A dataset of synthetic, maturation-informed magnetic resonance images of the human fetal brain

1
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
a da ase o syn he ic, ma u a ion-
in o med magne ic esonance
images o he human e al b ain
Hélène Lajous1,2 ✉ , and és Le Boeu Fló1,2,3, Ped o M. Go daliza1,2, Osca Es eban 1,
Fe an Ma ques3, Vincen Dune 1, Mé iam Koob1 & Me i xell Bach Cuad a 1,2
Magne ic esonance imaging (MRI) is a powe ul modali y o in es iga ing abno mal de elopmen al
pa e ns in u e o. Howe e , since i is no he i s -line diagnos ic ool in his sensi i e popula ion,
da a emain sca ce and he e ogeneous ac oss scanne s and hospi als. o add ess his, we p esen a
no el da ase o syn he ic images ep esen a i e o eal e al b ain MRI. Ou da ase comp ises 594
wo-dimensional, low- esolu ion se ies o 2-weigh ed images co esponding o 78 de eloping human
e al b ains be ween 20.0 and 34.8 weeks o ges a ional age. Da a a e gene a ed using a new e sion
o he Fe al B ain MR acquisi ion Nume ical phan om (FaBiAN) o accoun o local whi e ma e
he e ogenei ies h oughou ma u a ion. Bo h heal hy and pa hological ana omies a e simula ed
wi h s anda d clinical se ings. wo independen adiologis s quali a i ely assessed he ealism
o he simula ed images. A quan i a i e analysis con i ms an enhanced ideli y compa ed o he
o iginal e sion o he so wa e, wi h u he alida ion h ough i s applicabili y o e al b ain issue
segmen a ion. he coho is publicly a ailable o suppo he con inuous endea o o de eloping
ad anced pos -p ocessing me hods as well as cu ing-edge a i icial in elligence models.
Backg ound & Summa y
The e is a g owing awa eness o he impo ance o ea ly b ain ma u a ion on heal h la e in li e as he unde lying
complex, in e connec ed s uc u al and unc ional p ocesses can be al e ed by a ious gene ic and en i onmen-
al ac o s1–9. Accu a e cha ac e iza ion o in u e o de elopmen is he e o e c i ical.
Magne ic esonance imaging (MRI) is an eme gen adjunc o ul asound (US) in cases o diagnos ic ambi-
gui y, and o comp ehensi e diagnos ic, p ognos ic, and pos na al managemen planning10. MRI is adequa e
o explo ing he de eloping e al b ain due o i s excellen con as in so issue while being minimally in a-
si e. Clinical guidelines ecommend he acquisi ion o T2-weigh ed (T2w) as spin echo (FSE) sequences
o wo ka ound unp edic able e al mo ion du ing he exam. In p ac ice, a leas h ee o hogonal se ies o
wo-dimensional (2D) hick slices a e acqui ed o p o ide in o ma ion on he whole b ain olume wi h su -
icien signal- o-noise a io (SNR)11. Despi e all coun e -measu es and op imiza ions, e al MRI emains chal-
lenging due o mo ion a e ac s and ela ed signal d ops, as well as low SNR in small s uc u es wi hin he
ma u ing b ain su ounded by he mo he ’s womb and he amnio ic luid. Since MRI is second-line and no he
e e ence-s anda d echnique (i.e, US) o he ollow-up o he e us du ing p egnancy, la ge-scale clinical da a-
se s a e ela i ely sca ce in his coho o sensi i e subjec s. Besides, he e is no s anda dized imaging p o ocol
ac oss si es, which has esul ed in la ge a iabili y be ween MR schemes ac oss scanne s, s udies12, and e en
mo e so be ween endo s. Such disc epancies may esul in highly a iable MR con as s and image quali y.
Indeed, mos o he da a a ailable oday a e hea ily pos -p ocessed and in eg a ed in o spa io- empo al MRI
a lases o he e al b ain, ei he heal hy10,13,14 o pa hological15. This enables a ine ep esen a ion o he de elop-
ing b ain h oughou ges a ion. Howe e , such a lases a e age b ain scans ac oss se e al e uses a a gi en ges a-
ional age (GA), hus esul ing in high- esolu ion (HR) images a om a ealis ic clinical se -up, wi h smoo hed
in e -indi idual he e ogenei ies and ea u es.
Recen ly, he Fe al Tissue Anno a ions (FeTA) da ase has been p oposed as a benchma k o au oma ed
mul i- issue e al b ain segmen a ion16–19. Howe e , only supe - esolu ion (SR) econs uc ions20,21 o he e al
1Depa men o Radiology, Lausanne Uni e si y Hospi al and Uni e si y o Lausanne, Lausanne, Swi ze land.
2ciBM cen e o Biomedical imaging, Lausanne, Swi ze land. 3Depa men o Signal heo y and communica ions,
Uni e si a Poli écnica de ca alunya, Ba celona ech, Ba celona, Spain. ✉e-mail: [email p o ec ed]
DA A DeScRIp OR
OPEN
2
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
b ain olume and hei associa ed semi-au oma ed anno a ions ha e been made publicly a ailable, bu no he
o iginal clinical acquisi ions. In ac , o da e, no da abase p o ides anno a ed low- esolu ion (LR) se ies o he
e al b ain.
Thanks o hei abili y o p o ide a lexible and con olled en i onmen ha acili a es accu a e, obus , and
ep oducible esea ch, compu e simula ions a e widely used o MR de elopmen s o mi iga e da a sca ci y and
pos -p ocessing complexi y22–28.
While ecen ad ances in gene a i e a i icial in elligence (AI, such as GANs, VAEs, and di usion models)
ha e shown p omising esul s in he syn hesis o medical images29,30, ou physics-guided simula ion app oach
was speci ically chosen o p o ide p ecise con ol o e he acquisi ion pa ame e s o MRI and issue p ope -
ies h oughou he ma u a ion o he e al b ain. Al hough condi ional gene a i e models could po en ially
be ained o handle such pa ame e s31, hey would equi e ex ensi e pai ed aining da a co e ing all possi-
ble pa ame e combina ions ac oss GA - da a ha is pa icula ly sca ce in his sensi i e popula ion32. In his
con ex , we demons a ed ha syn he ic, ye ealis ic da a can e icien ly complemen sca ce clinical da ase s,
p o iding aluable suppo o da a-demanding deep lea ning (DL) models o e al b ain MRI issue segmen-
a ion28,33,34, as well as he op imiza ion o ad anced econs uc ion echniques28,35–37. These explo a o y s udies
we e based on he i s Fe al B ain magne ic esonance Acquisi ion Nume ical phan om (FaBiAN) ha simula es
as closely as possible he FSE sequences used in clinical ou ine o e al b ain examina ion o gene a e ealis ic
T2w images o he e al b ain h oughou ma u a ion om a a ie y o segmen ed HR ana omical images o
heal hy and pa hological subjec s28. Despi e a good issue con as , he syn he ic T2w MR images used in his
wo k we e o iginally de i ed om a h ee-class model o he e al b ain (g ay ma e (GM), whi e ma e (WM),
and ce eb ospinal luid (CSF)) ha does no allow o cap u e key ma u a ion p ocesses and me abolic changes
occu ing in WM issues ac oss ges a ion.
In he wake o his i s p o o ype38, he p oposed da a desc ip o showcases a ull da ase o highly ealis ic
in silico da a composed o :
• 594 syn he ic T2w MR images co esponding o 78 de eloping e al b ains, de i ed om HR anno a ions o
SR- econs uc ed, eal clinical da a acqui ed on a ious MR scanne s and ollowing he di e en clinical p o-
ocols in place a Lausanne Uni e si y Hospi al (CHUV) and a Uni e si y Child en’s Hospi al Zu ich (Kispi);
• au oma ically-gene a ed b ain masks and e al b ain anno a ions o he LR se ies;
• a co esponding SR econs uc ion o e e y subjec .
This da ase is based on a nume ical model o he de eloping e al b ain ha accoun s o he p onounced
MR signal changes ep esen a i e o issue he e ogenei y wi hin WM s uc u es h oughou ma u a ion39. Ou
me hodological con ibu ion (FaBiAN 2.040) hus iden i ies local spa ial egions o a iable wa e con en
wi hin a WM mask o simula e mo e ealis ic in u e o MR images.
Figu e1 p o ides a schema ic o e iew o he s udy design p esen ed in his wo k. We speci ically alida e
he da a as ollows. Fi s , we epo a la ge-scale, independen , quali a i e assessmen o he ealism o he newly
DATASET
TECHNICAL VALIDATION
Syn he ic low- esolu ion (LR) T2-weigh ed
images o he de eloping e al b ain
N=78 subjec s, 594 o hogonal LR se ies
Label maps
au oma ically p opaga ed
by nea es -neighbo in e pola ion
Supe - esolu ion (SR) econs uc ion
N=78 subjec s
Recons uc ion me hod: MIALSRTK o SVRTK
(i) Quali a i e e alua ion o he ealism
o he simula ed LR se ies
N=29 subjec s
Simula ion amewo k: FaBiAN 1.0 and 2.0
O hogonal LR se ies: 3 o e e y subjec
Independen expe s: 2
(ii) Quan i a i e compa ison be ween
simula ed and o iginal clinical images
N=29 subjec s
Simula ion amewo k: FaBiAN 1.0 and 2.0
SR econs uc ion me hod: MIALSRTK o SVRTK
Me ics: mu ual in o ma ion
(iii) Le e aging in silico da a o au oma ed e al
b ain mul i- issue segmen a ion wi h deep lea ning
N=70 subjec s
Simula ion amewo k: FaBiAN 2.0
SR econs uc ion me hod: MIALSRTK o SVRTK
Me ics: Dice simila i y coe icien
1.0 2.0 2.0
Fig. 1 Schema ic o e iew o he da ase p omo ed h oughou his wo k and o he s udy design o alida e i s
ele ance o he e al b ain MRI communi y.
3
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
simula ed images (FaBiAN 2.0) compa ed o he o me implemen a ion (FaBiAN 1.2) by wo neu o adiol-
ogy and pedia ic adiology expe s. Second, we compa e quan i a i ely he SR olumes econs uc ed om
eal clinical cases and om hei co esponding syn he ic LR images gene a ed using FaBiAN 1.2 and 2.0.
Finally, we demons a e a po en ial use-case o he da ase wi h an in silico augmen a ion o aining samples o
au oma ed e al b ain mul i- issue segmen a ion o in u e o MR acquisi ions, h ough he simula ion o a b oad
a ie y o sequence pa ame e s using FaBiAN 2.0.
This s udy aims a demons a ing he ele ance o such an ex ensi e da ase o syn he ic, ye highly ealis ic
T2w MR images o he e al b ain h oughou ma u a ion o a communi y s uggling wi h da a sca ci y in his
sensi i e popula ion ha equi es comp ehensi e e hical o e sigh o acqui e new da a as well as expe ienced
MR echnologis s. As such, ha ing access o mul iple MR images gene a ed om a ious se ings and in di e en
clinical scena ios has a g ea po en ial euse alue o u he de elopmen s o ad anced pos -p ocessing me h-
ods as well as cu ing-edge a i icial in elligence models.
Me hods
Da a. The inpu ana omical models used o gene a e all he LR se ies o he e al b ain wi h local WM changes
ac oss ma u a ion o he Technical Valida ion and Reuse Po en ial sec ions we e ex ac ed om he publicly
a ailable FeTA Da ase wi h e ined anno a ions41,42. The FeTA da ase ga he s 90 subjec s in o al17, om which
wo we e excluded a e quali y con ol due o he bad quali y o he SR econs uc ion, hence he co esponding
segmen a ions ha e no been manually co ec ed in he e ined da ase 42. The emaining 88 subjec s (34 neu o yp-
ical and 54 pa hological) a e in he GA ange o 20.0 o 34.8 weeks (27.0±3.58 weeks). Diagnosis o he pa hol-
ogy, absen om he o iginal anonymized FeTA da ase elease, has been assessed by wo adiologis s and sha ed
along his upda ed da ase 42: common pa hological condi ions such as spina bi ida p e- and pos -su ge y (n =
36)and en iculomegaly (n = 8), as well as he e o opia (n = 5), e al in ac anial cys ic lesions (n = 3), high- low
du al sinus mal o ma ion (n = 1), aqueduc al s enosis (n = 1), and ce ebella hemo hage (n = 1) a e included.
In his s udy, we p o ide simula ed da a based on 78 subjec s om he FeTA Da ase wi h e ined anno-
a ions41,42. Fo he sake o ep oducibili y, he co esponding subjec IDs a e p o ided asSupplemen a y
In o ma ion. On he one hand, 29 subjec s (neu o ypical: 16, pa hological: 13) we e ini ially simula ed o he
echnical alida ion, which consis ed o a quali a i e e alua ion o he ealism o he simula ed LR se ies and
a quan i a i e compa ison be ween he simula ed and o iginal clinical images. These heal hy and pa hological
subjec s we e andomly selec ed o span a b oad ange o GA, om 20.1 o 34.8 weeks. On he o he hand, 70
subjec s (neu o ypical: 27, pa hological: 43; GA ange: 20.0-34.8 weeks) we e simula ed as o o m he aining
se o he Reuse Po en ial sec ion ha le e ages in silico da a o au oma ed e al b ain mul i- issue segmen a ion
wi h deep lea ning. Among hem, we e-used 21 subjec s om he echnical alida ion and simula ed 49 addi-
ional subjec s. The emaining 18 subjec s (neu o ypical: 7, pa hological: 11; GA ange: 20.9-33.1 weeks) ou o
he 88 subjec s om he FeTA da ase cons i u ed he pu e es ing se (no simula ed da a a e used o es ing).
Thus, ou da ase encompasses 78 simula ed subjec s o e all (neu o ypical: 31, pa hological: 47; GA ange: 20.0-
34.8 weeks). De ails peculia o he di e en subg oups used o he echnical alida ion and he euse po en ial
expe imen a e gi en in he co esponding sec ions.
The o iginal MR images we e acqui ed a Kispi a 1.5T (Signa Disco e y MR450, GE Heal hca e) o 3T (Signa
Disco e y MR750, GE Heal hca e), ei he using an eigh -channel ca diac coil o body coil. Mul iple se ies o 2D
hick slices we e scanned in o hogonal o ien a ions (axial, co onal and sagi al) wi h espec o he e al b ain
using a T2w Single-Sho FSE (SS-FSE) sequence (TR/TE, 2000–3500 ms/120 ms (minimum); lip angle, 90°;
sampling pe cen age, 55%; slice hickness, 3.00–5.00 mm; ield-o - iew, om 200×200 mm2 o 240×240 mm2;
acquisi ion ma ix, 1.5T: 256×224 oxels2, 3T: 320×224 oxels2; iso opic in-plane esolu ion o 0.5 mm)16.
Eigh issues a e segmen ed: WM, in a-axial CSF, ce ebellum, ex a-axial CSF, co ical GM, deepGM, b ains em,
and co pus callosum.
The da a used in his s udy we e acqui ed in ea lie s udies in acco dance wi h he ele an guidelines and
egula ions, unde he supe ision o E hics Boa ds composed o ep esen a i es a di e en le els (hospi als,
can ons, and ede al s a e). Mo he s o all e uses included in he cu en wo k p o ided w i en in o med con-
sen o he e-use o hei da a o esea ch pu poses.
Modeling o whi e ma e he e ogenei y and changes ac oss ma u a ion. Figu e2 gi es an o e -
iew o FaBiAN 2.0, ou la es de elopmen s o accoun o local WM he e ogenei ies h oughou e al b ain
ma u a ion. In pa icula , Fig.3 highligh s he majo changes implemen ed o modula e b ain issue elaxa ion
imes ac oss GA.
Due o he lack o cha ac e iza ions o no ma i e maps o T1 and T2 elaxome y measu emen s o e al b ain
s uc u es, all issues om he segmen ed e al b ain images a e me ged in o h ee main classes: GM, WM and
CSF. Acco ding o p e ious empi ical elaxome y alues p o ided in he li e a u e43–47, an a e age T1, espec-
i ely T2 alue is assigned o each o hese classes, wi hou conside a ion o he GA o spa ial a p io i on he ine
loca ion wi hin he b ain28 (see Fig.3(A)).
We ely on he Gaussian Hidden Ma ko Random Field (GHMRF) model, i ed using he
Expec a ion-Maximiza ion algo i hm, ollowing he app oach ou lined in FMRIB’s Au oma ed Segmen a ion
Tool (FAST48) o in eg a e local spa ial WM he e ogenei ies in ou nume ical ep esen a ion o he de eloping
e al b ain, and he e o e cap u e biophysical changes ha a ise ac oss ma u a ion. Conc e ely, we au oma i-
cally segmen a WM mask in o h ee classes using FAST 6.0.5.1 as illus a ed in Fig.3(B)): hyd a ed WM a eas
ha appea hype in ense on a T2w image, hypoin ense a eas ep esen ing dense WM ibe s, and a hi d class
in e p e ed as an in e media e be ween hyd a ed WM and ma u ing, dense WM ibe s. Pa ame e s we e se
heu is ically as ollows: 0.1 MRF egula iza ion weigh ; ou i e a ions o bias ield emo al; 20.0 mm ke nel in
bias ield smoo hing.
4
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
The segmen a ion ou pu consis s in 3D pa ial olume (PV) concen a ion maps o each oxel (i) wi hin
he h ee segmen ed classes. These PV maps a e used o ul ima ely weigh T1 and T2 elaxa ion imes locally.
Speci ically, he in e media e WM class is conside ed as a baseline: he posi i e di e ence be ween he PV maps
o hyd a ed, espec i ely dense WM and his baseline will be used o inc ease, espec i ely dec ease he a e age
e e ence T1 and T2 elaxa ion imes in e e y oxel o he WM mask. The co esponding weigh (w+, espec i ely
w−) compu ed in e e y oxel o he WM mask acco ding o Equa ion (1) can be displayed in so-called posi i e
and nega i e ewa d maps (Fig.3(B)).
Simula ed low- esolu ion se ies
Ana omical model
(3D high- esolu ion label map)
ges a ional age (GA)
pa hological s a us
Tissue elaxa ion imes
Tunable pa ame e s
2D FT
<
EPG simula ions
echo ime (ms)
signal in ensi y (a.u.)
T2decay in
e e y oxel
MR acquisi ion scheme
K-space sampling
pa ial
Fou ie
andom
Gaussian noise
igid mo ion
2D FT-1
TEe
α β β β β
T1(ms)
T2(ms)
modula ed bo h spa ially
and h oughou ges a ion
2.0
Whi e ma e (WM) he e ogenei y:
locally: WM segmen a ion in o h ee classes
ac oss GA: based on WM in ensi y dis ibu ion
2.0
accoun ing o WM changes h oughou de elopmen
2.0
MR scanne modeling
main magne ic ield s eng h
in ensi y non-uni o mi ies
Sequence pa ame e s
MR con as
geome y
esolu ion
signal- o-noise a io
accele a ion echnique
In e -slice mo ion
3D ansla ion and o a ion
STUPNIWOLFKROWTUPTUO
Fig. 2 Wo k low o simula ing as spin echo (FSE) acquisi ions o he e al b ain which be e cap u e
local WM changes h oughou ma u a ion (FaBiAN 2.0). The majo changes compa ed o he o iginal
implemen a ion o he so wa e a e highligh ed by he lag “ 2.0”.
5
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
=









=+ −
=− −∀∈
+
−
××
wwi PV PV
wi PV PV
[] 1max(0,)
[] 1max(0,)
i
(1)
hyd a edWM
iin e media eWM
i
in e media eWM
iWM ibe s
iWHD
whe e: w+ and w− a e espec i ely he posi i e and nega i e ewa d maps, PVhyd a edWM, PVin e media eWM, and
PVWM ibe s s and o he PV concen a ion maps o each o he h ee segmen ed WM classes, espec i ely he
hyd a ed WM, he in e media e WM class, and he dense WM ibe bundles. W, H, and D ep esen he wid h,
heigh , and dep h o he maps espec i ely.
The adjus men o he mean e e ence T1 and T2 elaxa ion imes should be app oached wi h cau ion hough
as subs an ial weigh s (w) may esul in un ealis ic elaxome ic p ope ies o he modeled e al b ain issues. To
add ess his conce n, a sigmoid modula ion cha ac e ized by a con ol alue α (see Fig.3(D)) is applied o he
o iginal T1 and T2 alues in o de o smoo hen local WM changes wi hin b ain s uc u es and ensu e ha he
co esponding issue p ope ies do no excessi ely de ia e om hei e e ence alues. Because he ange o WM
in ensi y changes is no uni o m ac oss ges a ion (see Fig.3(C)), α needs o be adjus ed acco ding o he GA. We
ely on he no ma i e spa io empo al MRI a las (STA) o he e al b ain which includes 18 subjec s spanning 21
o 38 weeks o GA10 o de e mine he op imal se ing ollowing:
_ = _ × , = 1,2
(D) Modula ion o e e ence T1and T2 alues o smoo h local WM changes ac oss GA
T1(ms) T2(ms)
GA: 28 weeksGA: 35 weeksGA: 21 weeks
GA: 21 weeks
T1(ms) T2(ms)
GA: 28 weeks
T1(ms) T2(ms)
GA: 35 weeks
T1(ms) T2(ms)
(B) Whi e ma e (WM) segmen a ion
in o h ee classes using FAST-FSL
hyd a ed WM in e media e WM dense ibe bundles
posi i e ewa d (w+)nega i e ewa d (w-)
--
p obabili y
mul iplica ion ac o w
= × 2
1+ −2∗ −1
(A)Re e ence T
1
and T
2
maps
h ee-class model in a ian o ges a ional age (GA)
(C)WM changes ac oss ma u a ion
GA (weeks)
con ol alue α
s anda d de ia ion o WM
in ensi y σWM
GA (weeks)
(+1)= 2− +1 × ( )
(E) Resul ing T1and T2 maps wi h a ying issue p ope ies wi hin he WM mask and ac oss GA
along wi h subsequen T2-weigh ed simula ions
Fig. 3 Modula ion o e e ence T1 and T2 alues in he WM o depic smoo h, local WM changes h oughou
e al b ain ma u a ion.

6
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
GA
GA
GA GA
GA GA
:[]
02 i 21
[1]2 []
[1]i {22, 23,,38},(2
)
GA WM
WM
αα ασ
σ
==











.=
−




−−





∈…
whe e σWM s ands o he es ima ed s anda d de ia ion o WM in ensi y (see Fig.3C). We empi ically se
α21=0.2, and i e a i ely compu e α co esponding o he GA in e al acco ding o Equa ion (2) o he 17
emaining subjec s o he STA. Fo con inous GA es ima ion, α(⋅) is de e mined by shape-p ese ing piece-wise
cubic in e pola ion o GA be ween 20.0 and 38.0 weeks. Gi en (2), we ob ain he new T1 and T2 (see Fig.3D)
om he e e ence Tk as:
TT wk() {1,2}, (3)
kk
GA
new=⋅ ∈
w
e
being()2
1
11,
(4)
GA GA w
GA
2
α=





+−






+
−α
Simula ed da ase s. High- esolu ion mul i- issue anno a ions om clinical expe s a e used as inpu mod-
els o he de eloping b ain o simula e ypical e al b ain FSE acquisi ions, ei he a 1.5T o 3T28. O e all, syn he ic
images o 78 subjec s a e p o ided ou o he 88 subjec s om41,42. These subjec s (n=78) we e andomly selec ed
o achie e a homogeneous dis ibu ion o disease condi ion and GA. Di e en subg oups a e used in he sec ions
Technical Valida ion and Reuse Po en ial: i) a quali a i e e alua ion o he ealism o he simula ed low- esolu ion
se ies and ii) a quan i a i e compa ison be ween simula ed and o iginal clinical images a e based on n=29 sim-
ula ions; he au oma ed e al b ain mul i- issue segmen a ion wi h deep lea ning expe imen is based on n=70
simula ions used o aining. Twen y-one subjec s a e sha ed in he simula ions o he h ee expe imen s. Fo
each subjec , T2w se ies o 2D hick slices a e simula ed in he h ee o hogonal o ien a ions (i.e., axial, co onal,
sagi al) wi h espec o he e al b ain posi ion. As ou inely pe o med in clinical p ac ice, wo pa ially o e lap-
ping LR se ies a e gene a ed in e e y o ien a ion o subsequen SR econs uc ion o he e al b ain olume. Since
he ealism o in e slice, 3D andom igid mo emen s o he e us encoded du ing k-space sampling achie ed
wi h FaBiAN 1.238 has al eady been alida ed28, he da ase showcased in his wo k is gene a ed wi hou addi-
ional mo emen no o induce any bias in he quali a i e e alua ion o he ealism o he gene a ed LR se ies by
he adiologis s, bu also o make i possible o align SR econs uc ions om he simula ed images o he co e-
sponding g ound u h label maps hey o igina e om.
Table1 epo s he ange o MR sequence pa ame e s ha de e mine he con as , he geome y, and he
esolu ion o ep esen a i e clinical low- esolu ion se ies acqui ed in clinical ou ine on di e en MR sys ems
(Hal -Fou ie Acquisi ion Single-sho Tu bo spin Echo, HASTE o Siemens, SS-FSE o GE scanne s), a ei he
1.5 o 3T, o sc een he in u e o de eloping b ain. These pa ame e s a e he e o e ep oduced o simula e he T2w
images o he e al b ain used h oughou his s udy.
Besides non-s anda dized MR sequence pa ame e s, only a ew s udies ha e in es iga ed T1 and T2 p ope -
ies o he de eloping b ain, ei he in u e o44,45,47 o in p e e m newbo ns43,46. We de e mined mean e e ence
T146,47 and T2 alues43–45 o GM and WM based on measu emen s pe o med a 1.5T, knowing ha bo h T1 and
T2 p ope ies dec ease ac oss b ain de elopmen (so om in u e o o pos na al li e) and ha T2 > T2*. Since
he biochemical composi ion o he CSF does no a y be ween childhood and adul hood, T1 and T2 alues o
his issue we e assumed o be simila in bo h popula ions, by ex ension also in e uses. Besides, we ex ended
hese e e ence alues o 3T by assuming ha T2 elaxa ion ime emains cons an ac oss clinical magne ic ield
s eng hs o a gi en s uc u e, while T1 elaxa ion ime inc eases by abou 25% in GM and 10% in bo h WM
and CSF49–53. As epo ed in Table2, we hen andomly gene a ed T1 and T2 alues using he RANDBETWEEN
unc ion o Mic oso Excel ( e sion 2108) o inc ease he di e si y in issue elaxa ion imes ac oss GA and
wi hin he h ee main classes modeled in he simula ed images, as well as o accoun o measu emen unce -
ain y in p e ious wo k. In p ac ice, he T2 alue was compu ed as a andom in ege numbe in he ange o he
mean e e ence T2 elaxa ion ime plus o minus i s s anda d de ia ion o bo h GM and WM. The T1 alue was
compu ed as he di e ence (i he T2 elaxa ion ime was se smalle han i s e e ence alue), espec i ely he
sum (i he T2 elaxa ion ime was se la ge han i s e e ence alue) be ween he a io be ween he es ima ed
and he e e ence T2 alues mul iplied by he e e ence T1 alue, and a andom in ege numbe be ween 0 and
he s anda d de ia ion o he e e ence T1 alue.
Da a Reco ds
The simula ed LR se ies o he 78 subjec s included in his s udy ha e been eleased on Zenodo54 oge he
wi h he co esponding b ain masks and SR econs uc ions. The LR mul i- issue anno a ions o he 29 sub-
jec s gene a ed o he quali a i e e alua ion as well as he esampled pa cella ions o he SR- econs uc ed sub-
jec s used as aining se in he euse po en ial expe imen a e also included. They a e p o ided as comp essed
NI TI images and o ganized in he B ain Imaging Da a S uc u e (BIDS) o ma 55. The Ma lab (Ma hWo ks,
R2019a) and Py hon sc ip s used o gene a e his da ase ha e also been made publicly a ailable (FaBiAN 2.040).
Mo eo e , he implemen a ion o FaBiAN 1.2 has been sligh ly modi ied o be con aine ized in o adocke
image (h ps://hub.docke .com/ /pe e mcgo / abian-docke ), and he e o e acili a e he simula ion o addi-
ional T2w MR images o he de eloping e al b ain om HR anno a ions. Wi h he pe spec i e o b oadly
dissemina ing his ool o he communi y, a simila docke image o FaBiAN 2.0 is a ailable in Docke Hub56.
7
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
echnical Valida ion
Quali a i e e alua ion o he ealism o he simula ed low- esolu ion se ies. Figu e4 illus a es
how locally adjus ing T1 and T2 alues wi hin WM issues enhances he esemblance o simula ed FSE images
wi h eal clinical MR acquisi ions o he e al b ain ac oss ma u a ion. I shows a compa ison be ween syn he ic
LR se ies gene a ed using bo h e sions o FaBiAN om he segmen ed, SR- econs uc ed e al b ain olumes o
h ee ep esen a i e subjec s and he co esponding clinical images acqui ed a CHUV a 1.5T, in bo h heal hy
and pa hological e uses o 21, 31, and 33 weeks o GA espec i ely. The modula ion o T1 and T2 p ope ies
acco ding o he wa e /myelin con en o WM issues esul s in local a ia ions o he MR con as allows o be e
depic he complexi y o WM and he unde lying ma u a ion p ocesses. Fo ins ance, he mig a ion o neu ons
om he ge minal ma ix o he co ex du ing he i s wo imes e s o ges a ion is e lec ed by a mul ilaye
aspec o WM, which is inely cap u ed in syn he ic images gene a ed using FaBiAN 2.0. In con as , WM is-
sues appea highly homogeneous in he LR se ies simula ed by FaBiAN 1.2 (see Fig.4- op ow, in he co onal
o ien a ion).
Expe imen al design. A pedia ic neu o adiologis and a neu o adiologis om CHUV wi h 17 and 14 yea s o
expe ience espec i ely, p o ided independen , blind, quali a i e assessmen o he ealism o he syn he ic T2w
LR se ies o he e al b ain gene a ed o 29 subjec s (16 neu o ypical and 13 pa hological) in he GA ange o
20.1 o 34.8 weeks (27.0±4.20 weeks) by bo h he o iginal e sion o he so wa e (FaBiAN 1.2)38 and ou new
nume ical WM model (FaBiAN 2.0)40. Young e uses diagnosed wi h spina bi ida p io su ge y we e excluded
om his e alua ion as he absence o ex a-axial CSF subs an ially al e s he quali y o he clinical acquisi ions
and subsequen simula ions. Simula ions we e un a 1.5T (25 subjec s) and 3T ( ou subjec s), he main di -
e ence being he highe spa ial esolu ion and highe SNR ha can be eached a 3T. High-quali y LR se ies
MR sequence HASTE SS-FSE
Magne ic ield s eng h (T) 1.5 3 1.5 3
Con as
E ec i e echo ime (ms) 82-90 101 116-124 117-123
Echo spacing (ms) 4.08 10 10 10
Echo ain leng h 134 126 224 224
Exci a ion lip angle (°) 90 90 90 90
Re ocusing pulse lip angle (°) 149-180 180 150-180 150-180
Geome y
Slice hickness (mm) 3 3 3 3
Slice gap (mm) 0.3 0 0 0
Phase o e sampling (%) 80 63 0 0
Shi o he ield-o - iew (mm) 0 0 ±1.6 ±1.6
Resolu ion
Field-o - iew in he eadou di ec ion
(mm)360 350 240-300 240-320
Base esolu ion ( oxels) 320 320 256 256
Phase esolu ion (%) 70 75 100 100
Recons uc ion ma ix ( oxels) 320 640 512 512
Ze o-in e pola ion illing none yes yes yes
Accele a ion echnique
Accele a ion ac o 2 2 1 1
Re e ence lines 42 24 0 0
Noise
Mean 0 0 0 0
S anda d de ia ion 0.05 0.004 0.03-0.07 0.002-0.006
Table 1. Acquisi ion pa ame e s a clinical magne ic ield s eng h (1.5 and 3T) o wo MR sequences (HASTE
and SS-FSE) commonly used o e al b ain examina ion, and ep oduced o simula e T2w images o he
de eloping e al b ain.
T1 (ms) T2 (ms)
a 1.5T a 3T a 1.5T and 3T
G ay ma e 2200±150 2750±150 182±10
Whi e ma e 2700±300 2970±300 285±15
Ce eb ospinal luid 4000 4400 2000
Table 2. Range o e al b ain issue p ope ies simula ed in his s udy a 1.5T and 3T.
8
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
we e simula ed wi h high SNR and wi hou mo ion no o bias he e alua ion o he adiologis s wi h co up ed
images. Visual inspec ion and na iga ion h oughou he di e en se ies o he e al b ain we e made possible
ia ITK-SNAP57, displaying simula ions in he same o ien a ion plane bu gene a ed by bo h e sions o he so -
wa e in wo di e en windows. Bo h expe ienced adiologis s we e asked o e alua e: i) which one o bo h se ies
appea s he mos ealis ic (Expe imen 1), ii) how ealis ic and close o MR images acqui ed in clinical ou ine
e e y o hese simula ed se ies looks, aken independen ly, wi h a special ocus on WM appea ance, based on a
i e-poin Like scale ( om 1: poo ly ealis ic, o 5: highly ealis ic) (Expe imen 2).
Expe imen 1: Compa ison be ween FaBiAN 1.2 and FaBiAN 2.0. This alida ion aims a compa ing which
e sion o ou nume ical phan om p o ides he mos ealis ic LR se ies o he e al b ain h oughou ma u-
a ion in e e y o hogonal o ien a ion. Figu e5 ( op ow) illus a es a which equency e e y expe a ed
bo h e sions o FaBiAN as p o iding he mos ealis ic LR se ies compa ed o each o he ac oss de elopmen .
Unanimously, FaBiAN 2.0 gene a ed mo e ealis ic LR se ies o e ges a ion han FaBiAN 1.2, o all subjec s
acco ding o ou neu o adiologis , espec i ely in mo e han 96.5% o he cases o ou pedia ic neu o adiolo-
gis . In e es ingly, he h ee ou o 87 se ies gene a ed by FaBiAN 1.2 ha hey assessed as mo e ealis ic han
he ones simula ed by FaBiAN 2.0 co esponded o images o young e uses (i.e., below 26 weeks o GA). A
his ea ly s age o de elopmen , he di e ences in he simula ed images be ween bo h e sions o FaBiAN may
be mo e sub le o dis inguish.
Expe imen 2: Independen assessmen . This expe imen aims o independen ly e alua e he ealism o he
LR se ies o he de eloping e al b ain simula ed by FaBiAN 1.2 and FaBiAN 2.0. The iolin plo s in Fig.5
(bo om ow) display he dis ibu ion o a ings by each expe o all he syn he ic LR se ies gene a ed using
bo h phan om e sions. They show dispe si y in he ealism o he simula ed images, al hough LR se ies gene -
a ed by FaBiAN 2.0 we e e alua ed as he mos ealis ic o e all. These independen a ings o all he syn he ic
images p o ided, wi hou compa ison o bo h e sions o he phan om o each o he , u he suppo he ind-
ings in e ed om he i s expe imen , namely ha ou la es de elopmen s (FaBiAN 2.0) make i possible o
gene a e e en mo e ealis ic images han he o iginal p o o ype (FaBiAN 1.2) by cap u ing local changes wi hin
WM issues h oughou ma u a ion.
Bo h o hese expe imen s alida e ha he p oposed da ase gene a ed using FaBiAN 2.0 p o ides highly
ealis ic T2w MR images o he b ain h oughou in u e o de elopmen , and no only in olde e uses whe e
ma u a ion p ocesses a e u he ad anced.
Quan i a i e compa ison be ween simula ed and o iginal clinical images. This sec ion aims a
quan i ying he simila i y be ween simula ed (FaBiAN 1.2, espec i ely FaBiAN 2.0) and eal clinical da a.
GA: 21 weeks
neu o ypical
axial iew
GA: 31 weeks
pa hological
sagi al iew
GA: 33 weeks
neu o ypical
co onal iew
Real clinical
acquisi ions
Simula ed T2-weigh ed low- esolu ion se ies
FaBiAN 2.0FaBiAN 1.0
Fig. 4 Compa ison be ween syn he ic low- esolu ion se ies gene a ed in he di e en o hogonal o ien a ions
using FaBiAN 1.0 (le column) and 2.0 (middle column) and h ee ep esen a i e clinical as spin echo
acquisi ions ( igh column) a 1.5T in bo h heal hy and pa hological e uses o 21 ( op line), 31 (middle line),
and 33 weeks (bo om line) o ges a ional age (GA) espec i ely scanned a CHUV. Red a ows highligh a eas
whe e whi e ma e he e ogenei ies esul in MR con as a ia ions ha a e ep esen a i e o he ongoing
ma u a ion p ocesses. These whi e ma e changes a e well cap u ed by FaBiAN 2.0 simula ions, which
he e o e look mo e ealis ic and close o clinical acquisi ions compa ed o images gene a ed by FaBiAN 1.2.
9
Scien i ic Da a | (2025) 12:602 | h ps://doi.o g/10.1038/s41597-025-04926-9
www.na u e.com/scien i icda a
www.na u e.com/scien i icda a/
Since SR econs uc ions a e o g ea alue o u he quan i a i e analysis o he de eloping e al b ain, and
because manual anno a ions o hese da a a e a ailable, we compa e he simila i y be ween SR econs uc ions
om simula ed and eal cases (n=29 subjec s), wi h a special ocus on WM issues.
Pa ially-o e lapping o hogonal LR se ies ( wo in each acquisi ion plane) o he same 29 subjec s inspec ed
du ing quali a i e e alua ion by he adiologis s we e gene a ed and combined o econs uc an SR olume o
he e al b ain o he co esponding clinical cases using wo di e en pipelines, he Image Regis a ion Toolki
SVRTK20 unde Licence om Ixico L d., o MIALSRTK21), as in16. O e all, 21 subjec s (nine neu o ypical and
12 pa hological in he GA ange o 20.1 o 34.8 weeks (26.4±4.06 weeks) we e econs uc ed using SVRTK,
eigh subjec s (se en neu o ypical and one pa hological in he GA ange o 21.2 o 33.4 weeks (28.5±4.2 weeks)
wi h MIALSRTK espec i ely. Fo e e y subjec , mu ual in o ma ion (MI) was compu ed be ween he SR econ-
s uc ions om syn he ic da a (FaBiAN 1.2 agains FaBiAN 2.0), and be ween he SR econs uc ions om
simula ed (FaBiAN 1.2 o FaBiAN 2.0, espec i ely) and eal clinical da a, as a measu e o simila i y be ween
hem, wi h a ocus on he WM mask. The highe he MI be ween wo images, he close hey a e.
Expe imen 1: Compa ison be ween FaBiAN 1.0 and FaBiAN 2.0
Expe imen 2: Independen assessmen
O e all equency:
FaBiAN 1.0:3 se ies (1 axial, 2 co onal) 0 se ies
FaBiAN 2.0:84 se ies (> 96.5%) 87 se ies (100%)
Fig. 5 Quali a i e e alua ion o he ealism o he low- esolu ion (LR) se ies simula ed using FaBiAN 1.2
and FaBiAN 2.0 o 29 subjec s, in he h ee o hogonal o ien a ions, by wo independen expe s in pedia ic
neu o adiology (Expe #1) and in neu o adiology (Expe #2). Expe imen 1 ( op ow) consis ed in compa ing
images gene a ed o e e y subjec by bo h implemen a ions o he simula ion amewo k. The cumula i e
coun o he mos ealis ic syn he ic LR se ies is ep esen ed o e he ges a ional age, acco ding o each expe ,
showing a ne p e e ence o FaBiAN 2.0 (be ween 96% and 100% o he images e alua ed as he mos ealis ic
by Expe 1 and Expe 2 espec i ely) o e FaBiAN 1.2 h oughou e al b ain ma u a ion. Expe imen 2
(bo om ow) u he aimed a e alua ing he ealism o he simula ed LR se ies independen ly. The iolin plo s
show he dis ibu ion o he a ings (be ween 1-poo ly ealis ic and 5-highly ealis ic) o e e y image by each
expe . O e all, he LR se ies simula ed by FaBiAN 2.0 we e a ed as mo e ealis ic han he ones gene a ed
using FaBiAN 1.2.