rt-me-fMRI: A task and resting state dataset for real-time, multi-echo fMRI methods development and validation

A multi-echo fMRI dataset (N=28 healthy participants) with four task-based and two resting state runs was collected, curated and made available to the community. Its main purpose is to advance the development of methods for real-time multi-echo functional magnetic resonance imaging (rt-me-fMRI) analysis with applications in neurofeedback, real-time quality control, and adaptive paradigms, although the variety of experimental task paradigms supports a multitude of use cases. Tasks include finger tapping, emotional face and shape matching, imagined finger tapping and imagined emotion processing. This work provides a detailed description of the full dataset; methods to collect, prepare, standardize and preprocess it; quality control measures; and data validation measures. A web-based application is provided as a supplementary tool with which to interactively explore, visualize and understand the data and its derivative measures: https://rt-me-fmri.herokuapp.com/. The dataset itself can be accessed via a data use agreement on DataverseNL at https://dataverse.nl/dataverse/rt-me-fmri. Supporting information and code for reproducibility can be accessed at https://github.com/jsheunis/rt-me-fMRI.


Background and summary
Real-time functional magnetic resonance imaging (fMRI) is a brain imaging method where 3 functional brain signals are acquired, processed, and used during an ongoing scanning session. 4 Applications include real-time data quality control (Dosenbach et al., 2017), adaptive experimental 5 paradigms (Hellrung et al., 2015), and neurofeedback (Sitaram et al., 2017). Neurofeedback is a 6 cognitive training method where the real-time feedback signal is presented back to the participant 7 to allow self-regulation of their blood oxygen level-dependent (BOLD) signal, prompting 8 researchers to investigate it as an intervention for patients with neurological or psychiatric 9 conditions. Work by Ros et al. (2020) and Haugg et al. (2020) show an absence of standardisation 10 in experimental design and outcome reporting restricts the synthesis of evidence to determine the 11 efficacy of fMRI neurofeedback. Further, it remains a major challenge to delineate the sources of 12 variance in the brain and in neurofeedback signals and their eventual effects on neurofeedback 13 training outcomes. Similar challenges exist for separating BOLD and non-BOLD variations and 14 their influences on data quality, and subsequently on all real-time fMRI applications. 15 16 In recent work (Heunis et al., 2020a) we investigated the available acquisition and processing 17 methods for improving real-time fMRI signal quality, and identified an absence of methodological 18 denoising studies and a need for community-driven quality control standards. Here, we aim to 19 advance this process by curating a multi-echo fMRI dataset (rt-me-fMRI). It builds on known 20 benefits of multi-echo fMRI for increasing BOLD sensitivity both in resting state and task fMRI 21 2016), but real-time multi-echo processing methods remain underexplored. By releasing the rt-25 me-fMRI dataset, we aim to facilitate a community effort to advance the development of methods 26 and standards in this domain. 27 28 The rt-me-fMRI dataset includes multi-echo resting state and task-based fMRI data from 28 29 healthy participants. Fig.1 provides an overview, including the task types: finger tapping, emotion 30 processing, imagined finger tapping, and imagined emotion. Several factors influenced the 31 experimental and acquisition protocols: Multi-echo fMRI: To facilitate the development of real-time multi-echo methods, all functional 34 acquisitions have multiple echoes. The first resting state run allows calculation of quantitative 35 multi-echo parameters such as baseline T2* or S0 maps, which can in turn be used for echo 36 combination during subsequent runs.  6 7 Task and resting state: The motor cortex, amygdala, and visual system were selected as 8 representative regions based on frequency of studies in fMRI and neurofeedback literature 9 (Thibault et al., 2018), and tasks were selected to elicit appropriate BOLD responses. The 10 fingerTappingImagined and emotionProcessing tasks respectively allow investigations into 11 mental imagery and visual shape/face processing. Since these structures are located at distinct 12 anatomical regions that experience different levels of noise (e.g. the amygdala suffers from more 13 severe image dropout and physiological noise; Boubela et al., 2015), this allows investigation of 14 spatially distinct effects of real-time denoising. Resting state scans allow comparison of the effects 15 of processing steps in the absence and presence of a task.

17
Template data: In real-time fMRI applications, anatomical and functional scans are typically 18 acquired before the main session to generate registration, segmentation, and localisation 19 templates. This assists real-time realignment and extraction of region-based signals, and 20 minimises per-volume processing time.

22
No neurofeedback: To keep the setup applicable to a range of real-time scenarios without 1 introducing additional confounds, no neurofeedback was provided. Instead, to approximate 2 similar mental states, the second functional set of scans were structured as imagined versions of 3 the first functional set. This is a common approach in neurofeedback training: amygdala 4 neurofeedback participants have been asked to think about an emotional event in their past (e.g. asked to think about performing physical exercises (e.g. Subramanian et al., 2011). 7 8 Physiology data: To facilitate the development and exploration of real-time physiological 9 denoising methods and their relation to multi-echo-derived data, cardiac and respiratory signals 10 were acquired.

12
The rt-me-fMRI is available in BIDS format via the DataverseNL repository: All participants provided informed and written consent to participate in the study and for their 26 maximally de-identified data (also referred to as limited data) to be shared publicly under specific 27 conditions (see GDPR considerations below). Participants were provided with an electronic 28 version of a "Participant Information Letter" which contained, in addition to standard information 29 about the study protocol, clear information about their personal data privacy and the risks and 30 benefits involved in sharing maximally de-identified versions of their data. They were asked to 31 read it thoroughly and to discuss it with friends and family if they wished to do so. They were 32 granted an opportunity to discuss any questions or concerns about their voluntary participation in 33 the study with the lead researcher, both via email and in person. If they decided to continue with 34 participation, participants signed the consent form and were provided with an electronic copy.

36
The dataset was collected, processed and shared in accordance with the European Union's 37 General Data Protection Regulation (GDPR) as approved by Data Protection Officers (DPOs) at 38 Kempenhaeghe Epilepsy Center (Heeze, NL) and the Eindhoven University of Technology. Of 1 particular note is the procedure that was followed to enable sharing of the dataset under specific 2 conditions that allow personal data privacy to be prioritised while adhering to FAIR data standards 3 ("findable, accessible, interoperable, reusable"; see Wilkinson et al., 2016), with this being the 4 first documented implementation. It followed from the collaborative effort of the Open Brain 5 Consent Working Group (Pernet et al., 2020), a group of researchers, data experts, and legal 6 practitioners that aim to provide globally standardised templates for informed consent and data 7 privacy statements that allow for brain research data to be shared while prioritising personal data 8 privacy.
Steps to accomplish this include following best practices to de-identify brain images (e.g. 9 removing personally identifiable information from image filenames and metadata and removing 10 facial features from T1-weighted images), converting the data to BIDS format, employing a Data 11 Use Agreement, and keeping participants fully informed about each of these steps and the 12 associated risks and benefits. The Data Use Agreement can be accessed in this manuscript's 13 GitHub repository: https://github.com/jsheunis/rt-me-fMRI. 14

15
The rt-me-fMRI dataset consists of MRI and physiology data from 28 healthy, right-handed (self-16 report) adults recruited from the local student population: 20 male, 8 female; age = 24.9 ± 4.7 17 (mean ± standard deviation). During recruiting, possible participants were excluded if they 18 reported prior or current (at the time of the study) indications of neurological or psychiatric 19 conditions, or any other standard contraindications for MRI scanning. 31 participants were initially 20 recruited for the dataset, but three were excluded because of technical and administrative 21 challenges. All anatomical scans were inspected by a trained radiologist and no incidental findings 22 were reported.  To minimise participant motion during scans so as to improve spatial and temporal image quality, 32 participants were asked to remain as still as possible inside the scanner. Additionally, a length of 33 tape was fixed across the participants' foreheads to the stationary part of the head coil. All functional scans have 210 volumes and exactly the same sequence parameters. All task scans 6 follow a block design with 10 volumes (i.e. 20 s) per block, and with blocks alternating between 7 control and task conditions. All task designs start and end with a control condition. These block 8 design aspects are depicted in Fig. 2 below for all task runs. Take note that the depictions do not 9 necessarily agree with the exact stimuli as seen by the participants, as the depictions below are 10 purely illustrative. 11 12 For the fingerTapping task, participants were instructed to execute finger tapping with their right 13 hand by steadily tapping the tip of the thumb to the tip of each other finger in succession, reversing 14 the tapping order until the end of the task block is reached. For the fingerTappingImagined task, 15 participants were instructed to imagine doing exactly the same as in the actual finger tapping task, 16 but without actually moving their right fingers. For the control condition during the 17 fingerTappingImagined task, participants were asked to count backwards in multitudes of 7.

19
The emotionProcessing task was an adapted "Hariri" task from the emotion processing task used that resembled the top one, by pressing a left or right button (2 s duration). The inter-trial interval 25 was 1 s duration (see Fig. 3). Each 20 s block had 6 trials. The same design timing was used for 26 the control condition blocks, i.e. matching shapes, as for the trial condition blocks depicted in Fig.  27 3. Participants used an MRI-compatible button box with their right hand to complete the task. 28 Participants were asked to press the left button with their right index finger if selecting the bottom 29 left image (shape or face) on the screen, and to conversely press the right button with their right

MRI acquisition parameters
23 MRI data was acquired on a 3 Tesla Philips Achieva scanner (software version 5.1.7) and using 24 a Philips 32-channel head coil. 25 1 Note: for the majority of participants, the presentation timing for the emotionProcessing task was delayed by tens of milliseconds for each trial (planned versus actual timing). This resulted in the full task presentation running on for about 5 s after the scan acquisition stopped. This is not deemed a problem, mainly since the exact presentation time was captured and is available in the BIDS dataset. However, users should take note not to use the planned timing parameters as that would ignore the delay that occured. 210 (excluding 5 dummy volumes discarded by the scanner); total scan time = 7:00 min (excluding 8 5 dummy volumes); flip angle = 90˚; field of view = 224×224×119 mm 3 ; resolution = 3.5×3.5×3.5 9 mm 3 ; in-plane matrix size = 64×64; number of slices = 34; slice thickness = 3.5 mm; interslice gap 10 = 0 mm; slice orientation = oblique; slice order/direction = sequential/ascending; phase-encoding 11 direction = A/P; SENSE acceleration factor = 2.5; parts of the cerebellum and brainstem were 12 excluded for some participants to ensure full motor cortex and amygdala coverage.

14
The echo times, spatial resolution, and SENSE factor were tuned with the aim of improving spatial 15 resolution and coverage while limiting the TR at maximum 2000 ms, including a maximum number 16 of echoes, and keeping the SENSE factor low to prevent SENSE artefacts. 17 around the participant's upper abdomen. Heart rate was recorded using a pulse oximeter fixed to 20 the participant's left index finger. Both of these recording devices were wired directly to the 21 scanner, sampled at 500 Hz, synchronized internally to the start/stop pulses of each functional 22

Physiology data acquisition parameters
scan, and data were written to Philips's standard "scanphyslog" log file type. involved the use of several software packages and custom scripts to assist in file format 27 conversion and data structuring, as detailed below. A Jupyter notebook containing Python code 28 and descriptions for each of the steps below can be accessed at the project's code repository: and date stamps and any identifiable information related to the acquisition location or system from 7 the files output from bidsify. 8 9 Since PAR/REC files do not contain slice timing information, the converted NIfTI files did not 10 contain it either. Slice timing information was calculated using available parameters and added 11 with a script to the BIDS-specific JSON sidecar files. 12 2.7.2. Physiology data 13 Heart rate and breathing traces were converted from the Philips "scanphyslog" format to BIDS 14 format using the Python package scanphyslog2bids (v0.1; 15 https://github.com/lukassnoek/scanphyslog2bids). fMRwhy toolbox (see "Code Availability" for details). The basic anatomical and functional 23 preprocessing pipeline applied to all data is depicted in Fig. 4 below.

25
As a first step, the T1-weighted anatomical image was coregistered to the template functional 26 image (task-rest_run-1_echo-2, volume 1) using SPM12's coregister/estimate functionality, which 27 maximizes normalised mutual information to generate a 12 degree-of-freedom transformation 28 matrix. Before resampling to the functional resolution, this coregistered T1-weighted image was 29 segmented using tissue probability maps and SPM12's unified segmentation algorithm 30 (Ashburner and Friston, 2005). This yielded subject-specific probability maps for gray matter, 31 white matter, CSF, soft tissue, bone and air in the subject functional space. All of these probability 32 maps were then resampled (using coregister/write) to the subject functional resolution. Masks 33 were generated for gray matter, white matter, CSF, and the whole brain (a combination -logical 34 OR after thresholding -of the previous three masks). These were overlaid on the coregistered 35 and resampled T1w image below, to allow visual inspection of segmentation and registration 36 quality. 37 space to the subject functional space using SPM12 normalise/write, as well as the inverse 6 transformation field that was saved as part of the segmentation procedure mentioned above. The 7 regions of interest for this study include the left motor cortex (for the motor processing tasks), the 8 bilateral amygdala (for the emotion processing tasks) and the fusiform gyrus (for the 9 emotionProcessing task). These ROIs are overlaid on the coregistered and resampled T1-10 weighted image, to allow visual inspection of normalisation quality.

22 23
Functional data were preprocessed, starting with estimating realignment parameters for each 1 functional time series using SPM12's realign/estimate, which performs a 6 degree-of-freedom 2 rigid body transformation that minimizes the sum of squared differences between each volume 3 and the template volume. Realignment parameters were estimated for the second-echo time 4 series of each run. Then, slice timing correction was done with SPM12, which corrects for 5 differences in image acquisition time between slices. Each echo time series of all functional runs 6 were slice time corrected. 3D volume realignment followed, which applied spatial transformation 7 matrices derived from the previously estimated realignment parameters to all echo time series of 8 all functional runs. Both raw time series and slice time corrected time series were realigned. 9 Lastly, all echo time series of all functional runs were spatially smoothed using a Gaussian kernel 10 filter with FWHM = 7 mm (i.e. double the voxel size). Smoothing was performed on raw, slice time 11 corrected and realigned time series data. 12 13 Next, several signal time series were calculated or extracted for use as possible GLM regressors 14 in functional task analysis, or for quality control. (of all functional runs), signals were extracted per voxel and spatially averaged within the 26 previously generated tissue masks to yield tissue compartment signals for gray matter, white 27 matter, cerebrospinal fluid (CSF) and the whole brain.

29
The last set of preprocessing steps included calculation of image quality metrics and 30 visualizations, using the BIDS-compatible fmrwhy_bids_workflowQC pipeline from the fMRwhy 31 toolbox. Operations on functional time series data were all done on detrended (linear and 32 quadratic trends) realigned data, except where otherwise specified. Temporal signal-to-noise ratio 33 (tSNR) maps were calculated for all runs by dividing the voxel-wise time series mean by the voxel-34 wise standard deviation of the time series. Tissue compartment averages were then extracted 35 from these tSNR maps. Percentage difference maps (from the time series mean) were calculated 36 per volume for use in carpet plots (or gray plots). 37

Data Records 1
The rt-me-fMRI dataset is available in BIDS format via the Dutch research data repository 2 DataverseNL at the following link: https://dataverse.nl/dataverse/rt-me-fmri. This repository 3 includes the raw BIDS data, descriptive metadata, and derivative data including quality reports.

5
Apart from the dataset README file, all core files are available in one of 3 formats: NIfTI, TSV 6 and JSON. Functional and anatomical data are stored as uncompressed NIfTI files (with the ".nii" 7 extension), which contain image and header data and can be handled/viewed by all major 8 neuroimaging analysis packages and programming languages. Tabular data such as participants, 9 task events, response timing and physiology data are stored in tab-separated value text files (with 10 the extension ".tsv", or if compressed ".tsv.gz") and can be handled by text or spreadsheet 11 reading/editing software on all major operating systems, or alternatively by all major software 12 programming languages. Metadata about the dataset, tasks, events and more are stored as key-13 value pairs in text-based JSON files (with the extension ".json") that can be handled/viewed using 14 all major software programming languages.

16
All data files are organised according to the BIDS convention for dataset participants, MRI data 17 type (anatomical or functional), and derivatives, as depicted in Fig. 5.  The expansion of "sub-001" (top right) shows subdirectories "anat" and "func", each with neuroimages and metadata 5 related to anatomical and functional scans, respectively. The expansion of the "derivatives" directory (bottom right) 6 shows subdirectories "fmrwhy-dash" and "fmrwhy-qc". The former contains all derivative data required to run the 7 interactive browser-based application accompanying this dataset. The latter includes a quality report per participant in 8 HTML format.

9
Each participant directory contains two subdirectories: "anat" and "func", respectively containing 1 all anatomical and functional images and metadata. Different data types can be distinguished 2 based on BIDS identifiers, e.g. "_bold" for functional and "_T1w" for anatomical MRI data. The full 3 list of data acquisitions with their data types, descriptions, and formats are provided below in 4 Table 1. Note that for functional data, each resting state and task run consists of three separate 5 image files, one per echo (i.e. "_echo-1_bold.nii", "_echo-2_bold.nii", and "_echo-3_bold.nii"). 6 JSON sidecar files accompany all BOLD and physiology data files on the participant level, while 7 the accompanying JSON sidecar files for the four types of task event files are on the dataset level. 8 Other files on the dataset level include the README, the dataset description (JSON) and the 9 participant list (TSV). 10 11 Apart from the core dataset, rt-me-fMRI also includes derivative data in two subdirectories 1 generated by the fMRwhy toolbox and related scripts: "fmrwhy-qc" and "fmrwhy-dash". The former 2 results from the fmrwhy_bids_workflowQC pipeline and contains a subdirectory per participant, 3 each in turn including subdirectories "anat" and "func". These directories contain NIfTI, TSV and 4 PNG files of quality control outputs, which are all required for the HTML quality report contained 5 in the "report_[yyyymmddhhmmss]" directory. These reports can be opened with all major Internet 6 browsers. The "fmrwhy-dash" derivative directory contains (as TSV files) all data required to yield 7 the interactive visualisations of the supplementary browser-based application provided with this 8 dataset: https://rt-me-fmri.herokuapp.com/. 9

Data quality assessment 20
Image and data quality of this dataset was assessed using the fMRwhy toolbox. This allowed 21 quality to be assessed for raw and minimally (pre)processed versions of the data, and also for 22 interim steps on which the validity of eventual study outcomes might depend. A BIDS-compatible 23 workflow in the fMRwhy toolbox, fmrwhy_bids_workflowQC, runs initial preprocessing and quality 24 control of the raw data and outputs a quality report per subject, which includes metrics and 25 visualizations for anatomical and functional MRI data and for peripheral data.  and framewise displacement. These plots are useful quality checking tools as they make 17 it easy to visualise wide scale signal fluctuations across voxels, which can then be related 18 visually to changes in physiological signals or subject movement.

19
# Checking the quality of the recorded cardiac and respiratory traces is made possible with 20 images generated by TAPAS PhysIO during the process of calculating RETROICOR, CR 21 and RVT regressors. Images include a plot of the temporal lag between derived heart 22 beats within thresholds for outliers, and a plot showing the breathing belt amplitude 23 distribution that can be inspected for unexpected shapes.

25
All functional quality metrics of the full dataset, generated by the fmrwhy_bids_workflowQC 26 workflow, are summarised in Table 2. This includes, per run, mean framewise displacement, total 27 framewise displacement, framewise displacement outliers (based on a conservative 0.2 mm 28 threshold, and a liberal 0.5 mm threshold), and mean tSNR in all tissue compartments (grey 29 matter, white matter, cerebrospinal fluid, whole brain). This allows possible data users to inspect 30 the quality measures and to set personalised thresholds and exclusion criteria. 31 32 Table 2: Functional quality metrics for the rt-me-fMRI dataset 33 (Online version: https://github.com/jsheunis/rt-me-fMRI/blob/master/data/sub-all_task-all_desc-allQCmetrics.tsv) 34 35 Fig. 6 below displays summarised quality metrics for the rt-me-fMRI dataset, and examples of 36 single-subject quality images. Individual quality reports can be downloaded together with the 37 dataset.  The slice timing corrected, 3D realigned and spatially smoothed Echo 2 time series of all task 12 runs underwent individual-and group-level statistical analysis using a general linear model with 13 SPM12. Task regressors included the main "ON" blocks for the fingerTapping, 14 fingerTappingImagined, and emotionProcessingImagined tasks, and both the separate 15 "SHAPES" and "FACES" trials for the emotionProcessing task. Regressors not-of-interest for all 16 runs included six realignment parameter time series and their derivatives, the CSF compartment 17 time series, and RETROICOR regressors (both cardiac and respiratory to the 2nd order, 18 excluding interaction regressors, selected based on common implementation procedures in 19 literature). Additional steps executed by SPM12 before beta parameter estimation include high-20 pass filtering using a cosine basis set and AR(1) autoregressive filtering of the data and GLM 21 design matrix.

23
Contrasts were then applied to the single task-related beta maps for the fingerTapping, 24 fingerTappingImagined, and emotionProcessingImagined tasks, and to the FACES, SHAPES, 25 and FACES>SHAPES beta maps for the emotionProcessing task. Statistical thresholding, 26 consisting of familywise error rate control with p < 0.05 and a voxel extent threshold of 0, was 27 then applied on a per-subject basis to identify task-related clusters of activity. Unthresholded 28 subject-level contrast maps were normalized to MNI152 space and then fed into a group-level 29 one-sided t-test, for which the t-statistic maps were subsequently thresholded at p < 0.001 and 30 an extent threshold of 20 voxels. Unthresholded individual-and group-level t-statistic maps can 31 be accessed as a NeuroVault collection: https://neurovault.org/collections/XWDGUJHD/. 32 33 Fig. 7 below shows the resulting thresholded group t-statistic maps for all four task runs. Fig. 7A  34 clearly shows activity clusters in the left motor cortex and right cerebellum, as expected for a 35 finger tapping task as well as a negative activation pattern in the default mode network. Fig. 7C  36 shows activation in the visual cortex commensurate with a face/shape matching task, specifically 37 in the left and right fusiform gyri. Additional clusters are found in the amygdalae and hippocampi, 38 as expected for an emotion processing task. For both imagined tasks, similar but weaker 39 activation clusters are found in the expected regions (respectively the motor cortex in Fig. 7B, and  40 amygdalae in Fig. 7D) but both wide scale activation patterns are consistent with mental tasks 41 including imagery and memory recollection. Additionally, Figs. 7B and 7D show negative 42 activation patterns in the dorsal attention network. The activation results in Fig. 7 are further 1 evidenced by the resulting highest correlated terms when decoding the unthresholded group t-2 statistic images with the web-based Neurosynth tool (www.neurosynth.org, Yarkoni et al., 2011). 3 Table 3 shows the resulting terms 2 . 4 5 6 7 2 Task names of the rt-me-fMRI dataset were selected based on the desired activation response for the given use cases, e.g. emotionProcessing to elicit a response in regions involving emotion processing, with the knowledge that the tasks might yield varied responses and have varied use cases. This can lead to the activation analysis and Neurosynth decoding process yielding patterns and terms that do not necessarily reflect the task name, e.g. activation of the fusiform face area and related terms ("face", "fusiform", "occipital") for the emotionProcessing task. 2 Images were generated with bspmview. Fig. 7A clearly shows activity clusters in the left motor cortex and right 3 cerebellum, as expected for a finger tapping task. Fig. 7C shows activation in the visual cortex commensurate with a 4 face/shape matching task, specifically in the left and right fusiform gyri. For both imagined tasks, similar but weaker 5 activation clusters are found in the expected regions (respectively the motor cortex in Fig. 7B, and amygdalae in Fig.   6 7D) but both wide scale activation patterns are consistent with mental tasks including imagery and memory recollection. 7 8 Table 3: Neurosynth-decoded terms 9 samples multiple T2*-weighted images at a range of echo times along the decay curve following 12 a single transverse magnetic excitation, which theoretically allows the optimum BOLD contrast to 13 be optimized for a range of baseline tissue T2* values. Subsequently, echo combination through 14 weighted summation or averaging is a typical processing step that generally increases temporal 15 signal-to-noise ratio and contrast-to-noise ratio and decreases signal drop-out in regions with high 16 susceptibility artefacts and signal dropouts (Menon et al., 1993;Posse et al., 1999;Posse et al., 17 2012). Echoes can be combined using a variety of weights, including baseline voxelwise tSNR 18 and T2* maps.

20
Figs. 8 and 9 illustrate that such combination procedures improve tSNR and signal dropout, hence 21 validating the use of multi-echo fMRI for improved quality data. Representative signal recovery is 22 demonstrated in the tSNR maps of Fig. 8   with no prior or current (at the time of the study) indications of neurological or psychiatric 3 conditions. They also had to report the absence of any other standard contraindications for MRI 4 scanning. 32 participants were initially recruited for the study, and the datasets of three 5 participants were excluded due technical and one due to administrative challenges. No further 6 datasets were excluded, even in cases of more than average or severe motion (e.g. sub-010 and 7 sub-021), since it was decided that such data could still be useful for future methods development 8 or related insights. Table 2 (also available in the project's GitHub repository 9 (https://github.com/jsheunis/rt-me-fMRI) provides a list of all functional quality metrics for all 10 participants and runs, which allows possible data users to inspect the quality measures and to set 11 personalised thresholds and exclusion criteria. 12

13
An interactive environment (https://rt-me-fmri.herokuapp.com/) was created alongside this study 14 to allow users to interactively explore summaries of the data derivatives and quality control 15 aspects.

17
All software scripts and self-developed tools used to prepare, preprocess and quality check the 18 data are openly available at the project's code repository https://github.com/jsheunis/rt-me-fMRI.

19
This includes instructions to download, extract, and understand the data; the data preparation 20 script; the preprocessing script; the quality reporting script; and the script to reproduce the figures 21 for this manuscript.

23
Dependent software and toolboxes/packages used for these preparation, preprocessing and 24 quality reporting steps include:  preprocessed and analysed the data, and wrote the article.

18
MB reviewed the article.

19
CC designed the protocol, contributed to data analysis, and reviewed the article.

20
LH designed the protocol, contributed to data analysis, and reviewed the article.

21
WH designed the protocol, contributed to data analysis, and reviewed the article 22 JJ designed the protocol, contributed to data analysis, and reviewed the article 23 RL designed the protocol, contributed to data analysis, and reviewed the article 24 SZ administered the project and reviewed the article.

25
BA administered the project and reviewed the article.