METHODEN - Am MIT spezialisierte KI-Forschung

METHODEN

Empirical evaluation of human fetal fMRI
preprocessing steps

Lanxin Ji1

, Cassandra L. Hendrix1, and Moriah E. Thomason1,2,3

1Department of Child and Adolescent Psychiatry, New York University School of Medicine, New York, New York, USA
2Department of Population Health, New York University School of Medicine, New York, New York, USA
3Neuroscience Institute, New York University School of Medicine, New York, New York, USA

Schlüsselwörter: Fetal fMRI, Vorverarbeitung, Normalization, Denoising, Smoothing, Funktional
Konnektivität

Keine offenen Zugänge

Tagebuch

ABSTRAKT

Increased study and methodological innovation have led to growth in the field of fetal brain
fMRT. An important gap yet to be addressed is optimization of fetal fMRI preprocessing. Rapid
developmental changes, imaged within the maternal compartment using an abdominal coil,
introduce novel constraints that challenge established methods used in adult fMRI. This study
evaluates the impact of (1) normalization to a group mean-age template versus normalization
to an age-matched template; (2) independent components analysis (ICA) denoising at two
criterion thresholds; Und (3) smoothing using three kernel sizes. Data were collected from 121
fetuses (25–39 weeks, 43.8% weiblich). Results indicate that the mean age template is superior
in older fetuses, but less optimal in younger fetuses. ICA denoising at a more stringent
threshold is superior to less stringent denoising. A larger smoothing kernel can enhance cross-
hemisphere functional connectivity. Gesamt, this study provides improved understanding of
the impact of specific steps on fetal image quality. Findings can be used to inform a common
set of best practices for fetal fMRI preprocessing.

EINFÜHRUNG

Understanding of human brain development has grown rapidly with the introduction of fetal
resting-state functional connectivity (RSFC) (van den Heuvel & Thomason, 2016). In 2011,
Veronica Schöpf and colleagues published the first fetal RSFC study (Schöpf et al., 2011),
demonstrating that it was possible to noninvasively image whole-brain functional systems
prior to birth by using MRI. Before this time, very few studies had measured fetal brain activity
(Anderson & Thomason, 2013). In der Tat, most of what was known about prenatal brain devel-
opment was the product of histological or structural analytic approaches in postmortem or
clinical samples (Chi et al., 1977; Dobbing & Sands, 1973), or was inferred from RSFC studies
conducted in preterm neonates (Doria et al., 2010; Fransson et al., 2007).

Fetal RSFC has enabled us to begin describing properties of typical development as well as
the role of the environment in shaping neural network development. Studies of typical devel-
opment have revealed that network connectivity patterns in utero precede and may guide
functional selectivity of certain brain regions, such as the fusiform face area (van den Heuvel
et al., 2018), and that macroscale characteristics of fetal functional networks share significant
overlap with adult networks (Turk et al., 2019). In addition to shedding light on the origins of

Zitat: Ji, L., Hendrix, C. L., &
Thomason, M. E. (2022). Empirical
evaluation of human fetal fMRI
preprocessing steps. Netzwerk
Neurowissenschaften, 6(3), 702–721. https://doi
.org/10.1162/netn_a_00254

DOI:
https://doi.org/10.1162/netn_a_00254

zusätzliche Informationen:
https://doi.org/10.1162/netn_a_00254

Erhalten: 30 November 2021
Akzeptiert: 9 Mai 2022

Konkurrierende Interessen: Die Autoren haben
erklärte, dass keine konkurrierenden Interessen bestehen
existieren.

Korrespondierender Autor:
Moriah E. Thomason
moriah.thomason@nyulangone.org

Handling-Editor:
Michael Cole

Urheberrechte ©: © 2022
Massachusetts Institute of Technology
Veröffentlicht unter Creative Commons
Namensnennung 4.0 International
(CC BY 4.0) Lizenz

Die MIT-Presse

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

typical neural development, fetal RSFC studies also inform our understanding of health risk.
Zum Beispiel, exposures like prenatal stress (Thomason et al., 2021A), cannabis (Thomason
et al., 2021B), and lead (Thomason et al., 2019) have been linked to altered fetal neuro-
Entwicklung, which has important implications for policy and intervention. Taken together,
this work highlights that fetal MRI is a crucial tool for understanding typical and atypical
human neurodevelopment and for uncovering the earliest origins of disease risk.

There is need to optimize fetal RSFC analytic pipelines so that this important work can be
conducted in a rigorous and reproducible manner. A number of methodological studies have
highlighted vulnerabilities in the processing and analysis of RSFC data in adults. Zum Beispiel,
data-driven approaches have revealed that traditional denoising techniques using linear
modeling may incorrectly classify intrinsic neural signal as noise (Bright et al., 2017). Weiter-
mehr, interlab variation in fMRI processing choices can lead to disparate results, even when
labs are analyzing the same data (Botvinik-Nezer et al., 2020). Because of the potent impact of
analytic choices on fMRI outcomes, there have been several efforts to create and distribute
centralized, robust preprocessing pipelines for adult fMRI data such as fMRI PREP (Esteban
et al., 2019) and the Human Connectome Pipeline (Glasser et al., 2013). Jedoch, these pre-
processing pipelines were not developed to manage the unique challenges inherent to imaging
the brain in utero, including high motion, encasement within the maternal compartment, beide
fetal and maternal sources of noise, and unique geometry of the large field of view and abdom-
inal coil array. There is need for development of fMRI processing pipelines suited to the devel-
opmental and methodological considerations specific to the fetus (Rajagopalan et al., 2021).

We elected to focus on three preprocessing steps that require particular attention in the
developing brain: normalization to standard space, denoising, and smoothing. One of the
largest challenges in fetal fMRI is excessive motion as introduced by both the fetus and by
the mother (z.B., breathing). Discarding periods of high motion or excluding subjects whose
motion exceeds a stringent threshold is often the first attempt to tackle the problem, but it
invariably leads to significant data loss. The balance between maximizing amount of data
and maximizing data quality is challenging and highly varied across datasets with different
motion profiles. As an additional step, the regression-based motion artifact removal strategy
is widely used to control the secondary intravolume effects induced by motion, such as arti-
facts related to partial voluming and magnetic field inhomogeneities (Friston et al., 1996;
Pruim et al., 2015). Typical regression models include 6 Zu 24 motion covariates derived from
the volume realignment (Friston et al., 1996; Yan et al., 2013), yet these covariates are highly
reliant on the algorithm used for the realignment and, Außerdem, the underlying intravolume
effects cannot be captured by the realignment parameters. Beyond motion parameter-based
Modelle, spatial independent component analysis (ICA) provides a powerful tool to separate
neural-related signal from different sources of noise, including the motion-related artifacts.
Applied to fMRI data, ICA decomposes data into a set of spatial independent components
and associated time courses (Beckmann & Schmied, 2004). Components presenting noise fea-
tures can subsequently be regressed out of the data. ICA-based denoising is well established
as a method for removing motion artifacts in adult imaging, but has yet been evaluated in fetal
Bildgebung. Daher, this study examines ICA-based data denoising in a large collection of fetal fMRI
scannt. Another challenge of the fetal brain is its unparalleled, rapid development across
gestation, which complicates the normalization process. Zum Beispiel, it is unclear whether
normalizing to a fetal template from a particular stage in gestation (z.B., 32 weeks) is adequate,
or if instead it is necessary to normalize to a template that is closely age matched to the fetus
(z.B., within a week). Endlich, the smaller size of fetal brains compared to adults may require
different recommendations regarding smoothing kernel size, which may influence the

Denoising:
A critical step of preprocessing to
remove noise and nonneuronal
Beiträge, such as motion-related
and physiological noise.

Normalization:
A step transforming brains to a
Vorlage, to ensure that each voxel
for each subject corresponds to same
brain parts.

Realignment:
A step of preprocessing to correct
head movements, by co-registering
all volumes in a time series to a
reference volume.

Independent component
Analyse (ICA):
A data-driven approach to
decompose fMRI data into a set of
statistically independent spatial maps
together with associated time
courses.

Netzwerkneurowissenschaften

703

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

likelihood of identifying significant associations (Botvinik-Nezer et al., 2020). The present
study addresses the effect of these key processing decisions during the preparation of fetal fMRI
data for second-level analyses in a large fetal dataset.

MATERIALS AND METHODS

Teilnehmer
Healthy mothers were recruited during routine obstetrical appointments at Hutzel Women’s
Hospital in Detroit, Michigan. Inclusionary criteria included maternal age ≥18 years old,
native English speaking, singleton pregnancy, and normal fetal brain anatomy as assessed
by ultrasound and MRI examination. MRI visits occurred when fetuses were between 22
Und 39 weeks gestational age (GA). This study included data from second- and third-trimester
fetuses from a larger ongoing project on fetal brain development who had manually seg-
mented and quality assured raw resting-state fMRI data available at the time of this analysis
(N = 165). Development of automated processes for fetal brain segmentation is an active area
of study (Rutherford et al., 2021), Aber, at present, manual tracing of the brain is the most pre-
cise approach. Additional exclusions were applied for fetuses subsequently born very preterm
or with low birth weight (<33 weeks GA, <1,800 g; n = 14), those scanned before 25 weeks GA (n = 9), and with fewer than 100 low-motion volumes or high segment-weighted average motion (1.5 mm max excursion, 0.5 mean; rotational>2°, rotation mean
>1°, n = 21), Daraus ergibt sich eine endgültige Probe von 121 fetuses (68 männlich; 53 weiblich). Included fetuses
had a mean GA of 32.89 weeks at scanning (range = 25.86–39.57; SD = 3.75) and were born,
on average, bei 39.08 (SD = 1.49) weeks gestation. More detailed characteristics of the sample
are provided in Table 1. Motion parameters were not correlated with demographic variables

Tisch 1.

Sample demographic characteristics (N = 121)

Maternal age, Jahre

Race/ethnicity, N (%)

Caucasian

African American

Latina

Asian American

Biracial

Not disclosed

Fetal sex, N (%)

Female

Male

Gestational age at scan, weeks

Gestational age at birth, weeks

Birth weight, Gramm

Netzwerkneurowissenschaften

Mean ± SD
25.34 ± 4.56

10 (8.26%)

99 (81.82%)

1 (0.83%)

5 (4.13%)

53 (43.80%)

68 (56.20%)

32.89 ± 3.75

39.08 ± 1.49

3,237.22 ± 510.73

704

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

including scan age and sex in the final sample (see Supporting Information, Figure S1). Alle
study procedures were approved by the Wayne State University Human Investigation
Committee.

Data Acquisition

Fetal MRI data were acquired on a Siemens Verio 70-cm open-bore 3T MR system using a
550 g abdominal 4-Channel Siemens Flex Coil (Siemens, München, Deutschland). Twelve minutes
of fetal resting-state fMRI data were acquired using the following gradient echo planar imaging
sequence: TR/TE 2,000/30 MS, flip angle 80°, 360 frames, axial 4-mm-slice thickness, voxel
Größe 3.4 × 3.4 × 4 mm3. The sequence was repeated when time permitted.

Preprocessing Pipelines

A full preprocessing workflow is shown in Figure 1. Time frames in the raw fMRI data corre-
sponding to periods of significant head motion were identified using FSL image viewer (FSL,

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Figur 1. Workflow of the fetal fMRI preprocessing pipeline. Key steps validated in this study are colored by blue boxes.

Netzwerkneurowissenschaften

705

Evaluation of fetal fMRI preprocessing steps

fMRI preprocessing:
A set of image processing steps to
clean and standardize fMRI data
before statistical analysis.

2018) and excluded. Brainsuite (Shattuck & Leahy, 2002) was used to manually generate fetal
brain masks around single reference images that were then applied to each resulting contin-
uous, low-motion 4D segment. After implementing volume-to-volume motion correction using
SPM’s ‘Realign’ function within each segment, we then evaluated the effects of different strat-
egies in several key fMRI preprocessing steps:

1. Normalization to age-matched versus 32-week template. To assess the influence of
different templates on the quality of normalization, we tested normalization from the
functional data directly to the standard template of a 32-week GA fetus (mean age for
the group) versus to the nearest week-specific template for a given subject (reichend
aus 25 Zu 37 weeks GA). Serag’s 2012 templates were used (Serag et al., 2012),
and normalization was conducted in SPM using nonlinear warping. The warping met-
rics were estimated with the first volume of each segment and were then applied to
remaining volumes within that segment. Two metrics (Calhoun et al., 2017) wurden benutzt
for comparison: (1) voxel-wise variability of the normalized images across subjects; (2)
mean and maximum absolute frame-to-frame displacement derived from performing
volume-to-volume realignment, a second time, across the full normalized,
concatenated time series. Voxel-wise variability provides a measure for mismatch
between fMRI data and the template across subjects. If a given voxel is on the edge
and varying constantly between being “in” and “out” of the brain, this voxel will tend
to have a high standard deviation. Speziell, the first normalized volume of subjects
at the same gestational age were concatenated along the fourth dimension to create a
single image file. We calculated the standard deviation of this file along the subject
dimension using Image Calculator of DPABI toolbox (Yan et al., 2016) implemented
in MATLAB. Measurements of absolute displacement of the brain from the original
position included total translational movement (maximum and mean difference in
position in millimeters) and total head rotation (maximum difference in rotation in
degrees; Van Dijk et al., 2012). Scans with more accurate normalization across seg-
ments are expected to show lower intersubject displacement (Calhoun et al., 2017).
Endlich, to explore possible effects resulting from the choice of normalization templates
gebraucht, we additionally evaluated normalization to alternative age-specific fetal tem-
plates (Gholipour et al., 2017). This was a secondary analysis and was thus performed
for one representative subject from each gestational age.

After normalization, segments were concatenated within each scan and potential
misalignments between segments were corrected using SPM’s realignment function.
For the following processing step comparisons, data resulting from normalization to
the 32-week template were used. One subject was excluded here due to low usable
frames (n = 120 for the following analyses).

2. Masking the full concatenated data. To repress background spurious signals, a next
step evaluated the utility of applying a dilated brain mask. We tested whether masking
at this step improved downstream processing.

3. Denoising at two stringency thresholds, based on ICA. Data were decomposed into
independent components using FSL’s MELODIC (multivariate exploratory linear opti-
mized decomposition into independent components; Beckmann & Schmied, 2004). Der
number of components was automatically estimated by MELODIC. Noise components
were manually labeled twice, once in a less stringent way and once in a more stringent
Weg (Griffanti et al., 2017). With the less stringent method, components showing
nonbiological spatial banding patterns, ring-like patterns on edges of the brain, AND
high-frequency peaks were labeled as noise; with the more stringent method,

Netzwerkneurowissenschaften

706

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

components showing banding patterns, ring-like patterns on edges of the brain OR
clusters mainly located in the white matter or cerebrospinal fluid OR time series with
sudden jumps (caused by segment concatenation) OR significant changes in oscillation
patterns OR high-frequency peaks were labeled as noise. Allgemein, the main differ-
ence between the two thresholds is whether noise is defined on the basis of temporal
or spatial features, allein, or in combination of both. In this study, components showed
abnormal temporal features due to segment concatenation, such as sudden jumps
(Figure 2A) and alterations of oscillation patterns (Figure 2E), are unique to fetal

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Figur 2. Strategy used for less and more stringent ICA elimination. Exemplar components observed in the fetal dataset are presented above.
Observation of a single failure in spatial, zeitlich, or frequency domains results in elimination of the component, but only at the more strin-
gent level. Less stringent correction only eliminates components if more than one failure is observed, Zum Beispiel, in both spatial and temporal
domains. Pass and fail examples are provided here, depicted with checkbox and cross-out, jeweils. As examples of single failures, com-
ponent B shows nonbiological banding patterns (positive/negative stripes), but shows acceptable time course, and component C shows a
typical spatial pattern, but shows high-frequency peaks, indicative of scanner-related artifacts. Examples A–C, were eliminated only at the
more stringent threshold. D and E show failures in two domains, and F and G pass both spatial and temporal analysis.

Netzwerkneurowissenschaften

707

Evaluation of fetal fMRI preprocessing steps

Spatial smoothing:
A step of preprocessing aiming to
enhance the signal-to-noise ratio, von
averaging data points with their
neighbors.

Voxel-mirrored homotopic
Konnektivität ( VMHC):
A voxel-wise measure of functional
connectivity between hemispheres
by computing correlations between
images and their left-right mirror
Ausführung.

Seed-based functional connectivity:
A voxel-wise correlation analysis
between a predefined region, das ist,
the seed, and all other voxels in the
Gehirn.

Bildgebung. Criteria that formed the basis of each exclusion level are depicted in Figure 2.
Noise components were removed using the fsl_regfilt function. As outlined in the
following section, we evaluated the effect of ICA denoising on RSFC measures resulting
from each approach.

4. Smoothing with different kernel sizes. We tested the effect of smoothing kernels of
0 mm (no smoothing), 2 mm, Und 4 mm full-width at half maximum (FWHM) mit
SPM. The chosen kernels equal 1 oder 2 times of our voxel size. We quantified the effects
of spatial smoothing by evaluating mean cross-hemisphere functional connectivity
strength across different kernel sizes.

Functional Connectivity Analysis

Cross-hemisphere functional connectivity. We examined RSFC between homotopic voxels in the
brain by using the voxel-mirrored homotopic connectivity ( VMHC) technique (Zuo et al.,
2010), which is a voxel-wise correlation analysis between the images and their left-right mirror
Ausführung. Preprocessing pipelines were compared on the basis of resultant summary measures
of homotopic functional connectivity for each fetus. As an additional means of evaluating the
above preprocessing pipeline, global mean VMHC was tested for correlation with the number
of frames and head motion parameters derived from the entire time course across subjects
using R software (Ausführung 4.0.5). Adequate removal of noise in preprocessing steps should
be reflected in lack of association of RSFC with frame count and motion.

Seed-based functional connectivity. Seeds were selected to represent regions both distal and
proximal to the midline. Seeds were defined manually as spheres with a 3-mm radius
(179 voxels), centered on MNI coordinates: (−20.6, 19.8, −8.6), (−7.7, −18.9, −26.6),
(−8.6, 13.7, −0.8), (−5.2, −21.4, 8.6), (−5.2, 33.5, −8.6), (−9.5, −5.2, 22.3), (−8.6, −5.2,
−4.3), Und (−3.4, −39.5, −4.3). These were constructed using Mango Multi-image Analysis
Software (https://ric.uthscsa.edu/mango/mango.html). Locations approximate the anterior
insular, cerebellum, putamen, precuneus, medial prefrontal cortex, supplementary motor
Bereich, thalamus, and the visual cortex, jeweils, in the 32-week fetal template (Serag
et al., 2012). Seeds were selected to approximate locations used in prior research in preterm
and term newborns (Smyser et al., 2010) and because functional neural networks related to
these seeds are evidenced to be sensitive to early brain development (Thomason et al.,
2015). These left hemisphere masks were duplicated for the right hemisphere, resulting
in a total of 16 seed regions. Seed regions of interest (ROIs) are represented in Figure S2
von
the Supporting Information and files themselves are available online at www
.brainnexus.com.

Seed-to-voxel whole-brain analyses were performed on the denoised data in DPABI tool-
box. For each subject, the mean time course was extracted from each seed region and corre-
lated with the time course of each voxel throughout the whole brain, yielding individual RSFC
maps for each seed region. All RSFC maps were converted to z-scores for post hoc analyses.
Seed-based RSFC maps of less versus more stringently denoised data were compared using
paired two-sample t test. Clusters were corrected for whole-brain multiple comparisons by
using false discovery rate (FDR) P < 0.05. Group mean RSFC maps were estimated by one- sample t tests testing the z-transformed values against 0, with threshold at p < 0.00001 FDR corrected. In addition to the above voxel-wise Group-level ICA as a validation of the proposed pipeline. and seed-based functional connectivity analyses used for evaluation of preprocessing Network Neuroscience 708 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . t / / e d u n e n a r t i c e - p d l f / / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d t . f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Evaluation of fetal fMRI preprocessing steps strategies, we conducted a brain network analysis. To extract group-level intrinsic connec- tivity networks, we performed spatial ICA implemented in group ICA of Functional MRI Toolbox (GIFT v3.0b, https://trendscenter.org/software/gift/). Optimally preprocessed fMRI data (more stringently denoised, 4-mm FWHM smoothed) were decomposed into 35 spatial components, each of which exhibited a unique time course profile based on the Info- max algorithm. The number of components was estimated based on the image quality by using a minimum description length approach (Rissanen, 1978). A higher order ICA approach was applied to improve functional parcellation (Kiviniemi et al., 2009). Reliability and stability of the algorithm was ensured using ICASSO by repeating the component esti- mation 20 times (Himberg et al., 2004). Subject-specific spatial maps and time courses were obtained using the back-reconstruction approach (GICA; Calhoun et al., 2004) and con- verted to z-scores. RESULTS Assessment of Spatial Normalization to Age-Matched Versus 32-Week Template When plotting the distribution of 8,000 voxels with the highest standard deviation (Figure 3), we observe that intersubject alignment was improved (SD reduced) when using the 32-week template for fetuses older than 32 weeks. However, the reverse was true for younger fetuses; there, an age-matched template was associated with reduced intersubject alignment and better normalization performance. We also observe that the areas of greatest variability are identified at the edges of the brain (see Figure S3 in Supporting Information). This finding may reflect a combination of greater displacement associated with distance from origin and also stronger BOLD signal in cortex compared to white matter and cere- brospinal fluid. We did not find a significant difference in mean and maximum subject-to- subject displacement between the templates at any fetal age (see Figure S4 in Supporting Information). Furthermore, we did not observe a marked difference in normalization per- formance when an alternate fetal anatomical template set was used (see Figure S5 in Supporting Information). Assessment of Individual-Level Masking With unmasked data, we detected a number of noise components located outside of the brain. We found that masking before denoising reduced the number of ICA-derived components (Figure 4) for most fetal subjects; the total number of independent components across all subjects decreased from 4,043 to 3,623 after reapplying a brain mask. The individual-level masking removed most of the ring-like noise components in the following ICA analysis. This reduction of components alleviates the workload of manual inspection, which is currently needed for fetal data. The dilated mask is shown in Figure S6 of the Supporting Information, and is available online at https://www.brainnexus.com/. Assessment of ICA Denoising The less stringent approach to labeling noise components resulted in the identification of 5% of all ICA-derived components as noise. In contrast, the more stringent approach resulted in 60% of components being labeled as noise. More stringent denoising led to improvement by visual inspection; in particular, we observed reductions in spatial banding patterns, intensity inhomogeneity, and abnormal signal oscillations caused by rapid motion or segment Network Neuroscience 709 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / / t e d u n e n a r t i c e - p d l f / / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d . t f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Evaluation of fetal fMRI preprocessing steps l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / / t e d u n e n a r t i c e - p d l f / / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d t . f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Figure 3. Distributions of standard deviation values after normalization, by fetal age and by template used. Fetuses of different ages were normalized either to a 32-week template (mean for the group) or to a same-age template. Voxels on the edge of the brain have lower standard deviation if they are consistently characterized the same way. The 8,000 voxels with the highest standard deviation are plotted here. Review of observed distributions suggest that the 32-week template performs more optimally for fetuses older than 32 weeks, seen in a leftward shift of 32-week values. The reverse is noted for fetuses younger than 32 weeks, where the age-matched template corresponds to a leftward shift. Figure 4. Number of components estimated in FSL’s MELODIC by subjects with and without masking. Network Neuroscience 710 Evaluation of fetal fMRI preprocessing steps l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . Figure 5. Comparison of less stringent versus more stringent ICA denoising in a representative subject, Case 1. A single volume is shown for a case (35 weeks GA) presenting severe nonbiological banding patterns. (A) Raw data with different planes. (B) Less stringently denoised data with different planes (left), and corresponding cross-hemisphere RSFC (right). (C) More stringently denoised data with different planes (left) and corresponding cross-hemisphere RSFC (right). (D) Examples of ICA noise components related to the banding artifact. concatenation. Comparison of the two strategies in three representative subjects are shown below for qualitative inspection, following by group-level RSFC comparisons. / t / e d u n e n a r t i c e - p d l f / / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d t . f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Case 1: A representative subject with severe nonbiological banding patterns in the posterior part of the brain was selected as Case 1. As shown in Figure 5, the nonbiological banding patterns were detected with ICA (bottom row) and were slightly lessened with a less stringent denoising approach. The presence of this nonphysiological band- ing pattern is usually related to the MRI sequence (e.g., EPI susceptibility or multiband acceleration) or hardware artifacts (e.g., RF interference) or interactions of the acqui- sition with head motion (e.g., interleaved slice acquisitions) (Griffanti et al., 2017). In contrast, the more stringent denoising further improved homogeneity within the brain and was associated with reduced cross-hemisphere functional connectivity. It is pos- sible that banding patterns remain at the less stringent level of denoising, because even though the spatial pattern is atypical, the time series falls within the normal range. Case 2: When checking components, we noticed an unusual case with massive whole- brain intensity shift during the scan, which was likely due to an issue in the coil or other electronics. With the more stringent strategy, we labeled all components as noise for this subject, as all components shared the same time course that contained a sudden jump. The intensity shift was completely corrected in this approach. As could be expected, without this correction, the cross-hemisphere functional connec- tivity is biologically implausible across the whole brain (Figure 6). Network Neuroscience 711 Evaluation of fetal fMRI preprocessing steps Case 3: In the third subject, there was a sharp increase in intensity in the left parietal cortex in the middle of the scan that corresponded to one ICA-derived component (the bottom row of Figure 7). With more stringent ICA denoising, artifacts were removed (Figure 7) and cross-hemisphere functional connectivity increased. Strong intensity changes in the time course at the joint of two segments, observed in Figure 7D, results from both (1) normalization misalignment between segments and/or (2) fetal repositioning. These signal “jumps” thus reflect inconsistent segment-to-segment spatial alignment (imprecise normalization) or can reflect siz- able fetal repositioning with potential to change field geometry or interactions at tissue interfaces. Denoising effects on group-level cross-hemisphere functional connectivity. Global mean cross- hemisphere RSFC was correlated with number of frames (r = 0.29, p = 0.001) and with motion parameters (mean translational movements: r = 0.47, p = 6.3e−08; maximum translational movements: r = 0.49, p = 1.1e−08; mean rotations: r = 0.19, p = 0.03; maximum rotations: r = 0.19, p = 0.039) following less stringent denoising (top row of Figure 8A). After regressing out noise components by using the more stringent approach, the associations between RSFC, l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / t / e d u n e n a r t i c e - p d l f / / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d . t f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Figure 6. Comparison of less stringent versus more stringent ICA denoising in a representative sub- ject, Case 2. A single volume is shown for a case with noted intensity shift. (A) Raw data of two axial slices at the 1st and 188th volumes. (B) Less stringently denoised data at the 1st and 188th volumes (left) and corresponding cross-hemisphere RSFC (right). (C) More stringently denoised data at the 1st and 188th volumes (left) and corresponding cross-hemisphere RSFC (right). (D) Time series of an example noise component. Network Neuroscience 712 Evaluation of fetal fMRI preprocessing steps l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / / t e d u n e n a r t i c e - p d l f / / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d t . f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Figure 7. Comparison of less stringent versus more stringent ICA denoising in a representative sub- ject, Case 3. A single volume is shown for a case with high residual motion. (A) Raw data of one slice at the 1st, 50th, and the 100th volumes. (B) Less stringently denoised data at the 1st, 50th, and the 100th volumes (left) and corresponding cross-hemisphere RSFC (right). (C) More stringently denoised data at the 1st, 50th, and the 100th volumes (left) and corresponding cross-hemisphere RSFC (right). (D) The time course and the spatial map of the noise component corresponding to the artifact in left parietal cortex. Arrows indicate the volumes we showed in the above rows. frame count, and motion were no longer significant (bottom row of Figure 8A). Overall, a more stringent denoising strategy corresponded to a reduction in cross-hemisphere RSFC (Figure 8B). However, it is notable that decreased cross-hemispheric connectivity was not a ubiquitous feature of denoising; individual cases showed increases in cross-hemispheric connectivity when applying a more stringent ICA. For example, in Case 3 (Figure 7), after removing left- lateralized artifact, we observe increased cross-hemisphere RSFC, suggestive of unmasking underlying connectivity effects. Denoising effects on group-level seed-based functional connectivity. Use of different denoising levels was associated with changes in the pattern of RSFC across ROIs, as shown in Figure 9. Select regions showed increased or decreased connectivity at each denoising thresh- old. Overall, the pattern was such that less stringent denoising was associated with greater overall RSFC, demonstrated in Figure S7 of the Supporting Information. Assessment of Spatial Smoothing Generally, we observed a dose-dependent relationship between smoothing kernel size and cross-hemisphere functional connectivity, with a 4-mm kernel resulting in the greatest Network Neuroscience 713 Evaluation of fetal fMRI preprocessing steps l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / / t e d u n e n a r t i c e - p d l f / Figure 8. Group-wise comparison of different ICA denoising strategies. (A) Correlations of global mean voxel-mirrored homotopic connec- tivity ( VMHC) with frame counts and motion parameters (XYZ mean and XYZ max for translational movements; PYR mean and PYR max for rotations). (Top row) Less stringently denoised data; (bottom row) more stringently denoised data. (B) Group mean VMHC by age group with less versus more stringent denoising methods. Asterisks (*) in front of p values indicate significant correlations. / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d . t f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 cross-hemisphere functional connectivity. We also note that this effect does not interact with age (Figure 10), suggesting the impact of smoothing kernel size does not vary with fetal age. Resting-State Functional Networks in the Fetal Brain In an exploratory validation analysis, we examined the presence of fetal resting-state net- works in more stringent denoised ICA components (4-mm FWHM smoothed). Thirty-one of the 35 components (available online at www.brainnexus.com) were identified as signal com- ponents because their peak coordinates were located primarily in gray matter and their time courses were dominated by low-frequency fluctuations (Allen et al., 2011). We organized the signal components into nine functional networks based on the temporal correlation between the components and the anatomical locations, including the subgenual area, cerebellum, temporal regions, visual network, frontoinsular network, default mode network, temporopar- ietal network, motor network, and frontal pole areas (Figure 11). Examples of group-level ICA noise are shown in Figure S8 of the Supporting Information. The identified components spa- tially resembled those previously described in preterm neonates (Smyser et al., 2010) and fetuses (Thomason et al., 2015). Network Neuroscience 714 Evaluation of fetal fMRI preprocessing steps l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / t / e d u n e n a r t i c e - p d l f / Figure 9. Comparison of seed-based functional connectivity in data analyzed with more or less stringent denoising. One-sample t test was sued to compare more stringent ICA (blue) and less stringent ICA (red) (p < 0.00001, FDR corrected). Overlapping regions are shown in purple. / / / / 6 3 7 0 2 2 0 3 6 0 3 2 n e n _ a _ 0 0 2 5 4 p d . t f b y g u e s t t o n 0 7 S e p e m b e r 2 0 2 3 Figure 10. Group-wise comparison of different smoothing kernels. (A) Global mean VMHC changes without or with smoothing kernels of 2 mm and 4 mm. (B) Voxel-wise VMHC of a representative fetus of 37 weeks. Network Neuroscience 715 Evaluation of fetal fMRI preprocessing steps Figure 11. Fetal brain networks derived from ICA. Positive t maps threshold at t value > 6 are shown.

DISKUSSION

Our analyses confirm that different approaches to normalization, masking, denoising, Und
smoothing during fetal fMRI preprocessing have notable impacts on data quality. Specific ana-
lytic choices during preprocessing impact connectivity metrics derived from BOLD images at
the subject level, with implications for reliability and reproducibility of group-level effects.

Results indicate that choice of template relates to normalization variability in an age-
dependent manner. Speziell, using a 32-week template resulted in greater normalization
accuracy compared to using age-matched templates for fetuses 32 weeks or older. Umgekehrt,
using an age-matched template results in greater normalization accuracy for fetuses under
32 weeks gestational age. It is possible that either maturational changes within the brain,
variation in size and shape, or some combination of these contribute to this effect. For exam-
Bitte, wide difference in size between the source and template images leads to more scaling
transformations and increased interpolation. These data also show that when a large fetal
age range is studied, fetuses at extreme edges of the range will be most impacted by choice
of normalization template. Gesamt, these observations suggest that template choice may best
be determined by research objectives and characteristics of the sample. Zum Beispiel, Wenn
age-related development is not the goal of the study, it may be advisable to include gestational
age at scan as a nuisance covariate in second-level models and/or to exclude cases to restrict
the age range being studied. If the study objective(S) includes age-related development, mean
voxel-wise standard deviation can be considered as a regressor to correct for normalization
differences across age. In future work it would be valuable to evaluate alternative registration
algorithms that are not reliant on off-the-shelf tools.

In ICA-based denoising using FSL’s MELODIC, a high proportion of noise components is
usually expected in adults. With standard sequences at 3T, around 70% (Rummel et al., 2013)
Und 88% (Griffanti et al., 2014) of components may be reported as noise. In our fetal dataset,
even using a more stringent strategy, nur 60% of components were identified as noise, most of
which were motion related. Observing a lower proportion of noise components in fetal fMRI
data may be the result of either developmental processes, such as changes in cerebrovascular
Struktur (Reilly & Gutierrez, 2021), or altered noise characteristics unique to this context. Als
examples of the latter, abdominal versus head coil geometry or field inhomogeneity

Netzwerkneurowissenschaften

716

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

introduced by large field of view may contribute to differential effects. Another possibility is
that a smaller number of components arises simply from larger motion-induced artifacts wash-
ing out smaller, more subtle artifacts. Because ICA denoising may perform differently in fetal
Daten, it is advisable that comparisons between fetal and postnatal datasets take this possibility
into account.

We also observed that more stringent denoising removed the correlation between all
motion parameters and global cross-hemisphere RSFC and resulted in different seed-based
RSFC maps. These findings highlight the sensitivity of voxel-to-voxel and seed-to-voxel RSFC
to head motion. Jedoch, we also want to acknowledge that the more stringent denoising
does not guarantee “better” estimates of functional connectivity, considering the risk of poten-
tial inadvertent removal of signal. Selection of cross-hemispheric connectivity was the refer-
ence analysis used in the present study because it has been evidenced in previous studies a
sensitive measure of fetal development (Thomason et al., 2013) and because cerebral homo-
topy is a fundamental principle of brain organization (Toga & Thompson, 2003). Jedoch, Es
should be noted that this is one of many strategies that could be used to test the effects of
denoising on resultant RSFC. This method was useful in confirming that after denoising, ein
expected pattern in brain organization was observed and was no longer correlated with
motion parameters.

In fetal fMRI, motion-related artifact is a significant challenge; motion artifacts can be intro-
duced both by frequent and large-scale changes in fetal position in utero and by maternal res-
piration (Thomason, 2020). One previous attempt to remove motion-related artifacts in fetal
fMRI employed a combined approach of slice to volume registration and scattered data interpo-
lation with bias field and spin history corrections on a small sample of eight fetuses (Ferrazzi
et al., 2014). This approach avoids discarding frames, but requires additional scans to assist reg-
istration and estimations of field inhomogeneity. Im Gegensatz, ICA denoising can be implemented
without additional scans. At the single-subject level, ICA-based denoising has proven to be a
powerful tool for separating neural-related signal from different sources of noise, einschließlich
movement artifacts (Griffanti et al., 2017). This is the first study to verify the efficacy of ICA
denoising in fetal imaging in a large fetal cohort. Außerdem, the components manually labeled
here can be used as a training set for future automatic signal/noise classifiers in fetal imaging
Daten. Given ICA does not guarantee a uniform reconstruction of all frames, especially for
high-motion periods, combination of censoring, ICA-based denoising, and covariate regression
(z.B., Bewegung, fame count, usw.) may be advisable for mitigating noise in future fetal fMRI studies.

Spatial smoothing can improve signal to noise and reduce the effects of spatial normaliza-
tion misalignment (Lowe & Sorenson, 1997) at the expense of decreasing resolution. Previous
studies on adult brains suggest that the kernel size should be at least twice the size of the fMRI
acquisition voxel (Alahmadi, 2021). Jedoch, best practices for smoothing parameters in fetal
fMRI remain to be addressed. Given the significantly smaller size of fetal brains compared to
adult brains, the smoothing kernel recommended in adult imaging may not be appropriate. Wir
observed that using a larger smoothing kernel resulted in enhanced cross-hemisphere func-
tional connectivity in fetuses, fitting with likely improvement in signal to noise. Jedoch,
given the trade-off between sensitivity and spatial specificity (d.h., the resolution) of findings,
general advice in fetal RSFC could be to use a moderate value of ×1 to ×2 voxel size. Best
practices in kernel selection will necessarily vary with attributes of the data and with the ana-
lytic approach being applied.

While the present study examines key steps in preprocessing fetal fMRI data, it is important
to note that many questions remain to be addressed. Fetal data present novel challenges,

Netzwerkneurowissenschaften

717

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

especially higher and more complex motion and unique image artifacts (van den Heuvel &
Thomason, 2016). Methods tested in this article draw from parameters that are most standard
in published reports in the literature (Jakab et al., 2014; Thomason et al., 2019; Thomason
et al., 2021A; Turk et al., 2019). Jedoch, there are alternative emergent approaches,
such as that presented by Scheinost and colleagues that perform automatic censoring of
low-quality frames and aim to correct for both large and small motion, that are important to
explore with advancement of this field of study (Scheinost et al., 2018). It will be valuable for
future works to additionally evaluate the efficacy of alternative denoising strategies, wie zum Beispiel
CompCorr (Behzadi et al., 2007), confound regression and band-pass filtering (Yan et al.,
2013), and advanced slice level reconstruction (Ferrazzi et al., 2014), and to do so across
datasets with variable noise profiles and motion thresholds. Future studies will also benefit
from considering interaction between preprocessing steps, such as whether the choice of
normalization target would affect the ICA performance. Additional areas for future work are
to test the generalizability of preprocessing steps in data collected from different
scanners/vendors, and across in different populations, including clinical samples. Außerdem,
it will be useful to evaluate alternative fMRI acquisition techniques, such as multiecho–echo
planar imaging (ME-EPI), which lends itself to empirically informed strategies for denoising
data during postprocessing (Kundu et al., 2012). Außerdem, an additional acquisition
mapping the field may help to address the distortions of the image. Endlich, there is great prom-
ise in using deep learning algorithms to automate manual steps involved in fetal fMRI data
Verarbeitung, such as brain segmentation (Rutherford et al., 2021). One can imagine further
development of these, even for purposes of automated identification of noise and signal com-
ponents following ICA. Gesamt, fetal research MRI represents an extraordinary opportunity for
basic and clinical science, but it does require continued investment toward optimization and
transition to the mainstream. This is elaborated further in a recent commentary (Rajagopalan
et al., 2021). The present study addresses common decision points in fMRI data processing
and provides empirical comparisons of outputs achieved when applying different methods
at each step.

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

ACKNOWLEDGMENTS

The authors thank Jasmine Hect and Pavan Jella for their assistance in data acquisition and
thank Ava Palopoli and Amyn Majbri for assistance with data management and quality assur-
ance. Wichtig, the authors thank participant families who generously shared their time and
expressed interest in helping future babies to achieve their best possible health outcomes.

SUPPORTING INFORMATION

Supporting information for this article is available at https://doi.org/10.1162/netn_a_00254.
The data and code used in this study will be made available via https://ndar.nih.gov/ and/or
accessed upon direct request to M. E. Thomason (Daten) or L. Ji (Code).

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

BEITRÄGE DES AUTORS

Lanxin Ji: Konzeptualisierung; Datenkuration; Formale Analyse; Methodik; Software;
Visualisierung; Writing – original draft; Writing – review & Bearbeitung. Cassandra L. Hendrix:
Konzeptualisierung; Visualisierung; Writing – original draft; Writing – review & Bearbeitung. Moriah
E. Thomason: Konzeptualisierung; Akquise von Fördermitteln; Ressourcen; Aufsicht; Writing –
Rezension & Bearbeitung.

Netzwerkneurowissenschaften

718

Evaluation of fetal fMRI preprocessing steps

FUNDING INFORMATION

Moriah E. Thomason, Foundation for the National Institutes of Health (https://dx.doi.org/10
.13039/100000009), Award ID: MH110793. Moriah E. Thomason, Foundation for the
Nationale Gesundheitsinstitute (https://dx.doi.org/10.13039/100000009), Award ID:
DA050287. Moriah E. Thomason, Foundation for the National Institutes of Health (https://dx
.doi.org/10.13039/100000009), Award ID: MH122447. Moriah E. Thomason, Foundation for
the National Institutes of Health (https://dx.doi.org/10.13039/100000009), Award ID:
ES032294.

VERWEISE

Alahmadi, A. A. (2021). Effects of different smoothing on global and
regional resting functional connectivity. Neuroradiology, 63(1),
99–109. https://doi.org/10.1007/s00234-020-02523-8, PubMed:
32840683

Allen, E. A., Erhardt, E. B., Damaraju, E., Gruner, W., Segall, J. M.,
Silva, R. F., … Kalyanam, R. (2011). A baseline for the multi-
variate comparison of resting-state networks. Frontiers in Systems
Neurowissenschaften, 5, 2. https://doi.org/10.3389/fnsys.2011.00002,
PubMed: 21442040

Anderson, A. L., & Thomason, M. E. (2013). Functional plasticity
before the cradle: A review of neural functional imaging in the
human fetus. Neurowissenschaften & Bioverhaltensberichte, 37(9),
2220–2232. https://doi.org/10.1016/j.neubiorev.2013.03.013,
PubMed: 23542738

Beckmann, C. F., & Schmied, S. M. (2004). Probabilistic independent
component analysis for functional magnetic resonance imaging.
IEEE Transactions on Medical Imaging, 23(2), 137–152. https://
doi.org/10.1109/TMI.2003.822821, PubMed: 14964560

Behzadi, Y., Restom, K., Liau, J., & Liu, T. T. (2007). A component
based noise correction method (CompCor) for BOLD and perfu-
sion based fMRI. NeuroImage, 37(1), 90–101. https://doi.org/10
.1016/j.neuroimage.2007.04.042, PubMed: 17560126

Botvinik-Nezer, R., Holzmeister, F., Camerer, C. F., Dreber, A.,
Huber, J., Johannesson, M., … Adcock, R. A. (2020). Variability
in the analysis of a single neuroimaging dataset by many teams.
Natur, 582(7810), 84–88. https://doi.org/10.1038/s41586-020
-2314-9, PubMed: 32483374

Bright, M. G., Tench, C. R., & Murphy, K. (2017). Potential pitfalls
when denoising resting state fMRI data using nuisance regres-
sion. NeuroImage, 154, 159–168. https://doi.org/10.1016/j
.neuroimage.2016.12.027, PubMed: 28025128

Calhoun, V. D., Adalı, T., & Pekar, J. J. (2004). A method for com-
paring group fMRI data using independent component analysis:
Application to visual, motor and visuomotor tasks. Magnetic
Resonance Imaging, 22(9), 1181–1191. https://doi.org/10.1016/j
.mri.2004.09.004, PubMed: 15607089

Calhoun, V. D., Wager, T. D., Krishnan, A., Rosch, K. S., Seymour,
K. E., Nebel, M. B., … Kiehl, K. (2017). The impact of T1 versus
EPI spatial normalization templates for fMRI data analyses.
Kartierung des menschlichen Gehirns, 38(11), 5331–5342. https://doi.org/10
.1002/hbm.23737, PubMed: 28745021

Chi, J. G., Dooling, E. C., & Gilles, F. H. (1977). Gyral development
of the human brain. Annals of Neurology, 1(1), 86–93. https://doi
.org/10.1002/ana.410010109, PubMed: 560818

Dobbing, J., & Sands, J. (1973). Quantitative growth and develop-
ment of human brain. Archives of Disease in Childhood, 48(10),
757–767. https://doi.org/10.1136/adc.48.10.757, PubMed:
4796010

Doria, V., Beckmann, C. F., Arichi, T., Händler, N., Groppo, M.,
Turkheimer, F. E., … Nunes, R. G. (2010). Emergence of resting
state networks in the preterm human brain. Verfahren der
Nationale Akademie der Wissenschaften, 107(46), 20015–20020. https://
doi.org/10.1073/pnas.1007921107, PubMed: 21041625

Esteban, O., Markiewicz, C. J., Blair, R. W., Moodie, C. A., Isik,
A. ICH., Erramuzpe, A., … Snyder, M. (2019). fMRIPrep: A robust
preprocessing pipeline for functional MRI. Nature Methods, 16(1),
111–116. https://doi.org/10.1038/s41592-018-0235-4, PubMed:
30532080

Ferrazzi, G., Murgasova, M. K., Arichi, T., Malamateniou, C., Fuchs,
M. J., Makropoulos, A., … Aljabar, P. (2014). Resting State fMRI in
the moving fetus: A robust framework for motion, bias field and
spin history correction. NeuroImage, 101, 555–568. https://doi
.org/10.1016/j.neuroimage.2014.06.074, PubMed: 25008959
Fransson, P., Skiöld, B., Horsch, S., Nordell, A., Blennow, M.,
Lagercrantz, H., & Åden, U. (2007). Resting-state networks in
the infant brain. Proceedings of the National Academy of
Wissenschaften, 104(39), 15531–15536. https://doi.org/10.1073/pnas
.0704380104, PubMed: 17878310

Friston, K. J., Williams, S., Howard, R., Frackowiak, R. S., & Turner, R.
(1996). Movement-related effects in fMRI time-series. Magnetic
Resonance in Medicine, 35(3), 346–355. https://doi.org/10.1002
/mrm.1910350312, PubMed: 8699946

Gholipour, A., Rollins, C. K., Velasco-Annis, C., Ouaalam, A.,
Akhondi-Asl, A., Afacan, O., … Yang, E. (2017). A normative
spatiotemporal MRI atlas of the fetal brain for automatic segmen-
tation and analysis of early brain growth. Wissenschaftliche Berichte, 7(1),
1–13. https://doi.org/10.1038/s41598-017-00525-w, PubMed:
28352082

Glasser, M. F., Sotiropoulos, S. N., Wilson, J. A., Coalson, T. S.,
Fischl, B., Andersson, J. L., … Polimeni, J. R. (2013). The minimal
preprocessing pipelines for the Human Connectome Project.
N e u ro I m a g e , 8 0 , 1 0 5 – 1 2 4 . h t t p s : / / d o i . o rg / 1 0 . 1 0 1 6 / J
.neuroimage.2013.04.127, PubMed: 23668970

Griffanti, L., Douaud, G., Bijsterbosch, J., Evangelisti, S., Alfaro-
Almagro, F., Glasser, M. F., … Carone, D. (2017). Hand classifica-
tion of fMRI ICA noise components. NeuroImage, 154, 188–205.
https://doi.org/10.1016/j.neuroimage.2016.12.036, PubMed:
27989777

Netzwerkneurowissenschaften

719

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

Griffanti, L., Salimi-Khorshidi, G., Beckmann, C. F., Auerbach, E. J.,
Douaud, G., Sexton, C. E., … Mackay, C. E. (2014). ICA-based
artefact removal and accelerated fMRI acquisition for improved
resting state network imaging. NeuroImage, 95, 232–247. https://
doi.org/10.1016/j.neuroimage.2014.03.034 , PubMed:
24657355

Himberg, J., Hyvärinen, A., & Esposito, F. (2004). Validating the
independent components of neuroimaging time series via cluster-
ing and visualization. NeuroImage, 22(3), 1214–1222. https://doi
.org/10.1016/j.neuroimage.2004.03.027, PubMed: 15219593
Jakab, A., Schwartz, E., Kasprian, G., Gruber, G. M., Prayer, D.,
Schöpf, V., & Langs, G. (2014). Fetal functional imaging portrays
heterogeneous development of emerging human brain networks.
Grenzen der menschlichen Neurowissenschaften, 8, 852. https://doi.org/10
.3389/fnhum.2014.00852, PubMed: 25374531

Kiviniemi, V., Starck, T., Remes, J., Long, X., Nikkinen, J., Haapea,
M., … Zang, Y. F. (2009). Functional segmentation of the brain
cortex using high model order group PICA. Menschliches Gehirn
Mapping, 30(12), 3865–3886. https://doi.org/10.1002/ hbm
.20813, PubMed: 19507160

Kundu, P., Es sagte, S. J., Evans, J. W., Luh, W.-M., & Bandettini, P. A.
(2012). Differentiating BOLD and non-BOLD signals in fMRI time
series using multi-echo EPI. NeuroImage, 60(3), 1759–1770.
https://doi.org/10.1016/j.neuroimage.2011.12.028, PubMed:
22209809

Lowe, M. J., & Sorenson, J. A. (1997). Spatially filtering functional
magnetic resonance imaging data. Magnetic Resonance in
Medicine, 37(5), 723–729. https://doi.org/10.1002/mrm
.1910370514, PubMed: 9126946

Pruim, R. H., Mennes, M., van Rooij, D., Llera, A., Buitelaar, J. K., &
Beckmann, C. F. (2015). ICA-AROMA: A robust ICA-based
strategy for removing motion artifacts from fMRI data. Neuro-
Image, 112, 267–277. https://doi.org/10.1016/j.neuroimage
.2015.02.064, PubMed: 25770991

Rajagopalan, V., Deoni, S., Panigrahy, A., & Thomason, M. E.
(2021). Is fetal MRI ready for neuroimaging prime time? Ein
examination of progress and remaining areas for development.
Developmental Cognitive Neuroscience, 51, 100999. https://doi
.org/10.1016/j.dcn.2021.100999, PubMed: 34391003

Reilly, K., & Gutierrez, J. (2021). The embryological development
of the cerebrovascular system. In Pediatric vascular neurosurgery
(S. 1–5). Cham, Schweiz: Springer. https://doi.org/10.1007
/978-3-030-74749-7_1

Rissanen, J. (1978). Modeling by shortest data description. Automa-
tica, 14(5), 465–471. https://doi.org/10.1016/0005-1098(78)
90005-5

Rummel, C., Verma, R. K., Schöpf, V., Hauf, M., Abela, E., Zapata
Berruecos, J. F., & Wiest, R. (2013). Time course based artifact
identification for independent components of resting-state fMRI.
Grenzen der menschlichen Neurowissenschaften, 7, 214. https://doi.org/10
.3389/fnhum.2013.00214, PubMed: 23734119

Rutherford, S., Sturmfels, P., Angstadt, M., Hect, J., Wiens, J., Transporter
den Heuvel, M. ICH., … Thomason, M. (2021). Automated brain
masking of fetal functional MRI with open data. Neuroinfor-
matics. https://doi.org/10.1007/s12021-021-09528-5, PubMed:
34129169

Scheinost, D., Onofrey, J. A., Kwon, S. H., Cross, S. N., Sze, G.,
Ment, L. R., & Papademetris, X. (2018, April 4–7). A fetal fMRI

specific motion correction algorithm using 2nd order edge
Merkmale [Paper presentation]. 2018 IEEE 15th International
Symposium on Biomedical Imaging. https://doi.org/10.1109
/ISBI.2018.8363807

Schöpf, V., Kasprian, G., & Prayer, D. (2011). Funktionelle Bildgebung
in the fetus. Topics in Magnetic Resonance Imaging, 22(3),
113–118. https://doi.org/10.1097/ RMR.0b013e3182699283,
PubMed: 23558466

Serag, A., Kyriakopoulou, V., Rutherford, M. A., Edwards, A. D.,
Dämmerung, J. V., Aljabar, P., … Rueckert, D. (2012). A multi-channel
4D probabilistic atlas of the developing brain: Application to
fetuses and neonates. Annals of the BMVA, 2012(3), 1–14.

Shattuck, D. W., & Leahy, R. M. (2002). BrainSuite: An automated
cortical surface identification tool. Medical Image Analysis, 6(2),
129–142. https://doi.org/10.1016/S1361-8415(02)00054-3,
PubMed: 12045000

Smyser, C. D., Inder, T. E., Shimony, J. S., Hill, J. E., Degnan, A. J.,
Snyder, A. Z., & Neil, J. J. (2010). Longitudinal analysis of neural
network development in preterm infants. Hirnrinde, 20(12),
2852–2862. https://doi.org/10.1093/cercor/ bhq035, PubMed:
20237243

Thomason, M. E. (2020). Development of brain networks in utero:
Relevance for common neural disorders. Biologische Psychiatrie,
88(1), 40–50. https://doi.org/10.1016/j.biopsych.2020.02.007,
PubMed: 32305217

Thomason, M. E., Dassanayake, M. T., Shen, S., Katkuri, Y., Alexis,
M., Anderson, A. L., … Romero, R. (2013). Cross-hemispheric
functional connectivity in the human fetal brain. Science Trans-
lational Medicine, 5(173), 173ra124. https://doi.org/10.1126
/scitranslmed.3004978, PubMed: 23427244

Thomason, M. E., Grove, L. E., Lozon Jr., T. A., Vila, A. M., Ye, Y.,
Nye, M. J., … Yeo, L. (2015). Age-related increases in long-range
connectivity in fetal functional neural connectivity networks in
utero. Developmental Cognitive Neuroscience, 11, 96–104.
https://doi.org/10.1016/j.dcn.2014.09.001, PubMed: 25284273
Thomason, M. E., Hect, J. L., Rauh, V. A., Trentacosta, C.,
Wheelock, M. D., Eggebrecht, A. T., … Burt, S. A. (2019). Pre-
natal lead exposure impacts cross-hemispheric and long-range
connectivity in the human fetal brain. NeuroImage, 191, 186–192.
https://doi.org/10.1016/j.neuroimage.2019.02.017, PubMed:
30739062

Thomason, M. E., Hect, J. L., Waller, R., & Curtin, P. (2021A). Inter-
active relations between maternal prenatal stress, fetal brain
Konnektivität, and gestational age at delivery. Neuropsychophar-
macology, 46(10), 1839–1847. https://doi.org/10.1038/s41386
-021-01066-7, PubMed: 34188185

Thomason, M. E., Palopoli, A. C., Jariwala, N. N., Werchan, D. M.,
Chen, A., Adhikari, S., … Trentacosta, C. J. (2021B). Miswiring
das Gehirn: Human prenatal Δ9-tetrahydrocannabinol use associ-
ated with altered fetal hippocampal brain network connectivity.
Developmental Cognitive Neuroscience, 51, 101000. https://doi
.org/10.1016/j.dcn.2021.101000, PubMed: 34388638

Toga, A. W., & Thompson, P. M. (2003). Mapping brain asymmetry.
Nature Reviews Neurowissenschaften, 4(1), 37–48. https://doi.org/10
.1038/nrn1009, PubMed: 12511860

Turk, E., van den Heuvel, M. ICH., Benders, M. J., de Heus, R., Franx,
A., Manning, J. H., … van den Heuvel, M. P. (2019). Funktional
connectome of the fetal brain. Zeitschrift für Neurowissenschaften, 39(49),

Netzwerkneurowissenschaften

720

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

:
/
/

D
ich
R
e
C
T
.

ich
T
.

e
D
u
N
e
N
A
R
T
ich
C
e
–
P
D

F
/

6
3
7
0
2
2
0
3
6
0
3
2
N
e
N
_
A
_
0
0
2
5
4
P
D

B
j
G
u
e
S
T

Ö
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Evaluation of fetal fMRI preprocessing steps

9716–9724. https://doi.org/10.1523/JNEUROSCI.2891-18.2019,
PubMed: 31685648

van den Heuvel, M. ICH., & Thomason, M. E. (2016). Funktional
connectivity of the human brain in utero. Trends im kognitiven Bereich
Wissenschaften, 20(12), 931–939. https://doi.org/10.1016/j.tics.2016
.10.001, PubMed: 27825537

van den Heuvel, M. ICH., Turk, E., Manning, J. H., Hect, J., Hernandez-
Andrade, E., Hassan, S. S., … Thomason, M. E. (2018). Hubs in
the human fetal brain network. Developmental Cognitive Neuro-
Wissenschaft, 30, 108–115. https://doi.org/10.1016/j.dcn.2018.02
.001, PubMed: 29448128

Van Dijk, K. R., Sabuncu, M. R., & Buckner, R. L. (2012). Der
influence of head motion on intrinsic functional connectivity
MRT. NeuroImage, 59(1), 431–438. https://doi.org/10.1016/j
.neuroimage.2011.07.044, PubMed: 21810475

Yan, C.-G., Cheung, B., Kelly, C., Colcombe, S., Craddock, R. C., Aus
Martino, A., … Milham, M. P. (2013). A comprehensive assessment
of regional variation in the impact of head micromovements on
functional connectomics. NeuroImage, 76, 183–201. https://doi
.org/10.1016/j.neuroimage.2013.03.004, PubMed: 23499792
Yan, C.-G., Wang, X.-D., Zuo, X.-N., & Zang, Y.-F. (2016). DPABI:
Data processing and analysis for (resting-state) brain imaging.
Neuroinformatics, 14(3), 339–351. https://doi.org/10.1007
/s12021-016-9299-4, PubMed: 27075850

Zuo, X.-N., Kelly, C., Di Martino, A., Mennes, M., Margulies, D. S.,
Bangaru, S., … Castellanos, F. X. (2010). Growing together and
growing apart: Regional and sex differences in the lifespan
developmental trajectories of functional homotopy. Zeitschrift für
Neurowissenschaften, 30(45), 15034–15043. https://doi.org/10.1523
/JNEUROSCI.2612-10.2010, PubMed: 21068309