Abstract
Preprocessing fMRI data requires striking a fine balance between conserving signals of interest and removing noise. Typical steps of preprocessing include motion correction, slice timing correction, spatial smoothing, and high‐pass filtering. However, these standard steps do not remove many sources of noise. Thus, noise‐reduction techniques, for example, CompCor, FIX, and ICA‐AROMA have been developed to further improve the ability to draw meaningful conclusions from the data. The ability of these techniques to minimize noise while conserving signals of interest has been tested almost exclusively in resting‐state fMRI and, only rarely, in task‐related fMRI. Application of noise‐reduction techniques to task‐related fMRI is particularly important given that such procedures have been shown to reduce false positive rates. Little remains known about the impact of these techniques on the retention of signal in tasks that may be associated with systemic physiological changes. In this paper, we compared two ICA‐based, that is FIX and ICA‐AROMA, two CompCor‐based noise‐reduction techniques, that is aCompCor, and tCompCor, and standard preprocessing using a large (n = 101) fMRI dataset including noxious heat and non‐noxious auditory stimulation. Results show that preprocessing using FIX performs optimally for data obtained using noxious heat, conserving more signals than CompCor‐based techniques and ICA‐AROMA, while removing only slightly less noise. Similarly, for data obtained during non‐noxious auditory stimulation, FIX noise‐reduction technique before analysis with a covariate of interest outperforms the other techniques. These results indicate that FIX might be the most appropriate technique to achieve the balance between conserving signals of interest and removing noise during task‐related fMRI.
Keywords: noise‐reduction techniques, pain, systemic physiological changes, task‐related fMRI
Preprocessing fMRI data requires striking a fine balance between conserving signals of interest and removing noise. Preprocessing including FIX noise‐reduction technique conserves significantly more signal than a preprocessing protocol including CompCor noise‐reduction technique in both noxious heat and non‐noxious auditory stimulations, while removing only slightly less noise. These results suggest that FIX might be the most appropriate technique to achieve the balance between conserving signals of interest and removing noise.
1. INTRODUCTION
Functional MRI brain images, whether they result from resting‐state or task‐related sequences, are very noisy by nature (Fassbender et al., 2017; Network et al., 2013). The images need to be carefully preprocessed in order to filter out the noise while conserving the signal of interest (Fassbender et al., 2017; Johnstone et al., 2006; Lund et al., 2005; Oakes et al., 2005). Noise can emerge from different sources including motion, in particular head motion, MRI scanner artifacts, or physiological noise arising from normal cardio‐respiratory activity (Brooks et al., 2013; Fassbender et al., 2017; Lund et al., 2005; Network et al., 2013; Yoshikawa et al., 2020). In the case of task‐related MRI sequences, the task itself can add noise to the data (Fassbender et al., 2017). This additional noise can be due to the inherent nature of the task participants are asked to perform, such as any movement participants would have to complete in the task, or to any physical or physiological reaction to the performed task, in addition to noise introduced by equipment used during the task (Brooks et al., 2013; Epstein et al., 2007; Kasper et al., 2017; Yoshikawa et al., 2020).
Standard steps to preprocess the data typically include motion correction, slice timing correction, spatial smoothing, and high‐pass filtering. Motion correction is the process by which the head movements are corrected by realigning volumes across the scan (Jenkinson, 2002; Poldrack et al., 2009). Slice timing correction controls for the differences of acquisition time between slices of each volume (Poldrack et al., 2009). Spatial smoothing consists in the application of a filter to remove high‐frequency spatial noise, allowing better signal detection in large brain areas (Poldrack et al., 2009). Finally, high‐pass filtering removes the low‐frequency signals considered noise of physiological or scanner origin (Poldrack et al., 2009). In some cases, intensity normalization is also applied to the data, which results in a constant mean volume intensity over time (Coghill et al., 1994).
Numerous noise‐reduction techniques have been developed to further clean the data, which is essential to keep the false positive rate at a nominal level (Eklund et al., 2019). These techniques include Component Correction (CompCor) techniques and Independent Component Analysis (ICA) based noise‐reduction techniques, which are the techniques compared in this manuscript.
The CompCor‐based noise‐reduction techniques include anatomical CompCor (aCompCor) and temporal CompCor (tCompCor). aCompCor is widely used to denoise resting‐state and task‐related fMRI data. Noise is identified by performing a principal component analysis (PCA) on signals from the white matter and cerebrospinal fluid (CSF). These signals are then regressed out of the data by performing a general linear model (GLM) analysis (Behzadi et al., 2007). Its efficiency has been demonstrated in perfusion and blood‐oxygen level‐dependent fMRI sequences (BOLD) (Behzadi et al., 2007). tCompCor uses a similar approach: noise is identified by performing a PCA on signals from high‐variance voxels. The identified signals are then regressed out of the data. Typically, these voxels are found in the ventricles, edge regions, and vessels (Lund & Hanson, 2001).
The ICA‐based techniques used in this manuscript include the FMRIB's ICA‐based Xnoiseifier (FIX) and the ICA‐based strategy for automatic removal of motion artifacts (ICA‐AROMA). FIX was developed in the context of the Human Brain Connectome Project to filter out noise of various origins from resting‐state fMRI data (Salimi‐Khorshidi et al., 2014). FIX relies on the usage of a classifier to identify ICA components as noise or signal (Griffanti et al., 2014; Salimi‐Khorshidi et al., 2014). Usage of a classifier reduces the workload compared to manual labeling and cleaning of the data while preserving accuracy in the classification of the components. Although trained classifiers from the Human Connectome Project are available, it is recommended that the classifier is hand‐trained on one's dataset to further ensure accuracy of the classification (Salimi‐Khorshidi et al., 2014). Recent studies (Blasi et al., 2020; Mayer et al., 2019) highlighted the successful use of FIX to denoise task‐related fMRI data. ICA‐AROMA (Pruim, Mennes, van Rooij, et al., 2015) has been developed subsequently to FIX to avoid some of its drawbacks, such as the classifier training. It uses a set of four temporal and spatial features, which were defined based on previous literature, to identify motion components. These components are then filtered out of the data. ICA‐AROMA has been validated for denoising of resting‐state and task‐related data (Pruim, Mennes, van Rooij, et al., 2015).
Although the efficiency of these noise‐reduction techniques has been demonstrated on task‐related and resting‐state fMRI data, the tasks included in these tests were often associated with language or visual skills. These tasks have the advantage of producing clear, strong signals in well‐known areas, making them ideal to test noise‐reduction techniques. In addition, these tasks do not typically induce much additional noise.
There is a need to test further the efficiency of these techniques in tasks that induce global cerebral blood flow changes (Coghill et al., 1998; Zeidan et al., 2015) and/or have substantial physiological responses or movements associated with them (Brooks et al., 2013; Perlaki et al., 2015; Tousignant‐Laflamme et al., 2005), such as experimentally induced pain, alterations in breathing, or fear. For such tasks, it is essential to ensure that the noise‐reduction technique removes the noise without removing the signal of interest, even if this signal might be present in areas typically associated with noise, such as the white matter.
This paper aims to compare the performance of ICA‐based noise‐reduction techniques (FIX and ICA‐AROMA) with CompCor‐based noise‐reduction techniques (aCompCor and tCompCor) and define the most appropriate technique to conserve signals of tasks associated with substantial changes in the white matter and/or in global cerebral blood flow (Coghill et al., 1998; Zeidan et al., 2015). A standard preprocessing, which does not include any noise‐reduction technique, is used as the baseline to compare the two noise‐reduction techniques. Because FIX is designed to detect spatial and temporal features of noise specific to one's dataset, we hypothesize that the FIX noise‐reduction technique would be more sensitive to conserving signal of interest while removing the noise than other noise‐reduction techniques.
2. MATERIALS AND METHODS
2.1. Participants
A total of 143 healthy individuals (age range: 14–44‐years‐old) were enrolled in a neuroimaging study investigating the individual experience of pain. Data from 34 of these participants (Table 1; age: 28 ± 6.1, mean ± SD), which were not used in any prior analyses (Hoeppli et al., 2022), were used to train an automated classifier FSL FIX (FMRIB's ICA‐based Xnoiseifier, FSL, Oxford, UK) (Griffanti et al., 2014; Salimi‐Khorshidi et al., 2014) that was used to denoise the fMRI data. Four of these 34 participants had subtle incidental findings reported by a radiologist, that due to their nature and localization should not affect the training of the FIX classifier. As a consequence, these four participants were not excluded from the classifier training. Out of the remaining 109 participants, eight were excluded because of insufficient quality of the fMRI images or incidental findings of abnormalities on MRI. The remaining 101 healthy volunteers (Table 1; 43 males and 58 females, age: 28.5 ± 7.7, mean ± SD) were included in the analyses described in this manuscript.
TABLE 1.
Demographics of the training and test datasets.
| Training data (N = 34) | Test data (N = 101) | |
|---|---|---|
| Sex | ||
| Females | 22 (64.7%) | 58 (57.4%) |
| Males | 12 (35.3%) | 43 (42.6%) |
| Age | ||
| Mean (SD) | 28 (6.1) | 28.5 (7.72) |
Participants and parents/legal guardians of minor participants gave their written informed consent and minors provided written assent in accordance with the institutional review board of Cincinnati Children's Hospital Medical Center, which approved the study. Exclusion criteria included active neurological or psychiatric disorders that impacted the participant's ability to perform the tasks requested, the presence or history of chronic pain, medications that could interfere with quantitative sensory testing (QST) or brain function, positive screen for recreational drugs, any serious pathology, substantial uncorrected visual deficit, and any MRI contraindication, such as any metallic implant or braces.
2.2. General design
Participants in this study underwent two sessions: a QST session and an MRI session. The goal of the QST session was to familiarize participants with the stimuli and rating scale used in the MRI session. This session will not be further discussed in this manuscript, but a detailed description of this session can be found elsewhere (Hoeppli et al., 2022). The MRI session lasted approximately 90 min. During this session, participants experienced three types of stimuli: heat, cold, and auditory stimuli. After each stimulus, participants were asked to rate their pain intensity and unpleasantness sensations on two computerized visual analog scales. A complete description of the scale can be found in Hoeppli et al. (2022). Only the heat and auditory fMRI series will be detailed below; since analyses reported in this manuscript were performed on these stimulus modalities only. An in‐depth description of the cold stimulus modality is available in Hoeppli et al. (2022). We chose to focus on the heat pain modality, as it is one of the stimulus modalities most frequently used to experimentally induce pain and is associated with a physiological response that might impact the MRI response. The auditory stimulus modality was chosen as a control condition for the heat modality since it is typically not associated with the same level of white matter response as noxious heat stimulation.
A parent or legal guardian accompanied minor participants during both sessions to confirm eligibility to participate in the study (QST session) and MRI compatibility (MRI session). Once eligibility was confirmed, parents were asked to step out to avoid influence on participants' responses (McMurtry et al., 2010; Schinkel et al., 2017; Zohsel et al., 2006).
All participants were asked to turn their cell phones off during both sessions to avoid distractions and interruptions. During the MRI session, participants' cell phones were kept secure with their other ferromagnetic belongings in lockers.
2.3. MRI session
2.3.1. fMRI stimuli and Block‐design fMRI series
Heat fMRI series
Participants completed three fMRI series heat stimuli (Figure 1) followed by a 16‐second rating period and a 22‐second resting period. Participants received a total of 17 high‐intensity noxious stimuli (48°C) and 4 low‐intensity noxious stimuli (47°C). Two series included six high‐intensity and one low‐intensity noxious stimuli each, and one series included five high‐intensity and two low‐intensity noxious stimuli. Stimuli were delivered to the back of the lower left leg using a 16 × 16 mm thermode provided by a Medoc pathway system (Medoc, Ramat Yishai, Israel). To limit sensitization or habituation to the stimuli, the position of the thermode was slightly moved on the participant's calf between heat series. The repetition of the heat fMRI series was meant to increase statistical power at the individual level and decrease false positive rates.
FIGURE 1.
Experimental design of the three fMRI heat series. The time course of each fMRI heat series is displayed here. The orange line represents the target temperature for the high‐intensity stimuli (48°C). The green line represents the target temperature for the low‐intensity stimuli (47°C). Series 1 and 2 include 6 high and 1 low‐intensity stimuli, while series 3 includes 5 high and 2 low‐intensity stimuli. Once the target temperature was reached, the temperature plateaued for 10 s. The baseline temperature was set at 35°C. Rise and return temperature rates were set at 6°C/s. The design of the fMRI auditory series matched the design of the heat series.
For each stimulus, the temperature increased from a baseline of 35°C at a rate of 6°C/s, plateaued for 10 s, and returned to baseline at a rate of 6°C/s.
Participants were instructed to rate their perceived pain intensity and unpleasantness after each stimulus. Between the presentation of the scales, a black screen was displayed.
Auditory fMRI stimuli
Participants completed one auditory fMRI series. This series included five high‐intensity non‐noxious stimuli (90 dB) and two low‐intensity non‐noxious stimuli (80 dB). The auditory stimuli were designed as 900 Hz sawtooth waves with an overall time course similar to the heat stimuli. In the high‐intensity stimuli, the sound increased from silence to 90 dB in 2.2 s and plateaued for 10 s before returning to 0 dB at the same rate. In the low‐intensity stimuli, the sound increased to 80 dB in 2 s and plateaued for 10 s before returning to 0 dB at the same rate. After each auditory stimulus, participants were instructed to rate the intensity and unpleasantness of the sounds that they experienced.
2.3.2. MRI acquisition
During the MRI session, participants lay in a supine position in a Philips 3T Ingenia scanner with a 32‐channel head coil. During this session, all participants always underwent a T1 structural scan first. They then completed three BOLD fMRI series of heat stimuli, one fMRI series of cold stimuli and one fMRI series of auditory stimuli. Resting BOLD and arterial spin label (ASL) series were acquired but are not reported here. The order of the BOLD and ASL series was counterbalanced between participants.
A radiologist inspected the structural images of the participants for incidental findings.
T1 structural scan
The multi‐echo (four echoes) T1‐weighted series was acquired using the following parameters: repetition time (TR): 10 ms; echo times (TE): 1.8, 3.8, 5.8, 7.8; flip angle: 8; FOV: 256 × 224 × 200 mm; voxel size: 1 × 1 × 1 mm; slice orientation: sagittal. The total duration of this scan was 4 min 42 s.
BOLD fMRI
Each functional image series consisted of 193 volumes acquired using the following parameters: TR: 2 s; TE: 35 ms; voxel size: 3 × 3 × 4 mm; FOV: 240 × 240 × 136 mm; slice orientation: transverse; slice order: ascending; dummy scans: two. Each series lasted 6 min 26 s, after an 8 s pre‐scan time.
2.4. Statistical analyses on fMRI data
All MRI data were first inspected for motion and scanner artifacts. They were then preprocessed and analyzed with FSL (FMRIB Software Library, version 6.0.1 Oxford, UK).
2.4.1. Data preprocessing
Structural images were first corrected for bias using FMRIB's Automated Segmentation Tool (FAST) {Zhang:2001ve}. Images were then brain extracted using the Brain Extraction Tool (BET) {Smith:2002ef} and normalized into standard space MNI‐152 using FMRIB's Linear Image Registration Tool (FLIRT) (Jenkinson, 2002; Jenkinson & Smith, 2001). Finally, images were segmented into the different tissue types and white matter and CSF were masked at a probability threshold of 0.95.
fMRI data from 135 participants were used in these analyses. A total of 34 independent participants were included exclusively in the training dataset for a FIX classifier. Data from the remaining 101 participants were included in the test dataset to compare four preprocessing protocols with noise‐reduction techniques to a standard preprocessing protocol for fMRI data without noise‐reduction preprocessing (Figure 2). Two protocols used ICA‐based noise‐reduction techniques (FIX and ICA‐AROMA). Two protocols used CompCor‐based noise‐reduction techniques (aCompCor and tCompCor). In each preprocessing protocol, data were visually inspected after each preprocessing step to confirm that the correction was correctly applied to the data.
FIGURE 2.
Flowchart of the five preprocessing protocols, including standard preprocessing, preprocessing with aCompCor noise‐reduction technique, preprocessing with tCompCor noise‐reduction technique, preprocessing with ICA‐AROMA noise‐reduction technique, and preprocessing with FIX noise‐reduction technique, and the FIX classifier training protocol. PCA, principal component analysis; GLM, general linear model.
Standard preprocessing protocol
The test data were preprocessed using a standard preprocessing protocol (Figure 2). This protocol included the following steps: motion correction using Motion Correction FMRIB's Linear Registration Tool (MCFLIRT) (Jenkinson, 2002), slice timing correction, brain extraction (BET), spatial smoothing (FWHM: 5 mm) with SUSAN (Smith & Brady, 1997), high‐pass filter (cutoff: 100 s), and intensity normalization. No additional noise‐reduction technique was used to further remove noise from the data.
Preprocessing protocols with noise‐reduction techniques
All preprocessing protocols with noise‐reduction techniques underwent the following steps before the application of the noise‐reduction technique (Figure 2): registration to the structural scan and standard space MNI‐152 using FLIRT and FNIRT (FMRIB's Non‐Linear Image Registration Tool) (Andersson et al., 2007; Jensinson, 2002; Jenkinson et al., 2012; Smith et al., 2004; Smith & Brady, 1997; Woolrich et al., 2009), motion correction using MCFLIRT (Motion Correction FMRIB's Linear Registration Tool) (Jenkinson, 2002), slice timing correction, brain extraction (BET), spatial smoothing (FWHM: 5 mm) with SUSAN (Smith & Brady, 1997), and high‐pass filter (cutoff: 100 s).
Preprocessing protocol with FIX noise‐reduction technique (FIX preprocessing)
Preprocessing of the training dataset: Following the recommendations of Salimi‐Khorshidi et al. (2014) for optimal preprocessing with FIX, we used a training dataset to develop our own FIX classifier. The training dataset from 34 independent participants underwent the same preprocessing steps as described above. After these preprocessing steps and before training the FIX classifier, a probabilistic independent component analysis (PICA) using Multivariate Exploratory Linear Optimized Decomposition into Independent Components (MELODIC) (Beckmann & Smith, 2004) was performed to extract the components to classify.
To define the ideal number of components for our dataset to separate noise from the signal, MELODIC was performed using four predefined numbers of output‐independent components (ICs) (15, 20, 25, and 30 components) on each individual dataset of a random subsample of 10 participants with three fMRI heat series each, that is, 30 datasets. Spatial maps of the components were reviewed by three experts (MEH, MAG, and CKM). It was determined that PICA performed using MELODIC and resulting in 25 output IC was the most optimal number of output components, since it performed best at separating noise from signal in our data.
Thus, PICA performed using MELODIC and outputting 25 ICs was then performed on our complete training dataset. Components resulting from MELODIC were then manually categorized by our three experts as signal, noise, or unknown when the source of the activation could not be defined unambiguously. Categories were defined based on the spatial map, the power spectrum, and the time series of the IC. All the categories were defined according to Salimi‐Khorshidi et al. (2014). ICs identified as noise were then subcategorized into “Unclassified noise,” “Movement,” “MRI,” “Respiratory,” “White matter,” “Susceptibility motion,” “Sagittal sinus,” and “Cardiac.” A single IC identified as noise could be labeled with multiple subcategories of noise.
Once consensus was achieved between the three experts, categorized ICs of the 102 datasets (34 participants × 3 fMRI heat series) were used to train a FIX classifier, creating a trained‐weights file to be used in test datasets to automatically categorize ICA components (FMRIB's ICA‐based Xnoiseifier, FSL, Oxford, UK) (Griffanti et al., 2014; Salimi‐Khorshidi et al., 2014). Finally, a leave‐one‐out test was performed to assess the efficiency of the classifier to correctly identify components as signal and noise and define the optimal distinction threshold for our data. Data included in these training datasets were not used for any preprocessing comparison.
Preprocessing of the test datasets: Test datasets were preprocessed following the same steps as those described above, that is, motion correction, slice timing correction, BET, spatial smoothing (FWHM: 5 mm), and high pass filter (cutoff: 100 s). Next, PICA was performed on the preprocessed data using MELODIC and was set to output 25 ICs. Once ICs had been defined by MELODIC, the FIX classifier, which was trained on our training dataset, was applied to the ICs of the test datasets to categorize them into signal and noise and remove the components identified as noise. The threshold to separate noise and signal components was defined to preserve as much signal as possible, that is, 20 (Table 2). Before applying the classifier trained on the heat fMRI series to the auditory fMRI series, the accuracy of the classifier to distinguish noise from signal of interest in these data was confirmed by performing a leave‐one‐out test. The test confirmed the accuracy of the classifier on the auditory data (Table 2). Given the results of the leave‐one‐out test on the auditory data, the same classifier was used for the preprocessing of heat and auditory fMRI series. To perform FIX on the auditory data, a distinction threshold of 30 was chosen. Finally cleaned filtered images were corrected by intensity normalization.
TABLE 2.
Results of the leave‐one‐out tests performed on the heat and auditory fMRI series.
| Heat | ||||||||
|---|---|---|---|---|---|---|---|---|
| Thresholds | 1 | 2 | 5 | 10 | 20 | 30 | 40 | 50 |
| True positive rate (TPR) | 99.1 | 98.6 | 98.4 | 97.3 | 96.3 | 94.3 | 92.4 | 91.4 |
| True negative rate (TNR) | 78.1 | 80.9 | 85.5 | 90.2 | 93.7 | 95.5 | 96.6 | 97.7 |
| (3*TPR + TNR)/4 | 93.9 | 94.2 | 95.2 | 95.5 | 95.7 | 94.6 | 93.4 | 93 |
| Auditory | ||||||||
|---|---|---|---|---|---|---|---|---|
| Thresholds | 1 | 2 | 5 | 10 | 20 | 30 | 40 | 50 |
| True positive rate (TPR) | 100 | 100 | 100 | 99.1 | 99.1 | 99.1 | 98.1 | 98.1 |
| True negative rate (TNR) | 70.4 | 74.8 | 84.7 | 88.9 | 93.7 | 94.4 | 97.4 | 98.7 |
| (3*TPR + TNR)/4 | 92.6 | 93.7 | 96.2 | 96.5 | 97.7 | 97.9 | 97.9 | 98.3 |
Note: Results from these tests were used to define the distinction thresholds to optimize the separation of signal and noise components by FIX.
Preprocessing protocol with ICA‐AROMA denoising technique
ICA‐AROMA was implemented using the Python3 script developed by Pruim, Mennes, Buitelaar, & Beckmann (2015), and Pruim, Mennes, van Rooij, et al., (2015). ICA‐AROMA automatically identifies ICs related to motion and filters them out of the data. This technique relies on a similar classification technique as FIX without requiring the training of a classifier.
Preprocessing protocol with anatomical and temporal CompCor noise‐reduction technique (aCompCor and tCompCor preprocessing)
The test datasets were preprocessed using the same steps as described above: motion correction, slice timing correction, spatial smoothing, and high‐pass filtering.
The application of aCompCor noise‐reduction technique was performed using a custom‐made script including AFNI and FSL tools. It included a PCA and GLM analysis (Behzadi et al., 2007). The PCA was performed on the activation time series extracted from the white matter and CSF to create nuisance regressors. These regressors were then used to create the design matrix for the GLM analysis and filter out noise. This technique is based on two assumptions (Behzadi et al., 2007): (1) The activation present in the white matter and CSF is only physiological noise and not a signal of interest; (2) Noise in the gray matter is correlated to noise in the white matter and CSF and can be filtered out by regressing time series of noise in white matter and CSF. The denoised data underwent intensity normalization to complete the preprocessing protocol.
Python 3 and the Nipype package were used for the application of the tCompCor noise‐reduction technique. The threshold for selection of highest variance voxels was set at 0.02, that is, 2% of voxels with the highest variance were used (Behzadi et al., 2007).
2.4.2. Data analyses
Overview
Five analyses were performed to evaluate the impact of the five preprocessing protocols on the signal‐to‐noise ratio qualitatively and quantitatively. Qualitative comparisons included block‐design‐based analyses of the cerebral responses to high intensity heat stimulus and of the percentage of participants exhibiting the same brain activation (prevalence analysis). T‐tests using a block design were then performed to quantitatively compare brain activation differences between the preprocessing protocols. Contrast parameter estimates (copes) and variance of contrast parameter estimates (varcopes) were compared to investigate differences in the detected signal (copes) and noise (varcopes) between our preprocessing protocols. Finally, the T‐tests performed in the block‐design analysis were repeated using a single epoch design to compare differences in activation associated with a single epoch, that is, one noxious stimulus, between preprocessing protocols. For these tests, only three preprocessing protocols were used. These included the standard preprocessing protocol, the ICA‐based and CompCor‐based protocols showing the best results in the block‐design analyses, that is, FIX, respectively, aCompCor. Unlike the block‐design model, in the statistical model for the single‐epoch design, a regressor was then defined for each stimulus and two contrasts (positive and negative activation) were defined for the regressor of interest, that is, the first heat stimulus.
T‐tests using a block design and analyses of the parameter estimates and their variance were replicated in the auditory modality to investigate the effect of the five preprocessing protocols on a task with a lower level of white matter response. For all fMRI analyses, a clustering z threshold of 3.1 and a p threshold of 0.05 were used (Eklund et al., 2016).
Differences in brain activation associated with high intensity heat and auditory stimulation using a block design between the preprocessing protocols
To qualitatively compare the effect of the five preprocessing protocols on the data, GLM analyses were performed to define brain activation associated with high‐intensity heat and auditory stimuli.
First‐level GLM analyses were performed on each individual fMRI heat and auditory series preprocessed with each protocol using FEAT (Woolrich et al., 2001). One block‐design regressor of interest, that is, high‐intensity stimuli, and two block‐design regressors of no interest, that is, low‐intensity stimuli and rating periods, were defined. For the heat paradigm, which included three series, individual second‐level GLM analyses for each preprocessing protocol were performed on contrast parameter estimates (cope) images derived from the first‐level heat GLM analyses using FEAT (Woolrich et al., 2004), allowing combination of the individual heat series using a fixed‐effect statistical model. Finally, a group‐level GLM was performed with FEAT (Woolrich et al., 2004) using cope images from the individual second‐level heat analyses and the individual first‐level auditory analyses using mixed‐effect FLAME 1 and 2 statistical models.
To quantitatively identify the differential impact of the preprocessing protocols on brain activation associated with the high‐intensity heat stimulus (48°C), individual paired T‐tests were then performed on the individual second‐level copes using a fixed‐effect statistical model to test the difference in brain activation after each preprocessing protocols. The following comparisons were performed: FIX‐standard, FIX‐aCompCor, FIX‐tCompCor, FIX‐ICA‐AROMA, ICA‐AROMA‐standard, aCompCor‐standard, aCompCor‐ICA‐AROMA, tCompCor‐standard, tCompCor‐ICA‐AROMA, tCompCor‐aCompCor. Finally, 10 group‐level GLM analyses were performed on the individual differential copes using a mixed‐effect FLAME 1 and 2 statistical model to investigate differential effects of the preprocessing protocols on brain activation associated with high‐intensity heat stimuli at a group level.
Similarly, T‐tests were performed on the individual first‐level auditory copes, followed by similar group‐level GLM analyses as described above to investigate the differential effects of the preprocessing protocols in a task without important associated white matter response.
Relationship of reported stimulus intensity with brain activation in the heat and auditory series
The previous group‐level GLM analysis was repeated with individually averaged intensity rating of high‐intensity heat, respectively, auditory, stimuli as a regressor of interest.
Prevalence of activation across participants
The prevalence (percentage) of participants exhibiting activation in brain areas commonly activated in response to heat stimuli was calculated to assess differences in within‐individual activation between the processing protocols.
Individual second‐level block‐design copes of the heat paradigm were concatenated, resulting in a 4D dataset using fslmerge. This dataset was then binarized and averaged across time using fslmaths. This allowed us to examine the percentage of participants showing activation in brain areas relevant to the used paradigm.
Noise and signal detection following each preprocessing protocol
To further investigate the effect of each preprocessing protocol on the signal of interest and noise removal in the heat series, we isolated the parameter estimates of individual second‐level block‐design copes and varcopes averaged within areas of interest, including primary somatosensory cortex (S1), insula, anterior cingulate cortex (ACC), and within areas usually considered as sources of noise, that is, white matter and CSF. Values from the copes were used to assess signal of interest, while varcopes were used to assess remaining noise in the data after preprocessing (Eippert et al., 2017). We then compared these estimates within areas and between preprocessing protocols.
To define the clusters, we completed the following steps:
Clusters of each area were extracted from a 2 mm standard template using Harvard—Oxford and Juelich atlases.
If applicable, clusters from both atlases were added up.
Clusters were then thresholded and binarized.
Binarized and thresholded clusters were applied to the group‐level thresholded z statistical maps for each preprocessing protocol to create a functional mask for each protocol.
Individual averaged parameter estimates were extracted within area by performing featquery on the individual second‐level copes and varcopes using the functional masks defined in 4.
Resulting averaged parameter estimates in the heat series were compared by performing a three‐way repeated‐measures ANOVA [maps: 2 levels (cope/varcope) × brain areas: 5 levels (aCompCor, S1, insula, white matter, CSF) × preprocessing protocols: 5 levels (FIX, ICA‐AROMA, aCompCor, tCompCor, standard)] after outliers were excluded from the analyses. Outliers were defined as individual averaged parameter estimates that were greater than the third quartile + 1.5*interquartile range (Q3 + 1.5*(Q3‐Q1)) or that were smaller than the first quartile–1.5*interquartile range (Q1–1.5*(Q3‐Q1)). Results of each preprocessing for the outliers were visually inspected before their exclusion to ensure that they were not the results of a failed preprocessing.
This analysis was repeated on the auditory series. Parameter estimates were extracted from three areas: primary auditory cortex (A1), CSF, and white matter. After removal of the outliers following the same procedure as described above, a three‐way repeated‐measures ANOVA [maps: 2 levels (cope/varcope) × brain areas: 3 levels (A1, CSF, white matter) × preprocessing protocols: 5 levels (FIX, ICA‐AROMA, aCompCor, tCompCor, standard)] was performed.
For both paradigms, post‐hoc analyses were performed when appropriate and Bonferroni correction for multiple comparisons was applied when appropriate. A significance p threshold was defined at 0.05. These analyses were performed using Rstudio version 1.3.1093.
Differences in brain activation associated with a single high‐intensity heat stimulus using a single‐epoch design between preprocessing protocols
To further evaluate the differential impact of the preprocessing protocols on brain activation associated with high‐intensity heat stimulus, first‐level individual GLM analyses were repeated using a single‐epoch design, where each epoch represented one stimulus of the first heat fMRI series. Single‐epoch designs are more susceptible to noise than block designs (Koyama et al., 2003). Therefore, the impact of the preprocessing protocols should be more pronounced in a single‐epoch design. The first stimulus of that series was defined as a regressor of interest, while the other six stimuli and rating period were defined as regressors of no interest. Two contrasts were defined as positive (contrast 1) and negative (contrast 2) activation associated with the first stimulus. Group‐level GLM analyses were then performed for each protocol to identify positive and negative brain activation associated with that stimulus across participants.
Three second‐level individual T‐tests, which included similar comparisons as the ones done on the main block‐design analysis, were then performed on the individual first‐level copes to statistically compare brain activation associated with a single high intensity heat stimulus between the preprocessing protocols. Finally, three group‐level GLM analyses were performed on the individual differential copes using a mixed‐effect FLAME 1 and 2 statistical model to investigate differential effects of the preprocessing protocols on brain activation associated with a single high intensity heat stimuli across participants.
Correlations between aCompCor regressors and stimulus regressors
Given the reduced signal detection in our analyses after preprocessing with aCompCor, an additional analysis was run on this preprocessing protocol to investigate whether the application of aCompCor removed signal of interest due to its correlation with aCompCor regressors. Individual correlations were calculated between each CSF and WM regressor included in our aCompCor model and the stimuli for each heat and auditory series.
3. RESULTS
3.1. Differences in brain activation associated with high‐intensity heat stimulation using a block design between preprocessing protocols
Group‐level GLM analyses were performed to assess the BOLD signals associated with high‐intensity heat stimulation after each preprocessing protocol, that is, standard, FIX, ICA‐AROMA, aCompCor, and tCompCor noise‐reduction techniques.
Data preprocessed with all preprocessing protocols revealed changes in activations in similar areas (Figures 3, 4, 5). Increased activation was observed in gray matter areas encompassing the ACC, the primary sensorimotor cortex (S1M1), the insula, the basal ganglia, and the thalamus. Deactivation was observed primarily in gray matter areas including the amygdala and hippocampus, as well as in the posterior cingulate cortex (PCC) and precuneus.
FIGURE 3.
Brain activation associated with high‐intensity heat stimulus after preprocessing with ICA‐based noise‐reduction techniques is greater than after standard preprocessing. (1.a and 3.a) Following FIX preprocessing, noxious heat stimuli are associated with increased and decreased brain activation in gray matter areas previously associated with pain and in the white matter; (2.a and 3.b) similarly after ICA‐AROMA, increased activation associated with noxious heat was observed in pain‐related areas, while decreased activation was found in PCC and precuneus cortex and in the white matter. (1‐2.b) Brain activation associated with noxious heat stimuli after standard preprocessing was observed in similar areas as after FIX and ICA‐AROMA preprocessing. (1.c) The comparison of brain activation after preprocessing protocol with FIX noise‐reduction technique to brain activation after standard preprocessing shows greater positive and negative signal detection after FIX in gray matter areas previously associated with pain and in the white matter. (2.c) Similarly, the comparison of brain activation after ICA‐AROMA and standard preprocessing showed differences in pain‐related areas and in the white matter. (3.c) The comparison of brain activation following FIX preprocessing to brain activation following ICA‐AROMA preprocessing showed greater activation in S2, insula, ACC, and basal ganglia following FIX preprocessing. Little difference was found between these two preprocessing in terms of deactivation.
FIGURE 4.
Brain activation associated with high‐intensity heat stimulus after preprocessing with CompCor‐based noise‐reduction techniques is reduced compared to activation after standard preprocessing. (1.a and 3.a) After preprocessing with aCompCor (denoising techniques, activation was observed in gray matter areas previously associated with pain, including S1M1, S2, insula, and ACC. Deactivation is observed in the precuneus and PCC. (2.a and 3.b) After preprocessing including tCompCor, activation and deactivation was observed in similar areas as after preprocessing with aCompCor. In addition, increased activation was detected in the white matter. (1.b and 2.b) Activation and deactivation after standard preprocessing were detected in areas similar as after aCompCor preprocessing. In addition, deactivation was observed in the white matter. (1.c) The comparison of brain activation after preprocessing protocol with aCompCor noise‐reduction technique to brain activation after standard preprocessing protocol showed reduced positive and negative signal detection after preprocessing with aCompCor in pain‐related areas and in the white matter. (2.c) Similarly, reduced activation and deactivation were detected following preprocessing with tCompCor compared to standard preprocessing. (3.c) Little difference was found in brain activation and deactivation between preprocessing with aCompCor and preprocessing with tCompCor.
FIGURE 5.
Brain activation associated with high‐intensity heat stimulus after preprocessing with ICA‐based noise‐reduction techniques is greater than after preprocessing with CompCor‐based noise‐reduction techniques. (1‐2.a) Following FIX preprocessing, noxious heat stimuli are associated with increased and decreased brain activation in gray matter areas previously associated with pain (activation: S1M1, S2, insula, and ACC; deactivation: PCC, precuneus, and S1M1 associated with non‐stimulated body areas) and in the white matter; (3‐4.a) brain activation and deactivation associated with noxious heat stimuli after ICA‐AROMA was similar to the one detected after FIX preprocessing; (1.b and 3.b) after preprocessing with aCompCor, brain activation and deactivation was observed in brain areas previously associated with pain; (2.b and 4.b) brain activation and deactivation following preprocessing with tCompCor was similar to brain activation and deactivation after the other preprocessing protocols. (1‐4.c) The comparison of brain activation following preprocessing with ICA‐based noise reduction techniques (FIX and ICA‐AROMA) to brain activation after CompCor‐based preprocessing revealed greater brain activation and deactivation after ICA‐based preprocessing.
The comparison of preprocessing protocols with ICA‐based noise‐reduction techniques with the standard preprocessing (Figure 3‐1, 2) showed greater activation and deactivation in the areas mentioned above after preprocessing with noise‐reduction techniques. In addition, direct comparison of the ICA‐based noise‐reduction techniques revealed that FIX resulted in greater activations than ICA‐AROMA (Figure 3‐3). Few differences in deactivations were found between preprocessing with FIX and ICA‐AROMA (Figure 3‐3).
The comparison of preprocessing with CompCor‐based noise‐reduction techniques with standard preprocessing on brain activation associated with noxious heat (Figure 4‐1, 2) revealed less activation and deactivation in the areas mentioned above, that is, ACC, S1M1, insula, basal ganglia, thalamus, PCC, and precuneus following preprocessing with CompCor‐based noise‐reduction techniques. The comparison of the preprocessing protocols with CompCor‐based noise‐reduction techniques showed little difference between the two (Figure 4‐3).
When directly compared for brain activation associated with noxious heat, ICA‐based techniques revealed greater activations and deactivations than CompCor‐based techniques for the brain regions mentioned above (Figure 5).
Deactivation in the white matter was also observed following standard preprocessing and preprocessing with ICA‐based noise‐reduction techniques, that is, FIX and ICA‐AROMA (Figures 3, 4, 5). As expected, white matter changes were reduced after preprocessing with aCompCor. Surprisingly, white matter had a significant increase in activation after preprocessing with tCompCor.
Overall, the comparison of the different preprocessing protocols showed that the application of ICA‐based noise‐reduction techniques enhanced the magnitude of meaningful signal compared to standard or CompCor‐based noise‐reduction techniques. This enhancement was evident in gray matter areas previously associated with pain, including the S1M1, ACC, insula and basal ganglia, as well as PCC and precuneus.
3.1.1. Relationship of reported heat stimulus intensity with brain activation
No relationship was found between rating of intensity and brain activation associated with high‐intensity heat stimuli after standard, FIX and aCompCor preprocessing. This absence of a relationship has been addressed in detail previously (Hoeppli et al., 2022).
3.2. Prevalence of activation in the heat series across participants
Analyses were performed to assess the percentage of participants exhibiting brain activation in similar areas in response to high‐intensity heat stimulation following each preprocessing protocol.
These analyses revealed that the maximum percentage of participants showing brain activation and deactivation in similar areas was greater following standard preprocessing and preprocessing with ICA‐based noise‐reduction techniques than following preprocessing with CompCor‐based techniques (Table 3).
TABLE 3.
Prevalence of participants showing similar brain activation in response to a highly noxious stimulus is greater following preprocessing with ICA‐based or standard preprocessing than following preprocessing with CompCor.
| Preprocessing protocol | Increased activation | Decreased activation |
|---|---|---|
| FIX | 94% | 79% |
| ICA‐AROMA | 95% | 87% |
| Standard | 97% | 84% |
| aCompCor | 71% | 39% |
| tCompCor | 85% | 50% |
Note: Increased activation represents the prevalence for areas with increased activation, while decreased activation shows the prevalence for areas with decreased activation.
High prevalence of activation and deactivation, that is, activation/deactivation detected in more than 70% of participants, was detected differently depending on the preprocessing protocol (Figure 6; yellow/dark blue activation). Following FIX, ICA‐AROMA, and standard preprocessing, increased activation in the right S1M1 (leg area), the bilateral ACC, and the bilateral insula was detected at high prevalence (Figure 6a–c; yellow activation). Deactivations following FIX, ICA‐AROMA, and standard processing were detected in the bilateral precuneus and PCC at high prevalence (Figure 6a–c; dark blue deactivation). Moderate prevalence of activation and deactivation, that is, activation that was detected in 40%–70% of the participants, following ICA‐based and standard preprocessing was found in areas surrounding the high prevalence areas and in right S1M1 (hand area). Similarly, low prevalence activation and deactivation (de/activation in 10%–40% of participants) was observed in areas surrounding the previously found areas and in the white matter.
FIGURE 6.
Prevalence of activation, that is, percentage of participants with similar brain activation, is greater after ICA‐based preprocessing (FIX and ICA‐AROMA) compared to standard and CompCor‐based preprocessing protocols. Prevalence of increased activation is represented in yellow for high prevalence (activation detected in more than 70% of participants), in orange for moderate prevalence (40%–70% of participants), and in red for low prevalence (10%–40% of participants). Prevalence of decreased activation is represented in dark blue for high prevalence, in blue for moderate prevalence, and in light blue for low prevalence. (a), (b) Prevalence of increased activation after FIX (a) and ICA‐AROMA (b) preprocessing was high in areas including S1M1, the ACC, and the insula, moderate around these areas and low around the moderate areas. Decreased activation was observed with high prevalence in areas including the precuneus and PCC, with moderate prevalence around these areas and in the white matter, and with low prevalence in areas surrounding the moderate‐prevalence areas; (c) Compared to prevalence following ICA‐based preprocessing, high, moderate, and low prevalence of increased and decreased activation after standard preprocessing was observed in similar but smaller areas; (d) High prevalence of increased activation was only observed in a small areas of the right insula and S1M1 after preprocessing with tCompCor, while moderate‐ and low‐prevalence activations were observed in the insula, S1M1, and ACC. Decreased activation was observed at moderate and low prevalence in the hand area of S1M1, precuneus, and PCC. (e) Following preprocessing with aCompCor, activation in S1M1, S2, insula, putamen and ACC was only found at moderate and low prevalence, while deactivation was only found at low prevalence in the PCC and precuneus.
Only very small areas of the right S1M1 (leg area), S2, and insula showed high prevalence of increased activation following tCompCor preprocessing (Figure 6d; yellow activation). At moderate and low prevalence, activation following tCompCor was found in areas surrounding the one activated at high prevalence and in areas including the left S1M1, S2, insula, and ACC. Deactivation at these prevalence was observed in areas including the hand area of S1M1, the PCC, and the precuneus.
Following aCompCor (Figure 6e), only moderate and low activation were observed in areas including S1M1, S2, the insula, the putamen, and the ACC. Deactivation after aCompCor (Figure 6e) was only observed at a low prevalence in right S1M1 (hand area) and PCC.
3.2.1. Noise and signal detection in the heat series after each preprocessing protocol
A three‐way repeated‐measures ANOVA was performed to investigate the interactive effect between maps (cope/varcope), brain areas, and preprocessing protocols on averaged parameter estimates. Results of this ANOVA showed a significant interaction between maps, brain areas, and preprocessing protocols (F(3.54,137.87) = 86.853, p < .0001).
Given this significant result, two two‐way repeated‐measures ANOVA were performed to compare averaged parameter estimates between brain areas and preprocessing protocols within map. Results of these analyses showed significant interactions between brain areas and preprocessing protocols within map after Bonferroni correction (cope: F(4.69,305) = 86.1, p < .0001; varcope: F(1.41,73.4) = 84, p < .0001).
To further investigate these interactions, one‐way ANOVAs were performed to establish the effect of preprocessing protocols within map and brain areas. Results after Bonferroni correction showed a significant main effect of the preprocessing protocols on averaged parameter estimates in the cope map in all investigated areas: ACC (F(1.76,164) = 151, p < .0001), insula (F(2.39,229) = 137, p < .0001), S1 (F(2.25,205) = 69.7, p < .0001), CSF (F(2.38,203) = 113, p < .0001), and white matter (F(1.85,161) = 210, p < .001). A significant main effect of the preprocessing protocols was also found in the varcope map in all the brain areas studies (ACC: F(1.47,131) = 238, p < .0001; insula: F(1.54,137) = 329, p < .0001; S1: F(2.14,173) = 143, p < .0001; CSF: F(1.56,114) = 132, p < .0001; white matter: F(1.13,96.1) = 160, p < .0001).
Finally, pairwise paired T‐tests were performed within map and brain area to identify differences in parameter estimates between preprocessing protocols based on the significant main effects of the preprocessing protocols identified in the one‐way ANOVAs. In the cope image, Bonferroni‐corrected results (Figure 7, left panel; Table 4, left) showed significant differences in the S1 between ICA‐based protocols and standard protocol, between CompCor‐based and standard protocols, between aCompCor and tCompCor protocols, and between ICA‐based and CompCor‐based protocols (Table 4, left, row 3–12). In the insula, significant differences were found between ICA‐AROMA and standard preprocessing protocols, between CompCor‐based and standard preprocessing, between aCompCor and tCompCor, and between ICA‐based preprocessing protocols and CompCor‐based protocols (Table 4, left, row 13–22). Significant differences in ACC were found between ICA‐based protocols and standard protocol, between CompCor‐based and standard protocols, between aCompCor and tCompCor protocols, and between ICA‐based and CompCor‐based protocols (Table 4, left, row 23–32). Significant differences were found in the CSF between ICA‐AROMA and standard protocols, between FIX and ICA‐AROMA protocols, between tCompCor and standard protocols, between aCompCor and tCompcor, and between ICA‐based and CompCor‐based protocols (Table 4, left, row 33–42). Finally, significant differences in the white matter were observed between ICA‐AROMA and standard preprocessing, between FIX and ICA‐AROMA, between CompCor‐based and standard preprocessing, and between ICA‐based and CompCor‐based preprocessing (Table 4, left, row 43–52).
FIGURE 7.
ICA‐based noise‐reduction techniques conserve significantly more signal individually in areas of interest than CompCor‐based techniques, while removing more noise than standard preprocessing. Individual averaged parameter estimates in cope (contrast parameter estimates; left panel) and varcope (variance of contrast parameter estimates; right panel) images are represented for each preprocessing protocol (aCompCor (aCC): purple, ICA‐AROMA (AROMA): blue, FIX: teal; standard (std): green; tCompCor (tCC): yellow) for the following areas of interest: primary somatosensory cortex (S1), insula, anterior cingulate cortex (ACC), cerebrospinal fluid (CSF), and white matter (WM). Signal detection was significantly smaller after preprocessing with aCompCor than after the other preprocessing in all investigated gray matter areas. Signal detection was significantly greater in the CSF and WM after ICA‐AROMA preprocessing than after any other preprocessing protocol. Remaining noise in the images was always significantly greater following standard preprocessing and ICA‐based preprocessing than after CompCor‐based preprocessing in all areas investigated. Error bars: ± 1 standard error. Identical level of significance of pairwise comparisons are indicated by multiple square brackets with the reference preprocessing on the open end of the bracket. *p < .05; **p < .01; ***p < .001; ****p < .0001.
TABLE 4.
Results and significance of pairwise paired T‐tests in the heat series within brain area identifying differences in parameter estimates (cope) and their variance (varcope) between preprocessing protocols.
| Cope | Varcope | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Area | Preprocessing comparison | T statistic | df | p.adj | p.adj.signif | T statistic | df | p.adj | p.adj.signif |
| S1 | aCompCor‐ICA‐AROMA | −10.53459813 | 91 | 1.96E‐16 | **** | −13.514151 | 81 | 1.85E‐21 | **** |
| aCompCor‐FIX | −8.949849709 | 91 | 4.02E‐13 | **** | −13.361912 | 81 | 3.50E‐21 | **** | |
| aCompCor‐std | −12.12784772 | 91 | 1.04E‐19 | **** | −14.638799 | 81 | 1.83E‐23 | **** | |
| aCompCor‐tCompCor | −8.614911567 | 91 | 2.02E‐12 | **** | −9.2648542 | 81 | 2.38E‐13 | **** | |
| ICA‐AROMA‐FIX | 2.278817452 | 91 | 0.25 | ns | 1.91807264 | 81 | 0.586 | ns | |
| ICA‐AROMA‐std | −4.559060747 | 91 | 0.00016 | *** | −7.7709712 | 81 | 2.12E‐10 | **** | |
| ICA‐AROMA‐ tCompCor | 6.660558638 | 91 | 2.03E‐08 | **** | 10.7321481 | 81 | 3.16E‐16 | **** | |
| FIX‐std | −4.878604962 | 91 | 4.52E‐05 | **** | −7.6610001 | 81 | 3.48E‐10 | **** | |
| FIX‐tCompCor | 5.141289131 | 91 | 1.55E‐05 | **** | 9.4580803 | 81 | 9.89E‐14 | **** | |
| std‐tCompcor | 8.663494872 | 91 | 1.60E‐12 | **** | 14.7382118 | 81 | 1.23E‐23 | **** | |
| insula | aCompCor‐ICA‐AROMA | −14.22345195 | 96 | 2.28E‐24 | **** | −16.061573 | 89 | 4.94E‐27 | **** |
| aCompCor‐FIX | −16.52686045 | 96 | 7.90E‐29 | **** | −15.606048 | 89 | 3.29E‐26 | **** | |
| aCompCor‐std | −16.18539932 | 96 | 3.47E‐28 | **** | −20.842244 | 89 | 5.31E‐35 | **** | |
| aCompCor‐tCompCor | −5.317384775 | 96 | 6.85E‐06 | **** | −16.512636 | 89 | 7.74E‐28 | **** | |
| ICA‐AROMA‐FIX | −1.732030903 | 96 | 0.865 | ns | 8.05372313 | 89 | 3.37E‐11 | **** | |
| ICA‐AROMA‐std | −4.496402861 | 96 | 0.000193 | *** | −22.072654 | 89 | 7.31E‐37 | **** | |
| ICA‐AROMA‐tCompCor | 10.03709865 | 96 | 1.24E‐15 | **** | 6.77999411 | 89 | 1.27E‐08 | **** | |
| FIX‐std | −2.397545141 | 96 | 0.184 | ns | −18.61564 | 89 | 1.93E‐31 | **** | |
| FIX‐tCompCor | 11.33612061 | 96 | 2.08E‐18 | **** | −2.9880254 | 89 | 0.036 | * | |
| std‐tCompcor | 11.54156639 | 96 | 7.63E‐19 | **** | 20.8563465 | 89 | 5.05E‐35 | **** | |
| ACC | aCompCor‐ICA‐AROMA | −14.91079755 | 93 | 2.15E‐25 | **** | −15.565791 | 89 | 3.90E‐26 | **** |
| aCompCor‐FIX | −15.00485759 | 93 | 1.42E‐25 | **** | −15.175586 | 89 | 2.02E‐25 | **** | |
| aCompCor‐std | −14.84963568 | 93 | 2.82E‐25 | **** | −17.447841 | 89 | 1.81E‐29 | **** | |
| aCompCor‐tCompCor | −6.606628925 | 93 | 2.42E‐08 | **** | −15.702402 | 89 | 2.20E‐26 | **** | |
| ICA‐AROMA‐FIX | 0.319174987 | 93 | 1 | ns | 6.66084476 | 89 | 2.18E‐08 | **** | |
| ICA‐AROMA‐std | −5.005122785 | 93 | 2.63E‐05 | **** | −15.483142 | 89 | 5.51E‐26 | **** | |
| ICA‐AROMA‐tCompCor | 11.47496946 | 93 | 1.65E‐18 | **** | 8.39790179 | 89 | 6.60E‐12 | **** | |
| FIX‐std | −4.617364053 | 93 | 0.000124 | *** | −14.970179 | 89 | 4.86E‐25 | **** | |
| FIX‐tCompCor | 11.56385443 | 93 | 1.08E‐18 | **** | 3.59591238 | 89 | 0.005 | ** | |
| std‐tCompcor | 11.86913893 | 93 | 2.51E‐19 | **** | 16.1790465 | 89 | 3.04E‐27 | **** | |
| CSF | aCompCor‐ICA‐AROMA | −14.64024707 | 85 | 5.89E‐24 | **** | −12.152254 | 73 | 3.29E‐18 | **** |
| aCompCor‐FIX | 9.936121771 | 85 | 6.91E‐15 | **** | −9.9271812 | 73 | 3.48E‐14 | **** | |
| aCompCor‐std | 2.034918846 | 85 | 0.45 | ns | −20.125259 | 73 | 1.64E‐30 | **** | |
| aCompCor‐tCompCor | −7.045201194 | 85 | 4.52E‐09 | **** | −12.633335 | 73 | 4.76E‐19 | **** | |
| ICA‐AROMA‐FIX | 16.74281914 | 85 | 1.20E‐27 | **** | −9.0204801 | 73 | 1.70E‐12 | **** | |
| ICA‐AROMA‐std | 12.55218819 | 85 | 4.71E‐20 | **** | −19.22172 | 73 | 2.77E‐29 | **** | |
| ICA‐AROMA‐tCompCor | 12.58104807 | 85 | 4.14E‐20 | **** | 7.85068153 | 73 | 2.69E‐10 | **** | |
| FIX‐std | −2.124954777 | 85 | 0.365 | ns | −1.6966736 | 73 | 0.94 | ns | |
| FIX‐tCompCor | −12.64314133 | 85 | 3.15E‐20 | **** | 9.48243094 | 73 | 2.33E‐13 | **** | |
| std‐tCompcor | −3.815466113 | 85 | 0.003 | ** | 19.9907397 | 73 | 2.49E‐30 | **** | |
| WM | aCompCor‐ICA‐AROMA | −12.57703695 | 87 | 2.82E‐20 | **** | −17.244701 | 85 | 1.71E‐28 | **** |
| aCompCor‐FIX | 14.45291051 | 87 | 7.58E‐24 | **** | −17.008311 | 85 | 4.26E‐28 | **** | |
| aCompCor‐std | 11.38791278 | 87 | 6.31E‐18 | **** | −13.63059 | 85 | 4.25E‐22 | **** | |
| aCompCor‐tCompCor | −2.459376709 | 87 | 0.159 | ns | −16.562878 | 85 | 2.43E‐27 | **** | |
| ICA‐AROMA‐FIX | 25.71748749 | 87 | 2.00E‐41 | **** | 1.29324478 | 85 | 1 | ns | |
| ICA‐AROMA‐std | 19.30971287 | 87 | 3.31E‐32 | **** | −11.543259 | 85 | 4.32E‐18 | **** | |
| ICA‐AROMA‐tCompCor | 9.93747932 | 87 | 5.42E‐15 | **** | 15.8563778 | 85 | 4.03E‐26 | **** | |
| FIX‐std | 1.45789169 | 87 | 1 | ns | −10.816566 | 85 | 1.19E‐16 | **** | |
| FIX‐tCompCor | −13.65177682 | 87 | 2.41E‐22 | **** | 12.8405741 | 85 | 1.32E‐20 | **** | |
| std‐tCompcor | −11.01235865 | 87 | 3.58E‐17 | **** | 13.192706 | 85 | 2.83E‐21 | **** | |
Note: ACC, anterior cingulate cortex; CSF, cerebrospinal fluid, WM: white matter; S1, primary somatosensory cortex.
p < .05;
p < .01;
p < .001;
p < .0001.
In the varcope images (Figure 7, right panel; Table 4, right), significant differences were found between all preprocessing protocols except between both ICA‐based protocols in the S1 (Table 4, right, rows 3–12). Significant differences were found between all the preprocessing protocols in the insula (Table 4, right, rows 13–22) and in the ACC (Table 4, right, rows 23–32). In the CSF significant differences were found between all preprocessing protocols, except between FIX and standard protocols (Table 4, right, rows 33–42). In the white matter, significant differences were found between all preprocessing protocols except between both ICA‐based protocols (Table 4, right, rows 43–52).
Overall, the ICA‐based preprocessing protocols preserved more signal of interest at an individual level than the CompCor‐based preprocessing protocol, as shown by the greater parameter estimates in the cope maps while removing more noise as indicated by lower variance in the varcope maps (Figure 7). Comparing the standard preprocessing protocol to the CompCor‐based preprocessing protocol showed similar results. Finally, the standard preprocessing protocol preserved slightly more signal of interest at an individual level but also removed less noise than the ICA‐based preprocessing protocols. Surprisingly, the signal detected in the CSF following preprocessing with ICA‐AROMA noise‐removal technique was greater than following any other preprocessing protocol with or without noise‐removal techniques.
3.3. Differences in brain activation associated with a single high‐intensity heat stimulus using a single‐epoch design between preprocessing protocols
GLM analyses were performed to investigate the effect of the standard, FIX and aCompCor preprocessing protocols on brain activation associated with a single epoch, that is, a single high‐intensity heat stimulus.
Results from these analyses (Figure 8) showed increased brain activation after FIX, aCompCor, and standard preprocessing in areas such as ACC, S1M1, the insula, basal ganglia, and the thalamus. Deactivation was observed in the precuneus and PCC after preprocessing with all three protocols.
FIGURE 8.
Signal detection associated with a single high‐intensity heat stimulus after FIX preprocessing is greater than after aCompCor and standard preprocessing. (1.a and 3.a) Increased and decreased brain activation associated with a single noxious heat stimulus after FIX preprocessing is observed in gray matter areas consistent with pain processing and in the white matter; (1.b and 2.b) Increased and decreased brain activation associated with a single noxious heat stimulus after standard preprocessing was observed in areas similar to the ones detected after FIX preprocessing; (2.a and 3.b) Increased and decreased brain activation associated with a single noxious heat stimulus after aCompCor preprocessing was observed in gray matter areas consistent with pain processing. (1.c) The comparison of brain activation after FIX preprocessing and brain activation after standard preprocessing shows greater positive and negative signal detection after FIX preprocessing; (2.c) Compared to standard preprocessing, signal appears to be removed by the aCompCor preprocessing; (3.c) Similarly, signals appear to be removed by aCompCor preprocessing compared to FIX preprocessing.
After FIX preprocessing compared to standard preprocessing, an increased signal was detected at a single stimulus level (Figure 8‐1): greater activation was found in gray matter areas previously described and typically associated with noxious stimulation. Deactivation in gray matter areas, including the PCC and precuneus, and in the white matter was better detected following FIX preprocessing.
When comparing the preprocessing protocol including aCompCor noise‐reduction technique with the standard preprocessing protocol at a stimulus level, better signal detection can be observed after standard preprocessing (Figure 8‐2). Increased activation was detected after standard preprocessing in brain areas typically associated with noxious stimulation. Deactivation was also better detected after standard preprocessing in gray matter areas, including the PCC and the precuneus, and in the white matter.
Signal detection increased greatly after preprocessing including the FIX noise‐reduction technique compared to preprocessing with aCompCor noise‐reduction technique at a single stimulus level (Figure 8‐3). Increased activation was detected following preprocessing with FIX in gray matter areas associated with noxious stimulation. Deactivation was also better detected after preprocessing with FIX in gray matter areas, including the PCC and the precuneus, and in the white matter.
3.4. Differences in brain activation associated with high‐intensity auditory stimulation using a block design between preprocessing protocols
Non‐noxious auditory stimulation was used as a control modality to investigate the effect of the different preprocessing protocols on cerebral responses to stimulation that has no strongly associated systemic physiological response.
Analyses of brain activations associated with high‐intensity auditory stimulation following standard, ICA‐based (FIX and ICA‐AROMA), and CompCor‐based preprocessing protocols (aCompCor and tCompCor) showed increased activations in areas previously associated with auditory perception, including the primary auditory cortex (A1), ACC, and the insula (Figures 9, 10, 11). Deactivations were seen after all our preprocessing protocols in gray matter areas including S1M1, and in the white matter.
FIGURE 9.
Brain activation associated with the high‐intensity heat stimulus after preprocessing with ICA‐based noise‐reduction techniques is greater than after preprocessing with standard preprocessing. (1‐3.a and b) Following ICA‐based and standard preprocessing protocols, high‐intensity auditory stimuli were associated with increased and decreased brain activation in gray matter areas previously associated with auditory processing and in the white matter. (1.c) The comparison of brain activation after preprocessing protocol with FIX noise‐reduction technique to brain activation after standard preprocessing protocols mainly showed greater activation in the primary auditory cortex (A1) after FIX. (2.c) Similarly, greater A1 activation was detected following preprocessing with ICA‐AROMA compared to standard preprocessing. (3.c) No difference in brain activation was found after preprocessing with FIX compared to preprocessing with AROMA.
FIGURE 10.
Greater signal detection after CompCor‐based preprocessing compared to standard preprocessing. (1‐3.a and b) Following Compcor‐based and standard preprocessing protocols, high‐intensity auditory stimuli were associated with increased and decreased brain activation in gray matter areas previously associated with auditory processing (A1) and in the white matter. (1.c) The comparison of brain activation after preprocessing with aCompCor to brain activation after standard preprocessing shows greater activation in A1 and greater deactivation in the thalamus. (2.c) Differences in brain activation following tCompCor compared to activation following standard preprocessing show differences in gray matter areas such as S1M1, ACC, and in the white matter. (3.c) Little difference in brain activation associated with high‐intensity auditory stimuli were found when comparing signal detection after preprocessing with aCompCor to signal detection after preprocessing with tCompCor.
FIGURE 11.
Brain activation associated with high‐intensity heat stimulus after preprocessing with ICA‐based noise‐reduction techniques is greater than after preprocessing with CompCor‐based noise‐reduction techniques. (1‐4.a and b) Following ICA‐based and CompCor‐based preprocessing protocols, high‐intensity auditory stimuli are associated with increased and decreased brain activation in gray matter areas previously associated with auditory processing and in the white matter. (1.c) The comparison of brain activation after preprocessing protocol with FIX noise‐reduction technique to brain activation after aCompCor preprocessing protocol shows greater positive and negative signal detection after FIX in gray matter areas previously associated with auditory processing. (2.c) Greater activation in gray matter areas associated with auditory preprocessing and greater white matter deactivation were associated after FIX preprocessing compared to tCompCor preprocessing. (3.c) Small differences in the thalamus were found when comparing brain activation following ICA‐AROMA and aCompCor preprocessing. (4.c) Similarly, small differences in thalamus were found in the comparison of brain activation after ICA‐AROMA to tCompCor preprocessing. In addition, increased white matter deactivation was observed following ICA‐AROMA preprocessing.
After ICA‐based preprocessing compared to standard preprocessing, the detection of the signal of interest was increased in gray matter areas, including A1 and the insula (Figure 9‐1, 2). There was no difference in deactivation in areas of interest. In addition, there was no difference in activation or deactivation between FIX and ICA‐AROMA (Figure 9‐3).
The comparison of brain activation following aCompCor and standard preprocessing showed greater activation in A1 following aCompCor (Figure 10‐1). The comparison of brain activation following tCompCor and standard preprocessing revealed mostly increased activation in S1M1 and in the white matter following tCompCor (Figure 10‐2). Reduced activation was found in ACC following CompCor‐based protocols compared to standard protocol. Differences in deactivation between CompCor‐based and standard preprocessing revealed greater deactivation of the thalamus following CompCor‐based preprocessing (Figure 10‐1, 2). Finally, the comparison of both CompCor‐based protocol showed only a small difference in A1 in favor of aCompCor (Figure 10‐3).
The comparison of brain activation after ICA‐based preprocessing with CompCor‐based preprocessing showed significantly greater activation in A1, in the insula, in the basal ganglia, and in ACC after ICA‐based preprocessing (Figure 11). Greater deactivation was found in the white matter following ICA‐based preprocessing compared to tCompCor preprocessing (Figure 11‐2, 4).
3.5. Relationship of reported auditory stimulus intensity with brain activation
A significant correlation was found between ratings of perceived auditory intensity and brain activation associated with high‐intensity auditory stimuli following ICA‐based and standard preprocessing (Figure 12a–c). This included primarily the left A1, S2, and insula (positive correlation), as well as the ACC and PCC/precuneus (negative correlation). No such correlation was found following CompCor‐based preprocessing (Figure 12d, e). The level of signal of interest remaining in the data, which was extracted from the cope image, further supported this finding: values of parameter estimates following ICA‐based and standard preprocessing ranged between 0.3 and 0.4 (FIX: 0.3991; ICA‐AROMA: 0.3107; standard: 0.3566), while the values of parameter estimates following CompCor‐based preprocessing were below 0.1 (aCompCor: 0.0062; tCompCor: 0.08901).
FIGURE 12.
In the auditory series, individual ratings of perceived intensity are associated with brain activation following ICA‐based and standard preprocessing, but not following CompCor‐based preprocessing. Positive correlation (yellow‐red) was detected following FIX (a), ICA‐AROMA (b), and standard (c) preprocessing in an area including primarily the left A1, S2, and insula. Similarly, a negative correlation (blue) was found in the ACC and PCC/precuneus (not displayed). No similar correlation was found following tCompCor (d) or aCompCor (e).
3.6. Noise and signal detection in the auditory series after each preprocessing protocol
The analysis investigating remaining noise and signal following each preprocessing protocol was repeated in the auditory series. This three‐way repeated‐measures ANOVA (maps × brain area × preprocessing protocol) yield a significant interaction between maps, brain areas, and preprocessing protocols (F(1,53) = 291.77, p < .0001).
To further investigate this interaction, two‐way repeated‐measures ANOVAs within map were performed. Results of these analyses showed significant interactions between brain areas and preprocessing protocols within map after Bonferroni correction (cope: F(1,68) = 118, p < .0001; varcope: F(1,71) = 397, p < .0001).
One‐way ANOVAs were then performed to establish the effect of preprocessing protocols within map and brain area. Results after Bonferroni correction showed a significant main effect of the preprocessing protocols on averaged parameter estimates in the cope map in all the brain areas investigated: A1: F(1.9,175) = 37.7, p < .0001; CSF: F(1,79) = 118, p < .0001; white matter (F(2.58,217) = 32, p < .0001. Similarly, a significant main effect of the preprocessing protocols was found in the varcope map in all the brain areas studied: A1: F(1.74,150) = 250, p < .0001; CSF: F(1,78) = 387, p < .0001; white matter: F(1.12,93.8) = 150, p < .0001).
Finally, pairwise paired T‐tests were performed within map and brain area to identify differences in parameter estimates between preprocessing protocols. In the cope image, Bonferroni‐corrected results (Figure 13, left panel; Table 5, left) showed significant differences in the A1 between all preprocessing protocols, except between FIX and tCompCor (Table 5, left, rows 3–12). In the CSF, significant differences were found between ICA‐AROMA and standard preprocessing, between ICA‐based preprocessing protocols, between CompCor‐based and standard protocols, and between ICA‐based and CompCor‐based protocols (Table 5, left, rows 13–22). In the white matter, significant differences were found between FIX and standard preprocessing, between CompCor‐based and standard preprocessing, and between ICA‐based and CompCor‐based preprocessing protocols (Table 5, left, rows 23–32).
FIGURE 13.
In the auditory series similarly to the heat series, ICA‐based noise reduction techniques conserves significantly more signal individually in areas of interest than CompCor‐based techniques, while removing more noise than standard preprocessing. Individual averaged parameter estimates in cope (contrast parameter estimates; left panel) and varcope (variance of contrast parameter estimates; right panel) images are represented for each preprocessing protocol (aCompCor (aCC): purple, ICA‐AROMA (AROMA): blue, FIX: teal; standard (std): green; tCompCor (tCC): yellow) for the following areas of interest: primary auditory cortex (A1), cerebrospinal fluid (CSF), and white matter (WM). Signal detection was significantly smaller after CompCor‐based preprocessing protocols than after the other preprocessing in all the investigated areas. Remaining noise in the images was always significantly greater following standard preprocessing and FIX preprocessing than after CompCor‐based preprocessing in all areas investigated. Signal detection and remaining noise were exaggeratedly greater in the CSF and WM after ICA‐AROMA preprocessing than after any other preprocessing protocol, suggesting overcorrection. Error bars: ± 1 standard error. Identical level of significance of pairwise comparisons is indicated by multiple square brackets with the reference preprocessing on the open end of the bracket. *p < .05; **p < .01; ***p < .001; ****p < .0001.
TABLE 5.
Results and significance of pairwise paired T‐tests in the auditory series within brain area identifying differences in parameter estimates (cope) and their variance (varcope) between preprocessing protocols.
| Cope | Varcope | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Area | Preprocessing comparison | T statistic | df | p.adj | p.adj.signif | T statistic | df | p.adj | p.adj.signif |
| A1 | aCompCor‐ICA‐AROMA | −7.987919554 | 92 | 3.82E‐11 | **** | −16.424792 | 86 | 2.98E‐27 | **** |
| aCompCor‐FIX | −6.26373444 | 92 | 1.19E‐07 | **** | −16.986643 | 86 | 3.25E‐28 | **** | |
| aCompCor‐std | −8.525393165 | 92 | 2.88E‐12 | **** | −18.89148 | 86 | 2.35E‐31 | **** | |
| aCompCor‐tCompCor | −5.840326156 | 92 | 7.78E‐07 | **** | −16.984151 | 86 | 3.28E‐28 | **** | |
| ICA‐AROMA‐FIX | 3.170186567 | 92 | 0.021 | * | 4.47970302 | 86 | 0.000229 | *** | |
| ICA‐AROMA‐std | −3.809126548 | 92 | 0.003 | ** | −14.39569 | 86 | 1.26E‐23 | **** | |
| ICA‐AROMA‐tCompCor | 4.548143334 | 92 | 0.000165 | *** | 9.3576727 | 86 | 9.20E‐14 | **** | |
| FIX‐std | −4.719656182 | 92 | 8.41E‐05 | **** | −14.289143 | 86 | 1.98E‐23 | **** | |
| FIX ‐ tCompCor | 2.8310091 | 92 | 0.057 | ns | 6.83362313 | 86 | 1.13E‐08 | **** | |
| std‐tCompcor | 5.464642731 | 92 | 3.94E‐06 | **** | 17.1439874 | 86 | 1.76E‐28 | **** | |
| CSF | aCompCor‐ICA‐AROMA | −10.85178719 | 79 | 2.54E‐16 | **** | −19.676543 | 78 | 5.60E‐31 | **** |
| aCompCor‐FIX | −3.971260514 | 79 | 0.002 | ** | −8.4318433 | 78 | 1.37E‐11 | **** | |
| aCompCor‐std | −5.261546822 | 79 | 1.19E‐05 | **** | −10.890054 | 78 | 2.53E‐16 | **** | |
| aCompCor‐tCompCor | 1.28423209 | 79 | 1 | ns | 5.36357039 | 78 | 8.09E‐06 | **** | |
| ICA‐AROMA‐FIX | 10.86020989 | 79 | 2.45E‐16 | **** | 19.6770374 | 78 | 5.59E‐31 | **** | |
| ICA‐AROMA‐std | 10.85784666 | 79 | 2.48E‐16 | **** | 19.6771374 | 78 | 5.59E‐31 | **** | |
| ICA‐AROMA‐tCompCor | 10.84987082 | 79 | 2.56E‐16 | **** | 19.6766231 | 78 | 5.60E‐31 | **** | |
| FIX‐std | −2.062308301 | 79 | 0.425 | ns | −8.975147 | 78 | 1.21E‐12 | **** | |
| FIX‐tCompCor | 4.445991872 | 79 | 0.000282 | *** | 9.22699786 | 78 | 3.92E‐13 | **** | |
| std‐tCompcor | 5.493468905 | 79 | 4.65E‐06 | **** | 10.9894436 | 78 | 1.64E‐16 | **** | |
| WM | aCompCor‐ICA‐AROMA | −7.501232896 | 84 | 5.98E‐10 | **** | −14.553908 | 84 | 1.11E‐23 | **** |
| aCompCor‐FIX | −5.995922198 | 84 | 4.90E‐07 | **** | −12.482138 | 84 | 7.86E‐20 | **** | |
| aCompCor‐std | −7.451483225 | 84 | 7.51E‐10 | **** | −13.079757 | 84 | 5.78E‐21 | **** | |
| aCompCor‐tCompCor | −1.354666123 | 84 | 1 | ns | −9.0843688 | 84 | 4.01E‐13 | **** | |
| ICA‐AROMA‐FIX | 2.520431767 | 84 | 0.136 | ns | 8.91178255 | 84 | 8.93E‐13 | **** | |
| ICA‐AROMA‐std | −2.743870179 | 84 | 0.074 | ns | −10.933069 | 84 | 8.09E‐17 | **** | |
| ICA‐AROMA‐tCompCor | 5.958765735 | 84 | 5.75E‐07 | **** | 13.4447256 | 84 | 1.20E‐21 | **** | |
| FIX‐std | −3.607420173 | 84 | 0.005 | ** | −11.477924 | 84 | 6.87E‐18 | **** | |
| FIX‐tCompCor | 4.945812774 | 84 | 3.84E‐05 | **** | 5.94437822 | 84 | 6.12E‐07 | **** | |
| std‐tCompcor | 6.364276028 | 84 | 9.87E‐08 | **** | 13.0062448 | 84 | 7.95E‐21 | **** | |
Note: A1, primary auditory cortex; CSF, cerebrospinal fluid; WM, white matter.
p < .05;
p < .01;
p < .001;
p < .0001.
In the varcope image (Figure 13, right panel; Table 5, right), significant differences were found between all preprocessing protocols in A1 (Table 5, right, rows 3–12), in the CSF (Table 5, right, rows 13–22), and in the white matter (Table 5, right, rows 23–32).
Overall, these results are in line with the results of the same analysis performed in the heat series: More signal of interest and more noise remained in the images following ICA‐based preprocessing protocols compared to the CompCor‐based preprocessing protocols (Figure 13). The comparison of remaining signal and noise following the standard preprocessing protocol with the CompCor‐based preprocessing protocols showed similar results. In addition, the standard preprocessing protocol preserved more signal of interest and more noise than the FIX preprocessing protocol. Finally, remaining signal and noise following ICA‐AROMA preprocessing protocol in the CSF were extremely high. This result taken together with similar results in the heat series suggest that ICA‐AROMA might overcorrect the data and reintroduce noise.
3.7. Correlations between aCompCor regressors and stimulus time course
Individual correlations between aCompCor regressors and heat stimuli ranged from −0.77 to 0.68 with a mean of 0.0026 and standard deviation of 0.15. Correlations between aCompCor regressors and auditory stimuli covered a smaller range (−0.47 to 0.48) with a mean of 0.0013 and a standard deviation of 0.12. Thus, signal loss during aCompCor is likely related to correlation of signals with aCompCor nuisance regressors.
4. DISCUSSION
ICA‐based and CompCor‐based preprocessing protocols reduced noise in task‐related fMRI data. Specifically, all noise‐reduction techniques enhanced the ability to detect activation during an auditory stimulus. However, for tasks that elicit correlated white matter responses, such as noxious heat stimulation, ICA‐based noise‐reduction techniques increase signal detection while CompCor‐based techniques remove more signals of interest. Data preprocessed with ICA‐based techniques displayed robust activation in this task, while considerably less activation was detected in data preprocessed with CompCor‐based techniques. The comparison of the two CompCor‐based techniques (aCompCor and tCompCor) revealed very few differences, suggesting that they performed similarly. Interestingly, greater signal was detected after FIX preprocessing compared to ICA‐AROMA. Furthermore, the investigation of the association between individual auditory intensity ratings and brain activation revealed that different preprocessing protocols differed in their ability to detect relationships with covariates of interest. These findings underscore the benefits of data cleaning protocols on task‐related data, while emphasizing the fact that the wrong choice of a data cleaning protocol can have a deleterious impact on the ability to assess brain activation.
At a group level, analyses based on a block design showed that greater brain activation associated with high‐intensity heat stimulation was detected after FIX preprocessing than after all the other preprocessing protocols. Following FIX preprocessing, increased activation was detected in gray matter regions typically associated with noxious heat stimuli, including the ACC, the insula, and S1. Increased deactivation was detected in gray matter areas that are part of the default mode network and typically decreased in response to painful stimuli, including the precuneus and PCC.
To understand the impact of different noise‐reduction techniques at the level of the single participant, we examined how many individuals activated specific brain regions in a suprathreshold fashion (i.e., prevalence of activation). Results from this block‐design based prevalence analysis showed: (1) greater spatial extent within gray matter areas associated with response to noxious stimulation following ICA‐based preprocessing compared to standard preprocessing; (2) the maximum percentage of participants showing similar within‐voxel activation across the whole brain, as well as the prevalence of detection and spatial extent within gray matter areas usually associated with noxious stimulation, was lower following CompCor‐based preprocessing protocols compared to standard preprocessing; (3) when comparing the impact of the ICA‐based preprocessing and CompCor‐based preprocessing, the maximum percentage of participants, as well as the prevalence of detection and the spatial extent within gray matter areas typically associated with response to noxious stimulation were greater following ICA‐based preprocessing. Thus, the benefits of ICA‐based noise reduction are evident at the individual level.
The analysis of the individual averaged parameter estimates in areas of interest including the ACC, insula, and S1, showed that ICA‐based and standard preprocessing conserved more signal than CompCor‐based preprocessing (analysis of cope maps), while both ICA‐based and CompCor‐based preprocessing removed more noise than standard preprocessing (analysis of varcope maps). Similarly, FIX and standard preprocessing conserved more signal in the white matter than the other preprocessing, while ICA‐based and CompCor‐based preprocessing removed more noise. FIX and CompCor‐based preprocessing removed more noise than standard preprocessing in the CSF and CompCor‐based preprocessing removed more noise than FIX and standard preprocessing. In addition, a high amount of signal was detected in the CSF following ICA‐AROMA preprocessing, suggesting that this technique might overcorrect.
While the aCompCor preprocessing protocol removed the most noise at individual and group levels, it also removed signals of interest. This was shown by fewer activated areas and a lower magnitude of activation in our GLM analyses and by detecting similar patterns of brain activation in only a low percentage of participants. This is likely due to some signal of interest being temporally associated with white matter response in the noxious heat task. This is further supported by the correlations between heat stimuli and aCompCor regressors. The aCompCor noise‐reduction technique relies on the assumptions that all the signals present in the white matter and CSF are noise and that noise in the gray matter is temporally correlated with noise in the white matter and CSF. Given that our noxious heat task is associated with significant white matter response, by regressing out the signals detected in this area, temporally related signals of interest in the gray matter were removed as well. Unlike the common beliefs that responses in the white matter are the result of noise, our results suggest that in tasks including noxious stimulation, changes in white matter signal might be representative of some systemic response to the stimulus in line with recent publications (Grajauskas et al., 2019; Li et al., 2020). Unfortunately, the design of our study does not allow us to define the source(s) of this response. However, global changes in cerebral blood flow may potentially contribute to the white matter signal changes, as such changes have been previously identified during intense noxious stimuli (Coghill et al., 1998). Further studies are needed to delineate the mechanisms underlying these transient changes in white matter signal intensity.
The superior performance of the FIX noise‐reduction technique is further supported by the analyses performed on a single high‐intensity heat stimulus using a single epoch design. Single epoch designs are more sensitive to noise than the block design approaches used in the other analyses (Koyama et al., 2003). Greater activations and deactivations were detected in gray matter areas similar to those seen in the block design following FIX preprocessing in comparison with activations/deactivations detected after aCompCor and standard preprocessing. This clearly highlights the capacity of the FIX noise‐reduction technique to increase the signal‐to‐noise ratio during the cleaning process.
Interestingly, the effect of the preprocessing protocol, and in particular of the denoising technique, seems to lessen in tasks with less associated white matter response, as shown in the analyses of our control condition. The analyses performed on our auditory paradigm showed that, although ICA‐based, in particular FIX, preprocessing still increases signal detection compared to standard and CompCor‐based preprocessing, the difference in terms of signal of interest can be minimal (see the comparison between ICA‐based and CompCor‐based preprocessing, Figure 11). This suggests that the impact of the preprocessing protocols is lesser in tasks that are not associated with significant white matter response, such as our auditory task, compared to our noxious heat paradigm.
The effect of preprocessing protocols can also be seen in the investigation of brain analysis associated with a stimulus and a covariate of interest. We investigated the correlation between ratings of perceived intensity (covariate of interest) and brain activation associated with our high‐intensity auditory and heat stimuli. Results show that a significant correlation is found following ICA‐based and standard preprocessing in our auditory paradigm. No correlations were found after CompCor‐based preprocessing in the auditory paradigm. In addition, ratings of perceived heat intensity were not associated with brain activation following standard, FIX, and aCompCor preprocessing protocols (addressed in detail previously; Hoeppli et al., 2022).
These results clearly show that different preprocessing protocols impact signal‐to‐noise ratio differently depending on the tasks and additional covariates of interest. Furthermore, these results suggest that, among the preprocessing protocols tested in this study, the FIX noise‐reduction technique is the most appropriate data‐cleaning approach for task fMRI data associated with an important white matter response as well as when the investigation is interested in individual responses that may be related to a covariate of interest.
5. CONCLUSIONS
In conclusion, preprocessing fMRI data is always a delicate balance between conserving the signal of interest and removing the noise. The choice of preprocessing protocols can significantly influence the signal‐to‐noise ratio. The efficacy of these noise reduction techniques can vary dynamically according to the nature of the signals within the brain. Therefore, studies should carefully consider the choice of the preprocessing protocols based on the type of tasks that will be used. Our results suggest that in tasks with important associated white matter responses and tasks without such responses, a preprocessing protocol including a FIX noise‐reduction technique might be a good solution to achieve this balance.
CONFLICT OF INTEREST STATEMENT
The authors declare no conflict of interest.
ACKNOWLEDGMENTS
We would like to thank Eric Leon, Victor J. 2nd Schneider, Gregory R. Lee, Blaise V. Jones, Benjamin Hunter, and David L. Moore for their contribution to this study. This work is supported by the National Institute of Neurological Disorders and Stroke (R01 NS085391, R01 NS 039426).
Hoeppli, M. E. , Garenfeld, M. A. , Mortensen, C. K. , Nahman‐Averbuch, H. , King, C. D. , & Coghill, R. C. (2023). Denoising task‐related fMRI: Balancing noise reduction against signal loss. Human Brain Mapping, 44(17), 5523–5546. 10.1002/hbm.26447
DATA AVAILABILITY STATEMENT
Activation maps included in this manuscript and the code used to preprocess the data with each noise‐reduction technique are available on Github (https://github.com/coghill-painlab/preproc_comparison) and archived in Zenodo (Hoeppli et al., 2023).
REFERENCES
- Andersson, J., Jenkinson, M., & Smith, S. (2007). Non‐linear registration, aka Spatial normalisation. FMRIB technical report TR07JA2.
- Beckmann, C. F. , & Smith, S. M. (2004). Probabilistic independent component analysis for functional magnetic resonance imaging. IEEE Transactions on Medical Imaging, 23, 137–152. [DOI] [PubMed] [Google Scholar]
- Behzadi, Y. , Restom, K. , Liau, J. , & Liu, T. T. (2007). A component based noise correction method (CompCor) for BOLD and perfusion‐based fMRI. NeuroImage, 37, 90–101. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Blasi, B. D. , Caciagli, L. , Storti, S. F. , Galovic, M. , Koepp, M. , Menegaz, G. , Barnes, A. , & Galazzo, I. B. (2020). Noise removal in resting‐state and task fMRI: Functional connectivity and activation maps. Journal of Neural Engineering, 17, 046040. [DOI] [PubMed] [Google Scholar]
- Brooks, J. C. W. , Faull, O. K. , Pattinson, K. T. S. , & Jenkinson, M. (2013). Physiological noise in brainstem fMRI. Frontiers in Human Neuroscience, 7, 623. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Coghill, R. , Talbot, J. , Evans, A. , Meyer, E. , Gjedde, A. , Bushnell, M. , & Duncan, G. (1994). Distributed processing of pain and vibration by the human brain. Journal of Neuroscience, 14, 4095–4108. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Coghill, R. C. , Sang, C. N. , Berman, K. F. , Bennett, G. J. , & Iadarola, M. J. (1998). Global cerebral blood flow decreases during pain. Journal of Cerebral Blood Flow and Metabolism, 18, 141–147. [DOI] [PubMed] [Google Scholar]
- Eippert, F. , Kong, Y. , Jenkinson, M. , Tracey, I. , & Brooks, J. C. W. (2017). Denoising spinal cord fMRI data: Approaches to acquisition and analysis. Neuroimage, 154, 255–266. [DOI] [PubMed] [Google Scholar]
- Eklund, A. , Knutsson, H. , & Nichols, T. E. (2019). Cluster failure revisited: Impact of first level design and physiological noise on cluster false positive rates. Human Brain Mapping, 40, 2017–2032. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Eklund, A. , Nichols, T. E. , & Knutsson, H. (2016). Cluster failure: Why fMRI inferences for spatial extent have inflated false‐positive rates. Proceedings of the National Academy of Sciences, 113, 7900–7905. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Epstein, J. N. , Casey, B. J. , Tonev, S. T. , Davidson, M. , Reiss, A. L. , Garrett, A. , Hinshaw, S. P. , Greenhill, L. L. , Vitolo, A. , Kotler, L. A. , Jarrett, M. A. , & Spicer, J. (2007). Assessment and prevention of head motion during imaging of patients with attention deficit hyperactivity disorder. Psychiatry Res Neuroimaging, 155, 75–82. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Fassbender, C. , Mukherjee, P. , & Schweitzer, J. B. (2017). Minimizing noise in pediatric task‐based functional MRI; adolescents with developmental disabilities and typical development. Neuroimage, 149, 338–347. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Grajauskas, L. A. , Frizzell, T. , Song, X. , & D'Arcy, R. C. N. (2019). White matter fMRI activation cannot be treated as a nuisance regressor: Overcoming a historical blind spot. Front Neurosci‐Switz, 13, 1024. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Griffanti, L. , Salimi‐Khorshidi, G. , Beckmann, C. F. , Auerbach, E. J. , Douaud, G. , Sexton, C. E. , Zsoldos, E. , Ebmeier, K. P. , Filippini, N. , Mackay, C. E. , Moeller, S. , Xu, J. , Yacoub, E. , Baselli, G. , Uğurbil, K. , Miller, K. L. , & Smith, S. M. (2014). ICA‐based artefact removal and accelerated fMRI acquisition for improved resting state network imaging. NeuroImage, 95, 232–247. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Hoeppli, M. E. , Garenfeld, M. A. , Mortensen, C. K. , Nahman‐Averbuch, H. , King, C. D. , & Coghill, R. C. (2023). Denoising task‐related fMRI: Balancing noise reduction against signal loss: data repository. https://github.com/coghill-painlab/preproc_comparison [DOI] [PMC free article] [PubMed]
- Hoeppli, M. E. , Nahman‐Averbuch, H. , Hinkle, W. A. , Leon, E. , Peugh, J. , Lopez‐Sola, M. , King, C. D. , Goldschneider, K. R. , & Coghill, R. C. (2022). Dissociation between individual differences in self‐reported pain intensity and underlying fMRI brain activation. Nat Commun, 13, 3569. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Jenkinson, M. (2002). Improved optimization for the robust and accurate linear registration and motion correction of brain images. NeuroImage, 17, 825–841. [DOI] [PubMed] [Google Scholar]
- Jenkinson, M. , Beckmann, C. F. , Behrens, T. E. J. , Woolrich, M. W. , & Smith, S. M. (2012). FSL. NeuroImage, 62, 782–790. [DOI] [PubMed] [Google Scholar]
- Jenkinson, M. , & Smith, S. (2001). A global optimisation method for robust affine registration of brain images. Medical Image Analysis, 5, 143–156. [DOI] [PubMed] [Google Scholar]
- Johnstone, T. , Walsh, K. S. O. , Greischar, L. L. , Alexander, A. L. , Fox, A. S. , Davidson, R. J. , & Oakes, T. R. (2006). Motion correction and the use of motion covariates in multiple‐subject fMRI analysis. Human Brain Mapping, 27, 779–788. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kasper, L. , Bollmann, S. , Diaconescu, A. O. , Hutton, C. , Heinzle, J. , Iglesias, S. , Hauser, T. U. , Sebold, M. , Manjaly, Z.‐M. , Pruessmann, K. P. , & Stephan, K. E. (2017). The PhysIO toolbox for modeling physiological noise in fMRI data. Journal of Neuroscience Methods, 276, 56–72. [DOI] [PubMed] [Google Scholar]
- Koyama, T. , McHaffie, J. G. , Laurienti, P. J. , & Coghill, R. C. (2003). The single‐epoch fMRI design: Validation of a simplified paradigm for the collection of subjective ratings. Neuroimage, 19, 976–987. [DOI] [PubMed] [Google Scholar]
- Li, M. , Gao, Y. , Gao, F. , Anderson, A. W. , Ding, Z. , & Gore, J. C. (2020). Functional engagement of white matter in resting‐state brain networks. Neuroimage, 220, 117096. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lund, T. E. , & Hanson, L. G. (2001). Physiological noise reduction in fMRI using vessel time‐series as covariates in a general linear model. Neuroimage, 13, 191. [Google Scholar]
- Lund, T. E. , Nørgaard, M. D. , Rostrup, E. , Rowe, J. B. , & Paulson, O. B. (2005). Motion or activity: Their role in intra‐ and inter‐subject variation in fMRI. Neuroimage, 26, 960–964. [DOI] [PubMed] [Google Scholar]
- Mayer, A. R. , Ling, J. M. , Dodd, A. B. , Shaff, N. A. , Wertz, C. J. , & Hanlon, F. M. (2019). A comparison of denoising pipelines in high temporal resolution task‐based functional magnetic resonance imaging data. Human Brain Mapping, 40, 3843–3859. [DOI] [PMC free article] [PubMed] [Google Scholar]
- McMurtry, C. M. , Chambers, C. T. , McGrath, P. J. , & Asp, E. (2010). When “don't worry” communicates fear: Children's perceptions of parental reassurance and distraction during a painful medical procedure. Pain, 150, 52–58. [DOI] [PubMed] [Google Scholar]
- Network, F. B. R. , Greve, D. N. , Brown, G. G. , Mueller, B. A. , Glover, G. , & Liu, T. T. (2013). A survey of the sources of noise in fMRI. Psychometrika, 78, 396–416. [DOI] [PubMed] [Google Scholar]
- Oakes, T. R. , Johnstone, T. , Walsh, K. S. O. , Greischar, L. L. , Alexander, A. L. , Fox, A. S. , & Davidson, R. J. (2005). Comparison of fMRI motion correction software tools. Neuroimage, 28, 529–543. [DOI] [PubMed] [Google Scholar]
- Perlaki, G. , Orsi, G. , Schwarcz, A. , Bodi, P. , Plozer, E. , Biczo, K. , Aradi, M. , Doczi, T. , Komoly, S. , Hejjel, L. , Kovacs, N. , & Janszky, J. (2015). Pain‐related autonomic response is modulated by the medial prefrontal cortex: An ECG–fMRI study in men. J Neurol Sci, 349, 202–208. [DOI] [PubMed] [Google Scholar]
- Poldrack, R. A. , Nichols, T. , & Mumford, J. (2009). Handbook of functional MRI data analysis. Cambridge University Press. [Google Scholar]
- Pruim, R. H. R. , Mennes, M. , Buitelaar, J. K. , & Beckmann, C. F. (2015). Evaluation of ICA‐AROMA and alternative strategies for motion artifact removal in resting state fMRI. NeuroImage, 112, 278–287. [DOI] [PubMed] [Google Scholar]
- Pruim, R. H. R. , 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. NeuroImage, 112, 267–277. [DOI] [PubMed] [Google Scholar]
- Salimi‐Khorshidi, G. , Douaud, G. , Beckmann, C. F. , Glasser, M. F. , Griffanti, L. , & Smith, S. M. (2014). Automatic denoising of functional MRI data: Combining independent component analysis and hierarchical fusion of classifiers. NeuroImage, 90, 449–468. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Schinkel, M. G. , Chambers, C. T. , Caes, L. , & Moon, E. C. (2017). A comparison of maternal versus paternal nonverbal behavior during child pain. Pain Practice, 17, 41–51. [DOI] [PubMed] [Google Scholar]
- Smith, S. M. , & Brady, J. M. (1997). SUSAN—A new approach to low level image processing. International Journal of Computer Vision, 23, 45–78. [Google Scholar]
- Smith, S. M. , Jenkinson, M. , Woolrich, M. W. , Beckmann, C. F. , Behrens, T. E. , Johansen‐Berg, H. , Bannister, P. R. , De Luca, M. , Drobnjak, I. , Flitney, D. E. , Niazy, R. K. , Saunders, J. , Vickers, J. , Zhang, Y. , De Stefano, N. , Brady, J. M. , & Matthews, P. M. (2004). Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage, 23, S208–S219. [DOI] [PubMed] [Google Scholar]
- Tousignant‐Laflamme, Y. , Rainville, P. , & Marchand, S. (2005). Establishing a link between heart rate and pain in healthy subjects: A gender effect. Journal of Pain, 6, 341–347. [DOI] [PubMed] [Google Scholar]
- Woolrich, M. W. , Behrens, T. , & Beckmann, C. F. (2004). Multilevel linear modelling for FMRI group analysis using Bayesian inference. NeuroImage, 21, 1732–1747. [DOI] [PubMed] [Google Scholar]
- Woolrich, M. W. , Jbabdi, S. , Patenaude, B. , Chappell, M. , Makni, S. , Behrens, T. , Beckmann, C. , Jenkinson, M. , & Smith, S. M. (2009). Bayesian analysis of neuroimaging data in FSL. NeuroImage, 45, S173–S186. [DOI] [PubMed] [Google Scholar]
- Woolrich, M. W. , Ripley, B. D. , Brady, M. , & Smith, S. M. (2001). Temporal autocorrelation in univariate linear modeling of FMRI data. NeuroImage, 14, 1370–1386. [DOI] [PubMed] [Google Scholar]
- Yoshikawa, A. , Masaoka, Y. , Yoshida, M. , Koiwa, N. , Honma, M. , Watanabe, K. , Kubota, S. , Natsuko, I. , Ida, M. , & Izumizaki, M. (2020). Heart rate and respiration affect the functional connectivity of default mode network in resting‐state functional magnetic resonance imaging. Frontiers in Neuroscience, 14, 631. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zeidan, F. , Emerson, N. M. , Farris, S. R. , Ray, J. N. , Jung, Y. , McHaffie, J. G. , & Coghill, R. C. (2015). Mindfulness meditation‐based pain relief employs different neural mechanisms than placebo and sham mindfulness meditation‐induced analgesia. Journal of Neuroscience, 35, 15307–15325. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zohsel, K. , Hohmeister, J. , Oelkers‐Ax, R. , Flor, H. , & Hermann, C. (2006). Quantitative sensory testing in children with migraine: Preliminary evidence for enhanced sensitivity to painful stimuli especially in girls. Pain, 123, 10–18. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
Activation maps included in this manuscript and the code used to preprocess the data with each noise‐reduction technique are available on Github (https://github.com/coghill-painlab/preproc_comparison) and archived in Zenodo (Hoeppli et al., 2023).