Running QSIPrep

Most options have defaults that do not need to be changed. This page goes through the options in the order they appear in qsiprep --help and says what each one decides and when to change it. The full help text is on the Command-line reference page.

A minimal command

Suppose the BIDS input contains:

sub-1/ses-1/dwi/sub-1_ses-1_acq-multishell_run-01_dwi.nii.gz
sub-1/ses-1/dwi/sub-1_ses-1_acq-multishell_run-02_dwi.nii.gz
sub-1/ses-1/dwi/sub-1_ses-1_acq-multishell_run-03_dwi.nii.gz
sub-1/ses-1/fmap/sub-1_ses-1_dir-PA_epi.nii.gz

with the epi fieldmap’s IntendedFor naming the three DWI runs. One way to process these data is:

qsiprep /path/to/bids /path/to/output participant \
    --output-resolution 1.7 \
    -w /path/to/work

The three positional arguments follow the BIDS App convention. --output-resolution has no default: it is the isotropic voxel size, in mm, of the preprocessed DWI (see Output space and resolution). With these options the three runs are denoised separately, concatenated, motion and distortion corrected as one series with eddy and TOPUP, and written to /path/to/output/sub-1/ses-1/dwi/sub-1_ses-1_acq-multishell_space-ACPC_desc-preproc_dwi.nii.gz.

How to run the container versions of this command is on the Installation page.

Selecting subjects, sessions and files

--participant-label and --session-label take space-separated lists of labels, with or without the sub-/ses- prefix. --bids-database-dir points PyBIDS at a reusable index, which saves minutes on large datasets, and --skip-bids-validation skips the validator.

--bids-filter-file restricts which files are read. QSIPrep finds its inputs with one query per data type:

{
    "fmap": {"datatype": "fmap"},
    "t2w": {"datatype": "anat", "suffix": "T2w"},
    "t1w": {"datatype": "anat", "suffix": "T1w"},
    "roi": {"datatype": "anat", "suffix": "roi"},
    "dwi": {"datatype": "dwi", "suffix": "dwi"}
}

The filter file adds BIDS entities to these queries. This file selects session MR1 for the T1w and DWI queries, and runs 2 and 3 of the DWI query only:

{
    "t1w": {"session": "MR1"},
    "dwi": {"session": "MR1", "run": [2, 3]}
}

Values may be lists, and "regex_search": "true" inside a query turns its values into regular expressions ("acquisition": "(?i)mprage"). The list of entities is in the PyBIDS configuration. If you want some scans combined and others processed separately, run QSIPrep more than once with different filter files.

Workflow scope

--boilerplate-only writes the methods text (see Citation boilerplate) and exits. --reports-only rebuilds the HTML reports from an existing output directory without running anything.

--ignore switches off processing the data would otherwise trigger, and --force switches on processing the metadata would otherwise skip. Both take space-separated lists.

`--ignore` values
Value	Effect
`fieldmaps`	Skip the `fmap/` directory. Reverse phase-encoded DWI runs still drive distortion correction.
`pepolar-dwis`	Stop pairing DWI series with each other for PEPOLAR estimation. The series are still processed, corrected by a fieldmap they are linked to, by a fieldmap-less method, or not at all.
`sdc`	No susceptibility distortion correction of any kind.
`t2w`	Drop T2w images, including T2w-based distortion correction.
`phase`	Ignore `part-phase` images; denoise as magnitude data.
`shims`	Treat all `ShimSetting` values as compatible when grouping series.
`fov`	Concatenate series with differently oriented fields of view anyway. The distortion correction is then misapplied to some of them.
`gradwarp`	No gradient nonlinearity correction, including the gradient deviation map.
`jacobian`	No Jacobian intensity modulation by QSIPrep itself. `eddy`’s internal modulation is unaffected (see Intensity modulation).

`--force` values
Value	Effect
`sdc-anat-reference`	Use the `--sdc-anat-reference` image for every DWI series, instead of only for series no fieldmap reaches.
`gradwarp3D` / `gradwarp1D`	Apply the full 3D, or only the through-plane, gradient nonlinearity correction to every run regardless of `ImageType`. Mutually exclusive; both need `--gradient-file`.
`jacobian`	Also Jacobian-modulate the fieldmap-less T2Wreg correction, which TORTOISE leaves unmodulated.
`gre-sdc-after-eddy`	Apply a GRE fieldmap to `eddy`’s outputs, as QSIPrep did before 26.1, instead of handing it to `eddy`. For comparing the two on real data. Deprecated, and will be removed in a future release.
`no-csf-synthstrip`	Restrict the nonlinear registration to the template with a tighter SynthStrip mask (its `--no-csf` model) that excludes CSF and dura, for brains where the default mask is looser than the template’s. Saved brain masks and everything downstream of them are unaffected.

Anatomical processing

--anat-modality picks the anatomical reference, T1w (the default) or T2w. The reference is skull-stripped and segmented, defines the subject’s ACPC space, and is normalized to the template. Choose none to use no anatomical image at all; the b=0 reference is then rigidly aligned to the template instead. This is rarely a good idea, because it is rare to have dMRI data without a T1w or T2w of the same person.

--subject-anatomical-reference decides how many anatomical spaces a subject gets. first-lex (the default) aligns all sessions’ anatomical images to the first one, unbiased builds a mid-point template from them, and sessionwise gives every session its own space and its own report. See Anatomical processing for what each involves.

--anat-biascorrect runs N4 on the anatomical images (n4, the default), never (none), or only when ImageType does not contain NORM (auto). --anatomical-template has one choice, MNI152NLin2009cAsym; --infant replaces it with the infant template. --skip-anat-based-spatial-normalization skips the nonlinear registration to the template, which saves about twenty minutes; the template-space anatomical derivatives and the ACPC to template transform are then not written. --force no-csf-synthstrip masks that registration with a tighter brain mask when the default one lets the dura land on the template’s brain edge, as in atrophied brains (see Anatomical processing).

--infant swaps the template for the MNIInfant cohort matching the participant’s age in months, read from participants.tsv or the *_sessions.tsv file. It requires --subject-anatomical-reference sessionwise. The cohort selection follows Nibabies.

Per-series preprocessing

Each DWI series is denoised on its own before series are concatenated. The order is fixed: MP-PCA or patch2self denoising on the raw data, then Gibbs unringing, then, after resampling, B1 bias field correction and b=0 intensity harmonization.

	Denoising	Gibbs unringing	B1 bias field correction
What it removes	Random noise	Ringing at sharp edges	Smooth intensity non-uniformity
Tools	`dwidenoise` (MRtrix3), `dwidenoise2`, `patch2self` (DIPY), `svht` (`svht_denoise`)	`mrdegibbs` (MRtrix3), `rpg` (TORTOISE), `svht` (`svht_denoise`)	`N4BiasFieldCorrection` on the b=0 images, applied to the series
Default	`dwidenoise`	Off	On
Option	`--denoise-method`	`--unringing-method`	`--dwi-biascorrect`

--dwidenoise-window sets the dwidenoise patch size in voxels, an odd integer or auto (the default, derived from the number of volumes). dwidenoise2 sizes its patches from its own schedule, and svht from extent; both are set with --denoise-config (see below). dwidenoise and dwidenoise2 are licensed for non-commercial use only; svht is not restricted. Its settings are also given with --denoise-config. rpg unringing is the method for partial Fourier acquisitions; mrdegibbs assumes full Fourier sampling. svht unringing reads PartialFourier from the sidecar: it handles full Fourier sampling and 7/8 or 6/8 partial Fourier along the j axis, and stops with an error for any other partial Fourier acquisition. --dwi-biascorrect auto skips bias correction when every DWI’s ImageType contains NORM, and none skips it always. We recommend none for prescan-normalized data. --no-b0-harmonization skips the rescaling that matches b=0 intensities across series. --b0-threshold (default 100) is the b-value below which a volume counts as b=0.

With part-phase data, dwidenoise and dwidenoise2 denoise the complex signal. Whether the complex data are carried into unringing depends on --mrtrix-version: only the dev branch of MRtrix3 has a complex-valued mrdegibbs; with stable the data are reduced to magnitude after denoising. svht denoising rotates the magnitude and phase onto the real axis instead, and keeps the magnitude of the result. --ignore phase drops the phase images altogether.

Denoising settings

--denoise-config takes a JSON file with settings for the denoiser that --denoise-method selects, which must be dwidenoise2 or svht. The keys the file may set depend on that method. Every key is optional and unknown keys are an error. Settings the two tools share use the same key and spelling. The file is checked when the command line is parsed, and a copy is written to sub-<label>/log/<run uuid>/<method>.json in the output directory.

dwidenoise2

Each key sets the dwidenoise2 option of the same name, except filter_method, which sets -filter, and schedule.

Key	JSON value
`aggregator`	`"exclusive"`, `"gaussian"`, `"invl0"`, `"rank"` or `"uniform"`
`datatype`	`"float32"` or `"float64"`
`debias_anchor`	`"sample"` or `"group_mean"`
`decomposition`	`"bdcsvd"` or `"selfadjoint"`
`demean`	`"none"`, `"volume_groups"`, `"shells"` or `"all"`
`demod_axes`	a list of non-negative integers, such as `[0, 1]`
`demodulate`	`"none"`, `"linear"`, `"hann"` or `"apc"`
`estimator`	`"exp1"`, `"exp2"`, `"med"`, `"mrm2023"` or `"tbme2022"`
`filter_method`	`"optshrink"`, `"optthresh"` or `"truncate"`
`fixed_rank`	an integer of at least 1
`noise_dof`	an integer of at least 1
`noise_in`	a number of at least 0; noise-map files are not supported
`preserve_noise_bias`	`true` or `false`
`vst_method`	`"none"`, `"linear"`, `"foi"`, `"koay"` or `"mom"`
`schedule`	a list of iterations, or the name of a bundled schedule

-aggregator_fwhm, -rankpermm_in and the diagnostic export options cannot be set. demodulate other than "none" needs part-phase data. noise_in only seeds the variance-stabilizing transform.

schedule lists the iterations of the multi-resolution noise estimation. Each iteration is an object whose keys are the columns of a dwidenoise2 schedule file: spatial_subsample (an integer or a list of three), kernel ("aspect=2.0", "rmse=0.02", "rank", "radius=4", "voxels=100", "cuboid=1x" or "rank_fixed"), smooth_noise and update_noise (booleans), temporal_subsample (a number in (0, 1]), partitions and max_partition_size. The last iteration is the reconstruction pass.

{
  "estimator": "exp2",
  "decomposition": "bdcsvd",
  "schedule": [
    {"spatial_subsample": 8, "kernel": "aspect=2.0", "update_noise": true},
    {"spatial_subsample": [4, 4, 2], "kernel": "rmse=0.02", "update_noise": true},
    {"spatial_subsample": 2, "kernel": "rank", "update_noise": false}
  ]
}

schedule may instead name a bundled schedule: "default", "legacy" or "vlarge" (recommended by dwidenoise2 for more than 255 volumes). Without a schedule key, dwidenoise2 uses its default, except that fixed_rank selects the single-iteration fixedrank schedule and "vst_method": "none" a single iteration.

svht

Each key sets the svht_denoise option of the same name, except filter_method, which sets -filter, and demean.

Key	JSON value
`demean`	`"none"` (the default) or `"shells"`; `"shells"` is recommended only for shelled schemes
`extent`	an odd positive integer whose cube exceeds the number of volumes; omitted, `svht_denoise` picks the smallest
`filter_method`	`"optshrink"`, `"optthresh"` or `"truncate"`
`aggregator`	`"gaussian"`, `"uniform"` or `"exclusive"`
`aggregator_fwhm`	a number greater than 0; needs the `"gaussian"` aggregator
`shape`	`"sphere"` or `"cube"`
`stride`	`1` or `2`; `"exclusive"` needs `1`
`vst`	`true` or `false`
`noise_dof`	an integer from 1 to 64
`preserve_noise_bias`	`true` or `false`

Keys the file omits keep the svht_denoise defaults.

"demean": "shells" removes the mean of each b-value shell before PCA and restores it afterwards, and passes the series’ b-values to svht_denoise with -bval. Use it only for shelled schemes (single- or multi-shell acquisitions). svht_denoise groups b-values into shells by clustering, with neighbours less than 80 s/mm² apart sharing a shell. On a non-shelled scheme (DSI, CS-DSI), that turns nearly every distinct b-value into its own shell of a few volumes. Each such mean is restored without being denoised, so much of the signal, and its noise, bypasses the denoiser. Even with "shells", demeaning is skipped for a series where it would leave svht_denoise fewer than three volumes to estimate the noise from.

extent is the side of a cubic patch; a spherical patch holds at least extent cubed voxels. It is checked against each series’ volume count when the workflow is built. --dwidenoise-window does not apply to svht. vst, noise_dof and preserve_noise_bias describe the variance-stabilizing transform of magnitude data, so they are an error for a series with part-phase data, which svht_denoise rotates onto the real axis instead.

{
  "filter_method": "optthresh",
  "shape": "cube",
  "noise_dof": 4
}

Gradient nonlinearity correction

--gradient-file enables gradient nonlinearity correction; the accepted file formats are on the data preparation page. Two things are corrected: the spatial displacement of voxels, and the diffusion encoding, which is written as a voxelwise gradient deviation map (*_graddev.nii.gz). How much spatial correction a run gets is read from its ImageType field:

`ImageType` tag	Behavior
(no `DIS` tag)	Full 3D correction
`DIS2D`	Through-plane correction only; the scanner already corrected in-plane distortion
`DIS3D`	No spatial correction; the scanner already applied it

--force gradwarp3D or --force gradwarp1D overrides the tag for every run, and --ignore gradwarp disables both corrections. The deviation map is written for every run that is not ignored, DIS3D included, because no scanner corrects the encoding. It is not written for outputs assembled by --distortion-group-merge.

Grouping and concatenation

By default the DWI series of a session are concatenated into one output, so head motion correction sees as much data as possible. The rules are on the data preparation page; two options change them.

--separate-all-dwis processes every DWI series on its own, giving one output per input file. --distortion-group-merge decides what happens to the corrected halves of an output when they were corrected separately, as they are for a reverse phase-encoded pair with TORTOISE: concat (the default) appends them along the fourth dimension, average averages the images at the same q-space coordinate, and none writes each correction unit as its own output. Whether an output’s series are corrected together or separately depends on the head motion method: eddy with TOPUP corrects a reverse phase-encoded pair as one series, so there is nothing to merge.

Use average for HCP-style data, where the whole sampling scheme was acquired in both phase encoding directions:

--hmc-method tortoise --distortion-group-merge average

This gives one image per q-space sample, like the HCP pipelines. concat keeps both copies.

Head motion and eddy-current correction

--hmc-method picks the software:

Method	Works on	Corrects
`eddy` (default)	Single-shell and multi-shell schemes only	Head motion, eddy currents, and (with TOPUP or a GRE fieldmap) susceptibility distortion in one model. FSL `eddy`.
`tortoise`	Any scheme, including Cartesian DSI and random q-space sampling	Rigid head motion and 24-parameter quadratic eddy currents. TORTOISE DIFFPREP.
`shoreline`	Any scheme	Head motion only, by registering each volume to a signal-model prediction. Scheduled for removal; use `tortoise` for non-shelled data.

eddy refuses non-shelled data: the run stops with an eddy-requires-shelled error. What each method does, step by step, is in eddy (FSL), TORTOISE DIFFPREP and SHORELine.

Each method reads its settings from a JSON file, and a default is used when none is given:

--eddy-config: The arguments to eddy. The default is eddy_params.json. Two keys are worth knowing: "method" (jac or lsr) is eddy’s resampling method, which also decides whether eddy applies Jacobian modulation itself, and "estimate_move_by_susceptibility" lets eddy model how the field changes with head position when it has a field to work with.
--diffprep-config: The arguments to DIFFPREP. The default is diffprep_params.json. "correction_mode" selects "motion" (rigid head motion only), "quadratic" (the default) or "cubic" eddy-current models. "drbuddi_synth_shell_bval" is described with DRBUDDI in TORTOISE DIFFPREP.
--shoreline-config: "model" ("3dshore", "tensor" or "none"), "iters" (default 2) and "transform" ("Affine" or "Rigid"). Only valid with --hmc-method shoreline.

GPU acceleration

--gpu runs selected tasks on the GPU: eddy, diffprep, drbuddi, synthstrip, synthseg, or all. GPU memory is usually the limit, so tasks are picked individually; an 8 GB card typically runs eddy, diffprep and drbuddi but not the FreeSurfer deep-learning tools. --gpu overrides use_cuda in the eddy and DIFFPREP config files. GPU builds are not numerically identical to the CPU builds, so this changes results, not only run time. The GPU must be exposed to the container (docker run --gpus all or apptainer run --nv). The image ships CUDA 12.2. --tortoise-gpu-cpu-ratio tunes how DIFFPREP shares volumes between the GPU and the CPU threads; the help text gives the formula.

Susceptibility distortion correction

Which correction a series gets is decided from the data: a fieldmap linked to the series is used if there is one, and reverse phase encoding is preferred over a GRE fieldmap when both are linked. The options below change the tool and provide a fallback for series that no fieldmap reaches.

--sdc-method: The tool for reverse phase-encoded (PEPOLAR) data. auto (the default) is TOPUP for eddy and DRBUDDI otherwise. topup and topup+drbuddi (TOPUP inside eddy, then DRBUDDI refining its result) require --hmc-method eddy; drbuddi works with every method.
--sdc-anat-reference: A fieldmap-less correction for series that no fieldmap reaches. synb0 synthesizes a distortion-free b=0 from the T1w with SynB0-DISCO; t2w registers to the T2w (TORTOISE T2Wreg); invt1w registers to the inverted-contrast T1w with SyN, as fMRIPrep does; auto picks synb0 with a T1w, else t2w with a T2w, else nothing. The default none leaves such series uncorrected. Add --force sdc-anat-reference to use the anatomical reference for every series, fieldmap or not.

What runs for each kind of data:

Data available	`--hmc-method eddy`	`--hmc-method tortoise`	`--hmc-method shoreline`
Reverse phase-encoded DWI series or `epi` fieldmap	TOPUP inside `eddy`; or DRBUDDI after `eddy`; or both	DRBUDDI after DIFFPREP	DRBUDDI after SHORELine
GRE fieldmap	Inside `eddy`	Applied after DIFFPREP	Applied after SHORELine
T1w and `--sdc-anat-reference synb0`	Synthetic b=0 joins the TOPUP inputs (with `topup` in `--sdc-method`)	T2Wreg to the synthetic b=0	Not corrected
T2w and `--sdc-anat-reference t2w`	Not corrected	T2Wreg to the T2w	Not corrected
T1w and `--sdc-anat-reference invt1w`	SyN after `eddy`	SyN after DIFFPREP	SyN after SHORELine
Nothing	Not corrected	Not corrected	Not corrected

A selection that cannot correct a series is reported as a warning in the log and the report, or as an error if the fieldmap-less method was curated or forced. The tools themselves are described in Susceptibility distortion correction.

Starting DRBUDDI or T2Wreg from a GRE fieldmap

When a series is corrected by DRBUDDI or T2Wreg and a GRE fieldmap also lists it, the registration starts from the field the GRE fieldmap measured and refines it, instead of the GRE fieldmap going unused. There is no option to turn this on; it happens whenever the linkage is there, and --ignore fieldmaps or --ignore sdc turns it off. It applies to:

a reverse phase-encoded pair corrected by DRBUDDI: --hmc-method tortoise or shoreline, or --hmc-method eddy --sdc-method drbuddi (not topup+drbuddi, where DRBUDDI only refines TOPUP);
an anatomical reference forced with --force sdc-anat-reference, corrected by T2Wreg (--hmc-method tortoise).

The GRE fieldmap must list the series. For a reverse phase-encoded pair of DWI series, its IntendedFor is enough as long as nothing in the session uses B0FieldIdentifier/B0FieldSource; once anything does, link the pair and the GRE fieldmap that way:

sub-01/
  fmap/
    sub-01_magnitude1.json   {"B0FieldIdentifier": "gre", ...}
    sub-01_magnitude2.json   {"B0FieldIdentifier": "gre", ...}
    sub-01_phasediff.json    {"B0FieldIdentifier": "gre", ...}
  dwi/
    sub-01_dir-AP_dwi.json   {"B0FieldIdentifier": "pepolar",
                              "B0FieldSource": ["pepolar", "gre"], ...}
    sub-01_dir-PA_dwi.json   {"B0FieldIdentifier": "pepolar",
                              "B0FieldSource": ["pepolar", "gre"], ...}

For an epi fieldmap, have both it and the GRE fieldmap name the DWI series in their IntendedFor. In the qsiplan report the series is then corrected by: pepolar with the GRE fieldmap listed as also eligible and labeled (initializes DRBUDDI/T2Wreg). During the run, QSIPrep logs Initializing DRBUDDI for <output> with GRE fieldmap <id>, and the report’s distortion correction entry ends in (GRE-initialized).

Intensity modulation

Correcting a spatial distortion moves signal between voxels, so QSIPrep also rescales the corrected image by the local volume change, following TORTOISE. Which component applies it depends on the backend; the table is in Intensity modulation. --ignore jacobian disables only the modulation QSIPrep itself applies. eddy modulates its own eddy-current and susceptibility corrections internally whenever its resampling method is jac (the default), and QSIPrep cannot undo that; with "method": "lsr" in --eddy-config those corrections are not modulated at all. --force jacobian additionally modulates the T2Wreg correction.

Output space and resolution

All outputs are written in the subject’s ACPC space: aligned to the anatomical reference and rotated so that the origin is at the anterior commissure, with the same orientation convention as the MNI templates. The T1w cannot align white matter to a template accurately, so the preprocessed DWI is not written in template space; spatial normalization is done after models are fit, in QSIRecon. The transform from ACPC to MNI152NLin2009cAsym is written with the anatomical derivatives.

--output-resolution sets the isotropic voxel size of the preprocessed DWI. Pass the acquired resolution to keep it, or a smaller value to upsample; some downstream methods, such as fixel-based analysis, recommend at least 1.3 mm. Head motion correction, distortion correction, coregistration and resampling are combined so the data are interpolated as few times as possible (once with SHORELine, twice with eddy and DIFFPREP, which write their own corrected volumes). Resampling uses Lanczos windowed sinc interpolation, or linear interpolation when upsampling by more than 10%.

--dwi2anat-dof is the number of degrees of freedom of the DWI to anatomical registration, 6 (rigid, the default) or 12 (affine).

DWI reference construction

--dwiref-definition decides which image is registered to the anatomical reference. distortion-group (the default) registers each distortion group’s own b=0 reference. subject first builds one mid-point template from every group’s reference, registers that once, and has every group inherit the result, so that data from different groups end up on exactly the same alignment. --dwiref-construction-iters (at least 2) and --dwiref-construction-transform (Rigid, Affine, BSplineSyN or SyN) configure the template build. A subject with a single distortion group gets the same subject outputs: no template is built, and that group’s reference becomes the subject reference, with an identity transform. subject cannot be combined with --subject-anatomical-reference sessionwise, which processes each session on its own and so cannot build a reference that spans the subject’s sessions. Which transforms are written in each case is described in Transforms.

Resources and execution

--nprocs (also --nthreads or --n-cpus) caps the number of threads across all processes and --omp-nthreads the number per process; --mem (or --mem-mb) caps memory in MB, and accepts a G suffix. --low-mem trades disk in the working directory for memory. --use-plugin (or --nipype-plugin-file) points at a Nipype plugin file for cluster schedulers.

-w/--work-dir is where intermediate files go; a run that stops can be restarted with the same working directory and picks up where it left off. --report-output-level chooses where the HTML reports are written (root, subject, session, or auto, which is session for --subject-anatomical-reference sessionwise and root otherwise). --config-file loads the settings of a previous run (its qsiprep.toml is in the log directory); command-line options override it. --stop-on-first-crash stops at the first failed node instead of running everything that does not depend on it. --resource-monitor records memory and CPU use per node. --notrack opts out of usage reporting. -v/--verbose raises the log level; -vvv is debug. --version prints the version and exits.

--sloppy, --debug and --write-graph are for developing QSIPrep itself. --sloppy output is not fit for analysis.

Reading the log

With -v -v, the terminal shows lines like:

[Node] Setting-up "qsiprep_wf.single_subject_PNC_wf.dwi_finalize_acq_realistic_wf.transform_dwis_t1.final_b0_ref.b0ref_reportlet" in "/scratch/qsiprep_wf/single_subject_PNC_wf/dwi_finalize_acq_realistic_wf/transform_dwis_t1/final_b0_ref/b0ref_reportlet".
  201229-21:33:46,213 nipype.workflow INFO:
    [Node] Running "b0ref_reportlet" ("niworkflows.interfaces.registration.SimpleBeforeAfterRPT")
  201229-21:33:48,51 nipype.workflow INFO:
    [MultiProc] Running 2 tasks, and 3 jobs ready. Free memory (GB): 3.70/4.00, Free processors: 0/2.
                      Currently running:
                        * qsiprep_wf.single_subject_PNC_wf.dwi_finalize_acq_realistic_wf.transform_dwis_t1.final_b0_ref.b0ref_reportlet
                        * qsiprep_wf.single_subject_PNC_wf.anat_preproc_wf.mni_mask

These describe what is running. Here b0ref_reportlet and mni_mask run at the same time, and both of the two processors allowed by --nprocs 2 are busy. The quoted name is the node, and the dotted path before it is the chain of workflows it belongs to. Those workflow names appear in the graphs on the Processing pipeline details pages: mni_mask is part of the anatomical workflow, b0ref_reportlet of the DWI reference workflow, and both sit under the subject workflow.