Data harmonisation for information fusion in digital healthcare: A state-of-the-art systematic review, meta-analysis and future research directions

In o ma ion Fusion 82 (2022) 99–122
A ailable online 24 Janua y 2022
1566-2535/© 2022 The Au ho s. Published by Else ie B.V. This is an open access a icle unde he CC BY license (h p://c ea i ecommons.o g/licenses/by/4.0/).
Da a ha monisa ion o in o ma ion usion in digi al heal hca e: A
s a e-o - he-a sys ema ic e iew, me a-analysis and u u e
esea ch di ec ions
Yang Nan
a
,
*
, Ja ie Del Se
d
,
e
, Simon Walsh
a
, Ca ola Sch¨
onlieb
, Michael Robe s
,
g
,
Ian Selby
h
, Ki Howa d
i
, John Owen
i
, Jon Ne ille
i
, Julien Guio
j
,
k
, Benoi E ns
j
,
k
, Ana Pas o
l
,
Angel Albe ich-Baya i
l
, Ma ion I. Menzel
m
,
n
, Sean Walsh
o
, Wim Vos
o
, Nina Fle in
o
,
Jean-Paul Cha bonnie
p
, E a an Rikxoo
p
, A ishek Cha e jee
q
, Hen y Wood u
q
,
Philippe Lambin
q
, Leono Ce d´
a-Albe ich
, Luis Ma í-Bonma í
, F ancisco He e a
s
,
,
#
,
Guang Yang
a
,
b
,
c
,
#
,
*
a
Na ional Hea and Lung Ins i u e, Impe ial College London, London, No he n I eland UK
b
Ca dio ascula Resea ch Cen e, Royal B omp on Hospi al, London, No he n I eland UK
c
School o Biomedical Enginee ing & Imaging Sciences, King’s College London, London, No he n I eland UK
d
Depa men o Communica ions Enginee ing, Uni e si y o he Basque Coun y UPV/EHU, Bilbao 48013, Spain
e
TECNALIA, Basque Resea ch and Technology Alliance (BRTA), De io 48160, Spain
Depa men o Applied Ma hema ics and Theo e ical Physics, Uni e si y o Camb idge, Camb idge, No he n I eland UK
g
Oncology R&D, As aZeneca, Camb idge, No he n I eland UK
h
Depa men o Radiology, Uni e si y o Camb idge, Camb idge, No he n I eland UK
i
Clinical Da a In e change S anda ds Conso ium, Aus in, TX, Uni ed S a es o Ame ica
j
Uni e si y Hospi al o Li`
ege (CHU Li`
ege), Respi a o y medicine depa men , Li`
ege, Belgium
k
Uni e si y o Liege, Depa men o clinical sciences, Pneumology-Alle gology, Li`
ege, Belgium
l
QUIBIM, Valencia, Spain
m
Technische Hochschule Ingols ad , Ingols ad , Ge many
n
GE Heal hca e GmbH, Munich, Ge many
o
Radiomics (Onco adiomics SA), Li`
ege, Belgium
p
Thi ona, Nijmegen, The Ne he lands
q
Depa men o P ecision Medicine, Maas ich Uni e si y, Maas ich , The Ne he lands
Medical Imaging Depa men , Hospi al Uni e si a i i Poli `
ecnic La Fe, Valencia, Spain
s
Depa men o Compu e Sciences and A i icial In elligence, Andalusian Resea ch Ins i u e in Da a Science and Compu a ional In elligence (DaSCI) Uni e si y o
G anada, G anada, Spain
Facul y o Compu ing and In o ma ion Technology, King Abdulaziz Uni e si y, Jeddah 21589, Saudi A abia
ARTICLE INFO
Keywo ds:
In o ma ion usion
da a ha monisa ion
da a s anda disa ion
domain adap a ion
ep oducibili y
ABSTRACT
Remo ing he bias and a iance o mul icen e da a has always been a challenge in la ge scale digi al heal hca e
s udies, which equi es he abili y o in eg a e clinical ea u es ex ac ed om da a acqui ed by di e en scanne s
and p o ocols o imp o e s abili y and obus ness. P e ious s udies ha e desc ibed a ious compu a ional ap-
p oaches o use single modali y mul icen e da ase s. Howe e , hese su eys a ely ocused on e alua ion me ics
and lacked a checklis o compu a ional da a ha monisa ion s udies. In his sys ema ic e iew, we summa ise he
compu a ional da a ha monisa ion app oaches o mul i-modali y da a in he digi al heal hca e ield, including
ha monisa ion s a egies and e alua ion me ics based on di e en heo ies. In addi ion, a comp ehensi e checklis
ha summa ises common p ac ices o da a ha monisa ion s udies is p oposed o guide esea che s o epo hei
esea ch indings mo e e ec i ely. Las bu no leas , lowcha s p esen ing possible ways o me hodology and
me ic selec ion a e p oposed and he limi a ions o di e en me hods ha e been su eyed o u u e esea ch.
* Co esponding au ho s.
E-mail add esses: [email p o ec ed] (Y. Nan), [email p o ec ed] (G. Yang).
#
F ancisco He e a and Guang Yang a e co-las au ho s o his wo k.
Con en s lis s a ailable a ScienceDi ec
In o ma ion Fusion
jou nal homepage: www.else ie .com/loca e/in us
h ps://doi.o g/10.1016/j.in us.2022.01.001
Recei ed 24 Oc obe 2021; Recei ed in e ised o m 22 Decembe 2021; Accep ed 7 Janua y 2022
In o ma ion Fusion 82 (2022) 99–122
100
1. In oduc ion
Compu a ional biomedical esea ch aims o ad ance digi al heal h-
ca e and biomedical s udies by de eloping compu a ional models ha
imp o e he p ecise diagnosis o disease spec um, analysis o gene ex-
p essions o ime se ies da a (e.g., elec oencephalog ams and elec o-
ca diog ams). These models a e designed o disco e no el isk
bioma ke s, p edic disease p og ession, design op imal ea men s, and
iden i y new d ug a ge s o applica ions such as cance , pulmona y
disease, and neu ological diso de s. Whils a well-pe o med model
should ha e cha ac e is ics o high pe o mance, obus ness, explain-
abili y, and ep oducibili y, i aces he issue ha he bias o dis ibu ion
be ween di e en da ase s d ama ically inc eases he di icul y o
de eloping models om la ge-scale s udies. Al hough da a ha mo-
nisa ion is needed wi h almos any kind o medical da a, au oma ed
me hods ha e been ex ensi ely used o medical images, gene exp es-
sion analysis, wi h he es o he modali ies being igno ed o ha mon-
ised manually. S udies ha e shown ha machine lea ning based
app oaches, especially deep neu al ne wo ks, a e highly sensi i e o he
dis ibu ion o aining da a. The e o e, he e is an u gen need o
de elop app oaches ha can in eg a e he de ice/si e-in a ian in o -
ma ion om mul iple da ase s. To add ess his issue, esea che s
es ablished s anda d acquisi ion p o ocols [1 3] o de ini ions [4,5] o
help da a collec o s o glean s anda dised da a. Fo ins ance, Delbeke
e al. [2] ecommended an acquisi ion p o ocol o F-FDG Posi on
emission omog aphy/compu e ised omog aphy imaging (PET/CT),
and Simon e al. [3] p esen ed a s anda dised MR imaging p o ocol o
mul i-scle osis. Schmid e al. [4,5] mainly ocused on in eg a ing he
da a om ou ine heal h in o ma ion sys ems, including conduc ing
manual ha monisa ion and ule-based alignmen o elec onic da a.
Al hough hese acquisi ion p o ocols could e ec i ely educe he coho
bias (non-biological a iances in c oss-scanne /si e da a), hey we e
limi ed in assis ing p ospec i e s udies because mos s udies we e
e ospec i e and could no be e-acqui ed wi h he same s anda d. In
addi ion, a non-s anda dised acquisi ion p o ocol is needed o pe son-
alised digi al heal hca e some imes. The e o e, i is impe a i e o
explo e a compu a ional me hod o ha monise mul icen e da ase s.
Al hough some su eys o compu a ional da a ha monisa ion ha e
been eleased [6,7,10], such as MRI (magne ic esonance imaging) [11]
o CT (compu e ised omog aphy) ha monisa ion, hese su eys only
explo ed me hods o single modali y o applica ion and a ely ocused
on e alua ion me ics and esea ch guidance (shown in Table 1).
Mo eo e , he e is a lack o a checklis ha can summa ise he common
p ac ice and gi e guidance o me hodology selec ion and de elopmen
o compu a ional da a ha monisa ion s udies. This su ey summa ises
he compu a ional da a ha monisa ion s a egies o mul imodal da a in
he digi al heal hca e ield in e ms o me hodologies, e alua ions, and
applica ions. Ou pape co e s h ee main a eas (i.e., gene exp ession,
adiomics, and pa hology), wi h o e 96 quali ied pape s published
wi hin wo decades. This is he la ges and he mos comp ehensi e
explo a ion o he compu a ional da a ha monisa ion s a egies o he
bes o ou knowledge. To p o ide a be e scien i ic p ac ice o he
communi y wo king on da a ha monisa ion, a comp ehensi e checklis
wi h all he s eps is p oposed o guide he esea che s on epo ing hei
s udies mo e e ec i ely. Wi h his checklis , explo a ions (wha he
s a egy is) and ad ances (how well he model pe o ms) o he s udy
can be clea ly illus a ed by epo ing he i ems in model and e alua ion
sec ions. O e all, he main con ibu ions o his su ey can be sum-
ma ised as:
•A h ee- old axonomy ha desc ibes he me hodology, e alua ion
and applica ions o compu a ional da a ha monisa ion s a egies.
•A checklis wi h all he s eps ha can be ollowed in u u e da a
ha monisa ion s udies.
•The c i ique and limi a ions o he exis ing da a ha monisa ion
s a egies and po en ial s udies.
The es o he manusc ip is o ganised as ollows: (1) Sec ion 2 de-
sc ibes he de ini ion, mo i a ion, u ilisa ion and solu ion o
compu a ional da a ha monisa ion issues; (2) Sec ion 3 illus a es
how his su ey is conduc ed; (3) Sec ions 4,5, and6 demons a e he
h ee- old axonomy o ha monisa ion s a egies; (4) Sec ion 7 de-
sc ibes he esul s o he me a-analysis and p esen s a checklis o
da a ha monisa ion s udies; (5) Sec ion 8 p esen s he checklis o
ha monisa ion s udies and summa ises he c i iques and limi a ions
o cu en s a egies; and (6) Sec ion 9 concludes his su ey.
2. Compu a ional da a ha monisa ion: de ini ion, o igin, wha
o and how?
This sec ion illus a es he de ails o da a ha monisa ion, including
he de ini ion, o igin, pu pose and solu ions o compu a ional da a
ha monisa ion asks. To be e desc ibe hese cha ac e is ics, he e -
minology o compu a ional da a ha monisa ion is illus a ed in Table 2.
2.1. Wha ?
Da a ha monisa ion e e s o combining he da a om di e en
sou ces o one cohesi e da a se by adjus ing da a o ma s, e minologies
and measu ing uni s [12]. I is mainly pe o med o add ess issues
caused by noniden ical anno a ions o eco ds o di e en ope a o s o
sys ems, which equi es a s anda d p o ocol o manual adjus men . The
con en ional app oach o da a ha monisa ion is pe o med by manu-
ally se ing ules o e ms o in eg a e mul icen e da ase s om heal h
Table 1
Compa ison o exis ing da a ha monisa ion e iew s udies.
Su ey [6] [7] [8] [9] Ou s
Pe iod ~2020 ~2019 ~2020 ~2021 ~2021
# o e iewed s udies N/A 23 49 42 96
Domain Radiomics Radiomics Radiomics Radiomics Radiomics, Gene, Pa hology
Me ic ××××√
Checklis ××××√
Guidance ×√ × × √
Me a-analysis ×√ × × √
“# o e iewed s udies” indica es he numbe o included pape s in he su ey.
Table 2
Te minology o compu a ional da a ha monisa ion.
Te minology De ini ions
Coho A g oup o da a acqui ed by he same acquisi ion p o ocol and
de ices
Subjec s Pa ien s (objec s) in ol ed in he s udy
Ca ego y The classes ha we e in ol ed in he s udy, e.g., cance s. no mal
Cases Samples (a subjec can p oduce mul iple samples wi h di e en
acquisi ion p o ocols) in ol ed in he s udy
Coho bias The non-biological ela ed a iances caused by acquisi ion
p o ocols (also named as “ba ch e ec ” in gene exp ession s udies)
Sou ce coho s The coho ha needs o be ha monised om
Re e ence
coho
The coho ha needs o be ha monised o
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
101
in o ma ion sys ems. I equi es complex mapping o e minologies and
manual ha monisa ions.
Di e en om manual ha monisa ion ha elies on a s anda d
p o ocol and manual adjus men , compu a ional da a ha monisa ion in
digi al heal hca e aims o educe he coho bias (non-biological a i-
ances) gi en by di e en da a acquisi ion schemes. I applies
Fig. 1. Visualised di e ences in (a) adiomics and (b) pa hology images. (a) a lung umou cap u ed on he same CT scanne wi h 6 di e en acquisi ion p o ocols
(F om [13]). (b) H&E s ained issue images om di e en si es [14].
Table 3
Summa y o he ep oducibili y/ epea abili y s udies.
Re e ence In a- ep o In e - ep o Repea abili y Condi ion Va iables Objec Modali y
Jha e al. [23], 2021 30.7%
(332/1080)
14.3%
(154/1080)
82.2% (888/1080) ICC >0.90 Slice Sickness Phan oms CT
Emaminejad e al. [24], 2021 8.0%
(18/226)
/ / CCC>0.90 Recons uc ion Pa ien s CT
7.5%
(17/226)
/ / CCC>0.90 Radia ion Dose Pa ien s CT
Kim e al. [25], 2021 11.0%
(112/1020)
/ / CCC>0.85 Accele a ion Fac o s Pa ien s MRI
Ymashi a e al. [26], 2020 / 5.6%
(15/266)
/ CCC>0.90 Di e en Scanne s Pa ien s CECT
Fise e al. [27], 2019 / 22.6%
(398/1761)
/ ICC >0.90 Di e en Scanne s Pa ien s MRI
Saeedi e al. [28], 2019 20.5%
(8/39)
/ / CoV<5% Tube Vol age Phan oms CT
30%
(13/39)
/ / CoV<5% Tube Cu en Phan oms CT
Meye e al. [29], 2019 20.8%
(22/106)
/ / R2>0.95 Radia ion Dose Pa ien s CT
52.8%
(56/106)
/ / R2>0.95 Recons uc ion Pa ien s CT
39.6%
(42/106)
/ / R2>0.95 Recons uc ion Pa ien s CT
12.3%
(13/106)
/ / R2>0.95 Slice Sickness Pa ien s CT
Pe in e al. [30], 2018 24.8%
(63/254)
/ / CCC>0.90 Injec ion Ra es Pa ien s CECT
13.4%
(34/254)
/ / CCC>0.90 Resolu ion Pa ien s CECT
Midya e al. [31], 2018 11.7%
(29/248)
/ / CCC>0.90 Tube Cu en Phan oms CT
19.8%
(49/248)
/ / CCC>0.90 Noise Phan oms CT
63.3%
(157/248)
/ / CCC>0.90 Recons uc ion Pa ien s CT
Al azi e al. [32], 2017 21.5%
(17/79)
/ / Mean di e ence <25% Recons uc ion Pa ien s PET
Zhao e al. [13], 2016 11.2%
(10/89)
/ / CCC>0.90 Recons uc ion Pa ien s CT
/ / 69.7%
(62/89)
CCC>0.90 / Pa ien s CT
Hu e al. [33], 2016 / / 64.0%
(496/775)
ICC>0.80 / Pa ien s CT
Choe e al. [34], 2019 15.2%
(107/702)
/ / CCC>0.85 Recons uc ion Pa ien s CT
CCC: conco dance co ela ion coe icien ; ICC: in aclass co ela ion coe icien ; CoV: coe icien o a ia ion; R2: R-squa ed; CT: compu ed omog aphy; MRI:
magne ic esonance imaging; CECT: consecu i e con as -enhanced compu ed omog aphy; PET: posi on emission omog aphy.
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
102
compu a ional s a egies (such as machine lea ning, image/signal p o-
cessing) o in eg a e mul icen e da ase s and educe hei non-
biological he e ogenei y. Compa ed wi h da a cleansing, da a no mal-
isa ion, s anda disa ion, e c., da a ha monisa ion has a b oade de ini-
ion and is a e m ha ep esen s he s a egies o educing coho biases
(caused by di e en acquisi ion p o ocols and de ices). I can be con-
duc ed by emo ing ou lie s (da a cleansing), aligning he loca ion-and-
scale pa ame e s o coho s (da a no malisa ion), con e ing mul iple
da ase s in o a common da a o ma (da a s anda disa ion/ ans-
o ma ion, e e ing o manual ha monisa ion). I is o no e ha da a
ha monisa ion is no as same as s yle ans o ma ion (e.g., gene a ing T1
images using T2 in MRI, o gene a ing CT using X-Ray images), i only
ocuses on in a-modali y da ase s.
2.2. Why?
This sec ion i s illus a es he mo i a ion o compu a ional da a
ha monisa ion app oaches, hen desc ibes he sou ce o non-biological
a iances. Compu a ional me hods e e o he au oma ic analysis o
digi al heal hca e da a, using machine lea ning o ma hema ical
modelling algo i hms. I usually equi es he ex ac ion and usion o
da a-de i ed ea u es om he aw da a. Fo ins ance, he g ey le el co-
occu ence ma ix (GLCM), which is one o he mos commonly used
ex ual ea u es in adiomics, can be used as an independen p ognos ic
ac o ( ep esen ing in F-FDG PET/CT images he me abolic in a-
umo al he e ogenei y) in pa ien s wi h su gically ea ed ec al can-
ce [15]. Howe e , da ase s acqui ed om di e en si es p esen sig-
ni ican a iances (Fig. 1), which can hinde he e ec i eness o
ex ac ed ea u es and lead o uns able pe o mance o bo h compu a-
ional and manual diagnosis. In pa icula , Zhao e al. [13] ound a
conside able segmen a ion based inconsis ency o lung umou s while
conduc ing epea ed manual labelling by h ee adiologis s. This
inconsis ency could lead o a signi ican educ ion ( om 0.76 o 0.28) o
conco dance co ela ion coe icien s o ce ain adiomics ea u es.
The e o e, compu a ional da a ha monisa ion is p oposed o elimina e
o educe hese non-biological a iances in mul icen e da ase s o (1)
enhancing he obus ness and ep oducibili y o compu a ional mod-
ules; (2) p oducing he usion o knowledge cap u ed be o ehand wi h
knowledge cap u ed o e a new ask; (3) p omo ing he comp ehensi e
pe o mance o compu a ional modules.
The non-biological da a a iances a e mainly om ha dwa e (e.g.,
scanne s and pla o ms), acquisi ion p o ocols (e.g., signal/imaging
acquisi ion pa ame e s) and labo a o y p epa a ions (e.g., s aining and
slicing). These a iances may lead o he weak ep oducibili y o
quan i a i e bioma ke s and limi he ime-se ies s udies based on mul i-
sou ce da ase s, indica ing an u gen need o da a ha monisa ion s a-
egies o gene a e ep oducible ea u es [15,16].
2.2.1. He e ogenei y o acquisi ion de ices (in e -de ice a iabili y)
He e ogenei y o acquisi ion de ices leads o he a iance o mul i-
cen e da a, which is mainly disco e ed in signals, CT, MRI, and pa h-
ological images. This he e ogenei y is mainly b ough by di e en
de ec o sys ems o endo s, he sensi i i y o he coils, posi ional and
physiologic a ia ions du ing acquisi ion, and magne ic ield a ia ions
in MRI, amongs o he s. [17–20]. S udies ha e shown ha e en using a
ixed acquisi ion p o ocol o di e en b ands o scanne s, some adio-
mics ea u es a e s ill non- ep oducible. Fo ins ance, Be engue e al.
[21] explo ed he ep oducibili y o adiomics ea u es on i e di e en
scanne s wi h he same acquisi ion p o ocol and wi nessed la ge di -
e ences, anging om 16% o 85% o he adiomics ea u es we e
ep oducible. Sunde land e al. [22] explo ed he la ge a iance o
s anda d up ake alue (SUV) in di e en b ands o scanne s, wi nessing
a much highe maximum SUV o newe scanne s compa ed wi h old
ones.
2.2.2. He e ogenei y o acquisi ion p o ocols (in a-de ice a iabili y)
The di e en acquisi ion p o ocols a e he main easons o c oss-
coho a iabili y. They mainly include he scanning pa ame e s (e.g.,
ol age, ube cu en , he ield o iew, slice hickness, mic ons pe
pixel, e c.) and econs uc ion app oaches (e.g., di e en econs uc ion
ke nels) [35]. To in es iga e he in a/in e ep oducibili y o adiomics
ea u es, se e al s udies ha e been conduc ed by es - ese expe imen s
(Table 3). In Table 3, a good ep oducibili y/ epea abili y is de ined as
he high co ela ion coe icien (e.g., ICC, CCC, R2) o low di e ence (e.
g., mean di e ence, CoV) be ween wo ea u es. Fo ins ance, a ce ain
adiomics ea u e is conside ed ep oducible/ epea able when he CCC
be ween ea u es ex ac ed om wo epea ed scans is la ge han 0.90.
As shown in Table 3, he scanning pa ame e s no ably a ec he adio-
mics ea u es, making he s a is ical analysis di icul . Fo ins ance, only
15.2% o adiomics ea u es a e ep oducible when using so and sha p
ke nels du ing he econs uc ion [34]. This weak ep oducibili y
g ea ly hinde s he la ge-scale digi al heal hca e s udies and applica-
ions o compu a ional models. Al hough implemen ing s ic s anda d
p o ocol can educe non-biomedical a iances, he non-s anda d
acquisi ion p o ocol is needed by physicians o pe sonalised
cen e-based image quali y conside a ions. Fo ins ance, he hickness
and pixel size a e egula ly adjus ed on a case-by-case p inciple o
imp o e he da a quali y [36]. The e o e, he he e ogenei y o acquisi-
ion p o ocol is una oidable which equi es a gene al solu ion.
2.2.3. He e ogenei y o labo a o y p epa a ions (P epa a ion a iabili y)
All he gene exp ession, adiomics, and pa hological da a hea ily
su e om labo a o y a iances, including sample p epa a ion, assay,
slicing, and s aining. Fo single-cell RNA sequencing (scRNA-seq) and
mic oa ay da a, he e a e a ious analysis pla o ms wi h di e en
biases, making i di icul o in eg a e and compa e esul s om mul i-
cen e/ba ch o da a [37,38]. Fo adiomics da a, a iances such as in-
jec ion a e and adia ion dose may also a ec he da a quali y.
Conside ing he pa hology da a, a iances a e mainly om manual op-
e a ions [39,40] (e.g., biopsy sec ioning, sample ixa ion, dehyd a ion
and s ain concen a ion), all hese ac o s esul in he a ia ion o pixel
alues and s ain consis encies.
2.3. Wha o ?
La ge scale and longi udinal s udies. The challenges o in eg a ing
and u ilising mul icen e da ase s make esea che s ealise he impo -
ance o da a ha monisa ion when conduc ing la ge-scale s udies [41].
On he one hand, he in o ma ion usion wi hou ha monisa ion canno
achie e ep oducible esul s in la ge scale and longi udinal s udies [13,
31,42]. Some esea che s ha e ad ised ha he conclusions eached
mus be ea ed wi h cau ion since some ea u es can a y g ea ly
agains mino non-biomedical changes [43]. Da a ha monisa ion, on he
o he hand, is c i ical o pa ien s who a e moni o ed longi udinally and
imaged on di e en scanne s. Fo ins ance, he longi udinal PET canno
p o ide help ul in o ma ion i hey a e ga he ed om mul i-scanne s,
since he ela ionship be ween SUV and ou comes may ge concealed
[16].
T ans e abili y o compu a ional models. The uns able pe o -
mance has been ound when applying compu a ional models o mul i-
cen e da ase s [44]. To add ess his issue, ans e lea ning was
p oposed o enhance he obus ness o compu a ional models by holding
a p io i knowledge on he way da a can a y. I eeds he model wi h
u he da a which e lec s he a iabili y ha he model may encoun e
a in e ence ime. Howe e , ans e lea ning equi es ex a aining
samples o educe he unce ain y wi h espec o he a iabili y o da a
ha models can cope wi h. This could be inapplicable o p ospec i e
s udies in he digi al heal hca e ield. Di e en om ans e lea ning,
compu a ional da a ha monisa ion s a egies can p ocess he da a
wi hou ex a aining o ine- uning, which p o ide an applicable so-
lu ion o mul icen e s udies. Meanwhile, he e has been moun ing
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
103
e idence ha combining da a ha monisa ion wi h machine lea ning
algo i hms enables obus and accu a e p edic ions on mul icen e
da ase s [45].
2.4. How?
The deploymen o a compu a ional me hod includes p epa a ion
(acqui ing da ase s such as s aining, scanning), p e-p ocessing, model-
ling and analysing, while he da a ha monisa ion can be pe o med
h ough he p ocessing o images/signals/gene ma ices (i.e., sample-
wise) o alignmen o da a-de i ed ea u es (i.e., ea u e-wise). The
sample-wise ha monisa ion is usually conduc ed be o e modelling,
aiming o educe he coho a iance o all aining samples and use
mul icen e samples as a single da ase . I in ol es image p ocessing,
syn hesis and in a ian ea u e lea ning app oaches. A e acqui ing
coho -in a ian da a, a single model can be de eloped o clinical
ela ed asks. The ea u e-wise ha monisa ion aims o educe he bias o
ex ac ed ea u es, such as he GLCM, con ex hull a ea o he egion o
in e es . I is usually pe o med on ex ac ed ea u e ma ixes, elimi-
na ing he coho a iances h ough using he ex ac ed ea u es (shown
as he le bo om sub igu e Fig. 2, he ed and blue do s indica e samples
om di e en coho s). Bo h he sample-wise and ea u e-wise da a
ha monisa ion can e ec i ely educe he a iances and imp o e he
pe o mance o he analysis. Howe e , he ea u e-wise ha monisa ion
equi es se e al models o ex ac ea u es o in e es , leading o com-
plex model de elopmen . Mo eo e , when he numbe o samples in
each coho is small, i is ha d o de elop he co esponding models.
3. Me hods
3.1. Li e a u e sea ch and e iew
The li e a u e sea ch, selec ion and eco ding we e conduc ed
independen ly by wo esea che s wi h expe ience in compu e science
and biomedicine. The ag eemen was hen achie ed by a hi d e iewe
wi h he expe ise o biomedical da a analysis. All hese sea ches we e
pe o med on Scopus P e iew (Else ie ) da abase o publica ions up o
July 10, 2021. To in es iga e he s a egies o ha monisa ion o in o -
ma ion usion, we sea ched he li e a u e using he keywo d o ’ba ch
e ec emo al’, ’deep lea ning’ and ‘ha monisa ion’, ‘da a
Fig. 2. Wo k low o de eloping a compu a ional da a ha monisa ion me hod.
Fig. 3. Li e a u e selec ion p ocedu e.
Y. Nan e al.

In o ma ion Fusion 82 (2022) 99–122
104
ha monisa ion’, ‘no malisa ion’ and ‘ha monisa ion’, ‘colou no mal-
isa ion’, ‘ ep oducibili y’ and ‘ adiomics’, ‘image s anda disa ion’.
These ini ial keywo ds we e sea ched bo h independen ly and join ly o
co e mo e li e a u e. I is o no e ha bo h ‘no malisa ion’ and
‘s anda disa ion’ a e me hods o ha monisa ion. The p e-sc eening was
i s conduc ed by iewing he abs ac and i le o il e hose i ele an
a icles. The eligibili y was hen checked h ough ou c i e ia (gi en in
Sec ion 3.2) o emo e he unquali ied wo ks o ull- ex e iew.
A lowcha demons a ing he li e a u e selec ion p ocedu e is
p esen ed in Fig. 3. A e emo ing he i ele an and duplica ed a icles
by sc eening he i les and abs ac s, 238 a icles we e selec ed o ull-
ex sc eening. Based on eligibili y c i e ia, 139 publica ions we e
conside ed unquali ied, and 96 pape s we e included in his sys ema ic
e iew.
3.2. Inclusion and exclusion c i e ia
The en y c i e ia we e: (1) o iginal esea ch publica ions in pee -
e iewed jou nals o in e na ional con e ences; (2) ocus on he
compu a ional da a ha monisa ion o digi al heal hca e da a. The
excluded c i e ia we e: (1) s udies ha only applied exis ing ha mo-
nisa ion s a egies wi hou u he de elopmen ; (2) s udies ha ocused
on manual ha monisa ion such as egula ions; (3) e iew and li e a u e
su ey s udies; (4) s udies ha only explo e he ep oducibili y o s a-
bili y wi hou de eloping ha monisa ion app oaches.
3.3. Da a collec ion
De ails o pape s o quali y e iew we e manually summa ised in a
sp eadshee , including i le, modali y, me hodology, me ics, da a scale,
yea o publica ion, da a p ope y (e.g., p i a e o public), applica ions,
numbe o coho s, and numbe o cases.
4. Da a ha monisa ion s a egies o in o ma ion usion
In his sys ema ic e iew, da a ha monisa ion app oaches we e
di ided in o ou g oups, wi h he dis ibu ion based me hods, image
p ocessing, syn hesis, and in a ian ea u e lea ning. To be e illus a e
he basic idea and ela ionship o compu a ional app oaches, a axon-
omy is shown in Fig. 4, ollowed by a de ailed desc ip ion o ha mo-
nisa ion echniques.
4.1. Dis ibu ion based me hods
The dis ibu ion based me hods es ima e/calcula e he bias be ween
coho s om he la en space, hen ma ch/map he sou ce da a o he
a ge ones h ough a bias co ec ion ec o o alignmen unc ions.
4.1.1. Loca ion-scale me hods (LS)
The loca ion-scale me hods es ima e he loca ion-scale pa ame e s
(mean and a iance) o each coho and align all da a owa ds he same
loca ion-scale.
ComBa : ComBa [46] obus ly es ima ed bo h he mean and he
a iance o each ba ch using empi ical Bayes sh inkage, hen ha mon-
ised he da a acco ding o hese es ima es. The da a was i s s and-
a dised o ha e simila o e all mean and a iance, ollowed by he
empi ical Bayes es ima ion ia pa ame ic empi ical p io s. Wi h hese
adjus ed bias es ima o s, he da a could be ha monised by he
loca ion-scale model based unc ions [47 66]. Fo ins ance, Radua e al.
applied ComBa o add ess he he e ogenei y o co ical hickness, su -
ace a ea and subco ical olumes caused by a ious scanne s and se-
quences [53]. Whi ney e al. implemen ed ComBa o ha monise he
adiomic ea u es ex ac ed ac oss mul icen e DCE-MRI da ase s [54].
ComBa -seq: Resea che s ha e made mo e ex ensions based on he
o iginal ComBa ha monisa ion. Since he assump ion o Gaussian dis-
ibu ion in he o iginal ComBa made i sensi i e o ou lie s, Zhang
e al. p oposed ComBa -seq [67] by assuming he Nega i e Binomial
dis ibu ion, which could be e add ess he ou lie issues. The
Fig. 4. Taxonomy o compu a ional da a ha monisa ion s a egies.
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
105
ComBa -seq i s buil a nega i e binomial eg ession model and ob-
ained he es ima o s o coho bias, ollowed by he calcula ion o ‘ba ch
ee’ dis ibu ions o mapping o iginal da a.
BM-ComBa : Di e en om he o iginal ComBa ha shi ed sam-
ples o he o e all mean and a iance, an M-ComBa [68] was p oposed
o p o ide a lexible solu ion, ans e ing he da a o he loca ion and
scale o a p e-de ined “ e e ence”. Wi h hese e o s, Da-ano e al. [69]
p oposed a BM-ComBa by in oducing a pa ame ic boo s ap in
M-ComBa o obus es ima ion, aiming o p o ide a mo e lexible and
obus ha monisa ion s a egy.
QN
–
ComBa : Mülle e al. [70] applied a quan ile no malisa ion
be o e ComBa co ec ion in longi udinal gene exp ession da a o ach-
ie e be e pe o mance.
Dis ance-Weigh ed Disc imina ion (DWD): DWD [71] sea ched
he hype plane whe e he samples could be well sepa a ed and p ojec ed
he di e en ba ches on he DWD plane. The da a was hen ha monised
by sub ac ing he DWD plane mul iplied by he ba ch mean. I is o no e
ha DWD epea ed he ansla ions o samples om di e en coho s
un il hei ec o s we e o e lapped.
4.1.2. I e a i e clus e ing me hods (IC)
The i e a i e clus e ing me hods ha monise he coho bias by con-
duc ing mul iple bias co ec ion h ough epea ed clus e ings p oced-
u es. These me hods usually (1) pe o m clus e o all samples om
di e en coho s, and (2) compu e he co ec ion ec o s o ha mo-
nisa ion based on clus e cen oids.
C oss-pla o m no malisa ion (XPN): XPN [72] ook he combined
s anda dised sample and median cen al gene as inpu o emo e g oss
sys ema ic di e ences, ollowed by he clus e s, aiming o iden i y ho-
mogenous g oups o genes and samples wi h simila exp essions in
combined da a. The gene clus e s we e hen acqui ed by assignmen
unc ion, which was used o compu e es ima ed model pa ame e s ia
s anda d maximum likelihood.
Ha mony: Ha mony [73] i s employed p incipal componen s
analysis (PCA) o educe he dimension o all samples, and classi ied
hem in o se e al g oups (one cen oid pe g oup) h ough k-means
clus e ing. Wi h hese cen oids, he co ec ion ac o s o ha mo-
nisa ion we e calcula ed. The abo e clus e ing and co ec ion we e
epea ed un il he con e gence.
4.1.3. Nea es neighbou s me hods (NNM)
NNM me hods i s ound he mu ual nea es pai s, hen compu ed
he bias co ec ion ec o s based on pai ed samples and sub ac ed hese
ec o s om he sou ce coho . Di e ences in hese me hods mainly
e e o he geome y space when loca ing he mu ual nea es pai s.
Mu ual nea es neighbou s (MNN): MNN iden i ied nea es
neighbou s be ween di e en coho s and ea ed hem as ancho s o
calcula e he coho bias [74]. I i s p e-no malised he gene da a wi h
cosine no malisa ion, ollowed by he es ima ion o he bias co ec ion
ec o by compu ing he Euclidean dis ances be ween pai ed samples.
The bias co ec ion ec o was hen applied o all samples ins ead o he
pa icipa ed pai s. I equi ed ha all pa icipa ed ba ches mus sha e a
leas one common ype wi h ano he .
Scano ama: Simila o he MNN me hod, pano ama s i ching (Sca-
no ama) [75] aims a es ima ing coho bias om samples ac oss
ba ches. I i s educed dimensions o aw da a (o sou ce da a) using
singula alue decomposi ion (SVD). Then an app oxima e nea es
neighbou was adop ed o ind he mu ually linked samples ac oss co-
ho s. Di e en om MNN, Scano ama checked he p io i y o da ase
me ging wi hin all ba ches and acqui ed he me ged pano ama based on
he weigh ed a e age o ba ch co ec ion ec o s. A las , he ha mo-
nisa ion was pe o med wi h Scanpy [76] wo k low.
Ba ch balanced k-nea es neighbou s (BBKNN): Ini ially, BBKNN
[77] ound he nea es neighbou s in a p incipal componen space based
on Euclidean dis ances. Then i buil a g aph ha linked all he samples
ac oss coho s based on he neighbou in o ma ion. These neighbou
se s we e hen ha monised by uni o m mani old app oxima ion (UMAP)
[78] algo i hms.
S anda d CCA and mul i-CCA (Seu a ): Di e en om o he NNM-
me hods, Seu a [79] pe o med canonical co ela ion analysis o ac-
qui e he canonical co ela ion ec o s ha could p ojec mul i-da ase s
in o he mos co ela ed subspace. In his subspace, he mu ual nea es
pai s we e loca ed o compu e he bias co ec ion ec o s o guide he
da a in eg a ion. When p ocessing mul i-coho da ase s (numbe o
coho s la ge han wo), he i s ba ch would be se as he e e ence
ba ch o he co ec ion o he second ba ch. Then he ha monised sec-
ond ba ch would be appended o he e e ence ba ch. This epea ed
p ocedu e s opped when all he ba ches a e ha monised [38, 79].
4.1.4. Remo e unwan ed a ia ions (RUV)
These me hods assumed ha he coho bias was independen o
hose biases e e o biological a iances, which could be es ima ed as
“unwan ed a ia ions”. Fo ins ance, he bias o nega i e con ol genes
(p io known genes ha would no be a ec ed by biological changes o
in e es ) could be ega ded as coho bias. Based on his assump ion, he
aw da a could be ha monised by sub ac ing hose “unwan ed
a ia ions”.
Remo e unwan ed a ia ions, 2-s ep (RUV-2): Con ol a iables
we e used by RUV-2 o disco e he ac o s ela ed o coho bias [80].
The nega i e con ol (p obes ha should ne e be exp essed in any
sample) samples we e subjec ed o componen analysis, and he esul -
ing ac o s we e inco po a ed in o a linea eg ession model. Va ia ions
in he exp ession le els o hese genes hus we e conside ed undesi able.
To ex ac low-dimensional ea u es, Risso e al. [81] p esen ed an
ex ension o he RUV-2 wi h a ze o-in la ed nega i e binomial model
ha accoun ed o d opou s, disc e isa ion, and he coun cha ac e o
he da a. The coho bias was hen sub ac ed om he aw da a o
gene a e a gene exp ession ma ix ha is ha monised.
Singula alue decomposi ion ha monic (SVDH): By ac o ising
he exp ession ma ix o inpu da a and econs uc ing i while aking o
he elemen s ela ed o he coho bias, singula alue decomposi ion
(SVD) could be used o educe coho bias. Al e e al. [82] sugges ed
using SVD o ha monise he da a by il e ing away he eigena ays ha
lead o noise o expe imen al a e ac s.
scMe ge: scMe ge [83] i s cons uc ed a g aph ha connec ed
clus e ings be ween coho s by sea ching o mu ual nea es neighbou s.
The unwan ed ac o s we e hen es ima ed using s ably exp essed genes
as nega i e con ols. A las , an RUV model was used o collec and
emo e unwan ed di e ences be ween coho s.
Su oga e a iable analysis (SVA): SVA [84] aimed o ecognise
and es ima e he unwan ed a ia ions o da a om mul iple coho s. I
could be pe o med wi hou any coho in o ma ion. The mixed da ase
was i s di ided in o a collec ion o n su oga e a iables ia SVD, ol-
lowed by he clea ance o da a wi h la ge a iances. SVA coe icien s
we e hen calcula ed o ha monisa ion by using a linea eg ession
unc ion wi h su oga e a iables and aw di usion in ensi ies.
P in - ip loess no malisa ion (PLN): PLN [85] was ini ially p o-
posed o deal wi h mic oa ay da a. To elimina e he coho bias, PLN
employed a blocking e m o cons uc a linea model wi h he inpu
da a. The coho bias was sub ac ed om he o iginal da a o p oduce
he ba ch co ec ed exp ession ma ix.
Remo al o a i icial oxel e ec by linea eg ession (RAVEL):
RAVEL [86] sepa a ed he oxel alue in o unwan ed a ia ion pa s and
biological pa s. The unwan ed a ia ion ac o s we e es ima ed om
he egion o in e es by SVD, based on he p io knowledge o oxel
alues, which we e no ela ed o disease s a us.
4.1.6. Sphe ical ha monics (SH)
Sphe ical ha monics app oaches we e designed o ha monise MRI
da a, aiming o coo dina e all da a om di e en coho s o he same
sphe ical ha monic domain, by adjus ing he sphe ical a iables.
Ro a ion in a ian sphe ical ha monics (RISH): RISH was based
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
106
on mapping di usion-weigh ed imaging da a om sou ce coho s o
a ge coho s [17,66,87,88]. I s a ed wi h calcula ing he
o a ion-in a ian ea u es om he es ima ed sphe ical ha monics co-
e icien s (o a ge and sou ce samples, espec i ely). These o a ion
in a ian ea u es we e hen mapped om he sou ce coho s o a ge
coho s h ough egion-speci ic linea mapping, ollowed by he
upda ing o sphe ical ha monics coe icien s. The ha monised di usion
signal was calcula ed o each subjec in sou ce coho s using he la es
sphe ical ha monics coe icien s in a ge coho s o g adien di ec ions.
Sphe ical momen ha monics. Due o he insu icien adjus men
by loca ion-scale pa ame e s in some cases, esea che s p oposed he
sphe ical momen me hod (SMM), which u ilised he sphe ical momen s
o map he di usion-weigh ed images om sou ce coho s o e e ence
coho s [89,90]. SMM ma ches he sphe ical mean (M1) and sphe ical
a iance (C2) pe b- alue ( he di usion weigh ing) by M1[Tb] =
M1[ (Sb)] and C2[Tb] = C2[ (Sb)], whe e Tb,Sb a e da a om he a ge
and sou ce coho s unde b shell, espec i ely. The mapping pa ame e s
o ha monising da a om di e en coho s we e acqui ed by he linea
ans o m .
4.1.7. Dis ibu ion alignmen (DA)
Dis ibu ion alignmen me hods aim o ans o m he dis ibu ion o
he sou ce coho o ha o he e e ence coho , using cumula i e dis-
ibu ion unc ions o p obabili y densi y unc ions.
Cumula i e dis ibu ion unc ions alignmen (CDFA): CDFA [91]
was i s p oposed o mul isi e MRI da a ha monisa ion, which aligned
he sou ce oxel in ensi ies h ough an es ima ed non-linea in ensi y
ans o ma ion o ma ch he a ge cumula i e dis ibu ion unc ions.
The es ima ed in ensi y ans o ma ion de ined a one- o-one mapping
be ween he oxels in sou ce and a ge coho s.
Gamma cumula i e dis ibu ion unc ions alignmen (GCDF):
The oxel in ensi ies we e e-pa ame e ised using a mix u e model o
wo Gamma dis ibu ions ha i ed a e e ence his og am [92]. This
epa ame e isa ion was based on he CDF o he Gamma componen ,
which modelled he pa icula up ake, and cons ained he new ea u e
space o [0, 1].
P obabili y densi y unc ion ma ching: GENESHIFT [93] es i-
ma ed he empi ical densi y and measu ed he dis ance be ween p ob-
abili y densi y unc ions. GENESHIFT i s picked he common genes
om di e en coho s, hen es ima ed hei p obabili y densi y unc-
ions o ind he bes ma ching o se s. The ha monised da a would be
acqui ed by sub ac ing he es ima ed o se s om he sou ce coho s.
4.2. Image p ocessing
Image P ocessing employs digi al image p ocessing algo i hms o
ha monise mul i-coho da a, including image il e ing (also called
image con olu ion), egis a ion, esampling and no malisa ion.
4.2.1. Image il e ing (IF)
Image il e ing (also called con olu ion) is he p ocess ha mul iplies
wo a ays o p oduce a new a ay o he same dimension. The 2D
second-o de Bu e wo h low-pass il e was ound o be able o elim-
ina e coho bias be ween CT images wi h di e en oxel sizes [94],
while he local bina y pa e n il e ing could p oduce s able and
ep oducible adiomic ea u es [95].
4.2.2. Physical-size esampling (Resample)
S udies ha e shown ha physical size such as pixel/ oxel size, mpp
(mic ons pe pixel o le el 0 in digi al pa hology) can g ea ly a ec he
adiomic/pa hological ea u es. This bias can be educed using bilinea
esampling o equalise all he physical sizes [94].
4.2.3. S anda disa ion/no malisa ion (SN)
S anda disa ion/no malisa ion models we e designed o educe he
a ia ion and in e - a iabili y in di e en coho s by linea ans o m.
These me hods usually pe o med loca ion-scale shi s in image spaces
(e.g., HSV, RGB,
α
β, illumina ion spaces, e c.) o image his og ams.
Global colou no malisa ion (GCN) ans e s he colou s a is ics
om he sou ce o he a ge images by globally al e ing he image
his og am [96,97]. A ypical ep esen a i e o GCN is Z-sco e no mal-
isa ion, assumed he a iable om coho i, subjec j as Xij, z-sco e
no malisa ion is conduc ed h ough
Xij =Xij −
μ
i
σ
i
(1)
whe e
μ
i and
σ
i a e he mean and s anda d de ia ion o each coho .
Howe e , his global alignmen may lose some in o ma ion.
Local colou no malisa ion (LCN) ans e s he colou s a is ics o
he speci ic egions, e.g., igno ing he backg ound egions, om sou ce
o a ge images. In [98], he au ho s i s con e ed he sou ce and
a ge images om he RGB in o he l
α
β space, and hen conduc ed a
ans o ma ion o ha monise he sou ce image and e-con e ed i in o
he RGB space. I is o no e ha he luminance o backg ound egions is
no in ol ed du ing he p ocessing. This helped he ans o ma ion o
p ese e in ensi y in o ma ion wi hin he egion o in e es while
equi ing he p e-de ini ion o ce ain egions.
His og am ma ching (HM): HM is a me hod o con as adjus men
using he his og am o images [99]. I adjus s he dis ibu ion o images
by scaling he pixel alues o i he ange o speci ied his og am (i.e., he
a ge one):
(x,y) = ITmax −ITmin
ISmax −ISmin (IS−ISmin)+ITmin (2)
whe e IT indica es he a ge image and IS is he sou ce image. Gene ally,
ITmax and ITmin a e 0 and 255, espec i ely. Fo ins ance, Shah e al. [100]
in es iga ed he his og am no malisa ion on MRI images o ha monise
c oss-coho da a o mul iple scle osis lesion iden i ica ion.
Fuzzy based Reinha d colou no malisa ion (FRCN): To dec ease
he colou a ia ion, Roy e al. [101] applied uzzy logic o egula e he
con as enhancemen in l space o adjus he colou coe icien s wi hin
he
α
β space.
Ca ego y based colou no malisa ion (Ca ego yCN): To educe
he a iance o global colou no malisa ion, esea che s p oposed a
ca ego y based app oach o accu a e colou no malisa ion [102]. Ca -
ego yCN i s classi ied each pixel by unsupe ised app oaches om he
sou ce and a ge images, hen conduc ed colou no malisa ion based on
he di e en classes.
Comple e colou no malisa ion (CCN): The comple e colou no -
malisa ion included he no malisa ion o illumina ion and spec um, one
o ha monise he illuminan du ing imaging and ano he o educe
spec al a ia ion [39,103]. CCN es ima ed he illuminan and spec al
ma ices om he a ge coho , hen ma ched he sou ce illuminan and
spec al es ima ions o he a ge ones.
4.2.4. S ain sepa a ion me hods (SS)
S ain sepa a ion app oaches sepa a ed he inpu images in o dis inc
channels (e.g., he haema oxylin channel, eosin channel, and he back-
g ound channel o H&E-s ained images) o e alua e he s ain ea u e
ma ix and ma ch hese ea u es h ough ce ain ope a ions om sou ce
o a ge coho da a. The co e concep o s ain sepa a ion was based on
Lambe Bee ’s law [104] (in he RGB space, s ain concen a ions a e
nonlinea ly dependan ), shown as
IC=I0e−ODc(3)
whe e I0 was he alue o inciden ligh , and ODc was he alue o images
in op ical densi y (OD) space. Mos s ain sepa a ion me hods aimed o
ac o ise he OD alues in o wo ma ices as
ODc=log(I0
IC)=S∗D(4)
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
107
whe e S was he s ain dep h ma ix and D was he s ain colou appea -
ance (SCA) ma ix.
Colou decon olu ion (CD): These app oaches es ima ed he con-
cen a ion o s ains in pixel alues and no malised he spec al a ia ion
in sepa a ed s ains [105 108]. Fo example, es ima ion o he s ain
ma ix was i s gi en by e alua ing he p opo ion o RGB channels
wi hin di e en coho s, ollowed by colou decon olu ion [106,107].
The in e se o he s aining appea ance ma ix was mul iplied wi h he
op ical densi y space in ensi y alue o ge no malised s ain channels
using non-linea spline mapping.
S uc u ed-p ese ing colou no malisa ion (SPCN): SPCN
assumed ha mos issue egions we e cha ac e ised by he mos e ec-
i e s ain amongs he used s ains [109]. I i s con e ed a gi en RGB
image o op ical densi y using he Bee -Lambe Law. A e ha , SPCN
decomposed images in o se e al s ain densi y maps using spa se and
non-nega i e ma ix ac o iza ion (SNMF), ollowed by he combina ion
o he s ain densi y map and colou no malisa ion.
S ainCNNs: Inspi ed by SPCN, Lei e al. p oposed a deep neu al
ne wo k o s ain sepa a ion o educe he compu a ional consump ion
o SNMF [110]. The p oposed s ainCNNs app oach ook he sou ce im-
ages as inpu and lea ned o gene a e he s ain colou appea ance ma-
ix. I signi ican ly educed he p ocessing ime while e aining he high
quali y o he ha monised images.
Adap i e colou decon olu ion (ACD): ACD i s ans e ed he
inpu RGB images o op ical densi y space, hen pe o med s ain sepa-
a ion wi h adap i e colou decon olu ion ma ix o ob ain he hae-
ma oxylin (H) channel, eosin (E) channel and esidual channel [111]. A
las , he ha monised images we e ob ained h ough ecombining he H
and E componen s wi h a s ain colou appea ance ma ix o a ge
coho s.
Rough- uzzy ci cula clus e ing based s ain sepa a ion (RCCSS):
In RCCSS, s ain sepa a ion was ca ied ou using an image model based
on ansmission ligh mic oscopy [112]. Ini ially, each image was
ans e ed o OD space and hen decomposed o ob ain he SCA ma ix
and associa ed s ain dep h ma ix. Maji e al. [113] p esen ed a ci cula
clus e ing algo i hm o ind he ‘cen oid’, ‘a c isp lowe app oxima-
ion’, and he ‘ uzzy bounda y’, which could be in eg a ed by
sa u a ion-weigh ed hue his og am in he HIS colou space.
4.3. Syn hesis
The objec i e o syn hesis is o p ecisely ep oduce a sample ha
belongs o a missing modali y o domain, which ha monises he mul i-
coho da ase s. I elaxes ha monisa ion asks as s yle ans e and
conside s each coho as a ‘s yle’ and ans e s all samples o he same
‘s yle’. Based on he cha ac e is ics o he aining sample, syn hesis
me hods a e di ided in o pai ed syn hesis and unpai ed syn hesis.
4.3.1. Pai ed sample- o-sample syn hesis (P-s2s)
P-s2s me hods a e ained using pai ed samples gene a ed om he
same objec acqui ed using di e en p o ocols. These me hods aim o
lea n he da a ans e be ween sou ce and e e ence coho s, which
equi e he epea ed acquisi ion o he same subjec unde di e en
p o ocols. The e o e, hey can only be applied o adiomic da a since he
epea ed acquisi ion o he same subjec is impossible o gene
exp ession and pa hology.
Mul i-laye pe cep on ha monic (MLPH): In 2009, a pilo a chi-
ec u e o he au oencode - ela ed me hod was p oposed by Cheng e al.
[114] o gene a e he ha monised da a by lea ning he nonlinea
ans o m unc ion.
Sphe ical ha monic ne wo k (SHNe ): Golko e al. [115] p e-
sen ed a cascaded ully connec ed ne wo k ha employs ReLU and Ba ch
no malisa ion o ha monise he di usion MRI scans. Inspi ed by SHNe ,
Koppe s e al. [116] applied he esidual s uc u e o imp o e he
obus ness while a oiding o e i ing.
Deep o a ion in a ian sphe ical ha monics (Deep-RISH): Ka -
ayumak e al. [117] p oposed a deep lea ning based non-linea mapping
app oach ha u ilises RISH ea u es o map he aw signal (dMRI da a)
be ween scanne s wi h he same ib e o ien a ions. Deep-RISH was
composed o i e con olu ion laye s, which ook he 9 ×9 RISH ea u e
pa ches as he inpu .
DeepHa mony: DeepHa mony was p oposed o p oduce da a wi h
consis en con as wi hin di e en coho s [118]. I employed a U-Ne
based a chi ec u e, aking da a om he sou ce coho and p oducing
ha monised da a o he a ge coho .
Deep ha monics o di usion ku osis imaging (Deep HDKI):
Tong e al. [119] ca ied ou a concise a chi ec u e wi h h ee 3D-con o-
lu ion laye s o di usion ku osis images (DKI). The pai ed da a was
gene a ed using an i e a i e echnique called linea leas squa e and
we e non-linea ly egis e ed o di usion-weigh ed images acqui ed on
he a ge scanne using he compu a ional ools. Then he neu al
ne wo k was ained on he pai ed samples o ha monisa ion.
Deep ha monics o slice hickness (Deep HST): Pa k e al. [120]
s udied he ep oducibili y o adiomic ea u es in lung cance unde
di e en slice hicknesses and p oposed an end- o-end deep neu al
ne wo k o gene a e ha monised CT da a be ween 1-, 3-, and 5-mm slice
hickness.
Deep ha monics o econs uc ion ke nel (Deep HRK): Choe
e al. [34] explo ed he in luence o di e en econs uc ion ke nels on
adiomic ea u es and p esen ed a CNN wi h esidual lea ning o ans e
he da a om he so ke nel (B30 ) o he sha p ke nel (B50 ).
Dis ibu ion-ma ching esidual ne wo k (MMD-ResNe ): Sha-
ham e al. [121] p esen ed a comp ehensi e mul i-laye pe cep on o
ha monisa ion wi h esidual connec ion [122] and ba ch no malisa ion
[123] echniques. Gi en wo coho s o da a X[x1,x2,…,xm] ∈ D1 and
Y[y1,y2,…,yn] ∈ D2. The MMD-ResNe aimed o lea n a map φ:Rd→Rd
by minimising he maximum mean disc epancy [124] be ween φ(X)and
Y. I is o no e ha his was a ‘one-way s ee ’ dis ibu ion ma ching o
ha monisa ion and equi ed e- aining o in e se ans o ma ion.
Pulse sequence in o ma ion based con as lea ning on neigh-
bou hood ensembles (PSI-CLONE): PSI-CLONE [125] i s calcula ed
sequence pa ame e s ∅s om sou ce coho s, hen applied ∅s o he
e e ence coho s o p oduce he sou ce-s yle da a. By aining a
eg ession model o lea n he nonlinea mapping be ween syn hesised
sou ce-s yle da a and e e ence da a, he sou ce coho s could be
ha monised e ec i ely. Based on PSI-CLONE, Jog e al. [126] applied
he mul i-scale ea u e ex ac ion o imp o e he pe o mance.
4.3.2. Unpai ed sample- o-sample syn hesis (Up-s2s)
Up-s2s app oaches gene a e he ha monised da a by cycle-consis en
gene a i e ad e sa ial ne wo ks o condi ional a ia ional au oencode -
decode , which equi e su icien samples and coho labels om
di e en coho s o ne wo k aining.
Cycle-consis en gene a i e ad e sa ial ne wo ks (CycleGAN):
Mos syn hesis me hods o unpai ed sample- o-sample ansla ion we e
based on CycleGAN [127,128] and i s de i a i es [62,129,130]. In
[130], a CycleGAN wi h Ma ko ian disc imina o was applied o
ha monise he di usion enso da a, which was designed o u he
imp o e he abili y o cap u e local in o ma ion.
Condi ional a ia ional au oencode -decode (Condi ional
VAE): Va ia ional Au oencode (VAE) is commonly used in da a syn-
hesis, dimensional educ ion, and ea u e e inemen asks. I employs
an encoding ne wo k Eθ(z|x) o decompose he inpu high dimensional
da a x in o hidden ep esen a ion z, and a decoding ne wo k Dδ(x|z) o
econs uc he aw da a x, whe e θ and δ a e pa ame e s o E and D. The
condi ional VAE modi ies he decode o a condi ional decode Dδ(x|z,c)
ha akes he la en a iable z and speci ied coho c back o a
ha monised da a x. By in eg a ing Condi ional VAE wi h he ad e sa ial
module, coho ans e can be pe o med wi hou pai ed aining
samples. Se e al s udies ha e been p oposed using Condi ional VAE o
da a ha monisa ion, including:
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
114
size, de elopmen pla o m, e c.). Du ing he e alua ion, esea che s
should assess he ep oducibili y using new/independen da a o da a-
de i ed ea u es be o e and a e da a ha monisa ion by app op ia e
me ics. Meanwhile, he da a ha monisa ion pe o mance o p e ious
app oaches should be conside ed as compa isons o e lec he ad an-
ages o he p oposed me hod. A las , he no el y, s eng h, limi a ions
and u u e wo ks should be gi en in he discussion and conclusion
sec ions.
Table 5
Checklis o Compu a ional Da a Ha monisa ion in Digi al Heal hca e
(CHECDHA) c i e ia.
Ca ego y I em Explana ion Example
Mo i a ion Backg ound The applica ion ield
o he da ase (s)
In o ma ion usion
o DW-MRI da a
om di e en
scanne s
Impo ance Why his s udy is
conduc ed, how
impo an i is
D ama ically
inc ease he
s a is ical powe
and sensi i i y o
clinical s udies
Da a Common Da ase Wha he da ase (s) is
(a e), how i is ( hey
a e) collec ed (de ails
o acquisi ion
p o ocols, en y and
exi c i e ia)
How many
ca ego ies, coho s,
subjec s, and cases
a e included in he
s udies
m heal hy subjec s
unde n p o ocols
(m×n cases, n
coho s)
P o ocol 1: …
P o ocol 2: …
P ope y Whe he he da ase
(s) is (a e) in-house
o public, p o ide he
access link i
app op ia e
Public/In-house
P e-p ocessing How he da ase is
p e-p ocessed
Z-sco e
no malisa ion
G ound u h Wha he g ound
u h is and how i is
gene a ed
Coho x unde
p o ocol i
Pa i ion Fo machine
lea ning, how he
da ase is pa i ioned
in o aining,
alida ion, and
es ing subse s in
e ms o he numbe
o samples, pa ien s
7:2:1 o aining,
alida ion and es
Augmen a ion Fo machine
lea ning, how he
da ase is augmen ed
Randomized lip,
o a ion
Speci ic MRI sequence Wha he MRI
sequence is
Di usion-
weigh ed
Region Which egion(s) o
he body o he
subjec in he da ase
is (a e) co e ed
B ain
Slice size Wha he sizes o
each slice a e
512 ×512
Pixel/Voxel
size
Wha he physical
leng h o a pixel/
oxel is
0.25 mm/ 1mm3
WSI size Wha he sizes o he
whole slide images
a e
12,000 ×30,000
Pa ch size Wha he ex ac ed
image pa ches a e
256 ×256
mmp Wha he mic ons pe
pixel in he le el-
0 scan a e
–
Model Wo k low Wha he p ocedu es
o ain and in e ence
a e, illus a ed by he
low cha (s) i
app op ia e.
–
Lea ning
app oaches
Wha he lea ning
me hod is. e.g.,
supe ised lea ning,
un/semi-supe ised
lea ning
Semi-supe ised
lea ning
A chi ec u e Wha he s uc u e o
he p oposed neu al
ne wo k is, i
app op ia e
nnUNe
Table 5 (con inued)
Task The desc ip ion o
main asks conduc ed
on ha monised
da ase s, e.g., lesion
segmen a ion/
classi ica ion.
Tumou
Segmen a ion
Inpu domain Wha he inpu
modali y o he
p oposed me hod is
3-D images / 2D
ea u e ec o s
Inpu size The inpu sizes o he
model
n×w×h×c
Loss Wha he
op imisa ion
unc ions a e du ing
he aining.
Dice and c oss-
en opy loss
Open-sou ce Whe he he sou ce
code is a ailable o
no , p o ide he link
i app op ia e.
Open-sou ce code
www.gi hub.
com...
Pla o m The lea ning lib a y
used o build he
model
Tenso Flow 2.5.0
E alua ion S a is ical
Analysis
Wha he e alua ion
me hods o s a is ical
analysis a e
ANOVA- es
Me ic Wha indica o s a e
used o e alua e
ha monisa ion
pe o mance, e.g.,
he a io o he
ep oducible
ea u es, coe icien
o a ia ion, Pea son
co ela ion
coe icien .
In a-class
co ela ion
coe icien (>0.9 is
conside ed
ep oducible)
Compa ison Wha exis ing
app oaches a e used
o compa e he
pe o mance o he
p oposed me hod
s VAE
Visualisa ion Wha app oaches a e
used o isualise he
da a dis ibu ion
be o e and a e
ha monisa ion
s a egies
-SNE/UMAP/PCA
Resul Resul Wha he
quan i a i e alues o
e alua ion me ics
a e.
–
Time-
consuming
The compu a ional
ime o he p oposed
me hod and he
compa isons.
30 s pe case
Discussion No el y Wha he inno a ion
o he p oposed
me hod is.
–
S eng h The impo ance/
signi icance o he
issue add essed by
he p oposed
me hod.
–
Limi a ion Wha emained and
unsol ed issues a e.
–
Fu u e wo ks Whe he he e will be
po en ial s udies in
he u u e.
–
Y. Nan e al.

In o ma ion Fusion 82 (2022) 99–122
115
8.2. Guidance o da a ha monisa ion s a egies and me ics
S udies ha e shown ha implemen ing inaccu a e da a ha mo-
nisa ion s a egies may lead o signi ican bias, which esul s in mo e
inaccu a e p edic ions [168]. To guide he me hod selec ion, a lowcha
p esen ing possible ways o da a ha monisa ion is p esen ed in Fig. 13.
As he lowcha illus a es, he dis ibu ion based me hods can be well
pe o med on e ined ea u es o gene ma ices. Fo high dimensional
images, image p ocessing me hods a e ecommended when a
high-pe o mance GPU is no a ailable. The deep lea ning based
me hods (including in a ian ea u e lea ning and syn hesis) can be
applied o all kinds o modali ies, while i equi es su icien aining
samples. The in a ian ea u e lea ning me hods a e ecommended
when he main ask can be in eg a ed wi h he aining p ocess, since he
syn hesis may in oduce un ealis ic a e ac s o he da a.
Fo e alua ion, he selec ion o me ics can di ec ly a ec whe he
Fig. 7. Taxonomy o applica ions ha in ol ed compu a ional da a ha monisa ion s a egies.
Fig. 8. Numbe o publica ions and yea s in e ms o da a p ope ies and modali ies. The public da a is he open sou ce da a ha can be acqui ed, he in-house da a is
no a ailable om he in e ne . The pe cen age in he op le sub igu e is he a io o s udies ha we e conduc ed on he public da ase .
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
116
he esul s a e eliable o no . He e we summa ise and ecommend da a
ha monisa ion me ics based on di e en condi ions in Fig. 14. Visual-
isa ion is he mos in ui ional way o analyse da a ha monisa ion esul s,
which can be implemen ed by isualising he aw da a wi h -SNE/
UMAP/PCA o isualising he da a ha monised aw da a. Main ask
based e alua ion can di ec ly illus a e he e ec i eness o he da a
ha monisa ion s a egies, by compa ing he main ask pe o mance on
da a be o e and a e he da a ha monisa ion. I he ha monised g ound
u h is no a ailable, one can use dis ibu ion based me ics o assess he
deg ee o sample mix u e (al hough his may equi e he coho label).
When he ha monised g ound u h can be acqui ed, he alue based o
co ela ion based me ics can p ecisely p esen he da a ha monisa ion
pe o mance.
9. Conclusion
Compu a ional da a ha monisa ion has been p oposed o digi al
heal hca e esea ch s udies in decades. Howe e , b idging basic science
esea ch models and da a usion in o mul icen e, mul imodal and mul i-
scanne medical p ac ice and clinical ials can be challenging unless
Fig. 9. Ha monisa ion s a egies in e ms o di e en modali ies. ‘IFL’ indica es in a ian ea u e lea ning app oaches, “Img P o” e e s o image p ocessing ap-
p oaches. The pe cen age o sub-me hods is anno a ed wi h he abb e ia ions o sub-me hods in each pie cha .
Fig. 10. E alua ion me ics in e ms o di e en modali ies.
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
117
da a ha monisa ion can be pe o med e ec i ely. Fu he mo e, ans-
e / ede a ed/mul i ask lea ning and o he a eas whe ein knowledge is
exchanged amongs models only wo k unde ideal condi ions, whene e
he dis ibu ion shi is no la ge enough o he exchange knowledge o
emain cohe en ac oss models/cen es wo king o e di e en da a
sou ces. O he wise, da a ha monisa ion is needed. Un o una ely, i is
unclea which app oaches and me ics should be employed when
dealing wi h mul imodal da ase s. Mo eo e , he e lacks a ‘s and-
a dised’ s epwise design me hodology, which leads o poo ep oduc-
ibili y o he exis ing s udies.
To o e come hese issues, his pape summa ises and ca ego ises he
exis ing da a ha monisa ion s a egies and me ics based on di e en
heo ies, and subsequen ly p esen s he CHECDHA c i e ia. The p o-
posed CHECDHA c i e ia help esea che s o conduc da a ha mo-
nisa ion s udies in a s anda dised o ma , which can g ea ly ad ance
academic ep oducibili y and de elopmen . Mo eo e , da a ha mo-
nisa ion app oaches and e alua ion me ics in e ms o h ee modali ies
a e summa ised o help esea che s o selec app op ia e s a egies
(Fig. 7 and Fig. 8). In addi ion o summa ising he me hodologies,
guidance o me hod and me ics selec ion (Fig. 11 and Fig. 12) is also
p o ided acco ding o he di e en condi ions. Las bu no leas , limi-
a ions and di ec ions o di e en me hods a e illus a ed o u u e
Fig. 11. Scales o coho s in gene exp ession and adiomics s udies.
Fig. 12. Wo k low o conduc ing da a ha monisa ion s udies guided by he checklis .
Fig. 13. Flowcha o how o selec da a ha monisa ion s a egies.
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
118
wo ks.
Da a ha monisa ion, an impo an p ocess in la ge mul icen e
s udies, has d awn mo e and mo e a en ion in compu a ional biomed-
ical esea ch. I can be well adap ed o a ede a ed lea ning sys em o
p omo e he de elopmen o compu a ional modules and plays an
impo an ole in biomedical esea ch including adiomic, gene ic and
pa hological s udies. Due o he lack o c i e ia when epo ing esea ch
indings o ha monisa ion s udies, we s ongly appeal ha he e-
sea che s should ollow and expand he checklis p esen ed in his
su ey.
Au ho s a emen s
YN, JDS, FH and GY concei ed and designed he s udy, con ibu ed
o da a analysis, con ibu ed o da a in e p e a ion, and con ibu ed o
he w i ing o he epo . YN, JDS, SW1, CS, MR, IS, KH, JO, JN, JG, BE,
AP, AAB, MIM, SW2, WV, NF, JPC, EVR, AC, HW, PL, LCA, LMB, FH, and
GY con ibu ed o he li e a u e sea ch. YN con ibu ed o da a collec-
ion and pe o med da a cu a ion and con ibu ed o he ables and
igu es. YN, JDS, LCA, LMB, and GY con ibu ed o W i ing - Re iew &
Edi ing. SW1, CS, KH, JG, AAB, MIM, SW2, WV, EVR, PL, LMB and GY
con ibu ed o Funding acquisi ion. NF con ibu ed o P ojec adminis-
a ion. FH o e saw he s udy. GY supe ised he wo k. All au ho s
con ibu ed o he a icle and app o ed he submi ed e sion.
Decla a ion o Compe ing In e es
The au ho s decla e ha hey ha e no known compe ing inancial
in e es s o pe sonal ela ionships ha could ha e appea ed o in luence
he wo k epo ed in his pape .
Acknowledgemen
This s udy was suppo ed in pa by he Eu opean Resea ch Council
Inno a i e Medicines Ini ia i e (DRAGON
#
, H2020-JTI-IMI2
101005122), he AI o Heal h Imaging Awa d (CHAIMELEON
##
,
H2020-SC1-FA-DTS-2019–1 952172), he UK Resea ch and Inno a ion
Fu u e Leade s Fellowship (MR/V023799/1), he B i ish Hea
Founda ion (P ojec Numbe : TG/18/5/34111, PG/16/78/32402), he
SABRE p ojec suppo ed by Boeh inge Ingelheim L d, he Eu opean
Union’s Ho izon 2020 esea ch and inno a ion p og amme (ICOVID,
101016131), he Euskampus Founda ion (COVID19 Resilience,
Re . COn VID19), and he Basque Go e nmen (consolida ed esea ch
g oup MATHMODE, Re . IT1294–19, and 3KIA p ojec om he
ELKARTEK unding p og am, Re . KK-2020/00049).
#
DRAGON Conso ium:
Xiaodan Xing
a
, Ming Li
a
, Sco Wage s
b
, Rebecca Bake
c
, Cosimo
Na di
d
, B ice an Eeckhou
e
, Paul Skipp
, Pippa Powell
g
, Miles Ca oll
h
,
Alessand o Ruggie o
i
, Muhun han Thillai
i
, Judi h Baba
i
, E is Sala
i
,
William Mu ch
j
, Julian Hiscox
k
, Diana Ba alle
l
, Nicola S e zella i
m
##
CHAIMELEON Conso ium:
Ana Miguel Blanco
n
, Fuensan a Bell ís Ba alle
o
, Ma io Azna
p
,
Amelia Sua ez
p
, Se gio Figuei as
q
, Ka ha ina K ischak
, Monika Hie -
a h
, Yis oel Mi sky
s
, Yu al Elo ici
s
, Jean Paul Be egi
, Lau e Fou nie
,
F ancesco Sa danelli
u
, Tobias Penzko e
, Ka ine Seymou
w
, Nacho
Blanque
x
, Emanuele Ne i
y
, And ea Laghi
z
, Manuela F ança
aa
, Rica d
Ma inez
ab
a
Na ional Hea and Lung Ins i u e, Impe ial College London, Lon-
don, UK
b
BioSci Consul ing, Maasmechelen, Belgium
c
Clinical Da a In e change S anda ds Conso ium, Aus in, Texas,
Uni ed S a es
d
Uni e si y o Flo ence, Fi enze, I aly
e
Medical Cloud Company, Li`
ege, Belgium
TopMD, Sou hamp on, UK
g
Eu opean Lung Founda ion, She ield, UK
h
Depa men o Heal h, Public Heal h England, London, UK
i
Depa men o Radiology, Uni e si y o Camb idge, Camb idge, UK
j
Owls one Medical, Camb idge, UK
k
Uni e si y o Li e pool, Li e pool, UK
l
Uni e si y o Sou hamp on, Sou hamp on, UK
m
Uni e si y o Pa ma, Pa ma, I aly
n
Medical Imaging Depa men , Hospi al Uni e si a i i Poli `
ecnic La
Fe, Valencia, Spain
o
QUIBIM, Valencia, Spain
p
Ma ical Inno a ion, Mad id, Spain
Fig. 14. Flowcha o how o selec ha monisa ion me ics.
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
119
q
Bahía So wa e, A Co u˜
na, Spain
Eu opean Ins i u e o Biomedical Imaging Resea ch, Vienna,
Aus ia
s
Ben Gu ion Uni e si y o he Nege , Be’e She a, Is ael
Le Coll`
ege des Enseignan s en Radiologie de F ance, F ance
u
Resea ch Hospi al Policlinico San Dona o, Milan, I aly
Cha i ´
e – Uni e si ¨
a smedizin Be lin, Be lin, Ge many
w
Medexp im, Lab`
ege, F ance
x
Valencia Poly echnic Uni e si y, Valencia, Spain
y
Uni e si y o Pisa, Pisa, I aly
z
Sapienza Uni e si y o Rome, Rome, I aly
aa
The Cen o Hospi ala Uni e si ´
a io do Po o, Po ugal
ab
Uni e si y o Valencia, Valencia, Spain
Re e ences
1W.T. Cla ke, O. Mougin, I.D. D i e , C. Rua, A.T. Mo gan, M. Asgha , S. Cla e,
S. F ancis, R.G. Wise, C.T. Rodge s, Mul i-si e ha moniza ion o 7 esla MRI
neu oimaging p o ocols, Neu oimage 206 (2020), 116335.
2D. Delbeke, R.E. Coleman, M.J. Guibe eau, M.L. B own, H.D. Royal, B.A. Siegel, D.
W. Townsend, L.L. Be land, J.A. Pa ke , K. Hubne , P ocedu e guideline o umo
imaging wi h 18F-FDG PET/CT 1.0, J. Nucl. Med. 47 (2006) 885–895.
3J. Simon, D. Li, A. T aboulsee, P. Coyle, D. A nold, F. Ba kho , J. F ank,
R. G ossman, D. Pa y, E. Radue, S anda dized MR imaging p o ocol o mul iple
scle osis: conso ium o MS Cen e s consensus guidelines, Am. J. Neu o adiol. 27
(2006) 455–461.
4B.-.M. Schmid , C.J. Col in, A. Hohl eld, N. Leon, De ining and concep ualising
da a ha monisa ion: a scoping e iew p o ocol, Sys . Re . 7 (2018) 1–6.
5B.-.M. Schmid , C.J. Col in, A. Hohl eld, N. Leon, De ini ions, componen s and
p ocesses o da a ha monisa ion in heal hca e: a scoping e iew, BMC Med. In o m.
Decis. Mak. 20 (2020) 1–19.
6R. Da-Ano, D. Vis ikis, M. Ha , Ha moniza ion s a egies o mul icen e adiomics
in es iga ions, Phy. Med. Biol. 65 (2020) 24TR02.
7M.S. Pin o, R. Paolella, T. Billie , P. Van Dyck, P.-.J. Guns, B. Jeu issen, A. Ribbens,
A.J. den Dekke , J. Sijbe s, Ha moniza ion o b ain di usion MRI: concep s and
me hods, F on . Neu osci. 14 (2020) 396.
8S. Gi o, R. Cuocolo, D. Albano, F. Mo elli, L.C. Pesca o i, C. Messina, M. Imb iaco,
L.M. Scon ienza, CT and MRI adiomics o bone and so - issue sa comas: a
sys ema ic e iew o ep oducibili y and alida ion s a egies, Insigh s Imaging 12
(2021) 1–14.
9S.A. Mali, A. Ib ahim, H.C. Wood u , V. And ea czyk, H. Mülle , S. P imako ,
Z. Salahuddin, A. Cha e jee, P. Lambin, Making adiomics mo e ep oducible
ac oss scanne and imaging p o ocol a ia ions: a e iew o ha moniza ion
me hods, J. Pe s. Med. 11 (2021) 842.
10 T. Chen, M. Philip, K.-A.Lˆ
e Cao, S. Tyagi, A mul i-modal da a ha monisa ion
app oach o disco e y o COVID-19 d ug a ge s, B ie . Bioin o m. (2021).
11 C.M. Tax, F. G ussu, E. Kaden, L. Ning, U. Rud apa na, C.J. E ans, S. S -Jean,
A. Leemans, S. Koppe s, D. Me ho , C oss-scanne and c oss-p o ocol di usion MRI
da a ha monisa ion: a benchma k da abase and e alua ion o algo i hms,
Neu oimage 195 (2019) 285–299.
12 D.M. Hu chinson, E. Silins, R.P. Ma ick, G.C. Pa on, D.M. Fe gusson,
R. Haya bakhsh, J.W. Toumbou ou, C.A. Olsson, J.M. Najman, E. Sp y, How can
da a ha monisa ion bene i men al heal h esea ch? An example o he Cannabis
coho s esea ch conso ium, Aus alian New Zealand J. Psychia . 49 (2015)
317–323.
13 B. Zhao, Y. Tan, W.-.Y. Tsai, J. Qi, C. Xie, L. Lu, L.H. Schwa z, Rep oducibili y o
adiomics o deciphe ing umo pheno ype wi h imaging, Sci. Rep. 6 (2016) 1–7.
14 N. Kuma , R. Ve ma, S. Sha ma, S. Bha ga a, A. Vahadane, A. Se hi, A da ase and a
echnique o gene alized nuclea segmen a ion o compu a ional pa hology, IEEE
T ans. Med. Imaging 36 (2017) 1550–1560.
15 M. Ho a, R. Minamimo o, Y. Gohda, K. Miwa, K. O ani, T. Kiyoma su, H. Yano,
P ognos ic alue o 18 F-FDG PET/CT wi h ex u e analysis in pa ien s wi h ec al
cance ea ed by su ge y, Ann. Nucl. Med. (2021) 1–10.
16 M.V. Ma oli, M.L. Calcagni, S. Ta alli, L. Indo ina, B.S. Spo iswoode, A. Gio dano,
How o en do we ail o classi y he ea men esponse wi h [18 F] FDG PET/CT
acqui ed on di e en scanne s? Da a om clinical oncological p ac ice using an
au oma ic ool o SUV ha moniza ion, Mol. Imaging Biol. 21 (2019) 1210–1219.
17 H. Mi zaalian, L. Ning, P. Sa adjie , O. Pas e nak, S. Bouix, O. Michailo ich,
G. G an , C.E. Ma x, R.A. Mo ey, L.A. Flashman, In e -si e and in e -scanne
di usion MRI da a ha moniza ion, Neu oimage 135 (2016) 311–323.
18 T. Zhu, R. Hu, X. Qiu, M. Taylo , Y. Tso, C. Yiannou sos, B. Na ia, S. Mo i,
S. Ekholm, G. Schi i o, Quan i ica ion o accu acy and p ecision o mul i-cen e
DTI measu emen s: a di usion phan om and human b ain s udy, Neu oimage 56
(2011) 1398–1411.
19 J. Jo icich, M. Ma izzoni, B. Bosch, D. Ba ´
es-Faz, J. A nold, J. Benningho ,
J. Wil ang, L. Rocca aglia a, A. Picco, F. Nobili, Mul isi e longi udinal eliabili y o
ac -based spa ial s a is ics in di usion enso imaging o heal hy elde ly subjec s,
Neu oimage 101 (2014) 390–403.
20 P. Leo, G. Lee, N.N. Shih, R. Ellio , M.D. Feldman, A. Madabhushi, E alua ing
s abili y o his omo phome ic ea u es ac oss scanne and s aining a ia ions:
p os a e cance diagnosis om whole slide images, J. Med. Imaging 3 (2016),
047502.
21 R. Be engue , M.D.R. Pas o -Juan, J. Canales-V´
azquez, M. Cas o-Ga cía, M.
V. Villas, F. Mansilla Lego bu o, S. Saba e , Radiomics o CT ea u es may be
non ep oducible and edundan : in luence o CT acquisi ion pa ame e s, Radiology
288 (2018) 407–415.
22 J.J. Sunde land, P.E. Ch is ian, Quan i a i e PET/CT scanne pe o mance
cha ac e iza ion based upon he socie y o nuclea medicine and molecula imaging
clinical ials ne wo k oncology clinical simula o phan om, J. Nucl. Med. 56
(2015) 145–152.
23 A. Jha, S. Mi hun, V. Jaiswa , U. She khane, N. Pu anda e, K. P abhash,
V. Ranga ajan, A. Dekke , L. Wee, A. T a e so, Repea abili y and ep oducibili y
s udy o adiomic ea u es on a phan om and human coho , Sci. Rep. 11 (2021)
1–12.
24 N. Emaminejad, M.W. Wahi-Anwa , G.H.J. Kim, W. Hsu, M. B own, M. McNi -
G ay, Rep oducibili y o lung nodule adiomic ea u es: mul i a iable and
uni a iable in es iga ions ha accoun o in e ac ions be ween CT acquisi ion and
econs uc ion pa ame e s, Med. Phys., (2021).
25 M. Kim, S.C. Jung, J.E. Pa k, S.Y. Pa k, H. Lee, K.M. Choi, Rep oducibili y o
adiomic ea u es in SENSE and comp essed SENSE: impac o accele a ion ac o s,
Eu . Radiol. (2021) 1–14.
26 R. Yamashi a, T. Pe in, J. Chak abo y, J.F. Chou, N. Ho a , M.A. Koszalka,
A. Midya, M. Gonen, P. Allen, W.R. Ja nagin, Radiomic ea u e ep oducibili y in
con as -enhanced CT o he panc eas is a ec ed by a iabili ies in scan pa ame e s
and manual segmen a ion, Eu . Radiol. 30 (2020) 195–205.
27 S. Fise , M.L. Welch, J. Weiss, M. Pin ilie, J.L. Conway, M. Milose ic, A. Fyles,
A. T a e so, D. Ja ay, U. Me se , Repea abili y and ep oducibili y o MRI-based
adiomic ea u es in ce ical cance , Radio he . Oncol. 135 (2019) 107–114.
28 E. Saeedi, A. Dezhkam, J. Beigi, S. Ras ega , Z. Youse i, L.A. Mehdipou ,
H. Abdollahi, K. Tanha, Radiomic ea u e obus ness and ep oducibili y in
quan i a i e bone adiog aphy: a s udy on adiologic pa ame e changes, J. Clin.
Densi om. 22 (2019) 203–213.
29 M. Meye , J. Ronald, F. Ve nuccio, R.C. Nelson, J.C. Rami ez-Gi aldo, J. Solomon,
B.N. Pa el, E. Samei, D. Ma in, Rep oducibili y o CT adiomic ea u es wi hin he
same pa ien : in luence o adia ion dose and CT econs uc ion se ings, Radiology
293 (2019) 583–591.
30 T. Pe in, A. Midya, R. Yamashi a, J. Chak abo y, T. Saidon, W.R. Ja nagin,
M. Gonen, A.L. Simpson, R.K. Do, Sho - e m ep oducibili y o adiomic ea u es in
li e pa enchyma and li e malignancies on con as -enhanced CT imaging,
Abdominal Radiol. 43 (2018) 3271–3278.
31 A. Midya, J. Chak abo y, M. G¨
onen, R.K. Do, A.L. Simpson, In luence o CT
acquisi ion and econs uc ion pa ame e s on adiomic ea u e ep oducibili y, J.
Med. Imaging 5 (2018), 011020.
32 B.A. Al azi, G.G. Zhang, D.C. Fe nandez, M.E. Mon ejo, D. Hun , J. We ne , M.
C. Biagioli, E.G. Mo os, Rep oducibili y o F18-FDG PET adiomic ea u es o
di e en ce ical umo segmen a ion me hods, g ay-le el disc e iza ion, and
econs uc ion algo i hms, J. Appl. Clin. Med. Phy. 18 (2017) 32–48.
33 P. Hu, J. Wang, H. Zhong, Z. Zhou, L. Shen, W. Hu, Z. Zhang, Rep oducibili y wi h
epea CT in adiomics s udy o ec al cance , Onco a ge 7 (2016) 71440.
34 J. Choe, S.M. Lee, K.-.H. Do, G. Lee, J.-.G. Lee, S.M. Lee, J.B. Seo, Deep
lea ning–based image con e sion o CT econs uc ion ke nels imp o es adiomics
ep oducibili y o pulmona y nodules o masses, Radiology 292 (2019) 365–373.
35 A.N. P imak, C.H. McCollough, M.R. B uesewi z, J. Zhang, J.G. Fle che ,
Rela ionship be ween noise, dose, and pi ch in ca diac mul i–de ec o ow CT,
Radiog aphics 26 (2006) 1785–1794.
36 D.S. Gie ada, A.J. Bie hals, C.K. Choong, S.T. Ba el, J.H. Ri e , N.A. Das, C. Hong,
T.K. Pilg am, K.T. Bae, B.R. Whi ing, E ec s o CT sec ion hickness and
econs uc ion ke nel on emphysema quan i ica ion: ela ionship o he magni ude
o he CT emphysema index, Acad. Radiol. 17 (2010) 146–156.
37 P.-.Y. Tung, J.D. Blischak, C.J. Hsiao, D.A. Knowles, J.E. Bu ne , J.K. P i cha d,
Y. Gilad, Ba ch e ec s and he e ec i e design o single-cell gene exp ession
s udies, Sci. Rep. 7 (2017) 1–15.
38 T. S ua , A. Bu le , P. Ho man, C. Ha emeis e , E. Papalexi, W.M. Mauck III,
Y. Hao, M. S oeckius, P. Smibe , R. Sa ija, Comp ehensi e in eg a ion o single-cell
da a, Cell 177 (2019) 1888–1902, e1821.
39 S. Vijh, M. Sa aswa , S. Kuma , A new comple e colo no maliza ion me hod o
H&E s ained his opa holgical images, Appl. In ell. (2021) 1–14.
40 D.E. Chandle , R.W. Robe son, Bioimaging: cu en concep s in ligh and elec on
mic oscopy, (2009).
41 D. Sun, J. Wang, Y. Han, X. Dong, J. Ge, R. Zheng, X. Shi, B. Wang, Z. Li, P. Ren,
TISCH: a comp ehensi e web esou ce enabling in e ac i e single-cell
ansc ip ome isualiza ion o umo mic oen i onmen , Nucleic Acids Res. 49
(2021) D1420–D1430.
42 M. Sha iq-ul-Hassan, G.G. Zhang, K. La i i, G. Ullah, D.C. Hun , Y. Balagu una han,
M.A. Abdalah, M.B. Schaba h, D.G. Goldgo , D. Mackin, In insic dependencies o
CT adiomic ea u es on oxel size and numbe o g ay le els, Med. Phys. 44 (2017)
1050–1062.
43 D. Mackin, X. Fa e, L. Zhang, D. F ied, J. Yang, B. Taylo , E. Rod iguez-Ri e a,
C. Dodge, A.K. Jones, L. Cou , Measu ing CT scanne a iabili y o adiomics
ea u es, In es . Radiol. 50 (2015) 757.
44 G. Må ensson, D. Fe ei a, T. G anbe g, L. Ca allin, K. Oppedal, A. Pado ani,
I. Rek o o a, L. Bonanni, M. Pa dini, M.G. K ambe ge , The eliabili y o a deep
lea ning model in clinical ou -o -dis ibu ion MRI da a: a mul icoho s udy, Med.
Image Anal. 66 (2020), 101714.
45 S. Ra ho e, S. Bakas, H. Akba i, G. Shukla, M. Rozycki, C. Da a zikos, De i ing
s able mul i-pa ame ic MRI adiomic signa u es in he p esence o in e -scanne
Y. Nan e al.

In o ma ion Fusion 82 (2022) 99–122
120
a ia ions: su i al p edic ion o glioblas oma ia imaging pa e n analysis and
machine lea ning echniques, in: P oceedings o he Medical Imaging 2018:
Compu e -Aided Diagnosis, In e na ional Socie y o Op ics and Pho onics, 2018,
1057509.
46 W.E. Johnson, C. Li, A. Rabino ic, Adjus ing ba ch e ec s in mic oa ay exp ession
da a using empi ical Bayes me hods, Bios a is ics 8 (2007) 118–127.
47 U. Pandey, J. Saini, M. Kuma , R. Gup a, M. Ingalhalika , No ma i e baseline o
adiomics in B ain MRI: e alua ing he obus ness, egional a ia ions, and
ep oducibili y on FLAIR Images, J. Magn. Reson. Imaging (2020).
48 M. Ingalhalika , S. Shinde, A. Ka ma ka , A. Rajan, D. Rangap akash,
G. Deshpande, Func ional connec i i y-based p edic ion o Au ism on si e
ha monized ABIDE da ase , IEEE T ans. Biomed. Eng. (2021).
49 K. Wengle , C. Cassidy, M. an De Pluijm, J.J. Weins ein, A. Abi-Da gham, E. an
de Giessen, G. Ho ga, C oss-scanne ha moniza ion o neu omelanin-sensi i e MRI
o mul isi e s udies, J. Magn. Reson. Imaging (2021).
50 H. Beaumon , A. Iannessi, A.-.S. Be and, J.M. Cucchi, O. Lucida me,
Ha moniza ion o adiomic ea u e dis ibu ions: impac on classi ica ion o hepa ic
issue in CT imaging, Eu . Radiol. (2021) 1–10.
51 R. Ga cia-Dias, C. Sca pazza, L. Baecke , S. Viei a, W.H. Pinaya, A. Co in,
A. Redol i, B. Nelson, B. C espo-Faco o, C. McDonald, Neu oha mony: a new ool
o ha monizing olume ic MRI da a om unseen scanne s, Neu oimage 220
(2020).
52 J.C. Bee , N.J. Tus ison, P.A. Cook, C. Da a zikos, Y.I. Sheline, R.T. Shinoha a, K.
A. Linn, A.s.D.N. Ini ia i e, Longi udinal comba : a me hod o ha monizing
longi udinal mul i-scanne imaging da a, Neu oimage 220 (2020), 117129.
53 J. Radua, E. Vie a, R. Shinoha a, P. Kochuno , Y. Quid´
e, M.J. G een, C.S. Weicke ,
T. Weicke , J. B uggemann, T. Ki che , Inc eased powe by ha monizing s uc u al
MRI si e di e ences wi h he ComBa ba ch adjus men me hod in ENIGMA,
Neu oimage 218 (2020), 116956.
54 H.M. Whi ney, H. Li, Y. Ji, P. Liu, M.L. Gige , Ha moniza ion o adiomic ea u es o
b eas lesions ac oss in e na ional DCE-MRI da ase s, J. Med. Imaging 7 (2020),
012707.
55 M. Yu, K.A. Linn, P.A. Cook, M.L. Phillips, M. McInnis, M. Fa a, M.H. T i edi, M.
M. Weissman, R.T. Shinoha a, Y.I. Sheline, S a is ical ha moniza ion co ec s si e
e ec s in unc ional connec i i y measu emen s om mul i-si e MRI da a, Hum
B ain Mapp. 39 (2018) 4213–4227.
56 A. Espín-P´
e ez, C. Po ie , M. Chadeau-Hyam, K. an Veldho en, J.C. Kleinjans, T.
M. de Kok, Compa ison o s a is ical me hods and he use o quali y con ol samples
o ba ch e ec co ec ion in human ansc ip ome da a, PLoS ONE 13 (2018),
e0202947.
57 J.-.P. Fo in, N. Cullen, Y.I. Sheline, W.D. Taylo , I. Aselcioglu, P.A. Cook, P. Adams,
C. Coope , M. Fa a, P.J. McG a h, Ha moniza ion o co ical hickness
measu emen s ac oss scanne s and si es, Neu oimage 167 (2018) 104–120.
58 J.-.P. Fo in, D. Pa ke , B. Tunç, T. Wa anabe, M.A. Ellio , K. Rupa el, D.R. Roal , T.
D. Sa e hwai e, R.C. Gu , R.E. Gu , Ha moniza ion o mul i-si e di usion enso
imaging da a, Neu oimage 161 (2017) 149–170.
59 S. Ko ha i, J.H. Phan, T.H. S okes, A.O. Osunkoya, A.N. Young, M.D. Wang,
Remo ing ba ch e ec s om his opa hological images o enhanced cance
diagnosis, IEEE J. Biomed. Heal h In o m. 18 (2013) 765–772.
60 C.T. A end , D. Lei hne , M.E. Maye hoe e , P. Gibbs, C. Cze ny, C. A noldne ,
I. Bu ck, M. Leinung, Y. Tanyildizi, L. Lenga, Radiomics o high- esolu ion
compu ed omog aphy o he di e en ia ion be ween choles ea oma and middle
ea in lamma ion: e ec s o pos - econs uc ion me hods in a dual-cen e s udy,
Eu . Radiol. 31 (2021) 4071–4078.
61 A. Ib ahim, T. Re aee, R.T. Leijenaa , S. P imako , R. Hus inx, F.M. Mo aghy, H.
C. Wood u , A.D. Maidmen , P. Lambin, The applica ion o a wo k low in eg a ing
he a iable ep oducibili y and ha monizabili y o adiomic ea u es on a phan om
da ase , PLoS ONE 16 (2021), e0251147.
62 J. Lan, S. Cai, Y. Xue, Q. Gao, M. Du, H. Zhang, Z. Wu, Y. Deng, Y. Huang, T. Tong,
Unpai ed s ain s yle ans e using in e ible neu al ne wo ks based on channel
a en ion and long- ange esidual, IEEE Access 9 (2021) 11282–11295.
63 C. Wachinge , A. Rieckmann, S. P¨
ols e l, A.s.D.N. Ini ia i e, De ec and co ec bias
in mul i-si e neu oimaging da ase s, Med. Image Anal. 67 (2021), 101879.
64 J.J. Foy, H.A. Al-Hallaq, V. G ekoski, T. T an, K. Gu u adoo, S.G. A ma o Iii, W.
F. Sensako ic, Ha moniza ion o adiomic ea u e a iabili y esul ing om
di e ences in CT image acquisi ion and econs uc ion: assessmen in a cada e ic
li e , Phy. Med. Biol. 65 (2020), 205008.
65 M.-J.Sain Ma in, F. O lhac, P. Akl, F. Khalid, C. Nioche, I. Bu a , C. Malhai e,
F. F ouin, A adiomics pipeline dedica ed o B eas MRI: alida ion on a mul i-
scanne phan om s udy, magne ic esonance ma e ials in physics, Biol. Med. (2020)
1–12.
66 S.C. Ka ayumak, S. Bouix, L. Ning, A. James, T. C ow, M. Shen on, M. Kubicki,
Y. Ra hi, Re ospec i e ha moniza ion o mul i-si e di usion MRI da a acqui ed
wi h di e en acquisi ion pa ame e s, Neu oimage 184 (2019) 180–200.
67 Y. Zhang, G. Pa migiani, W.E. Johnson, ComBa -Seq: ba ch e ec adjus men o
RNA-Seq coun da a, NAR Genom. Bioin o m. 2 (2020) lqaa078.
68 C.K. S ein, P. Qu, J. Eps ein, A. Bu os, A. Rosen hal, J. C owley, G. Mo gan,
B. Ba logie, Remo ing ba ch e ec s om pu i ied plasma cell gene exp ession
mic oa ays wi h modi ied ComBa , BMC Bioin o ma ics 16 (2015) 1–9.
69 R. Da-Ano, I. Masson, F. Lucia, M. Do ´
e, P. Robin, J. Al ie i, C. Rousseau,
A. Me oye , C. Reinhold, J. Cas elli, Pe o mance compa ison o modi ied ComBa
o ha moniza ion o adiomic ea u es o mul icen e s udies, Sci. Rep. 10 (2020)
1–12.
70 C. Mülle , A. Schille , C. R¨
o hemeie , D.-.A. T ´
egou¨
e , C. P ous , H. Binde ,
N. P ei e , M. Beu el, K.J. Lackne , R.B. Schnabel, Remo ing ba ch e ec s om
longi udinal gene exp ession-quan ile no maliza ion plus ComBa as bes app oach
o mic oa ay ansc ip ome da a, PLoS ONE 11 (2016), e0156594.
71 M. Beni o, J. Pa ke , Q. Du, J. Wu, D. Xiang, C.M. Pe ou, J.S. Ma on, Adjus men o
sys ema ic mic oa ay da a biases, Bioin o ma ics 20 (2004) 105–114.
72 A.A. Shabalin, H. Tjelmeland, C. Fan, C.M. Pe ou, A.B. Nobel, Me ging wo gene-
exp ession s udies ia c oss-pla o m no maliza ion, Bioin o ma ics 24 (2008)
1154–1160.
73 I. Ko sunsky, N. Milla d, J. Fan, K. Slowikowski, F. Zhang, K. Wei, Y. Baglaenko,
M. B enne , P.- . Loh, S. Raychaudhu i, Fas , sensi i e and accu a e in eg a ion o
single-cell da a wi h Ha mony, Na . Me hods 16 (2019) 1289–1296.
74 L. Hagh e di, A.T. Lun, M.D. Mo gan, J.C. Ma ioni, Ba ch e ec s in single-cell RNA-
sequencing da a a e co ec ed by ma ching mu ual nea es neighbo s, Na .
Bio echnol. 36 (2018) 421–427.
75 B. Hie, B. B yson, B. Be ge , E icien in eg a ion o he e ogeneous single-cell
ansc ip omes using Scano ama, Na . Bio echnol. 37 (2019) 685–691.
76 F.A. Wol , P. Ange e , F.J. Theis, SCANPY: la ge-scale single-cell gene exp ession
da a analysis, Genome Biol. 19 (2018) 1–5.
77 K. Pola´
nski, M.D. Young, Z. Miao, K.B. Meye , S.A. Teichmann, J.-.E. Pa k, BBKNN:
as ba ch alignmen o single cell ansc ip omes, Bioin o ma ics 36 (2020)
964–965.
78 L. McInnes, J. Healy, J. Mel ille, Umap: uni o m mani old app oxima ion and
p ojec ion o dimension educ ion, a Xi p ep in a Xi :1802.03426, (2018).
79 A. Bu le , P. Ho man, P. Smibe , E. Papalexi, R. Sa ija, In eg a ing single-cell
ansc ip omic da a ac oss di e en condi ions, echnologies, and species, Na .
Bio echnol. 36 (2018) 411–420.
80 J.A. Gagnon-Ba sch, T.P. Speed, Using con ol genes o co ec o unwan ed
a ia ion in mic oa ay da a, Bios a is ics 13 (2012) 539–552.
81 D. Risso, F. Pe audeau, S. G ibko a, S. Dudoi , J.-.P. Ve , A gene al and lexible
me hod o signal ex ac ion om single-cell RNA-seq da a, Na . Commun. 9 (2018)
1–17.
82 O. Al e , P.O. B own, D. Bo s ein, Singula alue decomposi ion o genome-wide
exp ession da a p ocessing and modeling, P oc. Na l. Acad. Sci. 97 (2000)
10101–10106.
83 Y. Lin, S. Ghazan a , K.Y. Wang, J.A. Gagnon-Ba sch, K.K. Lo, X. Su, Z.-.G. Han, J.
T. O me od, T.P. Speed, P. Yang, scMe ge le e ages ac o analysis, s able
exp ession, and pseudo eplica ion o me ge mul iple single-cell RNA-seq da ase s,
P oc. Na l. Acad. Sci. 116 (2019) 9775–9784.
84 J.T. Leek, W.E. Johnson, H.S. Pa ke , A.E. Ja e, J.D. S o ey, The s a package o
emo ing ba ch e ec s and o he unwan ed a ia ion in high- h oughpu
expe imen s, Bioin o ma ics 28 (2012) 882–883.
85 G.K. Smy h, T. Speed, No maliza ion o cDNA mic oa ay da a, Me hods 31 (2003)
265–273.
86 J.-.P. Fo in, E.M. Sweeney, J. Muschelli, C.M. C ainiceanu, R.T. Shinoha a, A.s.D.
N. Ini ia i e, Remo ing in e -subjec echnical a iabili y in magne ic esonance
imaging s udies, Neu oimage 132 (2016) 198–212.
87 H. Mi zaalian, L. Ning, P. Sa adjie , O. Pas e nak, S. Bouix, O. Michailo ich,
S. Ka macha ya, G. G an , C.E. Ma x, R.A. Mo ey, Mul i-si e ha moniza ion o
di usion MRI da a in a egis a ion amewo k, B ain Imaging Beha . 12 (2018)
284–295.
88 H. Mi zaalian, A. de Pie e eu, P. Sa adjie , O. Pas e nak, S. Bouix, M. Kubicki, C.-.
F. Wes in, M.E. Shen on, Y. Ra hi, Ha monizing di usion MRI da a ac oss mul iple
si es and scanne s, in: P oceedings o he In e na ional Con e ence on Medical
Image Compu ing and Compu e -Assis ed In e en ion, Sp inge , 2015, pp. 12–19.
89 Y. Zhang, D.F. Jenkins, S. Manima an, W.E. Johnson, Al e na i e empi ical Bayes
models o adjus ing o ba ch e ec s in genomic s udies, BMC Bioin o ma ics 19
(2018) 1–15.
90 K.M. Huynh, G. Chen, Y. Wu, D. Shen, P.-.T. Yap, Mul i-si e ha moniza ion o
di usion MRI da a ia me hod o momen s, IEEE T ans. Med. Imaging 38 (2019)
1599–1609.
91 J. W obel, M. Ma in, R. Bakshi, P. Calab esi, M. Ellio , D. Roal , R. Gu , R. Gu ,
R. Hen y, G. Nai , In ensi y wa ping o mul isi e MRI ha moniza ion, Neu oimage
223 (2020), 117242.
92 A. Lle a, I. Hue as, P. Mi , C.F. Beckmann, Quan i a i e in ensi y ha moniza ion o
dopamine anspo e SPECT images using gamma mix u e models, Mol. Imaging
Biol. 21 (2019) 339–347.
93 C. Laza , J. Taminau, S. Meganck, D. S eenho , A. Cole a, D.Y.W. Solís, C. Mol e ,
R. Duque, H. Be sini, A. Now´
e, GENESHIFT: a nonpa ame ic app oach o
in eg a ing mic oa ay gene exp ession da a based on he inne p oduc as a
dis ance measu e be ween he dis ibu ions o genes, IEEE/ACM T ans. Compu .
Biol. Bioin o m. 10 (2013) 383–392.
94 D. Mackin, X. Fa e, L. Zhang, J. Yang, A.K. Jones, C.S. Ng, L. Cou , Ha monizing
he pixel size in e ospec i e compu ed omog aphy adiomics s udies, PLoS ONE
12 (2017), e0178524.
95 H. Mo admand, S.M.R. Aghami i, R. Ghade i, Impac o image p ep ocessing
me hods on ep oducibili y o adiomic ea u es in mul imodal magne ic esonance
imaging in glioblas oma, J. Appl. Clin. Med. Phy. 21 (2020) 179–190.
96 I. Pi as, Digi al Image P ocessing Algo i hms and Applica ions, John Wiley & Sons,
2000.
97 R.T. Shinoha a, E.M. Sweeney, J. Goldsmi h, N. Shiee, F.J. Ma een, P.A. Calab esi,
S. Ja so, D.L. Pham, D.S. Reich, C.M. C ainiceanu, S a is ical no maliza ion
echniques o magne ic esonance imaging, Neu oImage Clin. 6 (2014) 9–19.
98 E. Reinha d, M. Adhikhmin, B. Gooch, P. Shi ley, Colo ans e be ween images,
IEEE Compu . G aph. Appl. 21 (2001) 34–41.
99 C.P. Loizou, V. Mu ay, M.S. Pa ichis, I. Seimenis, M. Pan zia is, C.S. Pa ichis,
Mul iscale ampli ude-modula ion equency-modula ion (AM–FM) ex u e analysis
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
121
o mul iple scle osis in b ain MRI images, IEEE T ans. In . Technol. Biomed. 15
(2010) 119–129.
100 M. Shah, Y. Xiao, N. Subbanna, S. F ancis, D.L. A nold, D.L. Collins, T. A bel,
E alua ing in ensi y no maliza ion on MRIs o human b ain wi h mul iple scle osis,
Med. Image Anal. 15 (2011) 267–282.
101 S. Roy, S. Lal, J.R. Kini, No el colo no maliza ion me hod o Hema oxylin & Eosin
s ained his opa hology images, IEEE Access 7 (2019) 28982–28998.
102 M.D. Za ella, C. Yeoh, D.E. B een, F.U. Ga cia, An al e na i e e e ence space o
H&E colo no maliza ion, PLoS ONE 12 (2017), e0174489.
103 X. Li, K.N. Pla anio is, A comple e colo no maliza ion app oach o his opa hology
images using colo cues compu ed om sa u a ion-weigh ed s a is ics, IEEE T ans.
Biomed. Eng. 62 (2015) 1862–1873.
104 D.F. Swineha , The bee -lambe law, J. Chem. Educ. 39 (1962) 333.
105 T.A.A. Tos a, P.R. de Fa ia, L.A. Ne es, M.Z. do Nascimen o, Colo no maliza ion o
aded H&E-s ained his ological images using spec al ma ching, Compu . Biol. Med.
111 (2019), 103344.
106 A.M. Khan, N. Rajpoo , D. T eano , D. Magee, A nonlinea mapping app oach o
s ain no maliza ion in digi al his opa hology images using image-speci ic colo
decon olu ion, IEEE T ans. Biomed. Eng. 61 (2014) 1729–1738.
107 A.C. Rui ok, D.A. Johns on, Quan i ica ion o his ochemical s aining by colo
decon olu ion, Anal. Quan . Cy ol. His ol. 23 (2001) 291–299.
108 M.Z. Hoque, A. Keskina kaus, P. Nybe g, T. Sepp¨
anen, Re inex model based s ain
no maliza ion echnique o whole slide image analysis, Compu . Med. Imaging
G aph. 90 (2021), 101901.
109 A. Vahadane, T. Peng, A. Se hi, S. Alba qouni, L. Wang, M. Baus , K. S eige , A.
M. Schli e , I. Esposi o, N. Na ab, S uc u e-p ese ing colo no maliza ion and
spa se s ain sepa a ion o his ological images, IEEE T ans. Med. Imaging 35 (2016)
1962–1971.
110 G. Lei, Y. Xia, D.-.H. Zhai, W. Zhang, D. Chen, D. Wang, S ainCNNs: an e icien
s ain ea u e lea ning me hod, Neu ocompu ing 406 (2020) 267–273.
111 Y. Zheng, Z. Jiang, H. Zhang, F. Xie, J. Shi, C. Xue, Adap i e colo decon olu ion o
his ological WSI no maliza ion, Compu . Me hods P og ams Biomed. 170 (2019)
107–120.
112 P. Maji, S. Mahapa a, Rough- uzzy ci cula clus e ing o colo no maliza ion o
his ological images, Fundam. In o m. 164 (2019) 103–117.
113 P. Maji, S. Mahapa a, Ci cula clus e ing in uzzy app oxima ion spaces o colo
no maliza ion o his ological images, IEEE T ans Med Imaging 39 (2019)
1735–1745.
114 H.-.D. Cheng, X. Cai, R. Min, A no el app oach o colo no maliza ion using neu al
ne wo k, Neu al Compu . Appl. 18 (2009) 237–247.
115 V. Golko , A. Doso i skiy, J.I. Spe l, M.I. Menzel, M. Czisch, P. S¨
amann, T. B ox,
D. C eme s, Q-space deep lea ning: wel e- old sho e and model- ee di usion
MRI scans, IEEE T ans. Med. Imaging 35 (2016) 1344–1351.
116 S. Koppe s, L. Bloy, J.I. Be man, C.M. Tax, J.C. Edga , D. Me ho , Sphe ical
ha monic esidual ne wo k o di usion signal ha moniza ion, in: P oceedings o
he In e na ional Con e ence on Medical Image Compu ing and Compu e -Assis ed
In e en ion, Sp inge , 2019, pp. 173–182.
117 S.C. Ka ayumak, M. Kubicki, Y. Ra hi, Ha monizing di usion MRI da a ac oss
magne ic ield s eng hs, in: P oceedings o he In e na ional Con e ence on Medical
Image Compu ing and Compu e -Assis ed In e en ion, Sp inge , 2018,
pp. 116–124.
118 B.E. Dewey, C. Zhao, J.C. Reinhold, A. Ca ass, K.C. Fi zge ald, E.S. So i chos,
S. Saidha, J. Oh, D.L. Pham, P.A. Calab esi, DeepHa mony: a deep lea ning
app oach o con as ha moniza ion ac oss scanne changes, Magn. Reson. Imaging
64 (2019) 160–170.
119 Q. Tong, T. Gong, H. He, Z. Wang, W. Yu, J. Zhang, L. Zhai, H. Cui, X. Meng, C.
W. Tax, A deep lea ning–based me hod o imp o ing eliabili y o mul icen e
di usion ku osis imaging wi h a ied acquisi ion p o ocols, Magn. Reson. Imaging
73 (2020) 31–44.
120 S. Pa k, S.M. Lee, K.-.H. Do, J.-.G. Lee, W. Bae, H. Pa k, K.-.H. Jung, J.B. Seo, Deep
lea ning algo i hm o educing CT slice hickness: e ec on ep oducibili y o
adiomic ea u es in lung cance , Ko ean J. Radiol. 20 (2019) 1431–1440.
121 U. Shaham, K.P. S an on, J. Zhao, H. Li, K. Raddassi, R. Mon gome y, Y. Kluge ,
Remo al o ba ch e ec s using dis ibu ion-ma ching esidual ne wo ks,
Bioin o ma ics 33 (2017) 2539–2546.
122 K. He, X. Zhang, S. Ren, J. Sun, Deep esidual lea ning o image ecogni ion, in:
P oceedings o he IEEE con e ence on compu e ision and pa e n ecogni ion,
2016, pp. 770–778.
123 S. Io e, C. Szegedy, Ba ch no maliza ion: accele a ing deep ne wo k aining by
educing in e nal co a ia e shi , in: P oceedings o he In e na ional con e ence on
machine lea ning, PMLR, 2015, pp. 448–456.
124 A. G e on, K. Bo gwa d , M. Rasch, B. Sch¨
olkop , A. Smola, A ke nel me hod o he
wo-sample-p oblem, Ad . Neu al In . P ocess. Sys . 19 (2006) 513–520.
125 A. Jog, A. Ca ass, S. Roy, D.L. Pham, J.L. P ince, MR image syn hesis by con as
lea ning on neighbo hood ensembles, Med. Image Anal. 24 (2015) 63–76.
126 A. Jog, A. Ca ass, S. Roy, D.L. Pham, J.L. P ince, Random o es eg ession o
magne ic esonance image syn hesis, Med. Image Anal. 35 (2017) 475–488.
127 J.-.Y. Zhu, T. Pa k, P. Isola, A.A. E os, Unpai ed image- o-image ansla ion using
cycle-consis en ad e sa ial ne wo ks, in: P oceedings o he IEEE in e na ional
con e ence on compu e ision, 2017, pp. 2223–2232.
128 F. Zhao, Z. Wu, L. Wang, W. Lin, S. Xia, D. Shen, G. Li, Ha moniza ion o in an
co ical hickness using su ace- o-su ace cycle-consis en ad e sa ial ne wo ks, in:
P oceedings o he In e na ional Con e ence on Medical Image Compu ing and
Compu e -Assis ed In e en ion, Sp inge , 2019, pp. 475–483.
129 M. Ren, N. Dey, J. Fishbaugh, G. Ge ig, Segmen a ion- eno malized deep ea u e
modula ion o unpai ed image ha moniza ion, IEEE T ans. Med. Imaging 40
(2021) 1519–1530.
130 J. Zhong, Y. Wang, J. Li, X. Xue, S. Liu, M. Wang, X. Gao, Q. Wang, J. Yang, X. Li,
In e -si e ha moniza ion based on dual gene a i e ad e sa ial ne wo ks o
di usion enso imaging: applica ion o neona al whi e ma e de elopmen ,
Biomed. Eng. Online 19 (2020) 1–18.
131 D. Moye , G. Ve S eeg, C.M. Tax, P.M. Thompson, Scanne in a ian
ep esen a ions o di usion MRI ha moniza ion, Magn. Reson. Med. 84 (2020)
2174–2189.
132 N. Russkikh, D. An one s, D. Sh okalo, A. Maka o , Y. Vya kin, A. Zakha o ,
E. Te en ye , S yle ans e wi h a ia ional au oencode s is a p omising app oach
o RNA-Seq da a ha moniza ion and analysis, Bioin o ma ics 36 (2020) 5076–5085.
133 N. Johansen, G. Quon, scAlign: a ool o alignmen , in eg a ion, and a e cell
iden i ica ion om scRNA-seq da a, Genome Biol. 20 (2019) 1–21.
134 P. Haeusse , A. Mo d in se , D. C eme s, Lea ning by associa ion–a e sa ile semi-
supe ised aining me hod o neu al ne wo ks, in: P oceedings o he IEEE
con e ence on compu e ision and pa e n ecogni ion, 2017, pp. 89–98.
135 D. Wang, S. Hou, L. Zhang, X. Wang, B. Liu, Z. Zhang, iMAP: in eg a ion o mul iple
single-cell da ase s by ad e sa ial pai ed ans e ne wo ks, Genome Biol. 22 (2021)
1–24.
136 J. Mai al, F. Bach, J. Ponce, G. Sapi o, Online lea ning o ma ix ac o iza ion and
spa se coding, J. Machine Lea n. Res. 11 (2010).
137 S. S -Jean, P. Coup´
e, M. Desco eaux, Non local spa ial and angula ma ching:
enabling highe spa ial esolu ion di usion MRI da ase s h ough adap i e
denoising, Med. Image Anal. 32 (2016) 115–130.
138 T.A.A. Tos a, P.R. de Fa ia, J.P.S. Se a o, L.A. Ne es, G.F. Robe o, A.S. Ma ins, M.
Z. do Nascimen o, Unsupe ised me hod o no maliza ion o hema oxylin-eosin
s ain in his ological images, Compu . Med. Imaging G aph. 77 (2019), 101646.
139 C. Lu, J. Shi, J. Jia, Online obus dic iona y lea ning, in: P oceedings o he IEEE
Con e ence on Compu e Vision and Pa e n Recogni ion, 2013, pp. 415–422.
140 X. Li, K. Wang, Y. Lyu, H. Pan, J. Zhang, D. S ambolian, K. Susz ak, M.P. Reilly,
G. Hu, M. Li, Deep lea ning enables accu a e clus e ing wi h ba ch e ec emo al in
single-cell RNA-seq analysis, Na . Commun. 11 (2020) 1–14.
141 V.D. Blondel, J.-.L. Guillaume, R. Lambio e, E. Le eb e, Fas un olding o
communi ies in la ge ne wo ks, J. S a . Mech. Theo y Exp. (2008) P10008, 2008.
142 T. Wang, T.S. Johnson, W. Shao, Z. Lu, B.R. Helm, J. Zhang, K. Huang, BERMUDA: a
no el deep ans e lea ning me hod o single-cell RNA sequencing ba ch
co ec ion e eals hidden high- esolu ion cellula sub ypes, Genome Biol. 20 (2019)
1–15.
143 H. Guan, Y. Liu, E. Yang, P.-.T. Yap, D. Shen, M. Liu, Mul i-si e MRI ha moniza ion
ia a en ion-guided deep domain adap a ion o b ain diso de iden i ica ion, Med.
Image Anal. 71 (2021), 102076.
144 N.K. Dinsdale, M. Jenkinson, A.I. Nambu e e, Deep lea ning-based unlea ning o
da ase bias o MRI ha monisa ion and con ound emo al, Neu oimage 228
(2021), 117689.
145 S. Ge, H. Wang, A. Ala i, E. Xing, Z. Ba -Joseph, Supe ised ad e sa ial alignmen
o single-cell RNA-seq da a, J. Compu . Biol. 28 (2021) 501–513.
146 Z. Rong, Q. Tan, L. Cao, L. Zhang, K. Deng, Y. Huang, Z.-.J. Zhu, Z. Li, K. Li,
No mAE: deep ad e sa ial lea ning model o emo e ba ch e ec s in liquid
ch oma og aphy mass spec ome y-based me abolomics da a, Anal. Chem. 92
(2020) 5082–5090.
147 M. Bü ne , Z. Miao, F.A. Wol , S.A. Teichmann, F.J. Theis, A es me ic o
assessing single-cell RNA-seq ba ch co ec ion, Na . Me hods 16 (2019) 43–49.
148 Z. Wang, A.C. Bo ik, H.R. Sheikh, E.P. Simoncelli, Image quali y assessmen : om
e o isibili y o s uc u al simila i y, IEEE T ans. Image P ocess. 13 (2004).
149 A. Kolaman, O. Yadid-Pech , Qua e nion s uc u al simila i y: a new quali y index
o colo images, IEEE T ans. Image P ocess. 21 (2011) 1526–1536.
150 L. Zhang, L. Zhang, X. Mou, D. Zhang, FSIM: a ea u e simila i y index o image
quali y assessmen , IEEE T ans. Image P ocess. 20 (2011) 2378–2386.
151 J.-.F. Pamb un, R. Noumei , Limi a ions o he SSIM quali y me ic in he con ex o
diagnos ic imaging, in: P oceedings o he 2015 IEEE In e na ional Con e ence on
Image P ocessing (ICIP), IEEE, 2015, pp. 2960–2963.
152 L.G. Nyúl, J.K. Udupa, X. Zhang, New a ian s o a me hod o MRI scale
s anda diza ion, IEEE T ans. Med. Imaging 19 (2000) 143–150.
153 A. Albe , L. Zhang, A no el de ini ion o he mul i a ia e coe icien o a ia ion,
Biome . J. 52 (2010) 667–675.
154 P. Chi a, P. Leo, M. Yim, B.N. Bloch, A.R. Ras inehad, A. Pu ysko, M. Rosen,
A. Madabhushi, S.E. Viswana h, Mul isi e e alua ion o adiomic ea u e
ep oducibili y and disc iminabili y o iden i ying pe iphe al zone p os a e umo s
on MRI, J. Med. Imaging 6 (2019), 024502.
155 I. Law ence, K. Lin, A conco dance co ela ion coe icien o e alua e
ep oducibili y, Biome ics (1989) 255–268.
156 D. Liljequis , B. El ing, K. Ska be g Roaldsen, In aclass co ela ion–a discussion
and demons a ion o basic ea u es, PLoS ONE 14 (2019), e0219854.
157 F. O lhac, A. Lecle , J. Sa a o ski, J. Goya-Ou i, C. Nioche, F. Cha bonneau,
N. Ayache, F. F ouin, L. Du on, I. Bu a , How can we comba mul icen e a iabili y
in MR adiomics? Valida ion o a co ec ion p ocedu e, Eu . Radiol. 31 (2021)
2272–2280.
158 R. Mahon, M. Ghi a, G. Hugo, E. Weiss, ComBa ha moniza ion o adiomic
ea u es in independen phan om and lung cance pa ien compu ed omog aphy
da ase s, Phy. Med. Biol. 65 (2020), 015010.
159 G.S. Ioannidis, E. T i izakis, I. Me zakis, S. Papagiannakis, E. Lagoudaki, K. Ma ias,
Pa homics and deep lea ning classi ica ion o a he e ogeneous luo escence
his ology image da ase , Appl. Sci. 11 (2021) 3796.
Y. Nan e al.
In o ma ion Fusion 82 (2022) 99–122
122
160 M. Lo ollahi, F.A. Wol , F.J. Theis, scGen p edic s single-cell pe u ba ion
esponses, Na . Me hods 16 (2019) 715–721.
161 X. Jiang, G.-.B. Bian, Z. Tian, Remo al o a i ac s om EEG signals: a e iew,
Senso s 19 (2019) 987.
162 P. He, M. Kahle, G. Wilson, C. Russell, Remo al o ocula a i ac s om EEG: a
compa ison o adap i e il e ing me hod and eg ession me hod using simula ed
da a, in: P oceedings o he 2005 IEEE Enginee ing in Medicine and Biology 27 h
Annual Con e ence, IEEE, 2006, pp. 1110–1113.
163 P.S. Kuma , R. A umugana han, K. Si akuma , C. Vimal, Remo al o ocula a i ac s
in he EEG h ough wa ele ans o m wi hou using an EOG e e ence channel, In .
J. Open P oblems Comp . Ma h 1 (2008) 188–200.
164 H.J. Ae s, E.R. Velazquez, R.T. Leijenaa , C. Pa ma , P. G ossmann, S. Ca alho,
J. Bussink, R. Monshouwe , B. Haibe-Kains, D. Rie eld, Decoding umou
pheno ype by nonin asi e imaging using a quan i a i e adiomics app oach, Na .
Commun. 5 (2014) 1–9.
165 A.B. A ie a, N. Díaz-Rod íguez, J. Del Se , A. Benne o , S. Tabik, A. Ba bado,
S. Ga cía, S. Gil-L´
opez, D. Molina, R. Benjamins, Explainable A i icial In elligence
(XAI): concep s, axonomies, oppo uni ies and challenges owa d esponsible AI,
In o m. Fusion 58 (2020) 82–115.
166 G. Yang, Q. Ye, J. Xia, Unbox he black-box o he medical explainable ai ia mul i-
modal and mul i-cen e da a usion: a mini- e iew, wo showcases and beyond,
In o m. Fusion 77 (2022) 29–52.
167 A. Holzinge , M. Dehme , F. Emme -S eib, R. Cucchia a, I. Augens ein, J. Del Se ,
W. Samek, I. Ju isica, N. Díaz-Rod íguez, In o ma ion usion as an in eg a i e c oss-
cu ing enable o achie e obus , explainable, and us wo hy medical a i icial
in elligence, In o m. Fusion 79 (2022) 263–278.
168 F.J. L´
opez-Gonz´
alez, J. Sil a-Rod íguez, J. Pa edes-Pacheco, A. Ni˜
ne ola-Baiz´
an,
N. E himiou, C. Ma ín-Ma ín, A. Moscoso, ´
A. Ruibal, N. Ro´
e-Vell ´
e, P. Aguia ,
In ensi y no maliza ion me hods in b ain FDG-PET quan i ica ion, Neu oimage 222
(2020), 117229.
[169] Mongan, John, Linda Moy, and Cha les E. Kahn J . "Checklis o a i icial
in elligence in medical imaging (CLAIM): a guide o au ho s and e iewe s."
Radiology: A i icial In elligence 2.2 (2020): e200029.
[170] S. S -Jean, M.A. Vie ge e , A. Leemans, Ha moniza ion o di usion MRI da a se s
wi h adap i e dic iona y lea ning, Human b ain mapping 41 (16) (2020)
4478–4499.
Y. Nan e al.