Impact of thermal and physiological denoising on laminar functional connectivity

Abstract

The characterization of the direction of information flow between cortical areas via layer functional connectivity (FC) is of interest for many neuroscientific applications. Most laminar fMRI acquisitions are in the thermal-noise dominated regime due to spatial resolution requirements. However, laminar FC estimates using gradient-echo BOLD might be particularly biased by physiological sources of noise which are amplified by the draining vein effect. In this work, we aimed at assessing whether thermal and physiological denoising can be helpful in reducing biases in laminar functional connectivity studies. We tested NORDIC, RETROICOR, and aCompCor on a dataset acquired at 7T with a resolution of 0.8 × 0.8 × 1.5 mm³ and evaluated the following metrics on a laminar basis after each denoising step: temporal standard deviation (tSD), coefficient of variation (CoV), fluctuation amplitude (FA) and seed-based functional connectivity strength of the primary motor cortex (M1) with premotor (PM) and somatosensory (S1) regions. We found that NORDIC had the largest impact on the metrics considered, mostly in deeper laminae. The application of physiological denoising, especially aCompCor, following NORDIC had the largest effect on upper laminae, in line with the fact that they are more affected by physiological noise than the deeper laminae. The application of NORDIC and aCompCor resulted in laminar functional connectivity results pointing to a stronger connectivity of upper M1 laminae to S1 than lower M1 laminae to S1, and to a reduction in the upper-layer bias in the connectivity to PM. Therefore, both thermal and physiological denoising represent important steps to increase sensitivity and reduce the vascular bias of laminar fMRI data.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-37599-4.

Introduction

Layer fMRI

Advances in magnetic resonance imaging (MRI) acquisition schemes and hardware have made it possible to achieve unprecedented levels of resolution while keeping sufficient contrast to noise ratio for functional MRI (fMRI) applications. Over the past decade, this has enabled the investigation of the brain functional architecture at a laminar scale.

The human neocortex is organized into six cytoarchitectonic layers which can be considered, within each region, as distinct functional units. The thickness of each cortical layer varies from region to region but, in general, voxel sizes of about 0.8 mm have proven capable of discerning signal changes coming from different cortical depths. It has been experimentally observed that three cortical depths (laminae) are generally sufficient to distinguish input from output activity^1,2. In fact, the interest in imaging at laminar resolution is not only related to the increase in spatial specificity, but also to the additional information on directional information flow and hierarchical relationships between cortical areas, because feedforward and feedback projections selectively activate different layers^3–7. Despite the vast use of fMRI, layer connectivity in humans is not still fully characterized and relies largely on animal models⁸.

The increase in spatial resolution alone, however, is often insufficient to study layers: for instance, the gradient-recalled echo blood-oxygenation-level dependent (GE-BOLD) contrast, being mostly sensitive to large venous vessels, detects higher signal changes in upper layers and at the outer surface of gray matter (GM) due to the presence of pial veins, a phenomenon commonly referred to as “pial vein contamination”. GE-BOLD signal changes are, thus, weighted towards regions of higher venous blood density, which do not necessarily reflect higher neuronal activity are inherently blurred by the deoxyhemoglobin (dHb) drainage via intracortical veins.

Two main classes of approach have been developed to increase functional specificity: (i) reducing the macrovascular bias of GE-BOLD, (ii) using alternative functional contrasts that are more specific to the microvascular compartment. In the first case, strategies include phase regression^9,10, normalization with venous cerebral blood volume¹¹, normalization with resting-state fluctuation amplitude¹², signal division¹³, masking^14–16, deconvolution with a physiological point spread function¹⁷, signal subtraction¹⁸, or temporal decomposition¹⁹. In the second approach, alternative contrasts applied in laminar fMRI studies include arterial blood contrast (ABC)^20,21, vascular-space occupancy (VASO)^22,23, spin-echo BOLD (SE-BOLD)^11,24–27, perfusion^28–30, cerebral metabolic rate of oxygen consumption (CMRO₂) via calibration models^31–33, integrated VASO and perfusion contrast (VAPER)³⁴.

Noise in high-resolution datasets

Another important confounding factor in functional data is noise, which is partly system-related (thermal and instrumental) and partly subject-related (thermal, physiological, and motion)³⁵. Several denoising strategies have been proposed over the years³⁶ to address the problem of noise in fMRI at both high-resolution and lower resolutions. What changes with voxel volume (and tissue type) is the physiological to thermal noise ratio (σ_P/σ₀)³⁷. For GM at submillimeter voxel sizes at 7T, thermal noise dominates over physiological noise³⁸. In the regime of thermal-noise dominated voxels, tSNR scales linearly with image SNR³⁵, therefore fMRI with small voxel volumes benefits particularly from higher field strengths.

The shift to the thermal-noise dominated regime in fMRI has led to the development of algorithms devoted to its mitigation, such as noise reduction with distribution corrected principal component analysis (NORDIC PCA)^39,40, which suppresses zero-mean Gaussian distributed noise. Despite being recently introduced, NORDIC has already been applied to a relatively large number of fMRI studies^9,41–43.

Even if thermal noise is the main contributor to the total noise in the voxel, a few more considerations should be made. First, GM is not homogeneous, due to varying venous blood density and vessel sizes^44,45. Since the GE-BOLD contrast is heavily weighted towards large venous vasculature, any fluctuation in dHb concentration (neuronally-driven or not) will result in a stronger signal change where vessels are larger (i.e., in upper layers)⁴⁶. As a result, the critical voxel volume for which thermal noise dominates over physiological noise is not constant across cortical depth. Additionally, layer fMRI studies typically average the signal over many voxels, which could in turn increase the relative contribution of physiological noise.

Physiological noise reduction can be achieved through various techniques, some of which make use of physiological information obtained from external recording devices, such as RETROKCOR⁴⁷, RETROICOR⁴⁸, and correction for arterial CO₂ fluctuations⁴⁹. Other methods, such as ICA-AROMA⁵⁰, PESTICA⁵¹, and CompCor⁵² rely on data-driven approaches. These techniques have been tested and compared for GE-BOLD at standard-resolutions^53–55, while they have only been sparsely applied in high-resolution studies. A recent study compares several preprocessing steps for reducing noise in laminar fMRI data⁵⁶, although it does not include thermal denoising.

Aim of this study

This study aimed to assess whether the specificity of laminar functional connectivity estimates in the primary motor cortex using GE-BOLD at 7T can be improved by the application of thermal and physiological denoising. Despite the numerous drawbacks of GE-BOLD, it offers several advantages over other techniques, such as high sensitivity and speed (TR can be kept below 1 s), robustness, ease of acquisition and processing. For these reasons, GE-BOLD remains a popular choice of acquisition for research in laminar fMRI, and additional preprocessing strategies to improve specificity are hence beneficial to the community.

The denoising algorithms included in this study were NORDIC, RETROICOR, aCompCor, motion correction and motion regression. While NORDIC has been recently developed and applied to several high-resolution datasets, RETROICOR and aCompCor were proposed more than 15 years ago and were intended to denoise fMRI datasets acquired at lower resolutions. These two physiological denoising algorithms were chosen because they use physiological noise regressors obtained from different sources (external recording and fMRI data, respectively), are compatible with fMRI acquisitions that do not cover the whole brain, as in the present study, and have been tested and compared to other physiological denoising strategies^53,57. Therefore, the current study also aimed to validating these techniques for denoising high-resolution fMRI data.

The effect of denoising was evaluated using the following metrics calculated at a laminar level: the coefficient of variation (CoV), the fractional amplitude of resting-state fluctuations, and the seed-based functional connectivity strength. The analysis was performed on resting-state datasets capturing the primary motor cortex, somatosensory cortex and premotor area of the left hemisphere.

Materials and methods

Experimental setup and MRI acquisition

Ten healthy participants (4 males; age [mean ± std] = 23.4 ± 0.7 y) took part in the study after giving informed written consent. All experimental procedures were approved by the Ethics Committee at the Medical Faculty of the University of Leipzig (Reg.-No. 273–14-25,082,014). All methods were performed in accordance with the relevant regulations and guidelines⁵⁸. Participants were asked to refrain from coffee and alcohol consumption for at least two hours prior to the experiment. All scans were performed on a MAGNETOM classic 7T scanner (Siemens Healthineers, Erlangen, Germany) running Siemens software version syngo VB17 with SC72 (70 mT/m, 200 T/m/s) gradient coil, and using a single-channel transmit, 32-channel receive head coil (Nova Medical, Wilmington, MA, USA).

A gradient-echo BOLD sequence with a three-dimensional (3D) echo-planar readout⁵⁹ was used for the study. The nominal in-plane resolution was 0.8 × 0.8 mm², and the slice thickness was 1.5 mm, with a matrix size of 200 × 200 × 16. The volume repetition time (TR_vol) was 994 ms, and the echo time (TE) was 30 ms. Other acquisition parameters were: flip angle 16°, in-plane partial Fourier factor 6/8, in-plane GRAPPA acceleration factor 4, bandwidth 782 Hz/Px, echo spacing 1.44 ms, RF duration 2.56 ms, excitation pulse bandwidth time product 25. Missing k-space lines resulting from the use of partial Fourier were reconstructed using the vendor’s default zero-filling reconstruction. Physiological recordings of respiration and cardiac pulsation were made at a sampling rate of 50 Hz with the scanner physiological monitoring system using a vendor-provided respiratory bellow and pulse plethysmograph placed on the right index finger. Concurrently, a second set of physiological traces were sampled at a rate of 1 kHz with a respiratory pad and pulse plethysmograph placed on the left index finger (BIOPAC MP150, BIOPAC Inc., Goleta, CA, USA).

The acquisition was targeted at imaging the hand-knob region of the primary motor cortex. To reduce through-plane partial voluming artefacts caused by non-isotropic voxel dimensions, the imaging slab was oriented perpendicularly to the GM ribbon of the hand-knob in the left hemisphere, adapting to the local anatomy of each participant. The orientation was selected by using previously acquired MP2RAGE⁶⁰ images of each participant, as previously described²², resulting in subject-dependent slab orientations. The MP2RAGE sequence was acquired in a separate session with an isotropic resolution of 0.8 mm covering the whole brain.

The paradigm consisted of three consecutive resting-state runs. Each run lasted approximately 9 min (540 volumes), leading to a total resting-state experiment duration of 27 min. Each run was preceded by Larmor-frequency adjustment and update in order to account for the effect of low-frequency drift on the signal, which causes significant fluctuations in resting-state fMRI⁶¹. During the resting-state scan, participants were asked to stare at a fixation cross and not to fall asleep.

Preprocessing

Head motion was assessed by computing the Euclidean norm (Enorm)⁶² on the concatenated realigned dataset (details about the realignment are described below). Since the concatenation included three runs, the two pairs of transition volumes between them were discarded from the Enorm calculation, as they would otherwise introduce artefactual values due to discontinuities. After visual inspection of the realignment quality, fMRI series were retained only if the 99th percentile of the Enorm was below 0.4 mm⁶². This resulted in three subjects being excluded from the analysis due to excessive head motion. Therefore, the analysis was conducted on the remaining seven participants.

For each subject, the high-resolution 3D T1-weighted image acquired with MP2RAGE was segmented into probabilistic tissue maps of GM, WM, and CSF using CAT12⁶³. The reference volume, defined as the 265th volume of the first fMRI scan within the NORDIC-denoised (rN) dataset, was spatially coregistered to the T1-weighted image using a non-linear (diffeomorphic) spatial transformation computed with ANTs⁶⁴. Subsequently, the inverse of this transformation was applied to the tissue probability maps to coregister them to the rN space. The resulting coregistered maps were binarized with a threshold of 0.9 to reduce the probability of voxel misclassification, thereby generating GM, WM, and CSF masks in the reference volume space.

Denoising

Two denoising pipelines were tested, which differed in the way they dealt with physiological noise removal. They consisted of the following steps applied in the order reported:

1. RETROICOR pipeline (RMrN): NORDIC – realign – motion parameter and RETROICOR regression.

2. aCompCor pipeline (CMrN): NORDIC – realign – motion parameter and aCompCor regression.

Each of the applied steps is detailed below.

NORDIC (N)

NORDIC is a patch-based PCA technique which removes Gaussian distributed components, while preserving the structured (physiological) noise. In this study, it was applied to magnitude images using the NIFTI_NORDIC.m function (available at https://github.com/SteenMoeller/NORDIC_Raw), without including phase data or noise scans, which were not available for this dataset. All other parameters were left at their default values. To confirm that NORDIC could be applied on this dataset (magnitude-only) without losses in accuracy due to deviations from a Gaussian noise distribution, image SNR and skewness maps were calculated using the command LN_SKEW of the LayNii package⁶⁵.

Realign (r)

The volumes of all three runs were concatenated and realigned to the 265th volume of the first fMRI run, corresponding to about the half of the first fMRI run time length, using the MCFLIRT algorithm from FSL, with interpolation method spline and cost function normcorr (normalized correlation)⁶⁶. After realignment, the dataset were split back into three separate runs as in the original acquisition.

Regression for motion parameters (M)

Motion regression was performed using a general linear model (GLM), with twelve motion regressors: 6 realignment parameters and their 6 temporal derivatives. Motion regressors were included in the same GLM alongside 14 RETROICOR regressors in one case, and 10 aCompCor regressors in the other case (described below). Additionally, a constant and a linear term were added to each GLM.

RETROICOR (R)

Physiological traces were recorded with two different systems simultaneously. For building RETROICOR regressors, only the cardiac and the respiratory trace with the best quality were used, i.e., where peak identification done using the PhysIO Toolbox⁶⁷ had the fewest misses. For the cardiac trace, BIOPAC data was used for 4 subjects and Siemens data for 3 subjects, while, for respiration, BIOPAC was used for one subject and Siemens for 6 subjects. Fourteen RETROICOR regressors were computed using the PhysIO Toolbox⁶⁷: 4 cardiac, 4 respiratory, 4 multiplicative, the cardiac response (obtained by convolving the heart rate variability (HRV) with the cardiac response function (CRF)⁶⁷) and the respiratory response (obtained by convolving the respiratory volume per time (RVT) with the respiratory response function (RRF)⁶⁷). In total, 28 regressors were included in the GLM for the RETROICOR pipeline. Since the fMRI sequence used in this study had a 3D readout, no slice-specific RETROICOR regressors were built, but a single regressor for the entire volume was created.

aCompCor (C)

In this study, anatomical CompCor was used with 10 components regressed out: 5 taken from a WM region of interest (ROI) and 5 from a CSF ROI. To minimize partial volume effects or region-dependent biases, the CSF voxels and WM voxels used for component extraction were selected as follows: a probability tissue mask threshold of 0.9 was applied and, for CSF, the thresholded mask was manually edited to remove voxels still affected by partial volume effects or exhibiting poor coregistration accuracy to the functional space (typically at the outer brain borders). Manual editing was not necessary for the WM ROI. To maximize the efficiency of the GLM design, datasets were orthogonalized with respect to the motion parameter regressors before component extraction. The GLM for the aCompCor pipeline included a total of 24 regressors.

ROI definition

Layering in M1. Layering was performed directly in EPI space (Fig. 1A), exploiting the coregistered tissue masks obtained with CAT12 (Fig. 1B). The volumes were spatially resampled to a 0.2 mm in-plane resolution using the function 3dresample of AFNI⁶⁸ with nearest-neighbor interpolation. The hand-knob region of the primary motor cortex (M1) was identified based on its anatomical features, and the GM-WM and GM-CSF borders were manually delineated in the slice with the best through-plane orientation (Fig. 1C). A total of 20 equidistant laminae was generated using the LAYNII software⁶⁵.

Fig. 1.

Segmentation and layering for one representative participant. (A) Mean image of the realigned and NORDIC-denoised timeseries. (B) WM, GM, and CSF masks, used for ROI definition. (C) Equidistant laminae defined within the hand-knob region of the primary motor cortex (M1), ROI within the primary somatosensory cortex (S1, yellow), within the premotor cortex (PM, green) and within a control region (blue).

Premotor, somatosensory and control regions. For each subject, three ROIs were defined in order to evaluate their functional connectivity with different M1 depth bins. The premotor (PM) and somatosensory (S1) regions were identified within the GM mask based on anatomical features, similarly as in⁶⁹. A control region in GM was chosen close to the interhemispheric sulcus. The three ROIs had similar spatial extents within and across subjects, and were located in the same or adjacent slice as the M1 ROI (Fig. 1C).

Quality metrics

To assess the effectiveness of the applied denoising steps, a few quality metrics were evaluated on a depth-dependent basis. The metrics evaluated included temporal standard deviation (tSD) and coefficient of variation (CoV) for an indication in the time domain of the amplitude of the fluctuations; fluctuation amplitude (FA) for different frequency bands, extracted from the spectra; and the functional connectivity (FC) strength for a quantification of the connectivity between regions as calculated from the seed-based correlation approach.

Laminar standard deviation (tSD) and coefficient of variation (CoV)

A voxelwise temporal standard deviation and coefficient of variation, defined as the tSD divided by the mean, i.e., equal to 1/tSNR, was calculated after each denoising step. The resulting values were averaged over each lamina.

Laminar fluctuation amplitude (FA)

The calculation of the fluctuation amplitude was adapted from the fractional amplitude of low frequency fluctuation (fALFF)⁷⁰, with minor modifications: while fALFF considers only the low-frequency band (0.01–0.08 Hz), here, in order to observe the effect of denoising over the whole spectrum, fFA was calculated for five frequency bands, extending up to the maximum detectable frequency of 0.5 Hz: 0.01–0.1 Hz (band 1), 0.1–0.2 Hz (band 2), 0.2–0.3 Hz (band 3), 0.3–0.4 Hz (band 4), 0.4–0.5 Hz (band 5). The square root of the power at each frequency was calculated and the mean was taken for each band.

Laminar functional connectivity (FC) strength

Two seeds were selected for the functional connectivity analysis: one from lower laminae (‘lower M1’) which included the 6 M1 laminae closest to WM excluding the bottom-most lamina, and another from upper laminae (‘upper M1’) which included the 6 M1 laminae closest to CSF excluding the upper-most lamina. Therefore, from the initial 20 laminae in M1, the two extreme laminae (at the border with CSF and WM) were discarded, along with the central 6 laminae, to reduce partial volume effects. A voxelwise Pearson correlation analysis was performed separately for the lower M1 seed and the upper M1 seed. The z-Fisher transformed Pearson’s r values were then averaged within S1, PM and control regions to provide a measure of connectivity strength with these ROIs.

Additionally, for each denoising step, the timecourse of lower M1 and upper M1 entered a GLM (using FSL-FEAT) and were used to build a contrast of lower M1 against upper M1. The z-stat of the contrast was averaged within the PM ROI and the S1 ROI.

Statistical analysis

To test the effect of denoising on the different metrics, one-way ANOVA was performed across the denoising steps. For the laminar profiles of tSD, CoV and FA, the statistical analysis was performed for the two outermost laminae: the one closest to WM and the one closest to CSF. For the connectivity strength, it was conducted for the different regions. In all cases, a post-hoc multiple comparison using t-tests was performed (Bonferroni corrected). The differences in z-stats for the lower M1 > upper M1 contrast were tested by means of a t-test.

Results

The average image SNR (± std) of the raw timeseries was found to be 12.5 ± 1.5 in the brain, ranging from 9.3 to 14.5 across subjects and runs, and the skewness maps (Figure S5) show a uniform distribution, indicating that noise can be treated as Gaussian.

Throughout the figures and the text, the denoising steps applied are indicated by abbreviations (as described in the Denoising section). A sequence of letters represents all the denoising methods applied, with the order of application to be read from right to left: for example, RMrN indicates that NORDIC was applied first, followed by realignment, and then regression for motion parameters and RETROICOR.

Figure 2 shows the laminar tSD, CoV, and mean signal within the hand-knob region of M1. For tSD and CoV metrics, the undenoised, realigned timeseries (r) can be seen as the curve with the highest values over all cortical depths. All denoising steps had a tendency to reduce the laminar tSD and CoV across the cortex, but to varying degrees: NORDIC (rN) had a relatively stronger effect in deeper laminae, while motion regression (MrN) and physiological denoising with RETROICOR (RMrN) and aCompCor (CMrN) acted primarily in the upper laminae. For the upper layer (i.e., the one closest to CSF) in tSD, a significant reduction was found between r and all the other denoising steps (p <  < 0.001), and also between rN and MrN (p = 0.035), RMrN (p < 0.01), CMrN (p < 0.001). For the lower layer (i.e., the one closest to WM), a significant difference was found between r and all the other denoising steps (p <  < 0.001). For the upper layer in CoV, a significant reduction was found between r and all the other denoising steps (p < 0.001), and rN was found to be significantly different from CMrN (p = 0.025). The lower layer in r was significantly different from the lower layer in the other steps (p <  < 0.001). Maps of variance reduction and of absolute difference in tSD between the realigned dataset and the CMrN and RMrN datasets are shown in the Supplementary material (Figure S1) for two representative participants. The mean signal (Fig. 2C) exhibits a laminar profile with higher values near the cortical surface, which remained constant across denoising steps, as evidenced by the overlapping profiles.

Fig. 2.

Laminar tSD, CoV, and mean signal. Temporal standard deviation (A), coefficient of variation (B), and temporal mean (C) across the cortical depth after each denoising step. Each curve was calculated following the application of one or more denoising algorithms. Shades indicate the standard deviation across participants. r: motion-correction; N: NORDIC; M: regression of motion parameters; R: RETROICOR; C: aCompCor.

The laminar FA within the hand-knob region of M1 (Fig. 3), divided into five 0.1 Hz-wide bands, shows that NORDIC had the strongest effect for all frequency bands, reducing values all over the cortical depth. Regression for motion parameters and physiological denoising had a further tendency to reduce FA, only visible in band 1 (0.01–0.1 Hz) and almost exclusively in upper laminae. For bands 2, 3, 4 and 5, regression for motion and physiological denoising had a negligible effect. The shape of the profiles is slightly different for different frequency bands; in particular, the curves for band 1 appear steeper, i.e., higher values are reached towards the cortical surface, than the curves for higher frequency bands. For band 1 (0.01–0.1 Hz), a significant difference was found between the lower layer of r and all the other denoising steps (p < 0.001), while for the upper layer significant differences were found between r and MrN (p = 0.02), RMrN (p < 0.01), and CMrN (p < 0.01). For the four bands above 0.1 Hz, significant differences were found between the lower layer of r and all the other denoising steps (p < 0.001) and between the upper layer of r and all the other denoising steps (p < 0.01).

Fig. 3.

Laminar fluctuation amplitudes. Fluctuation amplitudes for the five considered frequency bands following each denoising step. Shaded areas represent the standard deviation across participants. r: motion-correction; N: NORDIC; M: regression of motion parameters; R: RETROICOR; C: aCompCor.

Laminar connectivity, presented as z-Fisher-transformed correlation coefficients, is shown in Fig. 4 for different regions and denoising steps. For the realigned, undenoised timeseries (r), both lower and upper M1 laminae showed some degree of correlation with all target areas (S1, PM and control). The correlation was always higher for upper M1 than for lower M1 (significantly higher only for the r timeseries, p < 0.05). NORDIC significantly increased the connectivity strength for S1 and PM (p < 0.05), slightly more for the lower M1 than for the upper M1. The subsequent physiological denoising steps tended to reduce connectivity strength to various degrees: regression for motion parameters (MrN) and RETROICOR (RMrN) only slightly (not significant), while aCompCor (CMrN) reduced it significantly (p < 0.05), almost restoring it to the original values (no significant differences). Comparing the correlations of the undenoised timeseries with the CMrN timeseries, although no significant differences were found, changes were more marked in the connections with PM and control regions. While the connectivity profile with S1 remained almost identical between the undenoised and CMrN timeseries, the connectivity of upper M1 with PM was reduced, making it similar in strength to the connectivity of lower M1 with PM. For the control region, the connectivity strength was closer to zero.

Fig. 4.

Laminar functional connectivity. Functional connectivity strength averaged across all participants expressed as z-Fisher transformed correlation values between upper or lower layers and the three ROIs (S1, PM, control). r: motion-correction; N: NORDIC; M: regression of motion parameters; R: RETROICOR; C: aCompCor.

The z-stats of the GLM contrast of lower M1 > upper M1 are shown in Fig. 5 for each denoising step. Each of the values for lower to S1, lower to PM and lower to control did not significantly change with added denoising steps. The difference between lower to S1 and lower to PM was found to be significant only for the CMrN step (p = 0.02).

Fig. 5.

Z-stats of the contrast lower M1 > upper M1 after each denoising step. r: motion-correction; N: NORDIC; M: regression of motion parameters; R: RETROICOR; C: aCompCor.

Discussion

NORDIC

In this study, thermal denoising was achieved using NORDIC on magnitude-only timeseries. As expected, thermal denoising had a significant impact on this dataset, improving tSD, CoV (Fig. 2) and FA (Fig. 3) much more than the other denoising steps, and increasing the correlation strength between different areas (Fig. 4). Although the current analysis could not prove that the amount of variance removed from the data was solely of thermal origin, the depth-dependent information suggests that data processed with NORDIC correspond to what would be expected following thermal noise filtering: all depths and all frequency bands were affected (Figs. 2 and 3) and connectivity values increased irrespective of the brain area considered (Fig. 4). Indeed, thermal noise should be equally present in all layers and ideally at all frequencies. However, the signal from deeper laminae was relatively more degraded by thermal noise than the signal from upper laminae, and this was reflected by a slightly stronger change in all quality metrics (Figs. 2–5) in the lower portion of the cortex following NORDIC. In a previous analysis, we compared the effect of NORDIC on laminar tSD when applied either before or after physiological denoising and found that it had a stronger and more homogeneous impact across laminae when applied after physiological denoising⁷¹. This different relative impact of thermal denoising —expected to be uniform across the brain— among layers is likely due to the non-uniform distribution of physiological noise across layers. In fact, the efficacy of NORDIC depends on the separability between thermal noise from signal (including physiological noise) and elevated physiological noise levels in upper layers may reduce denoising performance^46,72. The effect of NORDIC across different frequency bands additionally showed that the relative reduction in FA was lower for the 0.01–0.1 Hz band compared to higher bands in the upper portion of the cortex (Fig. 3). Frequencies below 0.1 Hz are expected to have a strong physiological noise component (Figures S2 and S3), which could again explain the reduced thermal denoising performance.

In this work, because of data limitations, NORDIC was applied directly to NIfTIs, without including phase data nor noise scans. Different denoising efficacy could have been achieved by adding more data or changing NORDIC parameters; this is indeed a topic of a recent study⁷³.

Regression of motion parameters

The regression of the realignment parameters tended to improve data quality for all metrics considered, although a significant effect was found only for the upper layer in tSD. It should be noted that the datasets used for the current analysis were not strongly affected by motion (Figure S4), thus this step might provide larger improvements for datasets more affected by motion. Although more realignment parameters could have been added to the model, we chose to limit the current model to include 12 regressors (6 realignment parameters and 6 derivatives) after observing that the additional inclusion of 12 squared terms did observably impact the results (data not shown). Moreover, the use of anisotropic voxel sizes introduces a variability in the precision of motion parameters estimation across image axes. Specifically, the higher in-plane relative to through-plane resolution in this study implies a higher estimation precision for the in-plane motion compared to the through-plane motion. Therefore, isotropic voxel dimensions would be preferable for a uniform motion estimation precision across all directions.

RETROICOR

Power spectra of RETROICOR cardiac and respiratory regressors (Figure S2) showed high power at low frequencies (below 0.15 Hz), and in the 0.2–0.3 Hz band, respectively. However, among the denoising steps applied, RETROICOR had the smallest effect on the metrics considered, in particular at frequencies above 0.1 Hz (Fig. 3). This might be related to the atypical resolution, TR and acquisition scheme of the dataset considered in this study. The majority of previous applications of RETROICOR were done on fMRI datasets acquired with a TR of 2–3 s, thus containing aliased cardiac fluctuations. In a study with a TR = 300 ms, the authors report an increase of tSNR in GM and WM of 1–2% after the application of RETROICOR⁷⁴.

The application of RETROICOR to 3D sequences was previously demonstrated^75,76. Compared to the 2D case, in 3D acquisitions each point is acquired over a longer period of time (several seconds vs. milliseconds for 3D and 2D sequences, respectively). The application of RETROICOR in the 3D case relies on the assumption that the cardiac and respiratory phases corresponding to the center of k-space are assigned to the entire volume, as most of the contrast comes from k-space center^36,76.

Concerning the resolution, in a 3T study at an isotropic voxel size of 1.5 mm, Lutti et al. demonstrated that RETROICOR significantly increased the sensitivity of 3D acquisitions⁷⁷. In the context of laminar fMRI, RETROICOR has been applied in a few BOLD studies^9,78 but there is no exhaustive report on its efficacy in such datasets; in a laminar study of hippocampus, aCompCor was applied and the authors reported indications of superior performance with respect to RETROICOR⁷⁹.

aCompCor

Physiological denoising with aCompCor proved to be effective in improving tSD, CoV, FA and functional connectivity. For tSD and CoV, aCompCor had the largest effect in upper layers, in accordance with the observed strong depth-dependent features of physiological noise⁴⁶.

The dataset used in this study covered a limited FOV with an in-plane extension of 16 × 16 cm², a thickness of 2.4 cm, and variable orientation based on each subject’s anatomical features. Therefore, the selection of the ROIs for aCompCor components was restricted by the portion of the brain available: in particular, it was not possible to use ventricles for the extraction of CSF components, which were then selected from other areas and might have increased the risk of partial volume contamination.

In terms of frequencies, physiological noise seemed to be an issue only for frequencies below 0.1 Hz, as RETROICOR and aCompCor did not have a noticeable effect at higher frequencies (Fig. 3). Given the TR of about 1 s used in this study and the typical heart rate of 50–80 beats/min, a substantial amount of power in the 0.01–0.1 Hz band of the aCompCor regressor spectra (Figure S3) might be due to aliased cardiac frequencies, especially for the components taken from the CSF, that are more affected by cardiac pulsatility than WM. Regressor power spectra, however, showed peaks also at higher frequencies but of lower amplitude (Figure S3). In fact, physiological noise is expected to be present at frequencies above 0.1 Hz because of aliasing and faster physiological processes (including respiration). The fact that the physiological denoising did not have an effect in this study might be, thus, related to the fact that those frequencies were almost entirely dominated by thermal noise, and the amplitude of physiological processes was not detectable. This might be explained by the inherently lower amplitude of BOLD responses following faster inputs than slower inputs⁸⁰, combined with the low tSNR of high resolution acquisitions.

Layer connectivity

In this study, laminar functional connectivity was evaluated using a seed-based approach by computing the correlation between deeper and upper layers with three regions of interest: S1, PM and control. The contrast of the lower M1 seed versus the upper M1 seed was further evaluated in the three targeted ROIs. Upper M1 layers should preferentially reflect input activity, while deeper M1 layers should reflect output activity⁸¹.

Of particular interest is the variation in FC strength profiles after denoising (Fig. 4): when comparing the original values (r) with those fully preprocessed with aCompCor (CMrN), the connectivity of M1 to S1 did not change much, both in terms of strength and profile shape (i.e., higher values were found for upper M1). This observation is in accordance with the fact that upper M1 layers are more connected to S1 than lower M1 layers, reflecting input activity⁶⁹. The connectivity of M1 to PM, however, although comparable in strength, qualitatively changed in shape, with upper M1 connectivity being relatively more suppressed and reaching values similar to lower M1 connectivity. The connectivity strength of lower M1 to PM was not higher than that of upper M1 to PM, but the z-stat of the contrast of lower M1 > upper M1 was significantly higher in PM compared to S1 in the CMrN data. This result might indicate that the denoising pipeline including NORDIC and aCompCor was particularly effective in improving the signal quality in lower layers.

Even with a careful denoising, the tendency of a stronger connectivity of lower M1 to PM compared to upper M1 to PM may be harder to see in BOLD data compared to VASO data⁶⁹ because the intrinsic large vessel bias and draining effect of the GE-BOLD signal could still be present after physiological denoising; therefore, the connectivity with upper layers might have been inflated by those effects. Moreover, while it is plausible that premotor cortex shows higher correlation to deeper layers, input to upper layers is still present⁸².

Bandpass filtering

In this study, volumes were acquired with a relatively high temporal resolution of about 1 s which allowed for the study of frequencies up to 0.5 Hz, and thus the observation of the effect of denoising at frequencies above 0.1 Hz. Often the TR_vol in fMRI acquisition is longer, and data are regularly bandpass filtered to remove certain frequencies or to isolate the frequency band believed to contain neuronally-driven oscillations, typically 0.01–0.1 Hz. The choice of this band is partly related to the fact that hemodynamic response function (HRF) timings would not allow for the resolution of frequencies higher than 0.15–0.25 Hz, according to early studies^83,84. More recent studies using different fMRI acquisition strategies have shown, however, that the HRF can vary substantially in timing and shape across brain regions, thus deviating from canonicity^85–87. It has furthermore been shown that spontaneous neuronal oscillations of frequency above 0.1 Hz can be detected^88,89. At high resolutions, restricting the analysis to the low frequency band could hence impose an additional upper-layer bias to GE-BOLD data, since the venous vasculature might exhibit slower dynamics due to the longer time-to-peak of venous vessels^80,90,91. In fact, our results indicate that low frequencies have higher amplitudes in upper cortical layers, as can be seen from the steeper curves in band 1 (Fig. 3) compared to the shallower curves in higher frequency bands (Fig. 3, bands 2–5), reflecting a more homogeneous spectral power distribution across cortical depths.

Possible improvements

In this work, we focused on common denoising steps that are routinely applied to fMRI data at lower resolution. Nevertheless, including more cleaning steps or using different denoising strategies in the preprocessing might improve detectability³⁶. For example, RETROICOR was applied here in a separate pipeline than aCompCor, to evaluate the effect of two different physiological denoising strategies. A recent study indicates that combining both methods could further improve data quality⁵⁷. Another possibility for circumventing some sources of noise in the frequency of interest would have been increasing the temporal resolution in order to resolve the cardiac peak⁹², which would come at the cost of spatial resolution and signal strength. Thermal and physiological denoising performances could also be better assessed by including more inferior brain regions than those considered in the current study, such as the insular cortex and subcortical areas, that are generally more affected by physiological noise due to the proximity to CSF and large vessels.

The functional connectivity analysis performed in this study only considered three target regions: S1, PM and control, and two depth bins in M1. The changes in FC strength after denoising were in line with what would be expected from the actual connectivity between those regions, but still do not represent a full picture of the effect of denoising and connectivity patterns. For example, Chen and colleagues showed that a whole-band linear nuisance regression can introduce spurious network structures in the high-frequency band (above 0.4 Hz)⁹³. Therefore, having a larger FOV, possibly including the whole brain, and including more cortical depths in M1 would allow for a more accurate connectivity assessment and could help identify patterns of feedforward and feedback activity.

Moreover, the quality of physiological recordings is essential for accurately characterizing physiological processes in fMRI timecourses using for example RETROICOR. Respiration is commonly recorded using bellows or pressure pads, as in this study. However, these devices can produce unstable recordings with substantial noise, compromising the estimation of respiratory regressors. Nasal cannulae offer a viable alternative, providing cleaner respiratory signals with minimal participant discomfort. When coupled with a gas analyzer for end-tidal CO₂ measurement, cannula-based systems provide additional information that can explain additional physiological variability⁴⁹. Pulse photoplethysmographs (PPG) for cardiac recording generally provide satisfactory signal quality but capture only the pulse waveform. Pulse oximeters return both pulse timing and oxygen saturation (SpO₂) information, that can be an additional source of variability in fMRI timecourses. However, both PPG and pulse oximetry measurements are susceptible to motion artifacts and confounding from contact pressure exerted on the participant’s finger, potentially introducing biases in the recorded signals. Accurate signal processing and physiological response function modeling are furthermore essential for incorporating respiratory and cardiac recordings into fMRI analyses³⁵.

Finally, this study focused on a resting-state analysis. The same approach could be tested on a task-based fMRI dataset, in which the signal modulation induced by the task could be used for further quality assessment⁷³.

Conclusion

This work examined the effects of common denoising techniques on a high-resolution resting-state fMRI dataset acquired with GE-BOLD at 7T. The effectiveness of the denoising strategies was evaluated on a depth-dependent basis in the hand-knob region of the primary motor cortex using various quality metrics.

Our findings suggest that the combination of thermal and physiological denoising improved data quality. In particular, thermal denoising was especially beneficial for deeper layers, while physiological denoising acted mostly on upper layers, in line with the fact that the ratio between these noise components changes with cortical depth. The inclusion of aCompCor, in particular, was able to suppress the spurious functional connectivity with the control region and reduce connectivity values of upper M1 layers with PM. Without accounting for noise or for the large venous bias, connectivity values could have been falsely interpreted as being related to stronger connections with upper layers.

Author contributions

Conceptualization: MG, GG, IM, HEM, DGN, FG; Data curation: MG, GG, IM; Formal analysis: MG, GG, FG; Funding acquisition: HEM, DGN, FG; Investigation: MG, GG, IM; Methodology: MG, DS, LK, LH, HEM, DGN, FG; Project administration: HEM, DGN, FG; Resources: LH, HEM, DGN, FG; Software: BAP, HEM; Supervision: BAP, LH, HEM, DGN, FG; Visualization: MG, GG, DS, LK; Writing – original draft: MG, GG; Writing – review & editing: MG, GG, DS, IM, LK, BAP, LH, HEM, DGN, FG.

Data availability

Data and code will be made available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.