e desc ibed wi h ex ac ion o 241 ea u es
(Supplemen a yTable1). These ea u es included local desc ip o s ela ed o image ex u e and dis ibu ion o
nuclei. Tex u e ea u es we e ex ac ed om local neighbo hoods ep esen ing dis inc le els o issue a chi-
ec u e and, hus, measu ed as di e en ea u es (e.g. Con as -H, NhoodCon as -H, ROI-BlockCon as -H,
ROICon as -H), and also om bo h hema oxylin and eosin channels (e.g. Con as -H, Con as -E). The neigh-
bo hoods o ex u e ea u es included bounding box o each segmen ed nucleus, 35 × 35 pixel neighbo hood
a ound each nucleus, non-o e lapping 50 × 50 pixel neighbo hoods wi hin e ec i e issue a ea wi h unwan ed
egions included and excluded (e.g. mROI-BlockCon as -H, ROI-BlockCon as -H), and he whole bounding
box image o he ROI. Nuclei dis ibu ion ea u es we e ex ac ed om a 100 × 100 pixel neighbo hood a ound
each nucleus. To ob ain a single ea u e ec o o he whole ROI a ea, mean ea u e alues we e calcula ed om
all blocks p esen ing one combina ion o ce ain ea u e and neighbo hood. De ails abou each ex ac ed ea u e
and he applied local neighbo hoods a e p esen ed in Supplemen a yTable1.
Tex u e ea u es. 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). P ope ies o he ex u e wi hin
each ROI we e also ex ac ed using local bina y pa e ns (LBP)28,29 and scale-in a ian desc ip o s ob ained ia
he Scale-in a ian ea u e ans o m (SIFT)30. Addi ionally, p ope ies o MSER egions we e used as ea u es.
VLFea 31 implemen a ions o SIFT and MSER we e used in his wo k.
Nuclei dis ibu ion ea u es. Fea u es ela ed o dis ibu ion o cell nuclei we e calcula ed om a nuclei
loca ion map, gene a ed by ma king he cen e poin o each segmen ed nucleus. Fea u es included desc ip o s
ela ed o in e -nuclei dis ance, nuclei loca ions wi h espec o each o he desc ibed wi h angula s a is ics, num-
be o nuclei wi hin a neighbo hood, and densi y ea u es. The densi y ea u es we e calcula ed om a Gaussian
il e ed nuclei loca ion map.
The angula s a is ic ea u es we e ex ac ed using Ci cS a oolbox32. Fo each nucleus, an angle o all i s
neighbou ing nuclei wi hin a 100 × 100 pixel block was calcula ed. Supplemen a yFigu e2 p esen s an example
pola his og am o hese calcula ed angles om bo h no mal epi helium sample and p eneoplas ic lesion sample.
The ea u es ela ed o angula s a is ics included p ope ies o his pola his og am, such as, a iance, s anda d
de ia ion, skewness, and ku osis.
Fea u e selec ion. To s udy he spa ial a ia ion in epi helium wi hin di e en lobula a eas o no mal is-
sue, as well as in he compa ison o no mal issue and p eneoplas ic mPIN lesions, we used all he ex ac ed 241
ea u es. Fo he compa ison o h ee g oups (no mal, P en+ /− , and Hi-Myc), a ea u e selec ion was pe o med
by s a is ical es ing be ween ea u es ex ac ed om bo h Hi-Myc and P en+ /− no mal p os a e epi helium sam-
ples. Two-sample Kolmogo o -Smi no es 33 (signi icance h eshold α = 0.05) was used o de e mine i he ea-
u e da a ex ac ed om hese wo no mal sample g oups we e om he same con inuous dis ibu ion. Al oge he
59 ea u es no showing s a is ically signi ican di e ece be ween he wo no mal popula ions we e included in
u he analysis (Supplemen a yTable2).
Classi ica ion. The ea u e ep esen a ions o ROI samples we e used o ain a andom o es model34.
Random o es algo i hm was chosen due o i s capabili y o handle bo h high da a dimensionali y and a ying
sample sizes in a compu a ionally e icien manne . Addi ionally, he algo i hm assigns weigh s o inpu ea u es
based on hei impo ance in he classi ica ion ask, p o iding an in e p e able classi ica ion model and addi ional
insigh in he con ibu ion o ea u es. Boo s ap agg ega ion, which is a machine lea ning algo i hm ha com-
bines mul iple e sions o decision ees in o a andom o es model, was used o imp o e he s abili y and accu-
acy o he model. Each decision ee e sion is cons uc ed om a andomly sampled da ase wi h eplacemen .
The implemen ed model was an ensemble o 50 decision ees.
Lea e-one-ou c oss- alida ion was used o es ima ing he classi ica ion pe o mance o o ou andom o -
es model. Fo example, when dis inguishing mPIN om no mal epi helium, we had 426 samples in o al, and
www.na u e.com/scien i ic epo s/
9
Scien i ic RepoR s | 7:44831 | DOI: 10.1038/s ep44831
he e o e, we ained 426 andom o es models. Fo each model, one sample was le ou om he aining phase
and hen he ained model was used o p edic he p obabili y o his excluded sample o belong o he g oup
wi h ea ly neoplas ic changes.
F om he LOOCV expe imen s, a e age ea u e impo ances and co esponding s anda d de ia ions
we e compiled. When dis inguishing mPIN om no mal epi helium, hese we e calcula ed om he ea u e
weigh s gi en by each o he ained 426 models (Supplemen a yFigu e1). When dis inguishing be ween
Hi-Myc-induced ea ly neoplasia, P en he e ozygous mPIN, and no mal epi helium, he a e age ea u e impo -
ances o he 582 models we e calcula ed (Supplemen a yFigu e5).
Re e ences
1. Yu, K.-H. e al. P edic ing non-small cell lung cance p ognosis by ully au oma ed mic oscopic pa hology image ea u es. Na u e
Communica ions 7 (2016).
2. Ba ke , J., Hoogi, A., Depeu singe, A. & Rubin, D. L. Au oma ed classi ica ion o b ain umo ype in whole-slide digi al pa hology
images using local ep esen a i e iles. Medical image analysis 30, 60–71 (2016).
3. Chak a a hi, B. V., Nepal, S. & Va ambally, S. Genomic and epigenomic al e a ions in cance . The Ame ican Jou nal o Pa hology
186, 1724–1735 (2016).
4. Roychowdhu y, S. & Chinnaiyan, A. M. T ansla ing cance genomes and ansc ip omes o p ecision oncology. CA: a cance jou nal
o clinicians 66, 75–88 (2016).
5. He, L., Long, L. R., An ani, S. & Thoma, G. R. His ology image analysis o ca cinoma de ec ion and g ading. Compu e me hods and
p og ams in biomedicine 107, 538–556 (2012).
6. Ko ha i, S., Phan, J. H., S okes, T. H. & Wang, M. D. Pa hology imaging in o ma ics o quan i a i e analysis o whole-slide images.
Jou nal o he Ame ican Medical In o ma ics Associa ion 20, 1099–1108 (2013).
7. 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. Jou nal o pa hology in o ma ics 7 (2016).
8. Diamond, J., Ande son, N. H., Ba els, P. H., Mon i oni, R. & Hamil on, P. W. The use o mo phological cha ac e is ics and ex u e
analysis in he iden i ica ion o issue composi ion in p os a ic neoplasia. Human Pa hology 35, 1121–1131 (2004).
9. Di C is o ano, A., Pesce, B., Co don-Ca do, C. & Pandol i, P. P. P en is essen ial o emb yonic de elopmen and umou supp ession.
Na u e gene ics 19, 348–355 (1998).
10. Ellwood-Yen, K. e al. Myc-d i en mu ine p os a e cance sha es molecula ea u es wi h human p os a e umo s. Cance cell 4,
223–238 (2003).
11. Maa en, L. . d. & Hin on, G. Visualizing da a using -sne. Jou nal o Machine Lea ning Resea ch 9, 2579–2605 (2008).
12. LeCun, Y., Bengio, Y. & Hin on, G. Deep lea ning. Na u e 521, 436–444 (2015).
13. Wang, D., Khosla, A., Ga geya, R., I shad, H. & Beck, A. H. Deep lea ning o iden i ying me as a ic b eas cance . a Xi p ep in
a Xi :1606.05718 (2016).
14. Chen, R., Jing, Y. & Jackson, H. Iden i ying me as ases in sen inel lymph nodes wi h deep con olu ional neu al ne wo ks. a Xi
p ep in a Xi :1608.01658 (2016).
15. Wang, P., Li, Z., Hou, Y. & Li, W. Combining con ne s wi h hand-c a ed ea u es o ac ion ecogni ion based on an hmm-s m
classi ie . a Xi p ep in a Xi :1602.00749 (2016).
16. Sadanandan, S. K., Rane all, P. & Wählby, C. Fea u e augmen ed deep neu al ne wo ks o segmen a ion o cells. In Eu opean
Con e ence on Compu e Vision, 231–243 (Sp inge , 2016).
17. G abowska, M. M. e al. Mouse models o p os a e cance : picking he bes model o he ques ion. Cance and Me as asis Re iews
33, 377–397 (2014).
18. Song, M. S., Salmena, L. & Pandol i, P. P. The unc ions and egula ion o he p en umou supp esso . Na u e e iews Molecula cell
biology 13, 283–296 (2012).
19. S ine, Z. E., Wal on, Z. E., Al man, B. J., Hsieh, A. L. & Dang, C. V. Myc, me abolism, and cance . Cance disco e y 5, 1024–1039
(2015).
20. C ose o, N., Bienko, M. & an Oudenaa den, A. Spa ially esol ed ansc ip omics and beyond. Na u e Re iews Gene ics 16, 57–66
(2015).
21. S åhl, P. L. e al. Visualiza ion and analysis o gene exp ession in issue sec ions by spa ial ansc ip omics. Science 353, 78–82 (2016).
22. Ka asalo, K., La onen, L., Visako pi, T., Nyk e , M. & Ruusu uo i, P. Benchma king o algo i hms o 3d issue econs uc ion. In
Image P ocessing (ICIP), 2016 IEEE In e na ional Con e ence on, 2360–2364 (IEEE, 2016).
23. Tuominen, V. J. & Isola, J. Linking whole-slide mic oscope images wi h dicom by using jpeg2000 in e ac i e p o ocol. Jou nal o
Digi al Imaging 23, 454–462 (2010).
24. Schneide , C. A., Rasband, W. S., Elicei i, K. W. e al. Nih image o imagej: 25 yea s o image analysis. Na me hods 9, 671–675 (2012).
25. O su, N. A h eshold selec ion me hod om g ay-le el his og ams. Au oma ica 11, 23–27 (1975).
26. Rui ok, A. C., Johns on, D. A. e al. Quan i ica ion o his ochemical s aining by colo decon olu ion. Analy ical and quan i a i e
cy ology and his ology 23, 291–299 (2001).
27. Ma as, J., Chum, O., U ban, M. & Pajdla, T. Robus wide-baseline s e eo om maximally s able ex emal egions. Image and ision
compu ing 22, 761–767 (2004).
28. Ojala, T., Pie ikäinen, M. & Mäenpää, T. G ay scale and o a ion in a ian ex u e classi ica ion wi h local bina y pa e ns. In
Compu e Vision-ECCV 2000, 404–420 (Sp inge , 2000).
29. 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 ion 33, 43–52
(2000).
30. Lowe, D. G. Dis inc i e image ea u es om scale-in a ian keypoin s. In e na ional jou nal o compu e ision 60, 91–110 (2004).
31. Vedaldi, A. & Fulke son, B. VLFea : An open and po able lib a y o compu e ision algo i hms. h p://www. l ea .o g/ (2008).
32. Be ens, P. e al. Ci cs a : a ma lab oolbox o ci cula s a is ics. J S a So w 31, 1–21 (2009).
33. Massey, F. J. Jou nal o he Ame ican S a is ical Associa ion 46, 68–78 (1951).
34. B eiman, L. Random o es s. Machine lea ning 45, 5–32 (2001).
Acknowledgemen s
We hank Ms. Päi i Ma ikainen, Ms. Ma ja Pi inen, and M s. Ma ika Vähä-Jaakkola o skill ul echnical
labo a o y assis ance. G an suppo has been ob ained om Finnish Funding Agency o Inno a ion (p ojec
Th 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, he Medical Resea ch Fund o Tampe e Uni e si y
Hospi al, Doc o al P og amme o Compu ing and Elec ical Enginee ing a Tampe e Uni e si y o Technology,
and Emil Aal onen Founda ion.