Left, right, or bilateral amygdala activation? How effects of smoothing and motion correction on ultra-high field, high-resolution functional magnetic resonance imaging (fMRI) data alter inferences
Introduction
The amygdalae are almond shaped nuclei embedded in the ventral medial temporal lobes anterior to the hippocampi (see Fig. 1). These small structures play an integral role in many processes that are involved in our survival and key psychological processes. The amygdalae are pivotal for threat assessments linked to phylogenetically preserved survival processes (Fox et al., 2015), and are responsible for integral affective (Hrybouski et al., 2016; Robinson et al., 2010) and cognitive processes such as memory (Guzmán‐Vélez et al., 2016), and learning (Farley et al., 2016). Furthermore, the amygdalae are implicated in many psychological disorders such as posttraumatic stress disorder (Prager et al., 2016), anxiety disorders (Li et al., 2016a), and depression (Connolly et al., 2017), to name a few. Given the breadth of involvement, and the necessity of the amygdala to healthy brain function, it is not surprising that it is one of the most well-studied structures in the human brain.
Commonly, functional magnetic resonance imaging (fMRI) reveals activation patterns associated with specific tasks, which allow investigators to theorize about structure-function relationships. As such, theories regarding lateralization of the amygdalae have become prominent in the literature, and supported by several studies (Robinson et al., 2010; Baeken et al., 2014). Specifically, some theories of amygdala function propose that the right amygdala is associated with response to animal stimuli (Mormann et al., 2011), positive picture encoding (Vasa et al., 2011), and anger (Fulwiler et al., 2012), while the left amygdala is more involved in fear or threat processing (Phelps et al., 2001), processing of emotional arousal and salience (Costanzo et al., 2015), and associative encoding processes (Killgore et al., 2000). Dyck et al. (2011), found evidence for amygdalae lateralization with the left amygdala involved with intentional mood control, and the right amygdala more involved with automatic emotional processing. Baas et al. (2004) provide evidence for functional lateralization of the amygdalae, with results suggesting the left amygdala is more associated with verbal and sustained emotional processing whereas the right amygdala is more associated with visual and dynamic emotional analysis. Robinson et al. (2010) performed a meta-analytic study, and observed distinct functional connectivity differences between the left and right amygdala. Together, these results provide strong evidence for potential lateralization effects.
While fMRI technology allows us to research the function of the amygdala and other structures, it is not without drawbacks. Spatial resolution of lower (i.e. 1.5 or 3 T) field strength scanners are not optimal for small brain structures like the amygdalae. Depending on the size of the voxel, the data may contain information from different types of tissue (i.e. grey matter, white matter, or vasculature) sometimes referred to as partial volume effects. The amygdala is particularly vulnerable to partial volume effects based on its location relative to the basal vein of rosenthal which may result in some spurious amygdala activations (Boubela et al., 2015). Regardless of voxel size, and the different types of tissues within, each voxel will also contain “noise” from the scanner, or physiological processes. Additionally, voxels may be large enough to incorporate more than one structure, especially in the case of smaller anatomical structures like the amygdalae. Further complicating this matter, it is standard to preprocess the data, which involves noise reduction steps that ultimate reduce resolution even more. Additionally, whole-brain data acquisition is required if one is concerned about global networks the amygdala may be involved with. As such, we sought to determine how the effects of preprocessing whole-brain fMRI data may effect inferences that can be drawn from data with regard to the amygdala.
The preprocessing of fMRI data converts measurements from each voxel into usable data by transforming, aligning, and correcting the data to allow for further analysis. Ashby (2011) refers to preprocessing as the analyses carried out on the data that are not driven by a specific hypothesis. Currently, there is no “gold standard” preprocessing pipeline (Aurich et al., 2015; Vergara et al., 2016). Many fMRI researchers have started to standardize some aspects of data management and reporting, such as reporting conventions (Nichols et al., 2016), but due to the complexity and variability of fMRI research paradigms there are no hard and fast rules or consensus on critical aspects of data processing.
Regardless of the preprocessing pipeline chosen, researchers often overlook how their choices affect the data. This issue can be exacerbated as technology allows us to increase the acquisition resolution while maintaining a high signal-to-noise ratio, allowing for smaller voxel sizes and more precise measurements of small structures such as amygdalae. Given the novelty of this methodology, it is unknown how such acquisitions respond to traditional preprocessing steps. To explore the effects of preprocessing at ultra-high resolution, we focus our attention on two steps that effect resolution the most, namely spatial smoothing and motion correction.
Spatial smoothing is the process of assigning each voxel a weighted average value based on neighboring voxels. This can be accomplished by applying a Gaussian kernel to the data resulting in a weighted average for each voxel. This weighted averaging is done mainly to reduce noise (internal and external) without removing the blood oxygen level dependent (BOLD) signal (Jenkinson, 2015), thus increasing signal to noise ratio, but also to allow the data to meet some statistical assumptions and account for slight anatomical differences between participants (Poldrack et al., 2011). However, there are notable drawbacks to smoothing as well. Increasing amounts of smoothing have been demonstrated to “move” activation patterns (Sacchet and Knutson, 2013) or cause clusters of activation to disappear, merge, split, or be created (Fransson et al., 2002; Mikl et al., 2008).
There are several published recommendations for FWHM kernel size. In order to meet statistical assumptions, the recommendations are twice (Poldrack et al., 2011) or three times voxel size (Yue et al., 2010). Mikl et al. (2008) recommended group inference FWHM to be set at 12 mm, sensorimotor cortex at FWHM of 8–10 mm, and smaller structures at 6 mm. The latter recommendation is consistent with a common recommendation that the smoothing kernel be no larger than the area of interest (Poldrack et al., 2011). Given the diversity of strategies for choosing an appropriate FWHM, it is no surprise that there is not a common consensus, and creating a hard and fast rule may not be appropriate. However, it is an important consideration that may have tremendous implications for data analysis.
Smoothing ultra-high-resolution data provides some unique challenges. Sub-millimeter resolution can increase the impact of functional and anatomical variance between participants (White et al., 2001). At lower resolutions, there is a high likelihood that active tissue, non-active tissue, and vasculature are all present in any single voxel. At higher resolutions, there is a greater likelihood that a voxel will contain a more homogenous tissue sample. Therefore, at sub-millimeter resolution, smoothing may average pure active tissue with non-active tissue and ultimately lose the “true” signal. However, there are some benefits to smoothing high resolution data. In a project conducted by Triantafyllou et al. (2006), the authors showed that collecting data at high resolutions and then smoothing the data to a lower resolution does improve the time-course signal to noise ratio.
Another common preprocessing step is motion correction. Motion correction is used to reduce the effect of participant movement while in the scanner. Movement can come from deliberate, conscious movement as participants look around or adjust how they lie; physiological processes like respiration, swallowing, and cardiac rhythm; or movement related to the task. The issue movement causes for fMRI analysis is that voxels are static locations in space, so when a participant moves, brain tissue can move from one voxel location to another. With increased spatial resolution, this can become exacerbated.
There are different methods for motion correction based on strategy and cost function applied to the data, but the overarching process involves computationally aligning volumes of data based on a single reference volume (Ashby, 2011). The various software packages commonly used within the field use different cost functions, but in a study conducted by Oakes et al. (2005) the top software packages produced similar results for motion correction despite different cost functions. In our study, we used motion correction in FMRIB’s Software Library (FSL) (Woolrich et al., 2009), which is accomplished by Motion Correction FMRIB’s Linear Image Registration Tool (MCFLIRT), with options for extended motion parameters (i.e., standard motion parameters plus their derivatives, and the square of their derivatives) (M. Jenkinson, Bannister, Brady, & Smith, 2002). Standard MCFLIRT uses three levels of a trilinear transformation starting with a coarse 8 mm search followed by two 4 mm searches with increasingly lower tolerances (M. Jenkinson et al., 2002).
Motion correction is an imperfect process. Regardless, motion must be accounted for to increase the accuracy of data. There are studies that examine the different methods of motion correct, such as a study by Parkes et al. (2018) who investigated the efficacy of 19 different approaches to motion correction. There exist various approaches to motion correction likely because of the vast number of variables that may influence the process (e.g., voxel size) and the different types of studies (e.g., resting state fMRI, task-based fMRI, clinical case studies). Additionally, motion correction may effect data differently depending on the type of experimental design (Johnstone et al., 2006). The primary aim of this paper is not to address all of these concerns, but rather to identify, at a basic level, how preprocessing may effect submillimeter fMRI data synergistically with spatial smoothing and how these effects may influence reported amygdala activations.
The preprocessing steps of spatial smoothing and motion correction are known to effect fMRI data, but to date, the extent of their effects on submillimeter fMRI data has not been fully explored. Therefore, we sought to examine the effects these steps have on amygdala activations using submillimeter fMRI data, using a common task known to activate the amygdala (Hariri et al., 2000). We examined the effects of various smoothing kernels and motion correction methods on resultant activation maps. We hypothesized that amygdala activation patterns would be sensitive to both smoothing and motion correction along with other neural structures.
Section snippets
Method
We performed high-resolution functional magnetic resonance imaging (hr-fMRI) during completion of a well-known face-matching task, that has had mixed results with regard to amygdala activation (Hariri et al., 2000). Thirty healthy individuals (26 right-handed, 12 males, age M ± SD = 21.3 ± 1.8) provided informed consent as approved by Auburn University’s Independent Review Board (IRB) and were scanned using an optimized EPI sequence (37 slices acquired parallel to the AC-PC line,
Results
Activation clusters varied with the different combinations of smoothing and motion correction. With no smoothing or motion correction (NM at 0 mm) there are only two significant clusters, one centered on the lingual gyrus and one local maximum in the cuneus (see Fig. 2). The combination of MC at 0 mm produces an activation cluster in the right fusiform gyrus that remains at all levels of motion correction at 0 mm, but the cuneus activation disappears at MC and MC + EV, then reappears at
Discussion
Here, we demonstrate the instability of fMRI activation patterns with heavy implications on inferences by varying two common preprocessing steps: motion correction and smoothing. The magnitude in, and nature of, the differences we found demonstrate that preprocessing choices can greatly influence the inferences drawn from fMRI data, particularly with small structures of interest such as the amygdalae. Our analyses also confirm the effects of spatial smoothing as previously published (Mikl et
Limitations
The major limitation of this study is that the true amygdala activation is unknown, making any specific recommendation for motion correction and spatial smoothing parameters to use with high-resolution data inadvisable. We also did not take any steps to process the data to focus only on the amygdala as one might if only a single region of interest (ROI) is the focus. Instead we examined preprocessing as one would for whole brain data commensurate with many studies that look for amygdala
Conclusion
The amygdalae are integral structures to many affective and perceptual process. The results from the current study should serve as a cautionary tale to researchers about the consequences fMRI data preprocessing choices have on amygdala activations as well as other small structures, which may profoundly affect the inferences drawn from a given data set. Furthermore, these results provide insight into the effects of preprocessing on submillimeter fMRI acquisition. Understanding the effects of
Context paragraph
This project was born out of questions within our lab on what preprocessing parameters should be used on the Siemens 7T data. The project began with taking a subsample of data and analyzing it with a couple different parameters. The results shocked us in that small changes resulted in large activation changes within the data. The next obvious step was to take a full dataset and systematically investigate the effects of preprocessing parameters, specifically looking at spatial smoothing and
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
References (56)
- et al.
Lateralization of amygdala activation: a systematic review of functional neuroimaging studies
Brain Res. Rev.
(2004) - et al.
General multi-level linear modelling for group analysis in FMRI
NeuroImage
(2003) - et al.
Resting-state functional connectivity of the amygdala and longitudinal changes in depression severity in adolescent depression
J. Affect. Disord.
(2017) - et al.
Hemispheric specialization in affective responses, cerebral dominance for language, and handedness: lateralization of emotion, language, and dexterity
Behav. Brain Res.
(2015) - et al.
Cognitive versus automatic mechanisms of mood induction differentially activate left and right amygdala
NeuroImage
(2011) - et al.
Extending the amygdala in theories of threat processing
Trends Neurosci.
(2015) - et al.
On the effects of spatial filtering—a comparative fmri study of episodic memory encoding at high and low resolution
NeuroImage
(2002) - et al.
Heart rate variability and its neural correlates during emotional face processing in social anxiety disorder
Biol. Psychol.
(2013) - et al.
Amygdala subnuclei response and connectivity during emotional processing
NeuroImage
(2016) - et al.
A global optimisation method for robust affine registration of brain images
Med. Image Anal.
(2001)
Improved optimisation for the robust and accurate linear registration and motion correction of brain images
NeuroImage
FSL
NeuroImage
A high performance 3D cluster-based test of unsmoothed fMRI data
NeuroImage
A voxelation-corrected non-stationary 3D cluster-size test based on random field theory
NeuroImage
The first step for neuroimaging data analysis: DICOM to NIfTI conversion
J. Neurosci. Methods
Effects of spatial smoothing on fMRI group inferences
Magn. Reson. Imaging
Comparison of fMRI motion correction software tools
NeuroImage
An evaluation of the efficacy, reliability, and sensitivity of motion correction strategies for resting-state functional MRI
NeuroImage
The publication and reproducibility challenges of shared data
Trends Cogn. Sci. (Regul. Ed.)
The functional connectivity of the human caudate: an application of meta-analytic connectivity modeling with behavioral filtering
NeuroImage
Spatial smoothing systematically biases the localization of reward-related brain activity
NeuroImage
Effect of spatial smoothing on physiological noise in high-resolution fMRI
NeuroImage
Enhanced right amygdala activity in adolescents during encoding of positively valenced pictures
Dev. Cogn. Neurosci.
Neural correlates of facial affect processing in children and adolescents with autism spectrum disorder
J. Am. Acad. Child Adolesc. Psychiatry
Anatomic and functional variability: the effects of filter size in group fMRI data analysis
NeuroImage
Robust group analysis using outlier inference
NeuroImage
Multi-level linear modelling for FMRI group analysis using Bayesian inference
NeuroImage
Statistical Analysis of fMRI Data
Cited by (13)
Maternal prenatal lead levels and neonatal brain volumes: Testing moderations by maternal depressive symptoms and family income
2024, Neurotoxicology and TeratologyThe Dichotomy of Threat and Deprivation as Subtypes of Childhood Maltreatment: Differential Functional Connectivity Patterns of Threat and Reward Circuits in an Adult Trauma Sample
2024, Biological Psychiatry: Cognitive Neuroscience and NeuroimagingThalamo-cortical circuits associated with trait- and state-repetitive negative thinking in major depressive disorder
2023, Journal of Psychiatric ResearchUltra-high field imaging of the amygdala – A narrative review
2023, Neuroscience and Biobehavioral Reviews“Don't judge me!”: Links between in vivo attention bias toward a potentially critical judge and fronto-amygdala functional connectivity during rejection in adolescent girls
2021, Developmental Cognitive NeuroscienceCitation Excerpt :Interestingly, similar prior research has implicated the right amygdala specifically (e.g., Gee et al., 2013; Price et al., 2016; Abend et al., 2020), which could suggest meaningful lateralization. However, lateralized effects could also be related to inadvertent effects from data analysis (Murphy et al., 2020). The regions of prefrontal cortex found to correlate with amygdala activity during rejection feedback were in more lateral portions of the anterior PFC (BA 10) and inferior frontal gyrus.