Correction of Physiological Artifacts in Multi‐Echo fMRI Data—Evaluation of Possible RETROICOR Implementations

ABSTRACT

The study evaluates the efficacy of RETROICOR (Retrospective Image Correction) in mitigating physiological artifacts within multi‐echo (ME) fMRI data. Two RETROICOR implementations were compared: applying corrections to individual echoes (RTC_ind) versus composite multi‐echo data (RTC_comp). Data from 50 healthy participants were collected using diverse acquisition parameters, including multiband acceleration factors and varying flip angles, on a Siemens Prisma 3T scanner. Key metrics such as temporal signal‐to‐noise ratio (tSNR), signal fluctuation sensitivity (SFS), and variance of residuals demonstrated improved data quality in both RETROICOR models, particularly in moderately accelerated runs (multiband factors 4 and 6) with lower flip angles (45°). Differences between RTC_ind and RTC_comp were minimal, suggesting both methods are viable for practical applications. While the highest acceleration (multiband factor 8) degraded data quality, RETROICOR's compatibility with faster acquisition sequences was confirmed. These findings underscore the importance of optimizing acquisition parameters and noise correction techniques for reliable fMRI investigations.

The study evaluates different implementations of the correction of physiological artifacts (RETROICOR) in multi‐echo fMRI data (before or after combination of the echoes) with several acquisition settings. Both implementations of RETROICOR can enhance signal quality, and the benefits are particularly notable in moderately accelerated acquisitions.

1. Introduction

Functional Magnetic Resonance Imaging (fMRI) stands as a cornerstone in the field of neuroscience, enabling non‐invasive investigations into the dynamics of brain activity with high spatial resolution. This modality has significantly advanced our understanding of human cognition, emotion, and behavior by mapping neural responses associated with a wide range of tasks and conditions (Biswal et al. 1995; Logothetis 2008). However, the utility of fMRI data is marred by the presence of various artifacts, which can introduce systematic distortions and confound the interpretation of results.

Artifacts in fMRI data encompass a diverse spectrum, including motion‐related, scanner‐related, and physiological sources (Power et al. 2012; Jezzard and Clare 1999). While each source poses distinct challenges, physiological artifacts, stemming from cardiac and respiratory processes, represent a persistent and pervasive issue. These artifacts manifest as high‐frequency fluctuations and spurious correlations within the BOLD (blood‐oxygen‐level‐dependent) signal, complicating the accurate assessment of true neural activity patterns (Birn et al. 2006). Consequently, addressing physiological noise is a critical concern in fMRI research.

Various methods have been developed to tackle the issue of physiological noise in fMRI data. These techniques aim to disentangle the true BOLD signal from the confounding influence of physiological fluctuations, enhancing the validity and sensitivity of neuroimaging studies (Birn 2012; Power et al. 2018). Prominent among these methods is RETROICOR (Retrospective Image Correction), which leverages concurrent physiological data, such as cardiac and respiratory signals, to model and remove the associated noise from fMRI time series data (Glover et al. 2000). RETROICOR has demonstrated efficacy in augmenting the signal‐to‐noise ratio (SNR) and facilitating the detection of subtle neural activations (Glover 2011).

While RETROICOR represents a pivotal advancement, it requires simultaneous acquisition of additional concurrent data, complicating the recording session. Multi‐echo fMRI involves the acquisition of multiple echoes for each image volume, yielding a set of images with varying contrast and sensitivity to physiological noise (Posse et al. 1999; Kundu et al. 2012). This approach has gained traction due to its potential to disentangle physiological noise from the neural signal, offering an elegant solution to the challenges posed by artifacts. It also offers several advantages, including increased SNR, improved sensitivity to BOLD signal, and the potential for quantifying multiple tissue properties such as T2* relaxation times. Additionally, multi‐echo fMRI provides opportunities for addressing physiological artifacts through data‐driven methods like Multi‐echo Independent Component Analysis (ME‐ICA) pioneered by Kundu et al. (2012) which differentiates BOLD and non‐BOLD signals in fMRI time series using multi‐echo EPI. While ME‐ICA has proven effective in isolating various sources of fMRI signal, we emphasize that its detailed exploration is beyond the scope of this work, as our primary focus lies in the investigation of acquisition parameter influences and the utility of RETROICOR in diverse fMRI sequences. The combination of multi‐echo acquisition and RETROICOR represents a compelling strategy to enhance the correction of physiological artifacts in fMRI data.

The influence of acquisition parameters on physiological artifacts has been a topic of extensive research. For instance, Gonzalez‐Castillo et al. (2011) investigated the effects of flip angle selection on physiological noise in BOLD fMRI. Their work highlighted the critical role of flip angles in modulating physiological artifacts, underscoring the need for careful optimization to enhance data quality. Similarly, Triantafyllou et al. (2011) conducted a thorough exploration of the dependence of physiological artifacts on TE (echo time), emphasizing the significance of TE selection in mitigating these confounds.

In parallel with parameter optimization, the field of fMRI has witnessed the emergence of advanced acquisition techniques designed to accelerate data collection. Among these, Simultaneous Multi‐Slice (SMS; Barth et al. 2016) imaging has emerged as a crucial innovation. SMS enables the simultaneous excitation and acquisition of multiple slices in the brain, leading to reduced scan times and potential improvements in temporal resolution. These advances in acquisition methodologies have motivated our specific focus on evaluating the influence of RETROICOR, a well‐established method for physiological noise correction, within a diverse set of multi‐echo fMRI sequences.

By leveraging the inherent differences in echo times, multi‐echo data can be processed using RETROICOR to remove physiological noise more effectively, thus improving the quality and reliability of fMRI results (Kundu et al. 2017).

In this article, we evaluate the synergy between multi‐echo fMRI acquisition and RETROICOR‐based noise correction. We present a comprehensive assessment of RETROICOR's efficacy in mitigating physiological artifacts across various multi‐echo fMRI sequences. Through a comprehensive analysis, we aim to shed light on the advantages and challenges of this innovative approach. By considering the influence of acquisition parameters, such as flip angles and TE, and recognizing the potential benefits of SMS, our study aims to provide valuable insights into optimizing fMRI data acquisition for more reliable and accurate neuroimaging investigations that can guide future research in optimizing fMRI data quality.

1.1. Aim

The primary objective of this study is to assess the efficacy of different methodologies for incorporating RETROICOR into the processing pipeline of multi‐echo (ME) data, specifically comparing the effects of applying RETROICOR to individual echo data versus applying RETROICOR to composite multi‐echo data.

Furthermore, the study aims to investigate the impact of incorporating RETROICOR into the multi‐echo data processing pipeline across various acquisition parameters, including parameters such as SMS acceleration and flip angle.

2. Methods

Data was obtained from a sample of 50 healthy individuals, consisting of 23 women and 27 men, aged 19 to 41 years. Subjects were screened for neurological, psychiatric, or mental disorder, showing no signs of them. Before the study, all participants underwent thorough familiarization with measurement procedures, received safety protocol briefings, and completed signed informed consent forms. Approval for the study protocol was granted by the Masaryk University Ethics Committee. Data collection utilized the Siemens Prisma 3T Mwhole‐body scanner with a 64‐channel head–neck coil, housed within the Laboratory of Multimodal and Functional Imaging at CEITEC Masaryk University.

The measurement protocol entailed seven MB (multiband) EPI (echo‐planar imaging) fMRI runs, each with varying levels of acceleration (multiband factor values), TR, and flip angles, as detailed in Table 1. Initially, high‐resolution anatomical images (T1‐MPRAGE) were obtained for precise brain area localization and to detect any potential abnormalities warranting subject exclusion. Subsequently, the protocol involved acquiring seven BOLD runs with different acquisition parameters. These fMRI protocols were based on the MB‐EPI BOLD sequence from the Centre for Magnetic Resonance Research at the University of Minnesota. Each run had a 6‐min acquisition time, though the number of images varied due to differing TR values. Both the field of view (FOV) (192 mm) and TE (17.00, 34.64, and 52.28 ms) remained constant across all fMRI runs. TE values were selected based on machine capabilities and recommendations from multi‐echo EPI review articles (Gonzalez‐Castillo et al. 2011; Gonzalez‐Castillo et al. 2016; Kundu et al. 2012; Kundu et al. 2017). The choice of TE values aimed for incremental increases while avoiding unnecessary TR prolongation. Flip angles were determined using Ernst angle calculations and modestly rounded down. Additionally, a flip angle from the next acceleration level was utilized (e.g., 45 in run 2 to match run 3). The order of measurements for runs 1–7 was counterbalanced to mitigate potential order effects.

TABLE 1.

Parameters of used runs.

fMRI run	Number of scans	Resolution (mm)	PAT factor	MB factor	TR (ms)	Flip angle (°)	Acq. matrix	No. of slices
Run1	120	3 × 3 × 3.5	2	1	3050	80	64 × 64	48
Run2	120	3 × 3 × 3.5	2	1	3050	45	64 × 64	48
Run3	450	3 × 3 × 3.5	2	4	800	45	64 × 64	48
Run4	450	3 × 3 × 3.5	2	4	800	20	64 × 64	48
Run5	600	3 × 3 × 3.5	2	6	600	45	64 × 64	48
Run6	600	3 × 3 × 3.5	2	6	600	20	64 × 64	48
Run7	900	3 × 3 × 3.5	2	8	400	20	64 × 64	48

Abbreviations: MB, multiband; TR, repetition time.

The fMRI task followed a block‐design structure with two alternating epochs. The initial epoch served as a baseline, during which participants were instructed to remain still, focus on a red cross displayed against a black background at the center of the screen, and refrain from intense thinking. This epoch lasted for 30.25 s. Subsequently, the second epoch commenced, wherein a series of red numbers (1, 2, 3, and 4) gradually appeared every second against a checkerboard background. Participants were instructed to press corresponding buttons simultaneously with the appearance of each number. This active measurement period lasted for 21.35 s, and the sequence of these two epochs was repeated eight times within a single fMRI run. We employed the same experimental task and acquisition parameters as described in our previous study (Kovářová et al. 2022). As part of the dataset overlaps between the two studies, we ensured consistency in task design and acquisition parameters to allow for comparative analysis.

2.1. Preprocessing

Acquired data underwent processing using SPM12 (rel. num. 6225) implemented in MATLAB (Figure 1). Each run's images were aligned using the SPM12 realign procedure, where images of the middle (i.e., second) echo were aligned to the first image of the middle echo, followed by applying the same translations and rotations to align the first and third echoes. Composite multi‐echo data were calculated using a contrast‐to‐noise ratio (CNR) weighted average. This means that the weights for each individual voxel and echo time were calculated as a temporal signal‐to‐noise ratio (tSNR) multiplied by the TE value and divided by the sum of these CNR values across the echoes in individual voxels. This CNR‐weighted approach was selected based on recommendations by Poser et al. 2006, who demonstrated its superior sensitivity compared to conventional processing techniques. Here is the equation for better clarity.

$$w{\left(\mathit{CNR}\right)}_n=\frac{{\mathit{SNR}}_n\times {\mathit{TE}}_n}{\sum _n{\mathit{SNR}}_n\times {\mathit{TE}}_n}$$

where n is the total number of echoes acquired at the different echo times TE _n and SNR is the temporal SNR calculated for each individual voxel and echo time.

FIGURE 1.

Scheme of data preprocessing.

Data integrity was assessed for dropouts and spatial abnormalities using the mask explorer tool (Gajdoš et al. 2016), while excessive movement was evaluated using the movement_info tool (https://www.nitrc.org/projects/movement_info), with control for the number of scans per subject exceeding FD (Framewise Displacement; Power et al. 2012) 0.5 and 1.5 mm. The commonly used 0.5 mm threshold served as a standard quality control benchmark, while the 1.5 mm threshold, which corresponds to half the voxel size in our data, was used to identify more pronounced motion. Our primary goal was to retain a complete and consistent dataset consisting of all seven runs from each of the 50 participants. Therefore, we tolerated occasional FD exceedances and chose not to apply motion scrubbing or censoring procedures. Instead, we examined the number and distribution of volumes exceeding each threshold across runs and participants.

For the 0.5 mm threshold, all participants but one (sub‐26, see details below) passed the criterion of having fewer than 20% of volumes exceeding this value in each run. Isolated exceedances of the 1.5 mm threshold were observed but occurred infrequently and without a consistent pattern across subjects. In run 1, one participant (sub‐26) exhibited more extensive motion. This subject exceeded FD 0.5 mm in 38% of volumes and FD 1.5 mm in 13 volumes. Other subjects in run1 with few volumes exceeding FD 1.5 mm are sub‐04 (3 volumes), sub‐12 (1 volume), sub‐37 (1 volume), sub‐39 (4 volumes). For run 2, the participants that exceeded 1.5 mm FD in two volumes are: sub‐26 (2 volumes), sub‐05 (3 volumes). In run 3, sub‐26 and sub‐04 exceeded the 1.5 mm threshold in two volumes each. In run 4, sub‐07 and sub‐12 had two and one volume(s) above 1.5 mm, respectively. In run 5, sub‐40 and sub‐43 each had one affected volume. In run 6, sub‐05, sub‐12, sub‐41, and sub‐42 each had approximately one volume above 1.5 mm. In run 7, only sub‐41 exceeded 1.5 mm in one volume. These instances were rare and dispersed, and no participant consistently exceeded the threshold across multiple runs.

To further mitigate the potential influence of motion on the data, we included an extended set of 24 motion parameters in the first‐level models. Given the minimal extent and lack of systematicity in high‐motion volumes, we did not exclude any participant from the dataset, and we did not apply additional censoring or scrubbing. This approach ensured data consistency while maintaining appropriate control over motion‐related variance.

The rest of the standard preprocessing involved co‐registration of anatomical images across all sequences and spatial normalization to the Montreal Neurological Institute template, followed by spatial smoothing with a 5 mm Full Width at Half Maximum Gaussian filter.

Two models were employed for utilizing RETROICOR within the preprocessing pipeline accompanied by a third implementation directly in SPM GLM analysis. In the first version, the algorithm was applied to individual echo realigned images and subsequently processed in a manner consistent with composite multi‐echo data. In contrast, the second model involved applying the RETROICOR algorithm to composite multi‐echo data. This allowed for the exploration of potential differences between the two approaches. Both implementations directly estimated and regressed out the physiological effects from fMRI data, and the procedure was applied slice by slice to allow precise modeling of the acquisition time of individual slices (i.e., the RETROICOR regressors modelling physiological artifacts were created for each individual slice).

Our RETROICOR method is implemented slice‐by‐slice, where we compute the delay individually by timing each slice. RETROICOR regressors (we used a total of 16 basis functions—8 for the breath, 8 for the heart; always sine and cosine with different multiples of the period between the R‐R and breath intervals, respectively) are created for each slice. This is a difference from the general linear model (GLM) implementation where the basic functions are identical for all slices.

Following this, spatial normalization and functional data smoothing were conducted. Single‐subject‐level statistical analysis employed a general linear model incorporating regressors for experimental stimulation responses, movement, and a constant term. To accommodate differing fMRI sampling rates, an advanced autoregression model in SPM (so called AR(fast); McDowell and Carmichael 2019) was utilized. Group analysis entailed evaluating the statistical models of individual subjects, including multi‐echo without RETROICOR and two multi‐echo RETROICOR models, using a random effect statistical model (one‐sample t‐test). To evaluate the possible interaction between the RETROICOR algorithm and GLM regressors, we additionally implemented the RETROICOR procedure directly to the GLM model as additional GLM nuisance regressors. In this case, the set of regressors was identical for all slices.

2.2. Global Metrics

We conducted a comparative analysis of multi‐echo, multi‐echo RETROICOR on individual echoes, and multi‐echo RETROICOR on composite multi‐echo data using several global metrics. These metrics included SNS (Signal to Noise Separation; Shirer et al. 2015) which evaluates the distance (using t‐values from two‐sample t‐test) between correlations of anatomical brain areas and random noise correlations. This metric assesses the ability to evaluate functional connectivity effectively.

Another global metric we used is SFS (Signal Fluctuation Sensitivity; DeDora et al. 2016), which is a metric used to assess the stability and reliability of fMRI signal measurements. It quantifies the sensitivity of the signal to fluctuations of non‐neural origin. SFS is particularly useful in evaluating the efficacy of preprocessing steps in reducing noise and improving the fidelity of the fMRI data.

We also calculated tSNR (temporal SNR; Krüger and Glover 2001) which was modified following the approach by Smith et al. (2013). Specifically, the tSNR value (the mean of individual time‐series divided by their standard deviation [SD]) was multiplied by the square root of the number of scans, denoted as tSNRn for clarity. This modification aims to balance the SNR drop in the images due to the large number of time points. The tSNRn was calculated for each voxel and subsequently averaged over defined regions, either across the entire gray matter or white matter for global metrics.

Other global metrics used for assessing data quality were the number of active voxels (p < 0.05 FWE corrected) and the average sum of residual mean squares from GLM activation models across gray matter. Additionally, we evaluated spatial smoothness of the data, a metric incorporated in the Human Connectome Project's quality control pipeline (Marcus et al. 2013).

The metrics described were analyzed using SPSS 27 through generalized mixed effect models. These models incorporated factors such as RETROICOR usage (without/RETROICOR on individual echoes/RETROICOR on composite multi‐echo data/RETROICOR GLM), slice accelerations (multiband factor), and flip angles (high vs. lower within the same acceleration level). This approach facilitated the assessment of how multi‐echo data, acceleration, and variations in flip angles impact the data, specifically comparing multi‐echo models with and without RETROICOR.

We conducted a parametric paired t‐test to evaluate individual differences between multi‐echo without RETROICOR and multi‐echo RETROICOR models, with significance assessed using a corrected p value of 0.05 and a more stringent 0.001 for multiple comparisons. The chosen significance level, adjusted with a Bonferroni correction, represents the likelihood that the null hypothesis—asserting that both tested options originate from the same distribution—is true. We assessed multiple global quality metrics, including temporal tSNRn and SFS in both white and gray matter. Additionally, we evaluated SNS, residual noise characteristics, spatial smoothness, and the number of active voxels as indicators of data quality and functional sensitivity.

2.3. Evaluation of Selected ROIs

To assess the statistical images and various parameters, we selected 10 Regions of Interest (ROIs) from the Automated Anatomical Labeling (AAL) atlas (Tzourio‐Mazoyer et al. 2002). These ROIs were chosen based on their relevance to the activation of our fMRI. Details of these selected regions are outlined in Table 2. We will show the results in ROI_1, Left precentral gyrus (L PCG).

TABLE 2.

Overview of AAL regions selected for ROI metrics.

ROI_ID	ROI description	AAL index
1	Left precentral gyrus (L PCG)	2
2	Left supplementary motor area (L SMA)	19
3	Right supplementary motor area (R SMA)	20
4	Left calcarine	43
5	Right calcarine	44
6	Left middle occipital gyrus (L MOG)	51
7	Right middle occipital gyrus (R MOG)	52
8	Left postcentral gyrus (L PCG)	57
9	Left pallidum	75
10	Right pallidum	76

Additionally, across all subjects, sequences, and ROIs, we computed β weights, residues, t‐values, percentage signal change (PSC; Luo and Nichols 2003), and task‐based SNR (variance of fitted BOLD response divided by variance of residuals). Within each ROI, t‐values were sorted by magnitude, and the 50 largest t‐values per ROI were selected for further analysis. All other parameters were subsequently extracted from the same positions as the 50 top t‐values to reveal characteristics more specific to the active part of the ROIs.

Furthermore, the average correlation of individual voxel time‐series with a representative signal within each ROI was calculated to assess homogeneity within the ROI (also referred to as repreCC). The representative signal for each ROI was determined as the mean of all individual time‐series within that ROI. A higher correlation value indicates greater uniformity of data within the area.

3. Results

As anticipated, the highest activation was seen in the visual cortex and in the motor areas (here mainly in the left hemisphere). The no RETROICOR SPM results showed higher activation in runs with slower TRs—runs 1 and 2 without multiband acceleration; and the level became equal in all models from the 3rd run on, except for run 7, as illustrated in Figure 2 (this is just illustrative picture of activations, see Figure 5 for quantitative results).

FIGURE 2.

Example of 1st level (single subject) activation maps comparing all three types of SPM results—no RTC, RTC_ind, RTC_comp—across all runs (multiband factor 1 to 8) showing axial and sagittal planes. The low threshold of t‐statistics “critical value”) was obtained using SPM12 and it is FWE corrected. The upper limit of the color scale is the maximum t‐statistics value in analyzed data. This subject was selected randomly for demonstration purpose.

FIGURE 5.

Comparison of global SPM metrics—number of active voxels, smoothness and variance of residuals in 4 models—without RETROICOR, RETROICOR on individual echoes, RETROICOR on composite multi‐echo data and RETROICOR in GLM. Solid line represents median, dotted line represents mean. GLM, general linear model; RTC, RETROICOR.

While first‐level activation maps (Figure 2) show a decrease of activation extent and decreased t‐values in the case of RETROICOR on individual echoes/RETROICOR on composite multi‐echo data compared to no RETROICOR, group‐level activation maps (Figure 3) provide more similar results. See Table S4 for a quantitative overview.

FIGURE 3.

Example of 2nd level activation maps comparing all three models—without RETROICOR and 2 RETROICOR models—across all runs (multiband factor 1 to 8) showing axial and sagittal planes. The low threshold of t‐statistics (“critical value”) was obtained using SPM12 and it is FWE corrected. The upper limit of the color scale is the maximum t‐statistics value in analyzed data.

3.1. Global Metrics

The comparison of all three models in mean tSNRn (temporal‐SNR), SNS (signal‐to‐noise separation) and SFS (Signal Fluctuation Sensitivity) is shown in Figure 4. Statistical evaluation of differences is in Table S5, and relevant percentual effects (differences among data with/without RETROICOR in preprocessing pipeline) are in Tables S1–S3.

FIGURE 4.

Comparison of global metrics—mean tSNRn in gray matter, SNS and mean SFS in gray matter in 3 models—without RETROICOR, RETROICOR on individual echoes and RETROICOR on composite multi‐echo data. Solid line represents median, dotted line represents mean. RTC, RETROICOR.

For mean tSNRn (in the upper part), we can see that the level of signal as well peaks at the 3rd run and then modestly decreases until run 7 where it drops strikingly with the highest multiband factor. Both multi‐echo RETROICOR models yield better results than multi‐echo without RETROICOR and these differences are statistically significant. The increase caused by RETROICOR is ranging from 2.5% for run 7 up to 17% for run 1. Furthermore, the differences between the two multi‐echo RETROICOR models are all statistically significant as well, but the practical effect is very small (less than 1%). However, the final run (7) is unsuitable for practical use due to the poor quality of the scanned data in terms of tSNRn and reflects the limits of acceleration when applied to physiological artifact correction.

The results for the SNS metric (in the middle part) show that the lower values in the RETROICOR models indicate a decrease in global correlations among brain regions. The trend is the same for all runs—the SNS values are higher for the multi‐echo data without RETROICOR than for both multi‐echo RETROICOR preprocessing implementations. The significant differences between the two RETROICOR models are only seen in the second run, suggesting that this metric is less sensitive to the choice of RETROICOR implementation but highly dependent on the multiband factor. The differences between each of the RETROICOR models and no RETROICOR are all statistically significant. The percentual change caused by any RETROICOR is about 4%–9%, and the difference between the two RETROICOR models is less than 0.5%.

SFS is a metric associated with enhanced sensitivity to both local and long‐range connectivity. In the lower part of Figure, we can see that both RETROICOR models exhibit higher SFS compared to the multi‐echo without RETROICOR model. The values are higher for both multi‐echo RETROICOR models than multi‐echo without RETROICOR model across all 7 runs. All these differences are statistically significant as well as the differences between the two multi‐echo RETROICOR models (except for the last run 7). The practical effects are in the same order as for tSNR metric.

The comparison of all four models (no RETROICOR, RETROICOR on individual echoes, RETROICOR on composite multi‐echo data, and RETROICOR in GLM) in the number of active voxels, smoothness and the variance of residuals is demonstrated in Figure 5. Statistical evaluation of differences is in Table S6.

The upper part of the figure represents results for the number of corrected voxels (FWE corrected p = 0.05) from individual first‐level models. Active voxels are generally lower in the RETROICOR models, reflecting the more stringent noise correction. The trend in the data tends to be the lowest for runs 1 and 2, then increases and decreases with the multiband factor towards run 7, which has almost as low values as the first two. Additionally, the ratio between multi‐echo without RETROICOR and multi‐echo RETROICOR models tends to be quite the same throughout all of the runs—the multi‐echo without RETROICOR model has a higher number of active voxels than both of the multi‐echo RETROICOR models. The differences between the two multi‐echo RETROICOR models are statistically significant in run 1, run 2, run 4, and run 5. The differences between no RETROICOR model and both RETROICOR on individual echoes and RETROICOR on composite multi‐echo data are significant for all the runs, except for run 7. The same applies to the two RETROICOR models versus RETROICOR in GLM.

Another metric, used to assess the data quality (Marcus et al. 2013), is smoothness. This is the average smoothness value which is calculated using SPM12. It is shown in the middle part of the figure. The differences between no RETROICOR model and the other three models are statistically significant in all runs. The differences between RETROICOR on individual echoes and RETROICOR on composite multi‐echo data are statistically significant in all runs, except for run 7. And RETROICOR in GLM versus the other two RETROICOR models showed significant differences in run 2, run 3, run 5, and run 6.

The lower part of Figure 5 demonstrates the comparison of variance of residuals for all four models in all runs. In runs 1 and 2, there are visible differences among all four models, which are all statistically significant. We can say that the differences for the four models across all runs are significant, except for the difference between RETROICOR on individual echoes and RETROICOR on composite multi‐echo data in runs 6 and 7.

Figure 6 is a visualization of the effects of some important factors (RETROICOR, multiband factor, flip angle) and interactions from the generalized mixed model computed by SPSS software. The results and significance of these factors are shown in Table 3 below.

FIGURE 6.

Overview of the effect of modeled factors in generalized mixed model statistics for global metrics—number of active voxels, smoothness, and variance of residuals. Estimated means (dots connected with lines) and standard deviations (whiskers) are presented for three types of comparisons. The first row demonstrates the effect of RETROICOR versus multiband factor. The second row demonstrates the effect of RETROICOR versus flip angle at two levels—high and low FA. The third row demonstrates the effect of multiband factor versus flip angle at two levels—high and low FA.

TABLE 3.

Results from generalized mixed effect models assessing the factors of RETROICOR usage (without/RTC_ind/RTC_comp), slice accelerations (multiband factor), and flip angles (higher vs. lower). The values in the table represent the significance (p value) of each individual factor or interaction between factors for each global metric from GLM‐based analysis.

Metric	Sig. of factor RTC	Sig. of factor MBF	Sig. of factor FA	Sig. of interaction RTC × MBF	Sig. of interaction RTC × FA	Sig. of interaction MBF × FA	Sig. RTC_ind × RTC_comp/RTC_GLM × noRTC
Active voxels	< 0.001	< 0.001	< 0.001	0.066	0.952	< 0.001	1.000/1.000
Smoothness	< 0.001	< 0.001	< 0.001	< 0.001	0.013	< 0.001	0.951/< 0.001
Residuals	0.202	< 0.001	< 0.001	0.999	1.000	< 0.001	1.000/0.274

Abbreviations: FA, flip angle; MBF, multiband factor; noRTC, data without RETROICOR; RTC_comp, RETROICOR on composite multi‐echo data; RTC_GLM, RETROICOR implemented in SPM GLM analysis; RTC_ind, RETROICOR on individual echoes.

We employed a generalized linear mixed model (GLMM) to investigate the statistical significance of factors influencing various metrics related to functional MRI (fMRI) data processing—active voxels, smoothness, and residuals. The results reveal significant effects of several factors on these metrics. First, the factor RETROICOR demonstrates significant impacts on active voxels and smoothness (p < 0.001), indicating its substantial influence on the spatial characteristics and signal quality of fMRI data. However, for residuals, the significance level for RETROICOR is marginally above the conventional threshold (p = 0.202), suggesting a weaker association with this metric. Second, factors MBF (Multiband Factor) and FA (Flip Angle) both exhibit highly significant effects on all three metrics (p < 0.001), highlighting the importance of slice accelerations and flip angles in determining data quality and processing outcomes. Additionally, interactions between RETROICOR and the multiband factor were observed (p = 0.066). However, interactions between RETROICOR and flip angle do not reach statistical significance (p > 0.05) suggesting limited interaction effects on the studied metrics. The last column says whether there is a significant difference between RETROICOR on individual echoes and RETROICOR on composite multi‐echo data or (after the dash) whether there is a difference between the results without RETROICOR and the results with RETROICOR implemented only within the GLM.

Figure 7 is a visualization of the effects of some important factors (RETROICOR, multiband factor, FA) and interactions from the generalized mixed model computed by SPSS software for global metrics. The results and significance of these factors are shown in Table 4 below.

FIGURE 7.

Overview of the effect of modeled factors in generalized mixed model statistics for global metrics—SNS, SFS, homogeneity (= repreCC), repreV (explained variability), and tSNRn for gray and white matter. Estimated means (dots connected with lines) and standard deviations (whiskers) are presented for three types of comparisons. The first row demonstrates the effect of RETROICOR versus multiband factor. The second row demonstrates the effect of RETROICOR versus flip angle at two levels—high and low FA. The third row demonstrates the effect of multiband factor versus flip angle at two levels—high and low flip angle.

TABLE 4.

Results from generalized mixed effect models assessing the factors of RETROICOR usage (without/RTC_ind/RTC_comp), slice accelerations (multiband factor), and flip angles (higher vs. lower). The values in the table represent the significance (p value) of each individual factor or interaction between factors for each global metric.

Metric	Sig. of factor RTC	Sig. of factor MBF	Sig. of factor FA	Sig. of interaction RTC × MBF	Sig. of interaction RTC × FA	Sig. of interaction MBF × FA	Sig. RTC_ind × RTC_comp
SNS	< 0.001	< 0.001	< 0.001	0.582	0.935	0.034	0.475
SFS	< 0.001	< 0.001	0.841	< 0.001	0.959	< 0.001	0.940
repreCC = homogeneity	< 0.001	< 0.001	< 0.001	0.627	0.768	< 0.001	0.876
repreV	< 0.001	< 0.001	< 0.001	0.072	0.791	< 0.001	0.843
tSNRn GM	< 0.001	< 0.001	< 0.001	< 0.001	0.272	< 0.001	0.765
tSNRn WM	< 0.001	< 0.001	< 0.001	< 0.001	0.024	< 0.001	0.866

Abbreviations: FA, flip angle; MBF, multiband factor; RTC, RETROICOR; SNS, signal‐to‐noise separation; tSNRn, temporal signal‐to‐noise ratio multiplied by number of scans.

The results of the generalized mixed effect models investigating the impact of RETROICOR usage, slice accelerations (multiband factor), and flip angles on various global metrics are presented in Table 4. Each cell of the table represents the significance (p value) of the corresponding factor or interaction between factors for a specific metric. Across all metrics, RETROICOR usage exhibited statistically significant effects, with p values consistently below 0.001. Similarly, the slice accelerations (multiband factor) also demonstrated significant effects on all metrics, with p values consistently below 0.001. However, the significance of flip angles (FA) varied across metrics. While flip angle showed significant effects on most metrics, there were instances where the p values exceeded the conventional threshold of 0.05, indicating non‐significant effects. Furthermore, the interactions between RETROICOR usage and slice accelerations (multiband factor) or flip angles (FA) were also examined. The interaction between RETROICOR usage and multiband factor showed significant effects on some metrics, particularly on SNS and repreV, with p values below 0.05. Conversely, the interaction between RETROICOR usage and flip angle demonstrated varying significance levels across metrics, with p values ranging from 0.024 to 0.959. The interaction between multiband factor and flip angle consistently exhibited significant effects across all metrics, underscoring the importance of considering these interactions in the analysis.

Generally, we did not observe many significant differences between RETROICOR on individual echoes and RETROICOR on composite echoes models. Some significant differences (based on paired t‐test) were observed in certain metrics and runs. But the practical significance was negligible (see Tables S1–S3 with percentual changes). We wanted to explore this in more depth because the RETROICOR application can affect the weights (it can directly change the data) so we created CNR‐weighted maps of all echoes to explore the effect of RETROICOR.

Figure 8 demonstrates the comparison between multi‐echo RETROICOR on individual echoes and no RETROICOR data models for CNR weights across the brain (i.e., CNR maps). We can see that value patterns are very similar with only subtle differences in the frontal lobe—for runs 1–3 in echo 1 and echo 3.

FIGURE 8.

Effect of RETROICOR on CNR weights for combination of echo time‐series. Maps of CNR weights for RETROICOR model (RTC_ind model, RETROICOR applied on individual echoes) and no RETROICOR in all runs and echoes.

3.2. ROI‐Based Evaluation

We further compared multi‐echo data, multi‐echo RETROICOR on individual echoes, and multi‐echo RETROICOR on composite multi‐echo data based on the evaluation of ROIs correlating with task‐related GLM outputs. This included the assessment of explained variability of RETROICOR and its influence on activation and the composite multi‐echo signal.

The ROI evaluating metrics were computed in all 10 selected ROIs for all 7 runs. To be more visual, the metrics are calculated for the top 50 voxels from each ROI. The results are demonstrated on ROI_1 (L PCG).

The comparison of all four models (no RETROICOR, RETROICOR on individual echoes, RETROICOR on composite multi‐echo data and RETROICOR in GLM) in mean t‐values in 50 most significant voxels, number of active voxels and SNR in 50 most significant voxels in ROI_1 is showed in Figure 9. Statistical evaluation of differences is in Table S8.

FIGURE 9.

Comparison of SPM metrics in ROI—mean t‐value in 50 most significant voxels, number of active voxels and SNR in 50 most significant voxels in 4 models—without RETROICOR, RETROICOR on individual echoes, RETROICOR on composite multi‐echo data and RETROICOR in GLM. Solid line represents median, dotted line represents mean. GLM, general linear model; RTC, RETROICOR.

The upper part of Figure 9 demonstrates the comparison of mean t‐values from the 50 most significant voxels around the peak activation in ROI_1. The first two runs (without multiband factor) yield significantly lower results than the rest. The differences between no RETROICOR and the RETROICOR on individual data and RETROICOR on composite multi‐echo data are all statistically significant as well as the differences between RETROICOR in GLM and the other two RETROICOR models. However, the differences between RETROICOR on individual echoes and RETROICOR on composite multi‐echo data were not significant in any run.

The number of active voxels in ROI_1 is displayed in the middle part. We can see that the lowest values are in run 1 and run 2, and then the number increases in run 3. The differences between no RETROICOR and the two RETROICOR models are all significant, except for run 7. Also, the differences between the two RETROICOR models and RETROICOR in GLM are significant, except for run 1. Finally, all the differences between RETROICOR on individual echoes and RETROICOR on composite multi‐echo data are not statistically significant.

Regional task SNR in the 50 most significant voxels is presented at the lower part of the figure. RETROICOR on individual echoes and on composite multi‐echo data provides decreased SNR compared to no RETROICOR and RETROICOR in GLM, and these changes are all statistically significant.

Figure S1 demonstrates the comparison of mean task effect in 50 most significant voxels, variance of residuals in 50 most significant voxels, and homogeneity of the region in ROI_1 in all four models (no RETROICOR, RETROICOR on individual echoes, RETROICOR on composite multi‐echo data and RETROICOR in GLM). Statistical evaluation of differences is in Tables S7 and S8.

The upper part of Figure S1 demonstrates the mean task effect (beta) in 50 most significant voxels. RETROICOR on individual echoes and on composite multi‐echo data provide decreased values compared to no RETROICOR and RETROICOR in GLM and these changes are all statistically significant, except for run 7.

The middle part of the figure demonstrates the comparison of the variance of residuals in 50 most significant voxels in ROI_1. The trend in data is increasing with lower TR and the version of each run with smaller flip angle always has a higher number of residuals. The last run 7 has the most residuals and is unsuitable for practical use because of the low data quality. There is only a minor (practically negligible) effect of RETROICOR on the variance of residuals. However, all the differences between RETROICOR in GLM and the other three models are statistically significant.

Homogeneity of the region is in the lower part of the figure. It is shown for no RETROICOR, RETROICOR on individual echoes, and RETROICOR on composite multi‐echo data. The trend in data seems to be decreasing with lower TR. The lowest results are, as expected, in run 7. Using RETROICOR caused a small decrease in homogeneity. The no RETROICOR model yielded modestly higher values than the two RETROICOR models in almost all runs, and the differences are statistically significant in runs 3–7. The differences between RETROICOR on individual echoes and RETROICOR on composite multi‐echo data are statistically significant for run 1, run 2, run 3, and run 7. Nevertheless, the practical difference is negligible (see Table S1).

While the previous figures represented ROI‐based characteristics for one selected ROI (left precentral gyrus—L PCG), Figures 10 and 11 accompanied by Tables 5 and 6 provide an overview of the global effects of individual factors across all ten ROIs (the region was modelled as one of the factors to see how the characteristics are regionally dependent). As for global GLM‐based metrics, here we added RETROICOR implemented directly in the GLM model in addition to RETROICOR versions applied directly within the preprocessing pipeline (Figure 10 and Table 5). There were no significant differences between RETROICOR on individual echoes and RETROICOR on composite multi‐echo data implementation. GLM RETROICOR provided activation characteristics similar to the data without RETROICOR. Therefore, RETROICOR during preprocessing removed part of the variability related to the task regressor.

FIGURE 10.

Overview of the effect of modeled factors in generalized mixed model statistics for local metrics (calculated in selected 10 ROIs)—number of active voxels, residuals in 50 top voxels, mean t‐value in top 50 voxels, and SNR in top 50 voxels. Estimated means (dots connected with lines) and standard deviations (whiskers) are presented for three types of comparisons. The first row demonstrates the effect of RETROICOR versus multiband factor. The second row demonstrates the effect of RETROICOR versus flip angle at two levels—high and low flip angle. The third row demonstrates the effect of multiband factor versus flip angle at two levels—high and low flip angle. The fourth row demonstrates the effect of RETROICOR versus FOI in all 10 ROIs. The fifth row demonstrates the effect of multiband factor versus ROI in all 10 ROIs.

FIGURE 11.

Overview of the effect of modeled factors in generalized mixed model statistics for local metrics (calculated in selected 10 ROIs)—SFS, homogeneity, repreV (explained variability) and tSNRn Estimated means (dots connected with lines) and standard deviations (whiskers) are presented for three types of comparisons. The first row demonstrates the effect RETROICOR versus multiband factor. The second row demonstrates the effect of RETROICOR versus flip angle at two levels—high and low flip angle. The third row demonstrates the effect of multiband factor versus flip angle at two levels—high and low flip angle. The fourth row demonstrates the effect of RETROICOR versus FOI in all 10 ROIs. The fifth row demonstrates the effect of multiband factor versus ROI in all 10 ROIs.

TABLE 5.

Results from generalized mixed effect models assessing the factors of RETROICOR usage (without/RTC_ind/RTC_comp/GLM RTC), slice accelerations (multiband factor), flip angles (higher vs. lower), and individual ROI (ROI_ID 1 to 10). The values in the table represent the significance (p value) of each individual factor or interaction between factors for each ROI metric.

Sig. of factor	Factor RTC	Factor MBF	Factor FA	Factor ROI	Interaction RTC × MBF	Interaction RTC × FA	Interaction RTC × ROI	Interaction MBF × FA	Interaction MBF × ROI	Interaction FA × ROI	Sig. RTC_ind × RTC_comp/RTC_GLM × noRTC
Metric
Active voxels	< 0.001	< 0.001	< 0.001	< 0.001	0.014	0.976	0.342	< 0.001	< 0.001	< 0.001	0.957/0.762
Residualsiin 50 voxels	0.008	< 0.001	< 0.001	< 0.001	0.613	0.991	0.998	< 0.001	< 0.001	< 0.001	0.968/0.077
Mean t‐value in 50 voxels	< 0.001	< 0.001	< 0.001	< 0.001	< 0.001	0.636	< 0.001	< 0.001	< 0.001	< 0.001	0.963/0.378
SNR in 50 voxels	< 0.001	< 0.001	< 0.001	< 0.001	< 0.001	0.629	< 0.001	0.023	< 0.001	< 0.001	0.998/< 0.001

Abbreviations: FA, flip angle; MBF, multiband factor; noRTC, data without RETROICOR; PSC, percent signal change of BOLD response estimated with general linear model; RTC, RETROICOR; RTC_comp, RETROICOR on composite multi‐echo data; RTC_GLM, RETROICOR implemented in SPM GLM analysis; SNR, signal‐to‐noise ratio based on the BOLD response estimated with general linear model versus standard deviation of residual: RTC_ind, RETROICOR on individual echoes.

TABLE 6.

Results from generalized mixed effect models assessing the factors of RETROICOR usage (without/RTC_ind/RTC_comp), slice accelerations (multiband factor), flip angles (higher vs. lower), and individual ROI (ROI_ID 1 to 10). The values in the table represent the significance (p value) of each individual factor or interaction between factors for each ROI metric.

Sig. of factor	Factor RTC	Factor MBF	Factor FA	Factor ROI	Interaction RTC × MBF	Interaction RTC × FA	Interaction RTC × ROI	Interaction MBF × FA	Interaction MBF × ROI	Interaction FA × ROI	Sig. RTC_ind × RTC_comp
Metric
tSNRn	< 0.001	< 0.001	< 0.001	< 0.001	< 0.001	< 0.001	0.982	< 0.001	< 0.001	< 0.001	0.584
repreVar	0.399	< 0.001	< 0.001	< 0.001	0.635	0.985	0.002	< 0.001	< 0.001	< 0.001	0.906
SFS2m	< 0.001	< 0.001	< 0.001	< 0.001	< 0.001	0.986	< 0.001	< 0.001	< 0.001	< 0.001	0.298
Homogeneity (repreCC)	< 0.001	< 0.001	< 0.001	< 0.001	< 0.001	0.150	0.410	< 0.001	< 0.001	< 0.001	0.845

Abbreviations: FA, flip angle; MBF, multiband factor; PSC, percent signal change of BOLD response estimated with general linear model; RTC, RETROICOR; SNR, signal‐to‐noise ratio based on the BOLD response estimated with general linear model versus standard deviation of residual.

4. Discussion

This study evaluates the efficacy of RETROICOR (Retrospective Image Correction; Glover et al. 2000) in mitigating physiological artifacts within multi‐echo multiband fMRI data. The comparison was conducted between two RETROICOR models integrated into the preprocessing pipeline and one model without RETROICOR, across different acquisition parameters, including SMS acceleration and variations in flip angle. Moreover, to enable comparison of RETROICOR's impact on activation‐based metrics, we additionally implemented RETROICOR within the GLM (as nuisance regressors) and used these data for some statistical comparisons. The RETROICOR method operates on each voxel separately and thus does not introduce artificial coupling of noise corrections across spatial regions. The key findings suggest that applying RETROICOR to fMRI multi‐echo data (either individual echoes or composite fMRI data) can help correct physiological noise without substantial signal loss, and in some cases, even leads to slight improvements, particularly in accelerated fMRI sequences.

RETROICOR was tested across seven fMRI runs with varying SMS acceleration factors (multiband factors) and flip angles. Our dataset consists of 50 subjects, making it relatively robust, and is freely available on Open Neuro for verification (Multi‐echo simultaneous multislice fMRI dataset: Effect of acquisition parameters on fMRI data. OpenNeuro, doi: 10.18112/openneuro.ds004499.v1.0.3). While our previous study (Kovářová et al. 2022) focused on the effects of multiband and multi‐echo acquisition on BOLD sensitivity and noise characteristics, this study extends that work by specifically examining the impact of RETROICOR for physiological noise correction. By isolating the influence of RETROICOR under different multiband settings, we aim to disentangle the contributions of acquisition parameters and post‐processing techniques to overall data quality and statistical outcomes.

The results demonstrate that RETROICOR improves several global signal quality metrics, including mean tSNRn (temporal signal‐to‐noise ratio), SFS (signal fluctuation sensitivity), and residual variance, with statistically significant gains compared to results without RETROICOR. Previous studies (e.g., Gonzalez‐Castillo et al. 2011) have shown that using flip angles lower than the Ernst angle can reduce physiological noise. Since higher flip angles increase physiological noise contributions, RETROICOR is expected to have a greater corrective effect under these conditions. Our findings support this, highlighting the importance of physiological noise correction in protocols using higher flip angles. The improvements from RETROICOR were most pronounced in runs with moderate acceleration (multiband factors 4 and 6), suggesting that the correction is effective for faster sequences without compromising spatial or temporal resolution. However, when acceleration was increased further (multiband factor 8), mean tSNRn dropped significantly, indicating that very high acceleration diminishes data quality, even with RETROICOR applied.

The observed decrease in SNS likely reflects the removal of signal components by RETROICOR that previously contributed to artificially elevated correlations between brain regions. As a result, correlations among regions (e.g., those defined by the AAL atlas) dropped slightly.

Regarding flip angles, our results suggest that higher flip angles for the same TR (i.e., those closer to Ernst angle, such as 45° for run3) yielded better performance in terms of signal sensitivity and reduction in residual physiological noise—especially when combined with moderate multiband factors. Lower flip angles for the same TR (e.g., 20° in run 4), which were used in some accelerated runs, produced higher levels of residual noise. This highlights the importance of carefully optimizing acquisition parameters in conjunction with RETROICOR application.

Two RETROICOR implementations were incorporated into the preprocessing pipeline: one that applied the correction to individual echoes, and another that applied the correction to composite multi‐echo data. Even though some statistical tests on the observed characteristics revealed significant differences between the two implementations, the practical effects (i.e., percentual differences) were negligible. Both implementations (in‐house solutions programmed in MATLAB) used the exact timing of individual slices, which was motivated by the effort to estimate the precise shape of the physiological response in individual brain regions. However, this should not affect the ability to effectively clean the data (which is our aim in this work), because the sets of periodic basis functions (sine and cosine) should be able to estimate the periodic response in the data even when the constant time shift is present. This is also the case for our third implementation of RETROICOR, where nuisance regressors are included within the SPM design matrix. Cleaning the fMRI data during preprocessing is necessary for the analysis of resting state data or for analyses of connectivity that do not use the SPM/GLM framework. Implementing denoising by regressing out artificial signal models can affect the degrees of freedom of the residuals—that is, the cleaned data (Kruggel et al. 2002). In the case of our two RETROICOR implementations during the early stage of preprocessing, the SPM routine used to estimate the number of degrees of freedom lacks information about previous denoising procedures and relies only on the actual design matrix and the estimation of temporal non‐sphericity in residuals. Despite the fact that SPM calculates the number of effective degrees of freedom, we did not observe any difference between the models with or without RETROICOR applied during preprocessing. This is probably due to the whitening process implemented in SPM12. While differences are observable in the non‐sphericity estimate (matrix SPM.xVi.V) and in the whitening matrix (SPM.xX.W), these differences diminish in the whitened and high‐pass filtered matrix SPM.xX.V, which serves as the input for the Satterthwaite approximation used to estimate the number of effective degrees of freedom in SPM12. This effect is a potential limitation of all GLM‐based denoising procedures. However, our results did not indicate an overestimation of p values related to activation analysis. On the contrary, we found weaker (i.e., lower t‐values) and a smaller extent of first‐level activations for data preprocessed with the RETROICOR procedure.

We assessed the efficacy of the two RETROICOR implementations by comparing them to each other as well as to the no‐RETROICOR model. Table S1 provides the mean and standard deviation (SD) for the observed metrics, comparing RETROICOR applied to individual echoes and RETROICOR applied to composite multi‐echo data across all seven runs. The mean tSNRn results show only modest improvements for RETROICOR on individual echoes compared to RETROICOR on composite multi‐echo data, particularly in early runs (1–3). SNS and SFS show similarly minor differences, but slightly favor RETROICOR on composite multi‐echo data. Therefore, the findings do not consistently support a difference for one RETROICOR implementation over the other.

Table S2 presents differences between no‐RETROICOR data and RETROICOR on individual echoes. The key findings are: tSNRn shows significant improvement in RETROICOR on individual echoes compared to no RETROICOR, especially in early runs; SFS also improves, suggesting modestly enhanced noise correction, particularly in early and middle runs. Residuals in ROIs also show better results for RETROICOR on individual echoes in early runs. However, in all other metrics, the no‐RETROICOR model outperformed the RETROICOR implementation on individual echoes.

Table S3 compares no RETROICOR and RETROICOR on composite multi‐echo data. The results are consistent with those in Table 2: RETROICOR on composite multi‐echo data improves mean tSNRn and SFS only.

Our hypothesis was that cleaning individual echoes first would be more robust. However, this was not supported by our data. In most cases, the results of both RETROICOR models are nearly equivalent, with only minor differences in either direction. Based on this, we conclude that both RETROICOR models appear equally suitable for practical use. However, both approaches share one limitation: they may partially explain variance attributable to the task. We demonstrated this as an undesirable side effect of implementing RETROICOR regressors directly within the GLM, where there is no longer a noticeable reduction in the statistical strength or extent of activations. Overall, RETROICOR on individual echoes yields only marginally better results than RETROICOR on composite multi‐echo data, particularly for metrics like mean tSNRn and SFS.

Interestingly, although both RETROICOR models improved global quality metrics compared to the no‐RETROICOR models, the differences between the two RETROICOR implementations were mostly non‐significant. This suggests that from a practical standpoint, both approaches are viable. Applying RETROICOR to individual echoes may offer potential benefits in scenarios involving more sophisticated echo combination procedures or advanced processing strategies (e.g., T2* fitting), although this lies outside the scope of our study.

For some metrics, we observed changes in the RETROICOR‐adjusted data that ran counter to expectations—for instance, fewer active voxels or reduced t‐values in GLM results. This can be partly explained by the removal of variance related to the task regressor when RETROICOR is implemented during preprocessing. While RETROICOR regressors do not typically correlate strongly with the stimulation time course, some portion of task‐related variability may be captured by chance when the ratio of regressors to the number of scans is high (Bright and Murphy 2015). The distribution of correlation coefficients between RETROICOR regressors and the task regressor, along with the coefficient of determination, is shown in Figures S2 and S3. This effect is especially apparent in single‐subject data and is more pronounced in data with no or lower acceleration. In group‐level activation analyses, differences between RETROICOR‐cleaned and uncleaned data become less apparent. Thus, RETROICOR should not significantly reduce the sensitivity of activation studies at the group level. Beyond the already noted reduction in SNS, we also observed a decrease in the homogeneity of individual regions (expressed as the average correlation between individual voxel time series and a representative signal—repreCC). This may be attributed to the broad impact of physiological artifacts across the brain, which can introduce artificial correlations. After artifact suppression, these areas may appear more variable.

In this study, we focused exclusively on processing multi‐echo data and did not acquire single‐echo data. However, previous studies by Bartoň et al. (2015, 2019) provide evidence that the single‐echo acquisition may yield similar results.

There are several other limitations to consider. Our dataset consists solely of task‐based fMRI data, so the findings may not fully generalize to resting‐state fMRI. Furthermore, we did not evaluate non‐physiological noise correction methods such as ME‐ICA (Kundu et al. 2017), which has shown robust performance in separating noise from neural signals without relying on physiological recordings. These aspects were beyond the scope of this study, but future research should explore integrating RETROICOR with such methods to further enhance noise correction in multi‐echo datasets.

5. Conclusion

In this study, we evaluated the efficacy of applying RETROICOR (Retrospective Image Correction) to multi‐echo fMRI data for mitigating physiological artifacts, comparing two implementations: correction on individual echoes and on composite multi‐echo data. Our findings demonstrate that both implementations can enhance signal quality, as evidenced by improvements in global metrics such as temporal SNR (tSNRn), Signal Fluctuation Sensitivity (SFS), and residual variance. These benefits were particularly evident in moderately accelerated runs (multiband factors 4 and 6) and with higher flip angles, consistent with prior research showing that higher flip angles are associated with increased physiological noise (Gonzalez‐Castillo et al. 2011). This supports the effectiveness of RETROICOR, especially when physiological fluctuations are more prominent due to faster acquisition sequences.

The minimal differences between the two RETROICOR implementations suggest that either approach is viable for practical applications. These results highlight the importance of optimizing acquisition parameters—such as acceleration factors and flip angles—in conjunction with noise correction strategies to ensure high‐quality and reliable fMRI data.

Ethics Statement

The study was approved by the local ethics committee of Masaryk University and all subjects signed the informed consent before entering the study.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1. Supporting Information.

Kovářová, A. , and Mikl M.. 2025. “Correction of Physiological Artifacts in Multi‐Echo fMRI Data—Evaluation of Possible RETROICOR Implementations.” Human Brain Mapping 46, no. 9: e70264. 10.1002/hbm.70264.

Data Availability Statement

The data used in this study is available at the openneuro.org website with dataset ID ds004499 (https://doi.org/10.18112/openneuro.ds004499.v1.0.3).