Imaging biomarkers of cortical neurodegeneration underlying cognitive impairment in Parkinson’s disease

ORIGINAL ARTICLE
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
h ps://doi.o g/10.1007/s00259-025-07070-z
20% o PD pa ien s (PD-MCI), and long- e m epidemiologi-
cal s udies ha e shown ha he la ge majo i y (~ 80%) o PD
pa ien s will e en ually de elop a PD-associa ed demen ia
synd ome (PDD) [2, 3]. The unde lying mechanisms o
cogni i e decline and demen ia associa ed wi h PD emain
unclea . Co ical in ol emen o Lewy body pa hology
In oduc ion
Cogni i e impai men (CI) is one o he mos common and
de as a ing non-mo o symp oms o Pa kinson’s disease
(PD) [1]. A disease onse , symp oms o mild cogni i e
impai men (MCI) a e al eady p esen in app oxima ely
Jesús Sil a-Rod íguez and Miguel Ángel Lab ado -Espinosa equally
con ibu ed o he cu en s udy and sha e i s au ho ship.
Ex ended au ho in o ma ion a ailable on he las page o he a icle
Abs ac
Pu pose Imaging bioma ke s bea g ea p omise o imp o ing he diagnosis and p ognosis o cogni i e impai men in Pa -
kinson’s disease (PD). We compa ed he abili y o h ee commonly used neu oimaging modali ies o de ec co ical changes
in PD pa ien s wi h mild cogni i e impai men (PD-MCI) and demen ia (PDD).
Me hods 53 cogni i ely no mal PD pa ien s (PD-CN), 32 PD-MCI, and 35 PDD unde wen concu en s uc u al MRI
(sMRI), di usion-weigh ed MRI (dMRI), and [18F]FDG PET. We ex ac ed g ey ma e olumes (sMRI), mean di usi -
i y (MD, dMRI), and s anda dized up ake alue a ios ([18F]FDG PET) o 52 co ical egions included in a neu oana-
omical a las. We assessed g oup di e ences using ANCOVA models and u he applied a c oss- alida ed machine lea ning
app oach o iden i y he modali y-speci ic b ain egions ha a e mos indica i e o demen ia s a us and assessed hei diag-
nos ic accu acy o g oup sepa a ion using ecei e ope a ing cha ac e is ic analyses.
Resul s In sMRI, a ophy o empo al and pos e io -pa ie al a eas allowed sepa a ing PDD om PD-CN (AUC = 0.77 ± 0.07),
bu diagnos ic accu acy was poo o sepa a ing PD-MCI om PD-CN (0.57 ± 0.10). dMRI showed mos p onounced di u-
si i y changes in he medial empo al lobe, which p o ided excellen diagnos ic pe o mance o PDD (AUC = 0.87 ± 0.06),
and a mo e modes bu s ill signi ican pe o mance o PD-MCI (AUC = 0.71 ± 0.09). Finally, [18F]FDG PET e ealed p o-
nounced hypome abolism in pos e io -occipi al egions, which p o ided he highes diagnos ic accu acies o bo h PDD
(AUC = 0.89 ± 0.05) and PD-MCI (AUC = 0.78 ± 0.05). In s a is ical compa isons, bo h [18F]FDG PET (p < 0.001) and dMRI
(p < 0.031) ou pe o med sMRI o de ec ing PDD and PD-MCI.
Conclusion Among he es ed modali ies, [18F]FDG PET was mos accu a e o de ec ing co ical changes associa ed wi h
cogni i e impai men in PD, especially a ea ly s ages. Di usion measu emen s may ep esen a p omising MRI-based
al e na i e.
Keywo ds Pa kinson’s disease · Cogni i e decline · [18F]FDG PET · MRI · DTI · Hypome abolism · A ophy · Mean
di usi i y
Recei ed: 13 June 2024 / Accep ed: 30 Decembe 2024
© The Au ho (s) 2025
Imaging bioma ke s o co ical neu odegene a ion unde lying
cogni i e impai men in Pa kinson’s disease
JesúsSil a-Rod íguez1,2,3· Miguel ÁngelLab ado -Espinosa2,3,4· Sand aCas o-Lab ado 1,2,3·
Lau aMuñoz-Delgado2,3· PabloF anco-Rosado2,3· Ana Ma íaCas ellano-Gue e o2· DanielMacías-Ga cía2,3·
Sil iaJesús2,3· As id D.Ada mes-Gómez2,3· Fá imaCa illo2,3· Juan F anciscoMa ín-Rod íguez2,3,5·
Da idGa cía-Solís6· Flo indaRoldán-Lo a7· PabloMi 2,3,9· Michel J.G o he1,2,3,8
1 3
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
is conside ed a key ea u e, bu mul iple mechanisms a e
likely in ol ed [4]. Thus, he iden i ica ion o bioma ke s
ha may iden i y b ain changes associa ed wi h PD- ela ed
CI could p o ide impo an insigh s in o he in ol ed pa ho-
logical p ocesses and imp o e accu a e diagnosis [5].
In ecen yea s, he e has been g owing in e es in using
neu oimaging o de ec ea ly co ical changes associa ed
wi h PD-CI [6, 7]. S uc u al magne ic esonance imaging
(sMRI) has been ex ensi ely employed o measu e g ey
ma e a ophy in PD-CI, ye he indings ha e been incon-
sis en ac oss s udies [7, 8]. While some s udies ha e iden-
i ied a ophy in a ious cogni ion- ele an a eas [9–11],
o he s ha e ailed o eplica e hese indings [12, 13]. Mo e-
o e , he speci ic co ical a eas epo ed o exhibi a ophy
ha e a ied widely ac oss di e en s udies [7]. By con as ,
[18F] luo odeoxyglucose ([18F]FDG) posi on emission
omog aphy (PET) has e ealed a mo e consis en pa e n
o pos e io -occipi al b ain hypome abolism associa ed wi h
PD-CI, which can al eady be obse ed in PD-MCI pa ien s
se e al yea s be o e p og ession o demen ia [14, 15].
While di ec compa isons be ween [18F]FDG PET and
MRI a e a e, a ecen me a-analysis ac oss uni-modal PET
and MRI imaging s udies concluded ha hypome abolism
on [18F]FDG PET is a mo e sensi i e and consis en imag-
ing ma ke o he neu odegene a i e changes in PD [16].
Simila ly, one p e ious mul imodal imaging s udy di ec ly
compa ing he pe o mance o sMRI and [18F]FDG PET
in PD-CI concluded ha changes in hypome abolism p e-
cede and exceed a ophy in he cou se o he disease [17],
especially in pos e io -occipi al egions. Mo e ecen ly, is-
sue di usion indices de i ed om di usion-weigh ed MRI
(dMRI) ha e shown p omise in de ec ing ea ly mic os uc-
u al al e a ions ha occu be o e appa en a ophy on sMRI
in he cou se o PD [13, 18, 19]. Conside ing he compa-
ably limi ed a ailabili y o [18F]FDG PET, dMRI may be
an a ac i e, MRI-based al e na i e o assessing PD-CI in
he clinic. Howe e , mul imodal imaging s udies allowing
a di ec compa ison be ween [18F]FDG PET and dMRI a e
s ill lacking.
The aim o he cu en s udy was o p o ide a di ec com-
pa ison o he sensi i i y and diagnos ic u ili y o sMRI,
dMRI, and [18F]FDG PET as neu oimaging bioma ke s o
PD-CI, using a ela i ely la ge coho o well-pheno yped
PD pa ien s spanning he clinical con inuum om cogni i e
no mali y (PD-CN) o e PD-MCI o PDD. Addi ionally, we
also e alua ed he po en ial o combining mul imodal imag-
ing in o ma ion o imp o e he de ec ion o pa hological
changes associa ed wi h CI in PD.
Me hods
Pa icipan s
Ou s udy sample included 120 pa ien s wi h PD p ospec-
i ely ec ui ed by he Mo emen Diso de s Uni a he
Uni e si y hospi al ‘Vi gen del Rocío’ in Se ille, Spain. PD
diagnosis was pe o med ollowing he Mo emen Diso -
de Socie y (MDS) Clinical Diagnos ic C i e ia [20], and
cogni i e pe o mance was e alua ed using he Pa kinson’s
Disease Cogni i e Ra ing Scale (PD-CRS). The sample was
in en ionally en iched o PD-CI du ing ec ui men , and
pa ien s we e diagnosed as ha ing PD wi h no mal cogni-
ion (PD-CN, PD-CRS > 81; N = 53), PD-MCI (81 ≥ PD-
CRS > 64; N = 32) [21], o PDD (PD-CRS ≤ 64 + con i med
unc ional impai men ; N = 35) [22]. The Mon eal Cog-
ni i e Assessmen (MoCA) scale was used as a comple-
men a y measu emen o global cogni i e pe o mance.
Cogni i e assessmen was pe o med while pa ien s we e
in he ON s a e. Pa ien s on choline gic ea men we e no
excluded, bu complemen a y analyses excluding his g oup
a e p esen ed. Mo o s a us was assessed using Pa III o
he Uni ied Pa kinson’s Disease Ra ing Scale (UPDRS-III)
and he Modi ied Hoehn and Yah scale (H&Y), bo h con-
duc ed while he pa ien s we e in he OFF s a e, de ined as
a pe iod o a leas 12 h wi hou dopamine gic medica ion.
The s udy p o ocol was app o ed by he E hics Commi -
ee o he Uni e si y hospi al ‘Vi gen del Rocío’ (Da e:
08/01/2021, app o al numbe : 2158-N-20) acco ding o he
guidelines o he Helsinki decla a ion, and w i en in o med
consen was ob ained om all s udy pa icipan s.
Image acquisi ion
All pa ien s unde wen a mul imodal MRI acquisi ion
in a Philips Ingenia 3T MRI scanne . sMRI images we e
acqui ed using a high- esolu ion 3D T1-weigh ed (T1W
3D TFE SENSE) sequence (TR = 8.2 ms, TE = 3.75 ms,
lip angle = 8 µ, acquisi ion ma ix = 256 × 256 × 180, oxel
size = 0.94 × 0.94 × 1 mm3), while dMRI da a was acqui ed
using a di usion-weigh ed spin-echo echo-plana -imaging
(DWI-EPI) pulse sequence wi h 32 non-collinea di usion
g adien o ien a ions (b = 1000 s/mm2) in addi ion o a non-
weigh ed (B0) image. Resul ing dMRI images had a ield o
iew (FOV) o 224 × 224 mm2 o e a 128 × 128 ma ix wi h
60 slices. Slice hickness was 2 mm, wi h no gaps.
[18F]FDG PET acquisi ions we e pe o med a a sepa a e
isi in empo al p oximi y o he MRI scan (Δ = 0.3 ± 0.5
yea s). Pa ien s we e scanned o 20 min, 45 min a e he
injec ion o ∼200 Mbq o [18F]FDG. Due o a ha dwa e
upg ade du ing he ec ui men phase, acquisi ions we e
pe o med on wo di e en scanne s, a Siemens BioG aph
1 3
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
HiRes (19 pa ien s: 9 PD-CN, 5 PD-MCI, 5 PDD) and a
GE Disco e y MI (101 pa ien s: 44 PD-CN, 27 PD-MCI,
30 PDD). Recons uc ion was pe o med using he 3D i e a-
i e econs uc ion me hods implemen ed in each scanne
(Siemens: FORE + OSEM-2D; GE: VPHD), including co -
ec ions o a enua ion, sca e , and andom coincidences.
Image p ocessing and analysis
sMRI images we e segmen ed in o g ey ma e (GM), whi e
ma e (WM), and ce eb ospinal luid (CSF) pa i ions, and
spa ially no malized o Mon eal Neu ological Ins i u e
(MNI) space using he s anda dized ou ines p o ided by
he Compu a ional Ana omy Toolbox (CAT12) o he S a is-
ical Pa ame ic Mapping so wa e (SPM12, h p s : / / w w w . i l
. i o n . u c l . a c . u k / s p m / ) . [18F]FDG PET images we e n o m a l i z e
d o MNI space using SPM12 and smoo hed o an iso opic
8-mm esolu ion by applying di e en ial smoo hing al-
ues o each scanne (Siemens: 4.5 mm, GE: 6.5 mm) [23],
which we e calcula ed using a p e iously alida ed esolu-
ion es ima ion me hod aimed a calcula ing e ec i e image
esolu ion di ec ly om b ain images ins ead o using dedi-
ca ed phan oms [24]. Finally, a p e iously alida ed da a-
d i en his og am-based in ensi y no maliza ion algo i hm
was applied [25, 26].
dMRI olumes we e i s p ocessed wi h FSL ( h p s : /
/ w w w . m i b . o x . a c . u k / s l ) o co ec o head mo ion and
eddy-cu en dis o ions, and o emo e oxels ou side o
he b ain. Regis a ion-based dis o ion co ec ion was pe -
o med by applying di eomo phic ans o ma ions be ween
he dMRI images and sMRI using he Ad anced No maliza-
ion Tools (ANTs, h ps://gi hub.com/ANTsX/ANTs) [27].
A e p e-p ocessing, images we e econs uc ed using he
compu a ional ou ines p o ided by Pas e nak e al. [28] o
ob aining ee wa e -co ec ed mean di usi i y (MD) maps
[19]. MD has been epo ed o be a sensi i e di usion index
o mic os uc u al co ical changes ha may p ecede olu-
me ic changes [29], and ee-wa e co ec ion may play an
impo an ole in e alua ing co ical di usi i y, as i helps
o emo e CSF con amina ion, a pa icula ype o pa ial
olume e ec ha occu s along he con ou lines o he en-
icles and a ound he pe ime e s o he b ain pa enchyma in
oxels sha ed by CSF and b ain issue [28]. Taking ad an-
age o his p ocessing, we also de i ed ee wa e -co ec ed
ac ional aniso opy (FA) alues and he ac ion o ee
wa e in he b ain’s ex acellula space (FW) as comple-
men a y dMRI me ics [30]. MD, FA, and FW maps we e
ans o med o MNI space using he ans o ma ions o he
co- egis e ed sMRI. No malized images we e hen masked
using he segmen ed GM de i ed om CAT12.
A e p ep ocessing, o each modali y quan i a i e al-
ues we e ex ac ed wi hin 52 co ical egions-o -in e es
(ROI) as de ined in he Ha a d-Ox o d neu oana omical
a las. Fo sMRI, he GM olume o each ROI was calcu-
la ed by summing up he modula ed GM oxel alues wi hin
he ROI. Values we e no malized by he o al in ac anial
olume (TIV), calcula ed as he o al sum o GM, WM, and
CSF pa i ions. Mean [18F]FDG PET s anda dized up ake
alue a ios (SUVR) we e calcula ed as he a e age signal
o he in ensi y no malized PET image ac oss oxels wi hin
each ROI. Simila ly, a e age MD was ob ained as he mean
alue wi hin each o he a las-de ined ROIs.
S a is ical analysis
Demog aphic, clinical, and bioma ke a iables we e com-
pa ed using wo-sample - es s o no mally dis ibu ed
con inuous a iables and Fishe ’s exac es s o ca ego ical
a iables.
B ain-wide g oup-le el analysis was pe o med using
an ROI-based app oach ha allowed an objec i e compa i-
son o he egional e ec s be ween modali ies. Independen
ANCOVA compa isons (PD-CN s. PD-MCI, PD-CN s.
PDD) we e pe o med o each o he ROIs using he p e i-
ously calcula ed GM, MD, and SUVR alues. Age and mo o
symp om se e i y (UPDRS-III) we e used as co a ia es o
isola e changes ela ed o cogni ion. Resul s a e epo ed
as e ec size (Cohen’s d) a an FDR-co ec ed h eshold o
p < 0.05. Fo con enience, Cohen’s d alues we e de ined o
be posi i e in he di ec ion o inc eased neu odegene a ion
o each modali y (i.e., dec eased GM, dec eased SUVR,
inc eased MD). In addi ion o he g oup compa isons, we
also pe o med complemen a y pa ial Pea son co ela ion
analysis be ween he PD-CRS and egional GM, MD, and
SUVR alues (also using age and UPDRS-III as co a ia es).
To assess he diagnos ic pe o mance o each neu oimag-
ing modali y in a clinical scena io, we ollowed a machine
lea ning app oach based on a p e iously de eloped ame-
wo k o compa ing egional neu oimaging bioma ke s
om mul iple modali ies [31]. Supplemen a y Fig. 1 shows
a schema ic o he used me hodology. B ie ly, we i e a i ely
ained a penalized logis ic eg ession model wi h an elas-
ic ne penal y and es ed i s diagnos ic pe o mance ac oss
independen aining- es spli s o he s udy sample (1000
i e a ions). Fi s , he PDD and PD-CN da a we e di ided a
each i e a ion in o 2/3 − 1/3 ain- es spli s, and he ain se s
we e used o es ima e he penalized eg ession model. Penal-
iza ion sh inks he less con ibu i e eg ession coe icien s
by imposing a penal y on he size o he co ela ion s eng h,
which can esul in he exclusion o he leas in o ma i e
a iables om he eg ession model by sh inking hei
espec i e eg ession coe icien s o ze o [31]. P ese ed
ea u es wi h non-ze o coe icien s may hus be conside ed
as hose ha a e ele an o he classi ica ion p oblem o
1 3
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
model combining only ea u es om sMRI and dMRI. The
la e holds pa icula in e es as bo h sequences can be
acqui ed in he same MRI imaging session.
S a is ical signi icance o di e ences be ween AUC al-
ues o he di e en classi ie s models was assessed using
DeLong es s [33]. All models we e implemen ed using he
sciki -lea n lib a y .1.3.0 unning on Py hon 3.10 (www.
sciki -lea n.o g) and included age and mo o symp om
se e i y (UPDRS-III) as co a ia es. All he main analy-
ses we e epea ed in sensi i i y analyses using subsamples
excluding pa ien s ha (a) we e unde choline gic ea men
a he s a o he s udy, o (b) we e acqui ed using an olde
scanne model (Siemens BioG aph HiRes). Fu he mo e,
we ha e also conduc ed complemen a y analyses explo ing
he b ain-wide g oup-le el di e ences in GM, SUVR, and
MD be ween PD-MCI and PDD pa ien s, as well as he pe -
o mance o modali y-speci ic classi ie s o his ask ha
we e ained using he p ocedu e desc ibed abo e.
Resul s
Demog aphics
Table 1 summa izes demog aphical and clinical da a o
ou coho . While he e we e no di e ences in sex be ween
g oups, bo h PD-MCI and PDD pa ien s we e signi ican ly
olde han PD-CN (p < 0.004) and also had an olde age o
disease onse (p < 0.02). Disease du a ion was longe o
PDD pa ien s compa ed o bo h PD-CN (d = 1.09, p < 0.001)
and PD-MCI (d = 0.81, p = 0.002), bu was simila be ween
PD-CN and PD-MCI. Nine pa ien s we e on i as igmine
ea men , all o hem pe aining o he PDD g oup (9/35).
Rega ding mo o symp oms (UPDRS-III), bo h CI g oups
we e signi ican ly mo e impai ed han PD-CN (PDD:
d = 1.29, p < 0.001; PD-MCI: d = 0.64, p = 0.007).
he gi en ain- es spli . In o de o a oid o e - i ing and
da a-leakage, o each i e a ion he hold ou es se was hen
used o an unbiased e alua ion o he pe o mance o he
model o dis inguishing be ween PD-CN and PDD as well
as o dis inguishing be ween PD-CN and PD-MCI (PD-
CN es se s. whole PD-MCI sample). Fo each ain- es
i e a ion o he PDD and PD-MCI classi ica ion asks, we
calcula ed he ela i e con ibu ion o each ROI o he gi en
model (model coe icien s no malized o pe cen age con i-
bu ion), he ecei e ope a ing cha ac e is ics (ROC) cu es
and co esponding a ea unde he cu e (AUC) alues, as
well as he maximum Youden Indices (Max(J)) and co e-
sponding sensi i i ies, speci ici ies, and accu acies. These
a iables we e hen a e aged ac oss he 1000 i e a ions o
ob ain obus and unbiased es ima es o he ROI con ibu-
ions and pe o mance me ics o each classi ica ion ask.
O no e, in ou main analysis we es ic ed he aining
phase o he PD-CN s. PDD classi ica ion ask o ensu e
ha we iden i ied hose neu oimaging ea u es ha a e
obus ly linked o demen ia, and we hen es ed whe he
hese abno mali ies can al eady be de ec ed and used o clas-
si y ea lie , p edemen ia s ages o CI in PD (i.e., PD-MCI).
The a ionale o his app oach lies in i s po en ial clinical
u ili y o he speci ic de ec ion o demen ia- ela ed imaging
abno mali ies in PD-MCI [15]. Howe e , PD-MCI is a he -
e ogenous en i y, including neu odegene a ion pheno ypes
ha may no be di ec ly ela ed o demen ia [32], and ocus-
ing on demen ia- ela ed imaging ea u es may po en ially
miss ele an in o ma ion o he iden i ica ion o MCI in
PD. Thus, in complemen a y analyses we ained and es ed
equi alen models di ec ly aimed a dis inguishing be ween
PD-MCI and PD-CN.
In addi ion o he modali y-speci ic models desc ibed
abo e, he same me hodology was applied o es he added
alue o a ully mul i-modal model including ea u es om
[18F]FDG PET, sMRI, and dMRI, as well as o a sepa a e
Table 1 Demog aphical and clinical da a o he di e en s udy g oups
PD-CN
(n = 53)
PD-MCI
(n = 32)
PDD
(n = 35)
PD-CN
s.
PD-MCI
PDD
s.
PD-CN
PDD
s.
PD-MCI
Age, y 60.5 ± 8.4 65.8 ± 7.3 71.7 ± 7.2 d = 0.66 p = 0.004(*) d = 1.41 p < 0.001(*) d = 0.81 p = 0.002(*)
Female, % 22.6 18.9 25.7 p = 0.787 p = 0.801 p = 0.566
PD onse age, y 55.1 ± 8.9 59.8 ± 8.2 61.0 ± 9.7 d = 0.53 p = 0.020(*) d = 0.63 p = 0.004(*) d = 0.14 p = 0.580
Disease Du a ion, y 5.3 ± 3.3 6.0 ± 4.6 10.7 ± 6.5 d = 0.17 p = 0.44 d = 1.09 p < 0.001(*) d = 0.81 p = 0.002(*)
UPDRS-III 17.6 ± 8.8 23.1 ± 8.7 33.4 ± 16.1 d = 0.63 p = 0.007(*) d = 1.29 p < 0.001(*) d = 0.78 p = 0.002(*)
H&Y 1.9 ± 0.5 2.2 ± 0.5 2.5 ± 0.8 d = 0.59 p = 0.01(*) d = 0.93 p < 0.001(*) d = 0.43 p = 0.08
MoCA, sco e 25.8 ± 2.6 21.7 ± 3.4 13.9 ± 4.2 d=-1.40 p < 0.001(*) d=-3.56 p < 0.001(*) d=-2.01 p < 0.001(*)
PD-CRS, o al sco e 100.0 ± 10.7 74.8 ± 3.9 46.8 ± 12.1 d=-2.83 p < 0.001(*) d=-4.64 p < 0.001(*) d=-3.00 p < 0.001(*)
PD-CRS, on o- subco ical sco e 70.7 ± 10.4 47.3 ± 4.3 24.1 ± 9.8 d=-2.69 p < 0.001(*) d=-4.54 p < 0.001(*) d=-2.96 p < 0.001(*)
PD-CRS, pos e io - co ical sco e 29.2 ± 1.0 27.5 ± 2.1 22.7 ± 4.1 d=-1.14 p < 0.001(*) d=-2.39 p < 0.001(*) d=-1.43 p < 0.001(*)
(*) signi ican ly di e en wi h p < 0.05. Abb e ia ions: PD, Pa kinson’s Disease; CN, cogni i ely no mal; MCI, mild cogni i e impai men ; PDD,
Pa kinson’s Disease Demen ia; UPDRS-III, Uni ied Pa kinson’s Disease Ra ing Scale, pa III (mo o sec ion); H&Y, Hoehn & Yah scale;
MoCA, Mon eal Cogni i e Assessmen scale; PD-CRS, Pa kinson’s Disease Cogni i e Ra ing Scale
1 3
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
Modali y-speci ic classi ica ion models
Figu e 2 shows he ROC cu es o he classi ie s o each
modali y, and Table 2 p o ides a mo e comple e desc ip ion
o he classi ica ion pe o mance o he modali y-speci ic
classi ie s. sMRI showed a easonable pe o mance o
classi ying be ween PD-CN and PDD (AUC = 0.77 ± 0.07,
Max(J) = 0.51±0.11), whe e he empo ooccipi al usi o m
gy us (18.1% a e age con ibu ion), he amygdala (18.1%),
he an e io pa o he middle empo al gy us (12.4%),
and he p ecuneus (11.3%) we e he ROIs ha mos con-
ibu ed o he models (Supplemen a y Fig. 7, op). How-
e e , he sMRI-based models only p o ided chance-le el
pe o mance o he disc imina ion be ween PD-CN and
PD-MCI (AUC = 0.57 ± 0.10, Max(J) = 0.38±0.13). Rega d-
ing dMRI, MD o he planum pola e (38.4% a e age con-
ibu ion), he hippocampus (19.1%), and Heschl’s gy us
(19.1%) showed he highes con ibu ions o disc imina ing
be ween PD-CN and PDD (Supplemen a y Fig. 7, Cen e ).
In he independen es da a, he esul ing models p o ided
an excellen classi ica ion pe o mance o PDD s. PD-CN
(AUC = 0.87 ± 0.06, Max(J) = 0.67±0.10), and a mo e mod-
es bu s ill signi ican disc imina ion be ween PD-MCI and
PD-CN (AUC = 0.71 ± 0.09, Max(J) = 0.47±0.12). DeLong
es s showed ha dMRI ou pe o med sMRI o he disc im-
ina ion o bo h PDD (p = 0.031) and PD-MCI (p = 0.009).
Finally, o [18F]FDG PET, he mos p ominen ea u es o
dis inguishing be ween PD-CN and PDD we e he p ecu-
neus (51.4% a e age con ibu ion) and he supe io pa
o he la e al occipi al co ex (19.4%) (Supplemen a y
Fig. 7, bo om). In he es da a, he [18F]FDG PET-based
classi ie s yielded an excellen pe o mance o disc imi-
na ing be ween PD-CN and PDD (AUC = 0.89 ± 0.05,
Max(J) = 0.75±0.10), and a no able pe o mance o disc im-
ina ing be ween PD-CN and PD-MCI (AUC = 0.78 ± 0.05,
Max(J) = 0.53±0.11). In DeLong es s, [18F]FDG PET sig-
ni ican ly ou pe o med sMRI (p < 0.001) o he classi ica-
ion o PDD, and bo h sMRI (p < 0.001) and dMRI (p = 0.03)
o he classi ica ion o PD-MCI.
Classi ica ion expe imen s o he PD-CN s. PDD ask
we e eplica ed excluding 9 PDD pa ien s ha we e on
choline gic ea men , p o iding almos iden ical esul s:
GM, AUC = 0.75 s. 0.77 using he ull PDD sample; MD,
AUC = 0.90 s. 0.87; [18F]FDG PET, AUC = 0.85 s. 0.89.
Simila ly, exclusion o he 19 pa ien s scanned on he Sie-
mens BioG aph HiRes did no no ably al e he esul s:
PD-MCI, 0.79 s. 0.78 using he whole sample; PDD: 0.88
s. 0.89 using he whole sample (Supplemen a y Fig. 8).
Addi ionally, classi ica ion models ained di ec ly o he
dis inc ion be ween PD-MCI and PD-CN yielded e y
simila indings as o he models ained o he dis inc ion
be ween PDD and PD-CN. Classi ica ion accu acies in he
ROI analysis
B ain-wide ROI-based analysis esul s a e isualized in
Fig. 1 and de ailed s a is ics a e epo ed in Supplemen a y
Tables 1, 2, and 3. In sMRI, PDD pa ien s showed signi i-
can a ophy compa ed o PD-CN in pos e io -pa ie al (no a-
bly he p ecuneus and he pos e io cingula e) and la e al
and medial empo al a eas. The mos a ec ed ROI was he
empo o-occipi al usi o m co ex (d = 0.80, p = 0.03). How-
e e , only e y mild changes we e obse ed in pa ien s wi h
PD-MCI compa ed o PD-CN, whe e no egion eached he
h eshold o s a is ical signi icance. In dMRI, PDD pa ien s
showed a signi ican b ain-wide inc ease in MD, which
was mos no able in he empo al pole (d = 0.93), pos e io
cingula e (d = 0.90), and angula gy us (d = 0.83). PD-MCI
pa ien s also showed signi ican al e a ions in MD, which
we e mos no able in he hippocampus (d = 0.83). Quali a-
i ely, dMRI p o ided la ge e ec sizes and mo e ex ended
e ec s compa ed o sMRI, especially o PD-MCI pa ien s.
Complemen a y di usion me ics, including FA and FW,
we e also assessed bu showed gene ally smalle e ec sizes
han MD o measu ing co ical changes (Supplemen a y
Fig. 2). Finally, [18F]FDG PET e ealed signi ican hypo-
me abolism in pos e io -occipi al egions in PDD, espe-
cially no able in he p ecuneus (d = 1.11), pos e io cingula e
(d = 1.00), and angula gy us (d = 0.92). In con as o dMRI
and especially sMRI, hese changes we e also p onounced
in PD-MCI pa ien s wi h a e y simila pa e n, espe-
cially a ec ing he angula gy us (d = 1.02) and p ecuneus
(d = 0.97). The e e se con as s (i.e., mo e neu odegene a-
ion in PD-CN s. PD-MCI o PDD) did no show any sig-
ni ican e ec o any o he modali ies. Findings emained
la gely unchanged when excluding he 9 PDD pa ien s on
choline gic ea men (Supplemen a y Fig. 3), o he 19
pa ien s scanned on he Siemens BioG aph HiRes (Supple-
men a y Fig. 4).
We also pe o med complemen a y b ain-wide ROI-
based analyses compa ing he PD-MCI and PDD g oups
di ec ly. Di e ences we e spa ially simila o hose o he
PDD s. PD-CN compa ison, bu wi h lowe egional e ec
sizes (Supplemen a y Fig. 5A). Finally, complemen a y pa -
ial Pea son co ela ion analysis be ween PD-CRS sco es
and he di e en egional imaging ea u es e ealed simila
esul s as he ca ego ical compa ison be ween diagnos ic
g oups, wi h simila modali y-speci ic egional pa e ns and
ma kedly highe e ec sizes (pa ial Pea son’s ) o [18F]
FDG PET compa ed o he MRI modali ies. (Supplemen a y
Fig. 6).
1 3

Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
Fig. 1 ROI-based analysis esul s o he di e en modali ies ( om op o bo om), o PD-MCI (le ) and PDD ( igh ). Colo scale ep esen s e ec
size (Cohen’s d), and he solid ba in he colo scale ep esen s he h eshold o p(FDR) < 0.05
1 3
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
Fig. 2 Recei e Ope a ing Cha ac e is ic (ROC) cu es o he di e en modali y-speci ic models (blue). G ey a eas p esen he s anda d de ia ion
o he a e aged ROC cu es
1 3
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
PD-CN and PD-MCI. None o hese models p o ided sig-
ni ican ly be e pe o mance compa ed o he models based
only on dMRI (p > 0.57).
Discussion
While se e al neu oimaging modali ies ha e shown some
p omise in de ec ing he neu odegene a i e p ocesses asso-
cia ed wi h he de elopmen o CI in PD, e y li le wo k
has been done compa ing di e en imaging modali ies o
his pu pose [16]. In addi ion o p o iding a be e unde -
s anding o he ela i e sensi i i ies o he a ailable imaging
bioma ke s, mul imodal s udies a e necessa y o unde -
s and i di e en modali ies may p o ide complemen a y
in o ma ion, hus po en ially inc easing he accu acy in he
e alua ion [8]. We used a mul imodal imaging app oach
o di ec ly compa e he pe o mance o h ee o he mos
widely used neu oimaging modali ies (sMRI, dMRI, [18F]
FDG PET) o de ec ing co ical changes associa ed wi h CI
in PD, including bo h ea ly (PD-MCI) and mo e ad anced
(PDD) s ages o CI.
Following p e ious e idence sugges ing ha a ec ed
b ain a eas may di e be ween imaging modali ies [7, 8, 16],
we i s conduc ed explo a o y b ain-wide analyses aimed
a measu ing he g oup-le el changes obse ed wi hin each
modali y o each s age o he PD cogni i e spec um (PD-
MCI, PDD) compa ed o a g oup o cogni i ely unimpai ed
PD pa ien s. In hese analyses, sMRI e ealed a ophy in he
medial and la e al empo al lobe as well as pos e io medial
a eas (pos e io cingula e, p ecuneus) o PDD pa ien s. This
a ophy dis ibu ion is simila o ha usually obse ed in
pa ien s wi h Alzheime ’s disease, and has been p e iously
ela ed o CI in PD [10, 34, 35]. In e es ingly, howe e ,
pa ien s wi h PD-MCI did no show signi ican changes
on sMRI compa ed o PD-CN, which aligns wi h p e i-
ous esea ch indica ing ha he pa hophysiologic p ocesses
unde lying ea ly cogni i e de ici s in PD may no be accu-
a ely cap u ed by a ophic changes on sMRI [12, 17, 36,
37]. While his is also in line wi h p e ious neu opa hologi-
cal indings sugges ing ha neu onal loss is a ela i ely la e
e en in PD pa hogenesis [38], some s udies ha e achie ed
highe accu acies wi h sMRI using mo e complex classi-
ie s [9, 39]. In e es ingly, dMRI-based di usion changes
in PDD showed a simila opological pa e n o abno mali-
ies compa ed o he a ophic changes on sMRI bu we e
gene ally mo e p onounced and widesp ead. Mo eo e ,
dMRI also e ealed signi ican mic os uc u al changes in
he medial empo al lobe o PD-MCI pa ien s. These esul s
a e in good ag eemen wi h p e ious s udies sugges ing ha
co ical mic os uc u al changes measu ed by dMRI may
p ecede a ophy in he cou se o he disease [19, 40]. The
es se we e highes o [18F]FDG PET (AUC = 0.81 ± 0.07),
ollowed by dMRI (AUC = 0.73 ± 0.10), and non-signi ican
o sMRI (AUC = 0.62 ± 0.12) (Supplemen a y Fig. 9A),
and classi ica ion models elied on simila imaging ea u es
(Supplemen a y Fig. 9B). Finally, classi ica ion models dis-
c imina ing di ec ly be ween he PD-MCI and PDD g oups
showed mode a e pe o mance o all modali ies (AUC:
0.69–0.75; Supplemen a y Fig. 5B) and he e we e no s a-
is ically signi ican di e ences be ween he modali ies o
his ask (p > 0.36).
Mul i-modali y models
Fo he mul imodal model including all modali ies, he
mos ele an ea u es we e he [18F]FDG PET SUVR o
he p ecuneus (26.2% a e age con ibu ion) and he [18F]
FDG PET SUVR o he empo ooccipi al pa o he in e-
io empo al gy us (25.6%) (Supplemen a y Fig. 10, op).
No ably, o he op en con ibu ing ea u es, se en came
om [18F]FDG PET and h ee came om dMRI, while
none came om sMRI. In e es ingly, classi ica ion mod-
els based on he combina ion o MRI and [18F]FDG PET
me ics did no show signi ican ly be e pe o mance (PDD
s. PD-CN: AUC = 0.86 ± 0.08; PD-MCI s. PD-CN: AUC
o 0.80 ± 0.06; Fig. 3, op) compa ed o he models based
on [18F]FDG PET alone (p > 0.19). In he model based only
on MRI ea u es (Fig. 3, bo om), he ea u es showing he
mos obus in ol emen we e he hippocampus MD (14.9%
a e age con ibu ion), empo al pole MD (14.8%), pa ie al
ope culum co ex MD (10.5%) and in e io on al gy us
pa s ope cula is MD (10.5%) (Supplemen a y Fig. 10, bo -
om). These MRI-only mul imodal models yielded an AUC
o 0.84 ± 0.07 o disc imina ing be ween PD-CN and PDD
and an AUC o 0.73 ± 0.08 o disc imina ing be ween
Table 2 Pe o mance me ics o he di e en classi ie s. Di e en
imaging modali ies a e p esen ed in di e en ows, and me ics a e
epo ed sepa a ely o he PD-CN s. PD-MCI (le ) and PD-CN s.
PDD asks ( igh )
AUC Max(J) Accu acy Sensi i i y Speci ici y
PD-CN s. PD-MCI
sMRI 0.57±0.10 0.38±0.13 0.63±0.08 0.73±0.13 0.66±0.13
dMRI 0.71±0.09 0.47±0.12 0.68±0.08 0.76±0.17 0.72±0.14
[18F]
FDG
PET
0.78±0.09 0.53±0.11 0.73±0.07 0.76±0.17 0.77±0.12
PD-CN s. PDD
sMRI 0.77±0.07 0.51±0.11 0.69±0.07 0.81±0.15 0.71±0.15
dMRI 0.87±0.06 0.67±0.10 0.77±0.07 0.86±0.13 0.81±0.12
[18F]
FDG
PET
0.89±0.05 0.75±0.10 0.81±0.07 0.89±0.09 0.87±0.10
Abb e ia ions: AUC, A ea unde he cu e; Max(J), Maximum o he
Youden index
1 3
Eu opean Jou nal o Nuclea Medicine and Molecula Imaging
assessmen o PD-CI, and se e al au ho s ha e highligh ed
he need o mo e di ec compa isons be ween di e en
modali ies [16, 43]. To he bes o ou knowledge, ou s udy
p o ides he i s di ec compa ison be ween dMRI and
[18F]FDG PET, and e y ew mul imodal s udies ha e con-
join ly s udied [18F]FDG PET and sMRI in PD-CI [17]. Ou
esul s la gely ag ee wi h hose by González-Redondo e al.
[17], who hypo hesized ha hypome abolism and a ophy
ep esen consecu i e s ages o he same neu odegene a-
ion p ocess. Howe e , simila ly o his p e ious s udy, we
also obse ed ema kable opological di e ences be ween
he hypome abolism and a ophy pa e ns. Conc e ely,
obse ed pa e n o inc eased co ical MD was also simila
o hose ound in p e ious s udies [19, 40], sugges ing ha
dMRI may p o ide mo e ep oducible esul s han sMRI in
he e alua ion o PD-CI [7]. Finally, o [18F]FDG PET we
obse ed a ypical pa e n o pos e io -occipi al hypome-
abolism in associa ion wi h CI, which has been ex ensi ely
desc ibed in p e ious s udies [15, 17, 41, 42]. In con as
o he MRI-based modali ies, his pa e n was al eady well-
de ined in PD-MCI, and e ec sizes we e gene ally highe
compa ed o bo h dMRI and sMRI (Fig. 1).
To da e, ew mul imodal imaging s udies ha e conduc ed
di ec compa isons be ween [18F]FDG PET and MRI o he
Fig. 3 Recei e Ope a ing Cha ac e is ic (ROC) cu es o he di e en mul i-modali y models (blue). G ey a eas p esen he s anda d de ia ion
o he a e aged ROC cu es
1 3