Methods of Analysis

Functional MRI for Presurgical

Planning
Kyung K. Peck, PhDa,b,*, Nicholas S. Cho, BAc, Andrei I. Holodny, MDd

KEY WORDS

Functional MRI (fMRI)
Methods of analysis
Pre-processing
Post-processing

Statistical analysis

KEY POINTS

Task instruction and rehearsal are the most important steps during an fMRI procedure.

Data processing steps must be taken to optimize the signal that is associated with specific func-
tional tasks and to minimize any noise-related signal.

Pre-processed data significantly affects statistical analysis, which has a great impact on image
interpretation.

Algorithms for head motion detection and correction, the theory and practicalities associated with
data processing remain complex and constantly evolving.

INTRODUCTION cortical areas adjacent to a tumor cannot be accu-

rately defined by simply examining a structural
Since its discovery in the early 1990s, functional brain image. In addition, one must be careful to
MRI (fMRI) has been used to study brain function acknowledge the existence of neurovascular
and to advance clinical care.1,2 One well- uncoupling (NVU) that may also influence the
established application of fMRI in the clinical blood-oxygen-level-dependent (BOLD) fMRI
setting is in neurosurgical mapping for patients signal and can lead to false negative results (see
with brain tumors near eloquent cortical areas.3–5 separate article elsewhere in this issue).14–19 The
The main goal of neurosurgery is to maximize tu- goal of clinical fMRI is to preoperatively identify
mor resection while sparing important brain struc- important functional areas associated with spe-
tures near the tumor, which relies on the proper cific functional tasks critical in the patients’ daily
identification of eloquent cortical areas.6–9 Howev- life, such as hand movement and language func-
er, interestingly, there is considerable anatomic tion.18,20,21 For example, when a lesion presents
variation in the human brain between individuals, in the vicinity of important areas like sensory motor
and the presence of a tumor can lead to cortical or language areas, surgeons may request an fMRI
plasticity that recruits cerebral areas for tasks scan to know how close the regions of activation
that are not originally associated with that re- are located near the lesion to best plan the
gion.10–13 As a result, the location of eloquent

Funding information: National Institutes of Health (NIH): NIH-NIGMS Training Grant GM008042 (N.S. Cho), NIH-
NIBIB R01 EB022720 (Makse and Holodny, principal investigators (PIs]), NIH-NCI R21 CA220144 (A.I. Holodny
neuroimaging.theclinics.com

and K.K. Peck, PI’s), NIH-NCI U54 CA137788 (Ahles, PI), NIH-NCI P30 CA008748 (Thompson, PI).
a
Department of Medical Physics, Memorial Sloan Kettering Cancer Center, 1275 York Avenue, New York, NY
10021, USA; b Department of Radiology, Memorial Sloan Kettering Cancer Center, 1275 York Avenue, New
York, NY 10021, USA; c Medical Scientist Training Program, David Geffen School of Medicine at UCLA, 885
Tiverton Drive, Los Angeles, CA 90024, USA; d Department of Radiology, Memorial Sloan Kettering Cancer
Center, 1275 York Avenue, New York, NY 10065, USA
* Corresponding author. Department of Medical Physics, Memorial Sloan Kettering Cancer Center, 1275 York
Avenue, New York, NY 10021.
E-mail address: [email protected]

Neuroimag Clin N Am 31 (2021) 23–32

https://doi.org/10.1016/j.nic.2020.09.006
1052-5149/21/Ó 2020 Elsevier Inc. All rights reserved.
24 Peck et al

Fig. 1. Examples of presurgical functional MRI scans in a 3D rendering map for a patient with a tumor in the left
hemisphere. (A) During a hand motor task, activation in the precentral gyrus is posterior to the lesion. (B) During
a language task, activation in Broca’s and Wernicke’s areas is inferior to the lesion.

surgery, including for intraoperative brain mapping account for its potential growth or spread. Based
or awake brain surgery (Fig. 1). on the lesion location, specific fMRI paradigms
are prescribed to elucidate the nearby and/or
WORKFLOW IN CLINICAL FUNCTIONAL MRI affected cortical areas via task function.
General Workflow On the day of the scan, the fMRI team begins by
instructing the patient regarding the functional
To obtain successful fMRI maps for presurgical
tasks that the patient will perform in the scanner.
planning, there are several important steps
During the scan itself, the patient’s performance
involved. The following diagram shows the typical
is monitored using the available real-time pre-pro-
workflow for a clinical fMRI procedure (Fig. 2).
cessing software in the MRI suite. Current modern
First, the neurosurgeon refers the patient to the
scanners incorporate some capability to perform
fMRI team. The presence of any neurologic deficits
t-test or Z-score analysis on a voxel-by-voxel level
should be noted to help guide the fMRI team with
in real-time. This quick, real-time review of the
the lesion location. The fMRI team will also deter-
fMRI activation will allow the fMRI team to have
mine the lesion location by examining any previous
knowledge about the quality of the data acquisi-
images of the patients’ lesion, if available, and
tion, guiding the team to either continue with the
scan or to re-conduct the scan. The completed
fMRI and anatomic dataset can then be trans-
ferred to an off-line computer for pre-processing
and statistical analysis. A neuroradiologist then re-
views the functional activation maps generated,
and a report is written for the neurosurgeon.

Task Instruction and Monitoring the Task

Performance
The importance of following 2 procedures before
any scanning is conducted cannot be overstated:
task instruction and monitoring task performance.
In clinical fMRI, the patient’s task instruction and
rehearsal before the scan are of utmost impor-
tance and more critical than the image processing
itself. The success of a fMRI scan also mainly de-
Fig. 2. Presurgical planning fMRI schema. pends on the subject’s task performance: even if
 Methods of Analysis 25

you have the highest quality MRI scanner, you assessment, head motion correction, spatial
cannot generate a good fMRI scan if the subject’s smoothing, and linear trend removal.
task performance is poor. Task performance is Normally, the fMRI data is processed using
especially important when the scan is used for pre- commercial or freely downloadable software (eg,
surgical planning because the patients may have FSL and AFNI22,23). Commercial software is
neurologic deficits related to their lesion. designed to be user-friendly and provides auto-
The subject’s task performance is also directly matic or manual pre-processing tools. However,
related to how well the subject understood the the fMRI data obtained from patients occasionally
task instruction. Before a subject is scanned, a needs more complicated and patient-specific im-
fMRI specialist instructs the subject about the age processing steps that cannot be handled by
functional tasks that will be performed in the the pre-determined automated processes found
scanner. During this procedure, the specialist in commercial software. Therefore, additional
should be able to gauge the subject’s perfor- advanced software may be used to tackle compli-
mance level as they rehearse the task before cated issues including randomly-occurring artifac-
going to the scanner. For example, hand or tual images, severe head motion, and incorrect
foot weakness can be evaluated when task performance.
rehearsing motor tasks. If the subject shows
very weak hand or foot motor strength, the Image Quality Assessment
fMRI specialist may need to make a decision to After the raw fMRI data are transferred for data
change the paradigm from a self-controlled processing, it is essential to carefully review the
task to an externally controlled stimulation for images and time courses, especially the areas
sensory motor task. For language function, where activation is expected. Any sources of arti-
several assessments including the Boston facts and noise-related abrupt signal change in
Naming Test can be performed to assess for the data should be identified and removed or mini-
any language deficits. If the subject has a lan- mized. Performing these image quality checks
guage deficit, then it would be better to apply consistently throughout the pre-processing steps
several different types of language paradigms significantly improves the final images used for
because one language paradigm could be easier clinical interpretation.
than another for the subject, and different para-
digms may elicit different brain activations that Functional MRI Artifacts
can be interpreted together as a whole for gen-
eral language function. Careful paradigm selec- The main artifacts in fMRI are caused by suscepti-
tion according to the patient’s condition, bility, which are due to local magnetic field inho-
proper instruction, and practice of the para- mogeneity. Common sources of these artifacts
digms prior to scanning also reduces difficulties are at air-tissue interfaces such as the ear and
with head motion artifacts and poor or inconsis- nasal canals, which can cause signal loss in the
tent task performance. These nontechnical con- auditory cortex and frontal lobe, respectively.
siderations during an fMRI scan can often be There are other potential sources of susceptibility
overlooked, but they remain vital for a successful artifacts including the sphenoid sinus in the inferior
clinical fMRI scan. frontal cortex and geometric distortion due to
scanner Bo field inhomogeneity. There are also
subject-related artifacts including those caused
by the movement of brain parenchyma and the
IMAGE PRE-PROCESSING subject’s motion due to heart rate, respiratory
Purpose of Functional MRI Data Processing rate, and bulk head motion. Cardiac pulsation arti-
After acquiring the fMRI data, fMRI data process- facts can appear as large as a few percent signal
ing is conducted to maximize the signal and filter change of BOLD contrast. Bulk head motion can
out the noise that is associated with the task per- cause voxels on the periphery of the brain to
formance (Fig. 3). This process involves reducing appear activated.24–26 Reviewing the images in
the artifactual signal in the voxel time course that “real-time mode” and checking temporal varia-
is, not part of the subject’s functional task perfor- tions in the time series often allows for identifying
mance. When this process is finished, there is problematic volumes or slices that should be
improved detectability of statistically significant removed before further processing steps (Fig. 4).
activation clusters that can be used to generate
Head Motion
functional maps for presurgical scanning. fMRI
data processing is generally conducted through Motion is a major problem for fMRI studies, for it is
several, essential steps including image quality the greatest cause of fMRI examination failure. In
26 Peck et al

Motion during fMRI can be of 2 kinds: randomly

timed motion or stimulus-correlated motion.
Randomly timed motion (that is not correlated
with the stimulus task) can block real activity by
increasing false negatives in the activation map.
False negatives may increase when the random
motion occurs during the baseline volumes where
no task-induced activation should be present. On
the other hand, stimulus-correlated motion creates
false positives, as voxels of higher signal intensity
can be shifted into surrounding areas of the acti-
vating region that are originally occupied by other
voxels of lower signal intensity. If 2 neighboring
voxels differ in intrinsic brightness by 15%, then a
Fig. 3. Pre-processing stream for fMRI data.
motion of 10% of a voxel dimension can result in
a 1.5% signal change which is comparable to the
clinical fMRI, patients tend to show more frequent BOLD signal percent change during stimulus acti-
unintended head and body movements due to vation at 1.5 T.25 Stimulus-correlated motion is
their underlying, neurologic condition than healthy particularly problematic because most software
control subjects. Motion is also more easily pre- programs do not have a tool to separate true activ-
vented than corrected. Therefore, patients must ity from activation due to head motion.26 Stimulus-
be specifically instructed how head motion can correlated motion can be especially severe in areas
degrade the image quality and may lead to a lying near the high-contrast boundaries in the brain
repeated scan. In the scanner, ensuring maximal (Fig. 5). Percent signal change of the voxels due to
patient comfort can decrease the chance of mo- stimulus-correlated motion is typically much higher
tion. In addition, vocal feedback and encourage- than that of the true activated voxels. The percent
ment right before each scan will help to minimize signal change could be used as a threshold to
motion artifacts. remove such false positive voxels.
As task-related BOLD signal percentage
changes are small (on the order of single-digit per-
Motion Correction
centage changes) and because data analysis is
based on signal variation in the voxel time series, Despite efforts to minimize motion, head motion
even small head motion can heavily confound sta- can still occur and must be corrected during pro-
tistical analyses.27 Motion can occur during cessing. During motion correction processing,
different paradigms and in different directions; we are assuming that the same voxel represents
leg motion for foot motor tasks can cause head the same location in the brain throughout the
movement in the inferior-to-superior direction, scans and the signal change in the same voxel be-
while the tongue motor task cause shifts in head tween volumes is mainly due to activation coupling
placement. with task performance. Many motion correction

Fig. 4. Artifactual image shown in the white outline and the associated abnormal signal fluctuation in the time
course in the region of interest.
 Methods of Analysis 27

for a time shorter than the length of the TR empha-

sizes the importance of having the patient hold as
still as possible. If such significant motion occurs in
one of stimulus blocks, it would be better to exclude
the volume in the following processing steps.

Spatial Smoothing
Generally, in fMRI data, high spatial frequencies
likely represent noise components whereas low
frequencies are changes produced by blood flow
and thus are more likely to represent the BOLD
Fig. 5. Artifactual activation from stimulus-correlated signal due to neurovascular coupling (NVC).
motion. Spatial smoothing increases the signal-to-noise
ratio (SNR) by suppressing high frequency noise
components and enhancing low frequency sig-
algorithms and software packages typically are nals. This is a helpful technique to use in data pro-
applied during pre-processing before statistical cessing because it enhances the likelihood of
analysis, with no single package producing determining eloquent cortices in the brain during
dramatically better results than others.27 Following functional tasks. By suppressing high frequencies,
each motion correction step, the fMRI data can be smoothing increases the area of the brain where
quality control checked by looking at the motion small signal changes from NVC can be detected,29
parameter plots and viewing the images in "cine thus allowing clinical decisions for a patient to be
mode" to check whether the mismatch between made from optimized functional information.
slices has improved. In spatial smoothing, voxels are averaged with
Generally, correction using rigid body motion (ie, their surrounding neighbors in an image, which
no change in the size or shape) by estimating 3- causes the blurring of sharp edges. Smoothing is
dimensional (3D) motion parameters (the set of 3 performed by convolving the image with a point
translocation and 3 rotation parameters) is used spread function using a Gaussian kernel with a
to minimize the difference between the reference width between 4 to 8 mm full width half maximum
volume and the other volumes. Usually, one can (FWHM). The FWHM provides an estimation of
choose the reference volume at the initial volumes smoothing of a Gaussian kernel as the standard de-
of the scan, and then register all remaining vol- viation(s) is calculated using the relation s 5 FWHM
umes against the reference volume. This is  0.425. After convolving with a point spread func-
because if 2D functional matching anatomic im- tion, a matrix of the same size is created. Then, the
ages are acquired before functional scan, then image is convolved with the matrix. By performing
presumably, the initial functional volumes would these steps, the signal from each voxel spreads
be matched the most with the 2D anatomic im- out to surrounding voxels. The optimal width of
ages. Next, the sum of absolute intensity differ- the smoothing filter determines the extent of blur-
ences between voxels in the reference volumes ring and is chosen to closely match the size of the
and the other volumes is calculated. The 6 motion region activated.30 A general rule of thumb for
parameters mentioned previously can be used to fMRI data is choosing an FWHM twice the voxel
detrend the data against correlation with the size. For example, if FOV 5 240 mm and 64  64
movements as a to minimize motion-correlated matrix with in-plane resolution 5 3.75  3.75 mm,
false positive activation.28 However, this approach the FWHM should be approximately 6 to 8 mm.
should be used with a caution because it can As FWHM increases, there is a tradeoff between
cause increased false negative voxels if the aver- better activation maps through increased SNR
aged time course associated with stimulus- and increased image blurring, as shown in Fig. 6.
correlated motion is similar to the time course Before smoothing, it is important to note the
associated with task-induced activation. task involved and the expected area of localization
Three-dimensional volume registration is gener- during imaging. Signal changes produced during
ally useful for intra- and intersession alignment. language and cognitive tasks are smaller than in
However, this approach cannot correct motions motor-sensory tasks. For example, BOLD re-
that occur within the time equal to or less than the sponses in the motor-sensory system are gener-
repetition time (TR) since this approach assumes a ally large (w2%–6%), robust, and within a
rigid movement of the entire volume. This limitation narrow area of possible activation (a few millime-
of being unable to correct for motion that occurs ters across). Thus, in presurgical mapping of the
28 Peck et al

Fig. 6. The effect of different size of Gaussian kernel applied to an image with in-plane resolution 2  2 mm. A
region of interest (red arrow) is placed in a hand motor gyrus and the associated time course is displayed.
Reduced high-frequency noise is obtained with increased kernel size.

motor cortex using fMRI, it is better to use narrow STATISTICAL ANALYSIS

or sometimes even no smoothing because an ac-
curate and robust activation map can often be pro- For statistical analysis, a hypothesis-driven statis-
duced. However, during language tasks, BOLD tical analysis is usually used to identify areas
responses are much less (w1%–3% signal showing increased response as compared with
change) in Broca’s and Wernicke’s Areas. To baseline conditions. For example, the t-test can
detect these smaller signal changes, a larger compare 2 conditions (for example, stimulus-off
kernel size for spatial smoothing is recommended. period and on period) and test if there is a mean
A larger kernel size increases the SNR, thereby difference. The analysis of variance test, as an
increasing the likelihood of retaining voxels extension of t-test, deals with more than 2 condi-
responding to functional tasks. tions. Generally, language and other cognitive
tasks produce activations that are widely distrib-
uted with less percentage signal change, which
Linear Trend Removal
makes the t-test approach not generally
Another frequently observed issue is a near- adequate.
linear baseline signal drift (0.0–0.015 Hz). This
temporal drift shows a linear trend and is likely
Analysis Methods: Correlation Analysis
due to scanner instabilities causing changes in
the local magnetic field. The low frequency The cross-correlation test is the most popular
signal drift issue affects statistical analysis and analysis method in clinical fMRI because the
power by altering the volume of activation.31 method addresses the changing time courses be-
Because the drifts are slowly rising and falling, tween stimulus-on and off periods (Fig. 7). For
the drifts can be effectively removed by using cross-correlation analysis, a boxcar reference
a high-pass filter. Several methods exist to waveform is built, representing the stimulus-on
remove the drift from analysis such as linear and off periods. When the patient is performing a
models, low-order polynomial models, and rest state and a task, a value of 0 and 1, respec-
spline models.25,27,28,32 tively, are set. The cross-correlation analysis
 Methods of Analysis 29

Fig. 7. Cross-correlation analysis. The green crosshairs show a voxel in the hand motor area. The voxel’s corre-
sponding time course is shown in black. The waveform convolved with hemodynamic response based on the
block paradigm is shown in red. The waveform correlates with BOLD signal across each voxel.

produces 2 results for each voxel. One is the always straightforward because in many situa-
magnitude of signal change and the other is statis- tions, data quality often varies between patients
tical significance of the magnitude calculated (cor- scanned. If our criteria are too conservative (ie, if
relation and the corresponding P value, our P value is too low), the power to detect mean-
respectively). ingful results will be low, and Type II errors (false
If a patient’s task performance shows delayed negative) will increase. If our criteria are too liberal
responses, there are 2 methods to address the (if P value is too high), the result will become
issue with correlation analysis. If the delay oc- contaminated by Type I errors (false positives).
curs consistently, then the designed ideal func- Ideally, the goal of setting a threshold is to maxi-
tion can be shifted to match the delay. If the mize the number of true positives while minimizing
delay occurs inconsistently, then the hemody- false positives.
namic response in the expected area can be
used as an ideal function in the method. In this Cluster analysis
method, motion parameters saved during head Another method to remove artifactually activated
motion correction can be used as a regressor, voxels is cluster analysis. In cluster analysis, a
and any voxels that are highly correlated with minimum statistical volume of a group of contig-
the motion profile should be removed. Motion- uous voxels, or a cluster, is set so that only the sig-
induced false positive activations can be mini- nificant clusters larger than the minimum volume
mized in this way. are visualized. Cluster analysis typically has good
sensitivity, but there is a risk to miss small but sig-
Setting up a threshold nificant voxels because the cluster threshold is
Corrected P<.01 remains the gold standard for based on the voxel size. Therefore, thresholds us-
general neuroimaging analysis, but setting a uni- ing both voxel level inference and cluster level
form statistical threshold in clinical fMRI is not inference should be properly applied to identify
30 Peck et al

voxels that pass both criteria. However, because for controlling FWE is Bonferroni correction where:
of the complexity of clinical fMRI data relative to a (corrected) 5 a/number of tests ja 5 desired a
neuroimaging studies with healthy subjects, it level ja (corrected) 5 FEW-corrected a level. For
would be recommended to carefully review the example, if the desired a 5 0.01 and there are
final activation map. 100,000 voxels to test, a (corrected) 5 0.01/
Despite the merit of the cluster analysis, cluster- 100,000 5 0.0000001. Therefore, the corrected a
based thresholding has limitations, including low value will be much lower than the desired level.
spatial specificity if the clusters are large.33 Also, However, the Bonferroni correction has several
the P value provided in cluster analysis does not limitations. The correction assumes that each
determine the statistical significance of voxel acti- test is independent, but in reality, voxels are not
vation within the cluster. Rather, the P value de- completely independent because time courses in
scribes the probability of obtaining a cluster of a adjacent voxels tend to be highly correlated and
given size or greater under the null hypothesis. because BOLD activity itself often spans large re-
gions. Also, spatial smoothing during pre-
Analysis Methods: Multiple Comparisons processing means that no voxels are truly inde-
An important issue in fMRI data analysis is setting pendent of its neighbors. As a result, the Bonfer-
an appropriate threshold for statistical activation roni correction overestimates the number of
maps. However, running the statistical analyses independent statistical tests.
separately for each voxel leads to the multiple Gaussian random field
comparisons issue. For instance, if the brain is As another FWE correction method, the Gaussian
divided into 100,000 voxels, then that means that Random Field takes into account the smoothness
voxel-level tests are repeated more than 100,000 of the data to reduce the overestimation in the
times. If the default P value threshold, P<.01, there Bonferroni correction. In the Gaussian Random
will be more than 1000 false positive voxels.34,35 Field correction, the data are smoothed to lower
As a result, the false positive rate across all voxels the resolution of the search area, thereby reducing
should be controlled and reduced. Otherwise, it the number of comparisons. Gaussian random
would be difficult to know if an activation result is field requires sufficient smoothness of data, such
truly positive. Therefore, the aim of correcting for as an FWHM w3 to 4 times the voxel size.
multiple comparisons is to identify areas of activity
that reflect true effects. In the actual clinical False discovery rate
setting, it is challenging to find an appropriate bal- The FWE mentioned previously ensures that false
ance between trying to minimize false positives positives will be controlled per all hypothesis tests.
(Type I error) while not being too omitting true ef- For example, a (FWE-corrected) 5 0.01 (1%) of
fects (Type II error). If the region(s) interested are contrast (across all voxels) will have a single false
known, the analysis can be limited to just those positive. In FDR, it controls the number of false
voxels with a prior hypothesis. In this way, the positives within the significant results obtained
number of voxels can be reduced and the multiple from a family of tests. If a (FDR-corrected 5 0.01),
comparison that should be corrected can be 1% of active voxels will be false positives. That
reduced. means the FDR will get more true active voxels,
Software used to analyze fMRI data can use but as a downside, it will have some false positive
various methods for multiple comparisons correc- voxels.
tion. For example, there is the Family-Wise Error
(FWE) correction that is based on the Bonferroni SUMMARY
correction or the Random Field Theory method.
The False Discovery Rate (FDR) correction method There are many technical and nontechnical steps
has also been used to control for multiple compar- involved in a successful clinical fMRI scan. The
isons.36 Image software tools automatically calcu- output from scanning and analysis can only be
late the corrected threshold while analyzing the as good as the input, so task instruction and
fMRI files based on the number of tested voxels rehearsal are the most important steps during an
and the number of voxels reporting significant ac- fMRI procedure. After fMRI data acquisition,
tivity before correction. proper data processing steps must be taken to
optimize the signal that is associated with specific
Family-wise error functional tasks and to minimize any noise-related
FWE is the probability of finding one or more false signal. Properly pre-processed data significantly
positives among all hypothesis tests. If affects statistical analysis, which has a great
FWE 5 0.01, the probability of finding one or impact on image interpretation. Even though there
more false positives is 1%. The simplest method is general agreement on how to process clinical
 Methods of Analysis 31

fMRI data, such as algorithms for head motion glioma: graph theory reveals the novel, adaptive
detection and correction, the theory and practical- network connectivity. Front Neurol 2019;10:702.
ities associated with data processing remain com- 13. Petrovich NM, Holodny AI, Brennan CW, et al. Iso-
plex and constantly evolving. lated translocation of Wernicke’s area to the right
hemisphere in a 62-year-man with a temporo-
parietal glioma. AJNR Am J Neuroradiol 2004;25:
DISCLOSURE
130–3.
K.K. Peck and N.S. Cho have nothing to disclose. 14. Chen CM, Hou BL, Holodny AI. Effect of age and tu-
AH is the Owner/President of fMRI Consultants, mor grade on BOLD functional MR imaging in pre-
LLC, a purely educational entity. operative assessment of patients with glioma.
Radiology 2008;248:971–8.
15. Del Ferraro G, Moreno A, Min B, et al. Finding influ-
REFERENCES
ential nodes for integration in brain networks using
1. Belyaev AS, Peck KK, Brennan NM, et al. Clinical optimal percolation theory. Nat Commun 2018;9:
applications of functional MR imaging. Magn Reson 2274.
Imaging Clin N Am 2013;21:269–78. 16. Fraga de Abreu VH, Peck KK, Petrovich-
2. Holodny AI, Shevzov-Zebrun N, Brennan N, et al. Brennan NM, et al. Brain tumors: the influence of tu-
Motor and sensory mapping. Neurosurg Clin N Am mor type and routine mr imaging characteristics at
2011;22:207–18, viii. BOLD Functional MR Imaging in the Primary Motor
3. Gabriel M, Brennan NP, Peck KK, et al. Blood oxy- Gyrus. Radiology 2016;281:876–83.
gen level dependent functional magnetic resonance 17. Mallela AN, Peck KK, Petrovich-Brennan NM, et al.
imaging for presurgical planning. Neuroimaging Clin Altered resting-state functional connectivity in the
N Am 2014;24:557–71. hand motor network in glioma patients. Brain Con-
4. Jacobs AH, Kracht LW, Gossmann A, et al. Imaging nect 2016;6:587–95.
in neurooncology. NeuroRx 2005;2:333–47. 18. Peck KK, Bradbury M, Petrovich N, et al. Presurgical
5. Petrella JR, Shah LM, Harris KM, et al. Preoperative evaluation of language using functional magnetic
functional MR imaging localization of language and resonance imaging in brain tumor patients with pre-
motor areas: effect on therapeutic decision making vious surgery. Neurosurgery 2009;64:644–52 [dis-
in patients with potentially resectable brain tumors. cussion: 652–3].
Radiology 2006;240:793–802. 19. Sun H, Vachha B, Laino ME, et al. Decreased hand
6. Brennan NP, Peck KK, Holodny A. Language map- motor resting-state functional connectivity in pa-
ping using fMRI and direct cortical stimulation for tients with glioma: analysis of factors including neu-
brain tumor surgery: the good, the bad, and the rovascular uncoupling. Radiology 2020;294:
questionable. Top Magn Reson Imaging 2016;25: 610–21.
1–10. 20. Kim MJ, Holodny AI, Hou BL, et al. The effect of prior
7. Gupta A, Shah A, Young RJ, et al. Imaging of brain surgery on blood oxygen level-dependent functional
tumors: functional magnetic resonance imaging and MR imaging in the preoperative assessment of brain
diffusion tensor imaging. Neuroimaging Clin N Am tumors. AJNR Am J Neuroradiol 2005;26:1980–5.
2010;20:379–400. 21. Li Q, Del Ferraro G, Pasquini L, et al. Core language
8. Holodny AI, Nusbaum AO, Festa S, et al. Correlation brain network for fMRI language task used in clinical
between the degree of contrast enhancement and applications. Netw Neurosci 2020;4:134–54.
the volume of peritumoral edema in meningiomas 22. Cox RW. AFNI: software for analysis and visualiza-
and malignant gliomas. Neuroradiology 1999;41: tion of functional magnetic resonance neuroimages.
820–5. Comput Biomed Res 1996;29:162–73.
9. Pan C, Peck KK, Young RJ, et al. Somatotopic orga- 23. Woolrich MW, Jbabdi S, Patenaude B, et al.
nization of motor pathways in the internal capsule: a Bayesian analysis of neuroimaging data in FSL.
probabilistic diffusion tractography study. AJNR Am Neuroimage 2009;45:S173–86.
J Neuroradiol 2012;33:1274–80. 24. Dagli MS, Ingeholm JE, Haxby JV. Localization of
10. Fisicaro RA, Jost E, Shaw K, et al. Cortical plasticity cardiac-induced signal change in fMRI. Neuroimage
in the setting of brain tumors. Top Magn Reson Im- 1999;9:407–15.
aging 2016;25:25–30. 25. Frank LR, Buxton RB, Wong EC. Estimation of
11. Holodny AI, Schulder M, Ybasco A, et al. Transloca- respiration-induced noise fluctuations from under-
tion of Broca’s area to the contralateral hemisphere sampled multislice fMRI data. Magn Reson Med
as the result of the growth of a left inferior frontal gli- 2001;45:635–44.
oma. J Comput Assist Tomogr 2002;26:941–3. 26. Thacker NA, Burton E, Lacey AJ, et al. The effects of
12. Li Q, Dong JW, Del Ferraro G, et al. Functional trans- motion on parametric fMRI analysis techniques.
location of Broca’s area in a low-grade left frontal Physiol Meas 1999;20:251–63.
32 Peck et al

27. Friston KJ, Ashburner J, Frith CD, et al. Spatial regis- 33. Woo C-W, Krishnan A, Wager TD. Cluster-extent
tration and normalization of images. Hum Brain based thresholding in fMRI analyses: pitfalls and
Mapp 1995;3:165–89. recommendations. NeuroImage 2014;91:412–9.
28. Jiang A, Kennedy DN, Baker JR, et al. Motion detec-
tion and correction in functional MR imaging. Hum 34. Bennett CM, Miller MB, Wolford GL. Neural corre-
Brain Mapp 1995;3:224–35. lates of interspecies perspective taking in the
29. Parrish TB, Gitelman DR, LaBar KS, et al. Impact of post-mortem Atlantic Salmon: an argument for mul-
signal-to-noise on functional MRI. Magn Reson Med tiple comparisons correction. NeuroImage 2009;47:
2000;44:925–32. S125.
30. Friston KJ, Holmes AP, Poline JB, et al. Analysis of fMRI 35. Nichols T, Hayasaka S. Controlling the familywise er-
time-series revisited. Neuroimage 1995;2:45–53. ror rate in functional neuroimaging: a comparative
31. Friston KJ, Williams S, Howard R, et al. Movement- review. Stat Methods Med Res 2003;12:419–46.
related effects in fMRI time-series. Magn Reson
Med 1996;35:346–55. 36. Genovese CR, Lazar NA, Nichols T. Thresholding of
32. Hajnal JV, Myers R, Oatridge A, et al. Artifacts due to statistical maps in functional neuroimaging using
stimulus correlated motion in functional imaging of the false discovery rate. Neuroimage 2002;15:
the brain. Magn Reson Med 1994;31:283–91. 870–8.