Quality measures

The quality check is for screening: a value far from the others of the same protocol is a reason to look at the image. Most measures depend on the scanner, the sequence and the field of view, so compare an image with the others of its protocol in your dataset (the In this dataset column of the quality panel), not with a fixed number. How to run the check is in Checking image quality; its methods and thresholds are settings (Settings).

Where a measure follows a definition of MRIQC, the explanation says how the two compare.

Anatomical images

How the check works: the brain is found with mindgrab and the tissues with robust_tissue (both from brainchop), and the image is registered to a template with niimath. Each runs in its own process; where one cannot run, a faster, more approximate method written in Python takes its place and the panel says so. The air around the head is measured only in images that are not defaced, not skull-stripped and not set to zero by the scanner.

Noise and contrast

How large the signal is against the noise, and how well grey and white matter are told apart. Compare only with images of the same protocol. Three families. The within-tissue SNRs (CSF, grey matter, white matter and their mean) compare each tissue's typical intensity with the spread inside that same tissue. The SNRs against the air compare tissue intensity with the noise in the background around the head. CNR and CJV measure how far apart grey and white matter are against their spreads.

SNR in white matter

White matter's typical intensity over the spread of intensities inside white matter. Higher is better; on a T1w it is the most informative of the three tissue SNRs.

In detail

What it measures: How large white matter's signal is compared with how much its intensity varies from voxel to voxel inside white matter. The spread holds thermal noise plus residual bias field, partial volume at the border with grey matter, real variation within the tissue and misclassified voxels: it is signal over everything that varies inside the tissue.
How it is computed: Bias-corrected intensities of the brain's voxels at the image's own resolution (up to a million sampled), each weighted by its probability of belonging to the tissue, from the class estimate made again at that resolution. The value is the weighted median over the weighted standard deviation, the SD multiplied by sqrt(n/(n-1)) with n the sum of the weights (a negligible correction at these sizes).
How to read it: Higher is better. On a T1w, white matter is bright, large and homogeneous, so its spread comes closest to the noise: of the three tissue SNRs this one tracks image quality best. On a T2w white matter is the darkest tissue, so its value is lower there and is read against other T2w images. Compare only with scans of the same contrast, sequence, field strength, coil and voxel size, which is what the 'In this dataset' column does.
What raises or lowers it: Thermal noise (small voxels, short scans, high acceleration), motion that blurs grey matter into white matter and adds ghosts, bias field the correction left behind, and white-matter lesions that widen the class. A segmentation that puts grey-matter or CSF voxels in the class lowers it too.
Caveats: An upper bound on the noise, not a pure noise measure: a perfect scan of a brain with variable white matter still reads lower. The memberships follow intensity at the borders, so voxels far from the class's typical value are partly given to another class, which trims the spread; the value depends on how the classes are split, and results from the networks and from the fast fallback should not be mixed. Smoothing at reconstruction or any resampling lowers the spread and raises the value at the cost of sharpness.
Compared with MRIQC: Same definition as MRIQC's snr_wm (median over SD with the n/(n-1) correction). Values differ from MRIQC's through the segmentation and the bias correction.
References: Esteban et al. 2017, PLoS One

SNR in grey matter

Grey matter's typical intensity over the spread inside grey matter. Higher is better, but it reads lower than white matter's for anatomical reasons, not because of noise.

In detail

What it measures: How large grey matter's signal is compared with the variation of intensity inside the grey-matter class: thermal noise plus partial volume, real variation within the tissue, residual bias and misclassified voxels.
How it is computed: Bias-corrected intensities of the brain's voxels at the image's own resolution (up to a million sampled), each weighted by its probability of belonging to the tissue, from the class estimate made again at that resolution. The value is the weighted median over the weighted standard deviation, the SD multiplied by sqrt(n/(n-1)) with n the sum of the weights (a negligible correction at these sizes).
How to read it: Higher is better. Expect it below white matter's on a T1w: grey matter is a ribbon a few millimetres thick where most voxels border white matter or CSF, and its intensity differs between the cortex and deep nuclei such as the thalamus and basal ganglia, so partial volume and anatomy inflate its spread. Read it beside white matter's: a fall in both points at noise or motion, a fall in grey matter's alone at blur, thick voxels or the segmentation.
What raises or lowers it: Noise and motion, as for white matter, and more than for white matter blur and thick voxels, which raise partial volume along the cortical ribbon. Deep grey nuclei, whose intensity differs from the cortex's, widen the class.
Caveats: Not a noise measure: in a sharp, quiet scan it is still limited by partial volume and anatomy. Voxels between grey and white matter are shared between the two classes by their probabilities, so the value depends on the segmentation and differs between the networks and the fast fallback.
Compared with MRIQC: Same definition as MRIQC's snr_gm; values differ through the segmentation and the bias correction.
References: Esteban et al. 2017, PLoS One

SNR in CSF

CSF's typical intensity over the spread inside CSF. On a T1w it is low even in an excellent scan and says little about quality.

In detail

What it measures: How large the CSF's signal is compared with the variation of intensity inside the CSF class.
How it is computed: Bias-corrected intensities of the brain's voxels at the image's own resolution (up to a million sampled), each weighted by its probability of belonging to the tissue, from the class estimate made again at that resolution. The value is the weighted median over the weighted standard deviation, the SD multiplied by sqrt(n/(n-1)) with n the sum of the weights (a negligible correction at these sizes). With the tissue network, CSF is everything inside the brain mask that the network does not call grey or white matter, held loosely so that intensity can move a border voxel to another class.
How to read it: Higher is better only against the same protocol. On a T1w CSF is the darkest class, so its median is small, and its mask is mostly thin sulci where nearly every voxel is mixed with grey matter, so its spread is large next to that median: a low value there is normal. On a T2w CSF is the brightest class and the value is higher, but flow and pulsation add variation of their own.
What raises or lowers it: Partial volume with grey matter (thick voxels, blur, motion); the extent of the brain mask, since extra-cerebral CSF, vessels or dura inside it widen the class; atrophy, which enlarges sulci and ventricles and changes the mix; and on a T1w the noise of magnitude images near zero intensity, which is not Gaussian.
Caveats: On a T1w it says more about anatomy, partial volume and the segmentation than about the scanner. Because it is low, it pulls the mean of the tissues down. Do not expect it to agree with the grey- or white-matter SNR.
Compared with MRIQC: Same definition as MRIQC's snr_csf; values differ through the segmentation and the bias correction.
References: Esteban et al. 2017, PLoS One

SNR (mean of the tissues)

The mean of the CSF, grey-matter and white-matter SNRs. Higher is better, but CSF's low value pulls it down.

In detail

What it measures: A one-number summary of the three within-tissue SNRs.
How it is computed: The plain mean of the three tissue SNRs; one that could not be computed is left out.
How to read it: Higher is better, against scans of the same protocol. It averages three measures that differ for anatomical reasons, so when it moves, look at the three to see which one moved. On a T1w, CSF's low SNR keeps it well below white matter's.
What raises or lowers it: Everything that lowers any of the three tissue SNRs: noise, motion, residual bias, partial volume and the segmentation.
Caveats: Two of its three parts (CSF and grey matter) are dominated by partial volume and anatomy, so it is a weaker quality measure than white matter's SNR alone. It weighs the three tissues equally, whatever their volume.
Compared with MRIQC: Same definition as MRIQC's snr_total (the mean of the three).
References: Esteban et al. 2017, PLoS One

SNR against the air (white matter)

White matter's typical intensity over the noise measured in the air around the head (Dietrich's method). Higher is better; needs real, untouched air.

In detail

What it measures: How large the white-matter signal is compared with the scanner's noise as the background shows it. Its denominator holds no anatomy, partial volume or bias field, so it is closer to the thermal noise than the within-tissue SNR.
How it is computed: 0.6551364 times white matter's (bias-corrected, probability-weighted) median, over the standard deviation of the intensities in the air: outside the head by a margin, above the plane through the glabella and the inion, without zero fill, at the image's own resolution. The factor is for magnitude images: where there is no signal the noise follows a Rayleigh distribution, whose SD is about 0.655 times the SD of the underlying Gaussian noise.
How to read it: Higher is better, against the same protocol. When its assumptions hold it is higher than the within-tissue SNR, because the air's spread holds no tissue variation. When the two disagree strongly, artefacts in the air (which raise its spread) or a filtered background (which lowers it) are the usual reasons.
What raises or lowers it: Noise (small voxels, short scans, acceleration) lowers it, and so does anything in the air that is not noise: ghosts, wrap-around, flow artefacts and ringing raise the air's SD.
Caveats: Valid only with real air: not computed on a defaced image (the sidecar's record, or more than 40 % of the face region exactly zero by default), on a skull-stripped image, with too little air in the field of view, or when more than 50 % of the air is exactly zero by default (both Settings > Quality control > Anatomical images). The 0.655 factor holds for single-channel magnitude noise. With multi-channel coils, parallel imaging (the noise then varies across the image, and the background no longer shows the brain's noise) or a vendor filter on the background, the value is unreliable.
Compared with MRIQC: Same method as MRIQC's snrd_wm. BIDS Manager leaves it empty where the air is not real, where MRIQC still reports a number.
References: Dietrich et al. 2007, J Magn Reson Imaging; Henkelman 1985, Med Phys

SNR against the air (mean)

Dietrich's SNR averaged over CSF, grey and white matter. Higher is better; needs real air.

In detail

What it measures: The tissues' typical intensities, averaged, over the noise measured in the air.
How it is computed: The mean of the three Dietrich SNRs, 0.6551364 times each tissue's median over the air's SD. The denominator is the same for all three, so it equals 0.655 times the mean of the three medians over the air's SD.
How to read it: Higher is better, against the same protocol. On a T1w the CSF median is small, so this reads below the white-matter value; on a T2w CSF is bright and raises it. Read it beside the white-matter value.
What raises or lowers it: As for the white-matter value: noise, and anything in the air that is not noise.
Caveats: The same conditions as the white-matter value: real, untouched air, and single-channel magnitude noise for the 0.655 factor to hold. Its tissue medians depend on the segmentation, so an unusual CSF class moves it.
Compared with MRIQC: Same method as MRIQC's snrd_total (the mean over the three tissues).
References: Dietrich et al. 2007, J Magn Reson Imaging

Contrast-to-noise ratio

How far apart grey and white matter are, against the noise of the air and the spread of both tissues. Higher is better; needs real air.

In detail

What it measures: The contrast between grey and white matter in units of the combined noise and tissue variation: whether the boundary between them stands out.
How it is computed: |median WM - median GM| / sqrt(SD_air^2 + SD_GM^2 + SD_WM^2). The tissue medians and SDs come from the bias-corrected, probability-weighted intensities; SD_air is the plain SD of the air's intensities, with no Rayleigh correction.
How to read it: Higher is better. The tissue SDs dominate the denominator and already hold noise, partial volume and tissue variation, so it moves with the within-tissue SNRs as well as with contrast. Compare only with scans of the same sequence and field strength: the grey-white contrast is set by the sequence (inversion time, flip angle, echo time) far more than by image quality.
What raises or lowers it: Noise, motion (which blurs the boundary and widens both classes) and residual bias lower it. Grey-white contrast also changes over the lifespan: it is very different in infants and falls in older adults.
Caveats: Not computed where the air is not real (defaced, skull-stripped, zero-filled). It mixes two kinds of spread (the raw air SD and the tissues' own), so read it as a relative measure within a protocol, not as a physical CNR.
Compared with MRIQC: Same formula as MRIQC's cnr.
References: Esteban et al. 2017, PLoS One

Coefficient of joint variation

The spread of grey and white matter over the distance between them (Ganzetti 2016). Lower is better.

In detail

What it measures: How much the grey- and white-matter intensity distributions overlap: their combined spread relative to the gap between their centres. Large when the two tissues are hard to tell apart.
How it is computed: (MAD_WM + MAD_GM) / |median WM - median GM| on the bias-corrected intensities at the image's resolution. Each MAD is the median absolute deviation scaled to an SD, over the voxels the class holds for sure (weight above half the class's largest); the medians are probability-weighted. It needs no air, so it is computed on defaced and skull-stripped images too.
How to read it: Lower is better. It works like the inverse of a contrast-to-noise ratio built from robust statistics: it rises with noise and blur and falls with stronger contrast. Compare with scans of the same sequence and field strength.
What raises or lowers it: Noise, motion (blur and ghosting widen both classes), residual intensity non-uniformity, and low grey-white contrast from the sequence or the participant's age.
Caveats: The bias field is removed before it is measured, so it shows only the non-uniformity the correction left, not the raw field (see Intensity non-uniformity). Like every tissue measure, it depends on the segmentation.
Compared with MRIQC: Same formula as MRIQC's cjv (MADs over the gap between medians).
References: Ganzetti et al. 2016, Front Neuroinform

Artefacts

Signal where there should be only noise: in the air around the head. Lower is better for all but FBER. Ghosts, ringing, motion and wrap-around put signal outside the head, where a clean scan has only noise. EFC, FBER, QI1, QI2 and the ghost ratio each look at that background in a different way.

Entropy focus criterion

How evenly the image's energy is spread over its voxels (Atkinson 1997). Lower is better: ghosting and motion blur spread energy into the background.

In detail

What it measures: The Shannon entropy of the voxel intensities, normalised: a sharp image concentrates its energy in the head, while ghosting and blurring spread it into the air.
How it is computed: Over every voxel except the zero fill, with x the intensities clipped at 0 and B = sqrt(sum x^2): E = sum (x/B) ln(x/B), divided by the value E takes when every voxel is equal. 0 means all the energy in one voxel, 1 an image of uniform intensity.
How to read it: Lower is better. Its value depends on the field of view and on the number of voxels, so compare only within one protocol, as the 'In this dataset' column does.
What raises or lowers it: Motion (ghosts and blurring), Nyquist or other ghosting, wrap-around, and noise in the background.
Caveats: Computed only on images whose air is real: a defaced, skull-stripped or zero-filled background changes the energy distribution for reasons that have nothing to do with the scan. Background noise raises it too, so a noisy but sharp image can read like a blurred one.
Compared with MRIQC: Same formula as MRIQC's efc; on the images compared during development the two agreed to within about 0.002.
References: Atkinson et al. 1997, IEEE Trans Med Imaging

Foreground to background energy

The head's typical energy over the air's (Shehzad 2015). Higher is better.

In detail

What it measures: How much brighter the head is than its background, in energy (squared intensity).
How it is computed: The median of the squared intensities inside the head mask over the median of the squared intensities everywhere outside it, zero fill excluded. Unlike the other air measures it uses all the background, below the face as well as above it.
How to read it: Higher is better. It is large on a clean scan with a dark background and falls when signal leaks into the air. Compare within one protocol: the head mask, the field of view and the scanner's background handling all move it.
What raises or lowers it: Ghosting, wrap-around, motion and noise in the air lower it, and so do flow artefacts that land outside the head.
Caveats: Not computed when the air is not real (defaced, skull-stripped, zero-filled) or when the air's energy is zero. A scanner or tool that filters or scales the background raises it without any change in the head.
Compared with MRIQC: Same formula as MRIQC's fber (medians of squared intensities). Where MRIQC returns -1 for an air of zero energy, BIDS Manager leaves the value empty.
References: Shehzad et al. 2015, Front Neurosci (Neuroinformatics 2015 conference abstract)

Artefacts in the air (QI1)

The share of the air above the face holding structured signal: ghosts, ringing, motion, wrap-around (Mortamet 2009). Lower is better.

In detail

What it measures: How much of the background holds signal that is not noise, as a percentage of the air measured.
How it is computed: On the working grid (about 2 mm), an air voxel counts when its intensity is more than 10 times the air's robust spread (its median absolute deviation scaled to an SD) and it lies further from the head than a tenth of the largest distance in the air. Isolated voxels are then removed by a morphological opening. The value is what remains as a percentage of the air above the face.
How to read it: Lower is better; 0 is usual for a clean scan. Above 0.5 % by default (Settings > Quality control > Anatomical images) the check reports structured signal in the air, and the 'Artefacts in the air' layer shows where it is.
What raises or lowers it: Motion ghosts along the phase-encoding direction, ringing, wrap-around (part of the head folded into the far side of the field of view), and objects that give signal, such as a fiducial marker.
Caveats: Only the air above the plane through the glabella and the inion is used: below it, folding and motion of the face, neck and shoulders are normal. Air close to the head is left out, so a ghost that overlaps the head itself is not counted. Not computed when the air is not real (defaced, skull-stripped, zero-filled).
Compared with MRIQC: MRIQC reports qi_1 with the same intent, but BIDS Manager's development notes record that MRIQC's artefact mask is always empty, so its qi_1 reads 0. This value follows Mortamet's definition as MRIQC meant it.
References: Mortamet et al. 2009, Magn Reson Med

Noise fit (QI2)

How poorly a chi-squared distribution fits the upper tail of the air's intensities (Mortamet 2009). Lower is better.

In detail

What it measures: Whether the background holds only noise. Pure noise in a magnitude image has a smooth distribution of a known family; structured signal in the air adds an upper tail that no such fit follows.
How it is computed: The air's positive intensities at the image's resolution (up to 100,000 sampled) are scaled so that their 99th percentile is 100, their density is estimated with a Gaussian kernel, and a chi-squared distribution is fitted to them. The value is the mean absolute difference between the density and the fit over the upper tail, beyond the last point where the density is above half its peak.
How to read it: Lower is better: small values mean the air looks like noise, larger ones that something else is there. Read it with QI1 and the 'Artefacts in the air' layer, which show where.
What raises or lowers it: Ghosts, ringing, wrap-around and motion artefacts in the air, and a background the scanner filtered or rescaled, which changes the noise's shape.
Caveats: Not computed with fewer than 1000 air voxels or when the air is not real. Multi-channel coils and parallel imaging change the noise distribution itself, so the fit can be poor without any artefact: compare within one protocol.
Compared with MRIQC: Same algorithm as MRIQC's qi_2; BIDS Manager estimates the density in numpy rather than with a library kernel estimator.
References: Mortamet et al. 2009, Magn Reson Med

Ghost to signal ratio

The extra signal where a ghost of the head would fall along the phase-encoding direction, over the head's mean. Lower is better.

In detail

What it measures: Ghosting: faint copies of the head displaced along the phase-encoding direction. It compares the air where a copy shifted by half the field of view would land with the rest of the air.
How it is computed: On the working grid, the head mask is shifted by half the field of view along the phase-encoding axis (PhaseEncodingDirection, else InPlanePhaseEncodingDirectionDICOM, from the sidecar) and intersected with the measured air. The value is (mean of that ghost region - mean of the rest of the air) / mean of the head.
How to read it: Lower is better. Near 0 the ghost region is no brighter than the rest of the air, and a slightly negative value is noise; a clear positive value means signal repeats at the half-field position.
What raises or lowers it: Nyquist (N/2) ghosting of echo-planar readouts, and motion or pulsation ghosts that reach the half-field position.
Caveats: Anatomical sequences are rarely echo-planar, and their motion ghosts repeat at spacings set by the motion, so only the part that reaches the half-field position counts; the 'Artefacts in the air' layer shows ghosts wherever they are. Not computed without a phase-encoding direction in the sidecar, with too little air at the ghost position, or when the air is not real.
Compared with MRIQC: No anatomical equivalent in MRIQC; its functional report has a ghost-to-signal ratio of the same idea (gsr_x, gsr_y).

Tissues

What the segmentation found: how the intensity range is used, how uniform the image is, the tissue shares, partial volume, and agreement with the template. These describe the tissue classes and the bias field rather than the noise. They flag a failed segmentation or registration, a strongly non-uniform image, and blur.

White matter to maximum

White matter's median over the top of the image's intensity range. Neither direction is simply better: very low values mean a few very bright voxels take up the range.

In detail

What it measures: Where white matter sits in the intensity range: whether the range is used by the brain or taken up by a long tail of bright voxels (vessels, fat, an artefact).
How it is computed: White matter's bias-corrected, probability-weighted median over the 99.9th percentile of all non-zero voxels of the stored image.
How to read it: Neither: read it against images of the same contrast and protocol. MRIQC's documentation gives about 0.6 to 0.8 as usual for a T1w, where fat in the scalp and marrow is brighter than white matter. Much lower values mean a long bright tail; values near or above 1 mean nothing is brighter than white matter, as in a skull-stripped or fat-suppressed image. On a T2w white matter is dark and the value is low by nature.
What raises or lowers it: Bright vessels (especially with a contrast agent), fat, a bright artefact or a hot spot near a coil element stretch the top of the range; intensity rescaling at conversion can change it too.
Caveats: The numerator is bias-corrected and the denominator comes from the uncorrected image, so a strong bias field moves the ratio. Needs the tissue classes.
Compared with MRIQC: Same idea as MRIQC's wm2max (white matter's median over a high percentile of the image); values differ through the segmentation and the exact ceiling.

Intensity non-uniformity

How much the intensity bias field varies over the brain: its 95th minus 5th percentile, the field normalised to 1. Lower is better; 0 is uniform.

In detail

What it measures: The slow, smooth change of brightness across the image that comes from the coil's sensitivity and the RF field rather than from the tissue.
How it is computed: The bias field is fitted with the tissue classes on the working grid (about 2 mm), N3's idea: a smooth fit, a third-order polynomial plus a Gaussian smoothing (sigma about 10 mm on the default grid) of what the polynomial leaves, to the part of the log image the classes do not explain. It is turned into a multiplier inside the brain and normalised to median 1; the value is its 95th percentile minus its 5th, so 0.2 means those two percentiles of the field lie 20 % of the median apart.
How to read it: Lower is better, within a protocol. Large values are common with multi-channel head coils and at higher field strengths when the scanner did not normalise the intensities, and smaller when it did.
What raises or lowers it: Receive-coil sensitivity (strong near the elements of a multi-channel array), transmit (B1) inhomogeneity, which grows with field strength, and the head's position in the coil.
Caveats: The smooth fit can also absorb real, slowly varying tissue differences (deep grey nuclei against the cortex, for example), so part of the value can be anatomy. The tissue measures are taken after this field is removed, so a large value does not by itself lower the SNRs: it says how much the correction had to do.
Compared with MRIQC: Not comparable with MRIQC's inu_range or inu_med, which read N4's field on its own scale.
References: Sled et al. 1998, IEEE Trans Med Imaging

CSF fraction

The share of the brain mask classified as CSF. Neither direction: a check on the segmentation and the anatomy.

In detail

What it measures: CSF's share of the segmented brain volume; the three fractions add up to 1.
How it is computed: The sum of the CSF probabilities over the brain on the working grid, over the sum of all three classes. With the tissue network, CSF is everything inside the brain mask that the network does not label grey or white matter.
How to read it: Neither: compare with the dataset's other images of the same contrast. It rises with age and atrophy (larger ventricles and sulci). An outlier more often means a brain mask or a segmentation that failed than an unusual brain: check the 'Tissue classes' layer.
What raises or lowers it: How far the brain mask reaches (extra-cerebral CSF, dura and vessels inside it count as CSF), the segmentation method, partial volume, and anatomy: age, atrophy, enlarged ventricles.
Caveats: Not a clinical volume: the brain mask is a quality-check mask, not an intracranial volume. Values from the networks and from the fast fallback differ.
Compared with MRIQC: Same idea as MRIQC's icvs_csf (each class's share of the total).

Grey matter fraction

The share of the brain mask classified as grey matter. Neither direction: a check on the segmentation and the anatomy.

In detail

What it measures: Grey matter's share of the segmented brain volume.
How it is computed: The sum of the grey-matter probabilities over the brain on the working grid, over the sum of all three classes.
How to read it: Neither: compare with the dataset's other images of the same contrast. It falls with age as the cortex thins, and depends strongly on where the segmentation puts the border with white matter. An outlier more often means a segmentation problem than unusual anatomy.
What raises or lowers it: The segmentation method, grey-white contrast (low contrast moves the border), partial volume, blur, and anatomy.
Caveats: Not a clinical volume. Values from the networks and from the fast fallback differ.
Compared with MRIQC: Same idea as MRIQC's icvs_gm.

White matter fraction

The share of the brain mask classified as white matter. Neither direction: a check on the segmentation and the anatomy.

In detail

What it measures: White matter's share of the segmented brain volume.
How it is computed: The sum of the white-matter probabilities over the brain on the working grid, over the sum of all three classes.
How to read it: Neither: compare with the dataset's other images of the same contrast. It changes over the lifespan (white matter grows through childhood and adolescence and shrinks in older age). An outlier more often means a segmentation problem than unusual anatomy.
What raises or lowers it: The segmentation method, grey-white contrast, partial volume, lesions, and anatomy.
Caveats: Not a clinical volume. White-matter lesions can be classed as grey matter or CSF and lower it. Values from the networks and from the fast fallback differ.
Compared with MRIQC: Same idea as MRIQC's icvs_wm.

Partial volume

The share of brain voxels that no tissue class holds with a probability of 0.9 or more. Lower is better.

In detail

What it measures: How many voxels are mixtures of tissues or undecided: voxels at the borders between tissues, and voxels that noise or blur leaves uncertain.
How it is computed: At the image's own resolution, on the sampled brain voxels, with the class probabilities of the estimate made again at that resolution (the intensity's likelihood under each class times a prior from the tissue network or the working-grid classes). A voxel counts when its largest probability is below 0.9 by default (Settings > Quality control > Anatomical images).
How to read it: Lower is better, against scans of the same voxel size and contrast. A rise shows blur, motion, thicker voxels or more noise, all of which leave more voxels undecided.
What raises or lowers it: Thick or large voxels, motion blur, reconstruction filtering or resampling, noise, and low grey-white contrast.
Caveats: Partly anatomy: the folded cortex always leaves many border voxels, and a brain with wider sulci has more. It depends on the segmentation's priors, so values from the networks and from the fast fallback differ.
Compared with MRIQC: Not MRIQC's rpve, which is not a fraction and is not comparable.

Overlap with the template (grey matter)

How well the grey-matter map matches the template's after an affine registration (fuzzy Jaccard index). Higher is better.

In detail

What it measures: Agreement between this image's grey-matter map and the template's (MNI 2009c) brought onto the image.
How it is computed: The template's tissue maps are carried onto the working grid by the affine registration (niimath -allineate, or BIDS Manager's own fit). The value is sum(min(p, q)) / sum(max(p, q)) over the image, p the image's grey-matter probabilities and q the template's.
How to read it: Higher is better, against the dataset. Values well below 1 are normal, since an affine registration cannot match individual folding. A clear outlier points at a failed registration or segmentation, or at anatomy far from the template.
What raises or lowers it: A registration that failed or is uncertain (reported as a finding), a wrong segmentation, a cut field of view, large ventricles, lesions, or a brain shape far from the adult template.
Caveats: It measures the registration and the segmentation together and is not an image-quality measure by itself. Not computed when the template is not registered (slabs thinner than 70 mm by default, Settings > Quality control > Anatomical images).
Compared with MRIQC: MRIQC's tpm_overlap_gm is the same fuzzy overlap; values differ through the registration and the segmentation.
References: Crum et al. 2006, IEEE Trans Med Imaging

Overlap with the template (white matter)

How well the white-matter map matches the template's after an affine registration (fuzzy Jaccard index). Higher is better.

In detail

What it measures: Agreement between this image's white-matter map and the template's (MNI 2009c) brought onto the image.
How it is computed: As for grey matter: sum(min(p, q)) / sum(max(p, q)) over the working grid, p the image's white-matter probabilities and q the template's, carried by the affine registration.
How to read it: Higher is better, against the dataset. White matter is a larger, smoother structure than the cortex, so it usually overlaps better than grey matter; a clear outlier points at a failed registration or segmentation.
What raises or lowers it: A failed or uncertain registration, a wrong segmentation, a cut field of view, enlarged ventricles, lesions, or anatomy far from the adult template.
Caveats: It measures the registration and the segmentation together. Not computed when the template is not registered (slabs thinner than 70 mm by default).
Compared with MRIQC: MRIQC's tpm_overlap_wm is the same fuzzy overlap; values differ through the registration and the segmentation.
References: Crum et al. 2006, IEEE Trans Med Imaging

Coverage and smoothness

Whether the field of view holds the whole brain, and how smooth the image is along each axis. The share of the brain that the field of view cuts off, from the registered template, and the smoothness along each voxel axis and its mean, from the image's own correlation between neighbours.

Smoothness along x

The full width at half maximum of the image's spatial correlation along x, in mm (Forman 1995). Lower means sharper, but noise lowers it too.

In detail

What it measures: How smooth the image is along one axis: how strongly neighbouring voxels resemble each other, expressed as the width of an equivalent Gaussian blur.
How it is computed: Inside the brain mask, at the image's own resolution, on the stored intensities: rho = 1 - var(differences between neighbours along x) / (2 var(intensities)), then FWHM = voxel size x sqrt(-2 ln 2 / ln rho). The axis is the image's own voxel axis, which is left-right, back-front or down-up only when the image is stored that way.
How to read it: Lower is sharper. It is in millimetres, so larger voxels give larger values: compare only within a protocol. The brain's own structure (grey against white matter) makes neighbours alike, so the value is several millimetres even for a sharp scan and is not the scanner's blur alone.
What raises or lowers it: Motion blur, reconstruction filtering, zero-filling interpolation and partial Fourier, any resampling (reorientation, registration, or a tool that rewrote the image), and larger voxels raise it.
Caveats: Noise makes neighbours less alike and lowers it, so a noisy image reads sharper: read it beside the SNRs. A difference between the axes is expected when the voxels or the resolution differ along one of them.
Compared with MRIQC: MRIQC reports fwhm_x (AFNI's 3dFWHMx) in voxels; BIDS Manager reports millimetres from Forman's estimator.
References: Forman et al. 1995, Magn Reson Med

Smoothness along y

The full width at half maximum of the image's spatial correlation along y, in mm (Forman 1995). Lower means sharper, but noise lowers it too.

In detail

What it measures: How smooth the image is along one axis: how strongly neighbouring voxels resemble each other, expressed as the width of an equivalent Gaussian blur.
How it is computed: Inside the brain mask, at the image's own resolution, on the stored intensities: rho = 1 - var(differences between neighbours along y) / (2 var(intensities)), then FWHM = voxel size x sqrt(-2 ln 2 / ln rho). The axis is the image's own voxel axis, which is left-right, back-front or down-up only when the image is stored that way.
How to read it: Lower is sharper. It is in millimetres, so larger voxels give larger values: compare only within a protocol. The brain's own structure (grey against white matter) makes neighbours alike, so the value is several millimetres even for a sharp scan and is not the scanner's blur alone.
What raises or lowers it: Motion blur, reconstruction filtering, zero-filling interpolation and partial Fourier, any resampling (reorientation, registration, or a tool that rewrote the image), and larger voxels raise it.
Caveats: Noise makes neighbours less alike and lowers it, so a noisy image reads sharper: read it beside the SNRs. A difference between the axes is expected when the voxels or the resolution differ along one of them.
Compared with MRIQC: MRIQC reports fwhm_y (AFNI's 3dFWHMx) in voxels; BIDS Manager reports millimetres from Forman's estimator.
References: Forman et al. 1995, Magn Reson Med

Smoothness along z

The full width at half maximum of the image's spatial correlation along z, in mm (Forman 1995). Lower means sharper, but noise lowers it too.

In detail

What it measures: How smooth the image is along one axis: how strongly neighbouring voxels resemble each other, expressed as the width of an equivalent Gaussian blur.
How it is computed: Inside the brain mask, at the image's own resolution, on the stored intensities: rho = 1 - var(differences between neighbours along z) / (2 var(intensities)), then FWHM = voxel size x sqrt(-2 ln 2 / ln rho). The axis is the image's own voxel axis, which is left-right, back-front or down-up only when the image is stored that way.
How to read it: Lower is sharper. It is in millimetres, so larger voxels give larger values: compare only within a protocol. The brain's own structure (grey against white matter) makes neighbours alike, so the value is several millimetres even for a sharp scan and is not the scanner's blur alone.
What raises or lowers it: Motion blur, reconstruction filtering, zero-filling interpolation and partial Fourier, any resampling (reorientation, registration, or a tool that rewrote the image), and larger voxels raise it.
Caveats: Noise makes neighbours less alike and lowers it, so a noisy image reads sharper: read it beside the SNRs. A difference between the axes is expected when the voxels or the resolution differ along one of them.
Compared with MRIQC: MRIQC reports fwhm_z (AFNI's 3dFWHMx) in voxels; BIDS Manager reports millimetres from Forman's estimator.
References: Forman et al. 1995, Magn Reson Med

Smoothness (mean)

The mean of the smoothness along the three voxel axes, in mm. Lower means sharper, but noise lowers it too.

In detail

What it measures: How smooth the image is overall: how strongly neighbouring voxels resemble each other, as the width of an equivalent Gaussian blur averaged over the three axes.
How it is computed: The mean of the three per-axis values (any that could not be computed is left out). Each is Forman's estimate inside the brain mask at the image's own resolution: rho = 1 - var(neighbour differences) / (2 var(intensities)), then FWHM = voxel size x sqrt(-2 ln 2 / ln rho).
How to read it: Lower is sharper. It is in millimetres, so larger voxels give larger values: compare only within a protocol. The brain's own structure (grey against white matter) makes neighbours alike, so the value is several millimetres even for a sharp scan and is not the scanner's blur alone.
What raises or lowers it: Motion blur, reconstruction filtering, zero-filling interpolation and partial Fourier, any resampling (reorientation, registration, or a tool that rewrote the image), and larger voxels raise it.
Caveats: Noise makes neighbours less alike and lowers it, so a noisy image reads sharper: read it beside the SNRs. A difference between the axes is expected when the voxels or the resolution differ along one of them.
Compared with MRIQC: MRIQC reports fwhm_avg (AFNI's 3dFWHMx) in voxels; BIDS Manager reports millimetres from Forman's estimator.
References: Forman et al. 1995, Magn Reson Med

Brain outside the field of view

The share of the template's brain that falls outside the image after registration: a cut top of the head or cerebellum. Lower is better.

In detail

What it measures: Whether the field of view covers the whole brain.
How it is computed: Every brain voxel of the template is carried through the registration into the image's voxel grid; the value is the percentage that lands outside it. The finding names the sides (top, bottom, front, back, left, right).
How to read it: Lower is better; 0 is a complete brain. Above 0.5 % by default the check warns and above 3 % it reports an error (Settings > Quality control > Anatomical images). A slab thinner than 120 mm by default covers part of the brain on purpose, and the same value is then reported as coverage.
What raises or lowers it: A field of view placed too low or too high, a small field of view, a large head, or a tilted head.
Caveats: Only as good as the registration: an uncertain registration (reported as a finding) makes the value uncertain too. It counts the template's brain, not the participant's own, so it depends on how well the fit holds at the edges. Not computed for slabs thinner than 70 mm by default, where no template is registered.

Saturated voxels

The share of the head at the image's maximum value: bright tissue clipped at the top of the stored range. Lower is better.

In detail

What it measures: Whether intensities were clipped: many voxels sharing exactly the highest value mean that the scanner or a conversion ran out of range.
How it is computed: At the image's resolution, inside the head mask: the percentage of voxels whose value equals the head's maximum.
How to read it: Lower is better; in an unclipped image it is a tiny fraction of a percent. Above 0.5 % by default (Settings > Quality control > Anatomical images) the check reports saturated voxels.
What raises or lowers it: A receiver or reconstruction scaling that overflowed the stored range, a conversion that rescaled and clipped, or very bright fat or vessels at the limit of the data type.
Caveats: Detail above the clipped value is lost and cannot be recovered. An image with very few distinct values (heavily quantised, or a mask) also has many voxels at its maximum.

Diffusion images

Gradient table

Whether the .bval and .bvec describe the series and cover the directions evenly. A wrong table ruins every analysis and no image measure shows it. The table itself, before any image: present, one entry per volume, unit vectors, at least one b=0, the shells, repeated directions, how evenly each shell covers the sphere, and (experimental) whether the vectors look flipped or swapped.

Largest gap between directions

The radius, in degrees, of the largest region of the sphere with no gradient direction of this shell. Lower is better.

In detail

What it measures: How evenly the shell's directions cover the sphere, either sign counting (a direction and its opposite measure the same thing).
How it is computed: 4000 random points on the sphere are each matched with the shell's nearest direction, either sign; the value is the largest of those angles, the radius of the biggest empty cap.
How to read it: Lower is better, but the number of directions sets a floor: few directions cannot cover the sphere closely however they are placed. Compare with shells of the same number of directions; a value far above a well-spread scheme of the same count means clustered directions or missing volumes.
What raises or lowers it: Few directions, a poorly designed scheme, and volumes dropped at conversion or by a scan stopped early.
Caveats: A property of the table, not of the image: it says nothing about whether the images match the table. It is estimated from random points, so it can be a little below the true largest gap.
References: Jones, Horsfield and Simmons 1999, Magn Reson Med

Motion

How the head moved during the series: from the b=0 volumes, the reliable record, and from every volume registered to what the tensor predicts for it. Displacement across the b=0 volumes, the largest step between them, volumes far out of place, the mean framewise displacement, eddy-current shifts, the largest rotation, and the drift of the b=0 signal.

Head motion across the b=0 volumes

The largest displacement of any b=0 volume from the first, in mm. Lower is better.

In detail

What it measures: How far the head moved over the scan, as the b=0 volumes record it.
How it is computed: Each b=0 volume is registered rigidly to the median b=0 image. Its position relative to the first b=0 is summed as the three absolute translations plus the three absolute rotations as arcs on a 50 mm sphere by default (Settings > Quality control > Diffusion); the value is the largest over the b=0 volumes.
How to read it: Lower is better. Above 1 mm by default (same section) the check reports head motion, as an error from three times that. With b=0 volumes spread through the series it tracks motion across the whole scan; with all of them at the start it says little about the rest.
What raises or lowers it: Head motion, including slow settling into the pillow over a long scan.
Caveats: With a single b=0 volume it reads 0, with nothing to compare. It is a displacement from the first b=0, not a frame-to-frame change, so slow drift and a single jump read alike: see the 'Displacement' plot.

Largest step between b=0 volumes

The largest change of head position between consecutive b=0 volumes, in mm. Lower is better.

In detail

What it measures: The biggest movement between one b=0 volume and the next.
How it is computed: Framewise displacement (Power 2012) between each b=0 volume and the next b=0, from the same registration: absolute changes of the three translations plus absolute changes of the three rotations as arcs on a 50 mm sphere by default (Settings > Quality control > Diffusion); the largest is reported.
How to read it: Lower is better, against the same protocol. Consecutive b=0 volumes can be many volumes apart, so this is the movement over that interval, not between two neighbouring volumes.
What raises or lowers it: Head motion between the b=0 volumes.
Caveats: Needs two b=0 volumes or more. It misses a movement that returns to the same position between two b=0 volumes.
References: Power et al. 2012, NeuroImage

Volumes far out of place

The number of volumes displaced far from the first b=0 and far from the rest of their shell. Lower is better; 0 is usual.

In detail

What it measures: Volumes during which the head probably moved.
How it is computed: A volume counts when its displacement from the first b=0 is above 2 mm by default (Settings > Quality control > Diffusion) and its robust z among the volumes of its own shell (or among the b=0 volumes) is above 5. A shell with fewer than four volumes is not judged.
How to read it: Lower is better. Each such volume probably moved during its own acquisition; the 'Displacement' plot marks them. When they are more than 5 % of the volumes the finding is an error rather than a warning.
What raises or lowers it: A sudden movement, a cough or a swallow during that volume; a registration that failed on a low-signal volume can look the same.
Caveats: Each volume is judged within its shell, so a scan in which every volume moved by the same amount shows none far out of place: read the motion across the b=0 volumes too.

Mean framewise displacement

The mean framewise displacement over the whole series, in mm. Lower is better, against scans of the same protocol.

In detail

What it measures: The average volume-to-volume movement of the head over the series.
How it is computed: Framewise displacement (Power 2012) between every pair of consecutive volumes, after the eddy-current part is removed, averaged. Each diffusion-weighted volume is registered to what the tensor predicts for its direction, each b=0 to the b=0 reference.
How to read it: Lower is better. It includes the registration's own noise on single low-signal volumes, larger at high b-values, so it reads higher than a BOLD run's FD for the same head motion: compare with other diffusion scans of the same protocol, not with a BOLD threshold such as 0.5 mm.
What raises or lowers it: Head motion, and registration noise that grows as the signal falls.
Caveats: Not the motion an analysis would correct (FSL eddy or similar). A shell with few volumes or a poor tensor fit raises the registration noise.
Compared with MRIQC: Shown beside MRIQC's fd_mean; the two come from different registrations, so compare within one tool.
References: Power et al. 2012, NeuroImage

Eddy-current shift

The largest shift of a diffusion-weighted volume explained by its gradient direction, in mm: eddy currents. Lower is better.

In detail

What it measures: Eddy currents from the diffusion gradients shift (and also stretch and shear) each diffusion-weighted image, mostly along the phase-encoding axis, by an amount that depends on the gradient's direction. A rigid registration reads the shift as head motion.
How it is computed: In each shell of at least 10 volumes, the translations are regressed on the gradient's three components; when the gradient explains at least 30 % of their variance by default (Settings > Quality control > Diffusion), the explained part is taken as eddy current and removed from the motion. The value is the largest absolute component of that part over every volume, and 0 when no shell passed.
How to read it: Lower is better. A shift of a voxel or more means distortion correction (FSL eddy or similar) is needed; that is routine for diffusion data and not a fault of the scan.
What raises or lowers it: Strong, fast-switching diffusion gradients, high b-values, and the scanner's eddy-current compensation.
Caveats: Only the shift is measured, not the stretch and shear eddy currents also cause. A shell with few volumes, or head motion that happens to follow the gradient direction, can be misread.
References: Andersson and Sotiropoulos 2016, NeuroImage

Largest rotation of the gradients

The largest rotation of the head against the first b=0, in degrees. Lower is better: the gradient directions turn with the head.

In detail

What it measures: How far the head turned, and so how far every gradient direction turned relative to the head.
How it is computed: The angle of each volume's rotation (the length of its rotation vector) relative to the first b=0, from the same registration; the largest is reported.
How to read it: Lower is better. Above 3 degrees the check adds a note: when motion is corrected, the b-vectors should be rotated with it, or the directions are wrong by up to that angle.
What raises or lowers it: Head rotation, most often nodding.
Caveats: From a screening registration: small values are within its noise.
References: Leemans and Jones 2009, Magn Reson Med

Signal drift over the scan

The change of the b=0 signal from the first volume to the last, in percent. Closer to zero is better; the sign says which way.

In detail

What it measures: A slow change of the scanner's signal over the series, usually from heating. Uncorrected, it biases every diffusion measure, because diffusion weighting is read as signal lost against the b=0.
How it is computed: The median brain signal of each b=0 volume over the first's; a straight line is fitted to the log of that ratio against the volume index, and the value is the change the fit gives from the first volume to the last: (exp(slope x (n - 1)) - 1) x 100.
How to read it: Closer to zero is better. Above 5 % by default (Settings > Quality control > Diffusion) the check warns: correct the drift before fitting. The 'b=0 signal' plot shows the points the fit went through.
What raises or lowers it: Gradient and shim heating over a long, gradient-intensive scan.
Caveats: Needs two b=0 volumes or more, and is only as good as their spread through the series: with every b=0 at the start, the fit is extrapolated over the rest. Drift is often not exponential, and motion also changes the median brain signal.
References: Vos et al. 2017, Magn Reson Med

Noise and contrast

How large the diffusion signal is against the noise: two noise estimates, the SNR in the corpus callosum per shell, and how much of the highest shell sits at the noise floor. The noise SD, by MP-PCA and from repeated b=0 volumes; the corpus callosum's signal over that noise at b=0 and, per shell, along and across its fibres; and the share of tissue signal at the highest shell too close to the noise to trust.

Noise from the b=0 repeats

The noise SD from the variation across repeated b=0 volumes, in scanner units. Neither direction on its own; motion adds to it.

In detail

What it measures: How much the same voxel varies between b=0 volumes that should be identical.
How it is computed: In the brain mask eroded by two voxels, each voxel's standard deviation (with n - 1) across the b=0 volumes; the value is the median over voxels. Needs three b=0 volumes or more.
How to read it: Neither on its own: it is in the scanner's units, so read it through the SNR measures, or against scans of the same protocol, where lower is better. Clearly above the MP-PCA estimate, it says the b=0 volumes differ by more than noise: motion, drift or pulsation.
What raises or lowers it: Thermal noise, plus motion between the b=0 volumes, signal drift, and cardiac pulsation (near the ventricles and the brainstem especially).
Caveats: An upper bound on the noise when the head moves. Few b=0 volumes give a noisy estimate. It is used for the SNRs only when MP-PCA could not be computed.

Noise by MP-PCA

The noise SD by Marchenko-Pastur PCA (Veraart 2016) on sampled patches of the largest shell. Neither direction on its own: it sets the SNRs.

In detail

What it measures: The thermal noise level, separated from the signal by the statistics of random matrices: the part of a patch's variation across volumes that behaves like pure noise.
How it is computed: Up to 300 patches (5 x 5 x 5 voxels, or 7 x 7 x 7 when the shell has 100 volumes or more) are sampled in the eroded brain mask, keeping those at least 80 % inside it. In each, the largest eigenvalues of the voxels-by-volumes matrix are set aside as signal, one by one, until the rest spread as the Marchenko-Pastur law allows for pure noise; the noise variance is the mean of the rest. The value is the median over patches; it needs six volumes or more in the shell.
How to read it: Neither on its own: it is in scanner units. It is the noise the callosal SNRs and the noise-floor measure use; a value below the b=0 estimate is expected, because motion between b=0 volumes does not enter it.
What raises or lowers it: Thermal noise: small voxels, high acceleration, few averages, low coil sensitivity.
Caveats: It assumes noise that is independent between voxels and the same within a patch: parallel imaging, partial Fourier, interpolation and filtering at reconstruction correlate the noise and bias the estimate. At low SNR, magnitude noise is not the Gaussian noise the method assumes.
Compared with MRIQC: Shown beside MRIQC's sigma_pca, a noise estimate of the same family; BIDS Manager samples 300 patches of one shell, so do not expect identical values.
References: Veraart et al. 2016, NeuroImage; Veraart, Fieremans and Novikov 2016, Magn Reson Med

SNR in the corpus callosum (b=0)

The mean b=0 signal in the corpus callosum over the noise SD. Higher is better.

In detail

What it measures: How far the b=0 signal of one well-defined white-matter tract stands above the noise.
How it is computed: The mean over the callosal voxels and over every b=0 volume, divided by the MP-PCA noise SD, or by the b=0 estimate when MP-PCA could not be computed.
How to read it: Higher is better, against the same protocol. It is the reference the diffusion-weighted SNRs fall from.
What raises or lowers it: Small voxels, long echo times, high acceleration, few averages, and low coil sensitivity in the centre of the head.
Caveats: Needs a tensor fit and at least 5 callosal voxels (FA above 0.4, principal direction within about 32 degrees of left-right, within 10 mm of the brain's midline). The corpus callosum is deep in the head, where a multi-channel coil is least sensitive, so its SNR is not the brain's best.
Compared with MRIQC: MRIQC reports a callosal SNR at b=0 too (snr_cc_shell0). BIDS Manager's notes record a shell-labelling defect in MRIQC's per-shell callosal SNRs, so compare with care.

SNR in the corpus callosum, along the fibres

The corpus callosum's signal in the shell's volume with the most left-right gradient, over the noise SD. Higher is better; this is the worst case.

In detail

What it measures: How far the most attenuated signal of the shell stands above the noise: diffusion is fastest along the callosal fibres, so a gradient along them attenuates the signal most.
How it is computed: One volume: the shell's volume whose gradient, in world terms, is closest to left-right. Its mean over the callosal voxels, over the noise SD (MP-PCA, else the b=0 estimate).
How to read it: Higher is better, against the same protocol. Near or below 2 (the threshold the noise-floor measure uses), the signal is at the noise floor of magnitude data, and anything that depends on it (axial diffusivity, high-b models) is biased.
What raises or lowers it: Higher b-values, smaller voxels, longer echo times, acceleration, few averages.
Caveats: From one volume, so it is noisier than the across value. If the table's axes are swapped, the volume picked is not the one along the fibres. Needs a tensor fit and at least 5 callosal voxels (FA above 0.4, principal direction within about 32 degrees of left-right, within 10 mm of the brain's midline). The corpus callosum is deep in the head, where a multi-channel coil is least sensitive, so its SNR is not the brain's best.
References: Gudbjartsson and Patz 1995, Magn Reson Med

SNR in the corpus callosum, across the fibres

The corpus callosum's signal in the shell's two volumes with the least left-right gradient, over the noise SD. Higher is better; the least attenuated.

In detail

What it measures: How far the least attenuated signal of the shell stands above the noise: a gradient across the callosal fibres meets the slowest diffusion.
How it is computed: The mean over the callosal voxels and over the shell's two volumes whose gradient, in world terms, has the smallest left-right component, over the noise SD (MP-PCA, else the b=0 estimate).
How to read it: Higher is better, against the same protocol. Read it beside the along value: the gap between them reflects the callosum's anisotropy, and a shell where both are low is limited by noise.
What raises or lowers it: Higher b-values, smaller voxels, longer echo times, acceleration, few averages.
Caveats: Needs a tensor fit and at least 5 callosal voxels (FA above 0.4, principal direction within about 32 degrees of left-right, within 10 mm of the brain's midline). The corpus callosum is deep in the head, where a multi-channel coil is least sensitive, so its SNR is not the brain's best.

Signal at the noise floor

The share of tissue signal at the highest shell below twice the noise SD, where magnitude data are biased upwards (the Rician floor). Lower is better.

In detail

What it measures: How much of the highest-b signal is too small to measure reliably. A magnitude image cannot go below the noise: where the true signal is comparable with the noise, what is stored is biased upwards.
How it is computed: Over brain voxels with an MD below 1.5 um^2/ms (tissue, not CSF, whose signal is meant to vanish at high b) and over every volume of the highest shell: the percentage of values below 2 times the noise SD (MP-PCA, else the b=0 estimate).
How to read it: Lower is better, against the same protocol. Some share is expected at a high b-value for directions along the fibres; a large share means models fitted to that shell (kurtosis, multi-compartment) are biased unless they account for the noise floor.
What raises or lowers it: High b-values, small voxels, long echo times, low coil sensitivity, few averages.
Caveats: Twice the noise SD is a reading aid, not a sharp limit: the bias grows smoothly as the signal falls toward the noise. Needs a tensor fit and a noise estimate.
References: Gudbjartsson and Patz 1995, Magn Reson Med

Artefacts

Artefacts found by comparing each slice and voxel with what the tensor predicts (dropout, interleave, spikes), plus EFC and FBER per shell. Slices that lost signal in one volume, volumes whose odd and even slices disagree, single voxels far above their prediction, and the ghosting and background of each shell's mean image.

Slices with signal dropout

The number of slices, across all volumes, whose signal falls far below what the tensor predicts. Lower is better; 0 is usual.

In detail

What it measures: Slices that lost signal in one volume and not in the shell's others.
How it is computed: Per slice and volume, the mean over the slice's brain voxels of (log observed - log predicted). Within each shell (and among the b=0 volumes) each slice's values become a robust z across the shell's volumes. A dropout has a z below -5 by default and is at least 10 % below the slice's typical value by default (both Settings > Quality control > Diffusion); the value counts (slice, volume) pairs.
How to read it: Lower is better. Each dropout is a slice of one volume to replace or exclude; when more than 5 % of the volumes hold one, the finding is an error. The blue cells of the 'Slice signal' plot are the dropouts; in a multiband (simultaneous multi-slice) acquisition, slices acquired together drop out together.
What raises or lowers it: Bulk motion during the diffusion encoding (cardiac pulsation in the brainstem and cerebellum, or the head moving), and table vibration with strong gradients.
Caveats: A dropout over part of a slice shows only if it lowers the slice's mean by the threshold. Slices with little brain (top and bottom) are not judged. The typical value for the 10 % test is the slice's median over every volume of every shell.
Compared with MRIQC: Not in MRIQC.
References: Andersson et al. 2016, NeuroImage

Volumes with an interleave artefact

The number of volumes whose odd slices disagree with their even slices far more than in the rest of the shell. Lower is better; 0 is usual.

In detail

What it measures: Motion between the two passes of an interleaved slice acquisition.
How it is computed: Per volume, the median residual against the tensor's prediction over odd slices minus that over even slices, as a robust z within the shell. A volume counts when |z| is above 5 by default (Settings > Quality control > Diffusion) and the difference is above 0.05 in log units (about 5 %).
How to read it: Lower is better. Such a volume shows stripes alternating slice by slice (a venetian-blind pattern) in a sagittal or coronal view.
What raises or lowers it: Head motion between the passes of an interleaved acquisition, which records odd and even slices at different times.
Caveats: Meaningful for interleaved acquisitions; with multiband or sequential ordering, odd and even slices are not two passes, and a flagged volume points at another slice-wise effect.
Compared with MRIQC: Not in MRIQC.

Spiking voxels

Voxels far above the tensor's prediction, per million brain voxels per volume, averaged over the volumes. Lower is better.

In detail

What it measures: Single voxels much brighter than they should be in one volume.
How it is computed: Each voxel's log residual (observed against predicted) in each volume of a shell is compared with its own median residual across that shell. A spike is more than 8 robust spreads above by default (Settings > Quality control > Diffusion), the spread floored at 0.05, and more than 0.5 in log units (about 65 % above). Counts per volume are scaled per million brain voxels and averaged.
How to read it: Lower is better; 0 is usual. The 'Spiking voxels' plot shows which volumes carry them.
What raises or lowers it: RF spikes (a spike in k-space shows as a striped pattern over the slice, parts of which cross the threshold), and motion that moves bright tissue into voxels that are usually dark at high b, such as CSF.
Caveats: An average over all volumes, so a single bad volume can look small: check the plot. Comparing each voxel with itself means the tensor's own misfit (CSF, crossing fibres) is not counted.
Compared with MRIQC: Not the same as MRIQC's spike measure: the module's notes record a defect in MRIQC's spike mask, so the two are not comparable.

Entropy focus criterion (per shell)

The entropy focus criterion of the shell's mean image (b=0: the median of the b=0 volumes). Lower is better.

In detail

What it measures: How evenly the energy of the shell's mean image is spread: ghosting and blurring spread it into the background.
How it is computed: As for anatomical images, over every voxel except the zero fill of the shell's mean image: E = sum (x/B) ln(x/B) with B = sqrt(sum x^2), divided by its value for a uniform image.
How to read it: Lower is better, against the same shell of scans of the same protocol. Higher b-values give darker images and different values, so compare shell with shell.
What raises or lowers it: Nyquist ghosts of the echo-planar readout, motion blur across the averaged volumes, and background noise.
Caveats: Not computed when the sidecar records defacing or the field of view holds fewer than 500 air voxels; the diffusion check does not detect a blank face or a skull-stripped image by itself. Averaging more volumes lowers the mean image's background noise and changes the value.
Compared with MRIQC: Same formula as MRIQC's efc.
References: Atkinson et al. 1997, IEEE Trans Med Imaging

Foreground to background energy (per shell)

The head's typical energy over the air's in the shell's mean image. Higher is better.

In detail

What it measures: How much brighter the head is than its background in the shell's mean image.
How it is computed: The median of the squared intensities inside the head mask (made on the b=0 image) over the median outside it by two voxels, zero fill excluded, on the shell's mean image (b=0: the median of the b=0 volumes).
How to read it: Higher is better, shell by shell, against the same protocol: diffusion weighting darkens the head, so FBER falls with b.
What raises or lowers it: Nyquist ghosts and any other signal in the background lower it.
Caveats: The same conditions as EFC. The air is all the background outside the head, face and neck included, because the diffusion check registers no template.
Compared with MRIQC: Same formula as MRIQC's fber.
References: Shehzad et al. 2015, Front Neurosci (Neuroinformatics 2015 conference abstract)

Diffusion

Sanity checks of the diffusion signal: the tensor's median FA and MD, how often the tensor is not physical, and whether neighbouring directions agree. FA and MD summarise the tensor over the brain; voxels with a negative eigenvalue count unphysical fits; the neighbouring-direction correlation says whether volumes with similar directions look alike.

Median FA in the brain

The median fractional anisotropy of the tensor over the brain mask. Neither direction: a sanity value to compare across the dataset.

In detail

What it measures: How directional diffusion is, from 0 (the same in every direction) to 1 (along one direction only), summarised over the brain.
How it is computed: A tensor is fitted in every brain voxel by weighted least squares (ordinary least squares on the log signal, then one step weighted by the squared prediction) on the b=0 volumes and b-values up to 1500 by default (Settings > Quality control > Diffusion). Negative eigenvalues are set to 0 before FA is computed; the value is the median over the brain mask.
How to read it: Neither: compare with the dataset's other diffusion scans of the same protocol. The median mixes white matter, grey matter and CSF, so it moves with the brain mask as much as with the tissue; an outlier points at the mask, the b-values, heavy noise or motion.
What raises or lowers it: Noise raises FA in tissue of low anisotropy; motion and artefacts disturb the fit; a looser or tighter brain mask changes the mix of tissues; age and pathology change the real anisotropy.
Caveats: A reversed or swapped b-vector table leaves FA unchanged, so this cannot detect one: see the flip check under Gradient table.
References: Basser and Pierpaoli 1996, J Magn Reson B

Median MD in the brain

The median mean diffusivity of the tensor over the brain mask, in um^2/ms. Neither direction: a sanity value.

In detail

What it measures: The average rate of diffusion over the three directions, summarised over the brain.
How it is computed: From the same tensor fit as FA, with negative eigenvalues set to 0: MD is their mean, multiplied by 1000 to read in um^2/ms (10^-3 mm^2/s when the b-values are in s/mm^2). The value is the median over the brain mask.
How to read it: Neither: compare across the dataset. Tissue reads about 0.7 to 0.9 and free water (CSF) about 3 at body temperature, so a whole-brain median sits a little above the tissue range when the mask holds some CSF. A value far off for every scan of a protocol points at the b-values: a wrong unit or scale in the .bval scales MD by the same factor.
What raises or lowers it: The brain mask's share of CSF, partial volume, atrophy, and errors in the b-values; signal at the noise floor lowers the apparent diffusivity.
Caveats: A reversed or swapped b-vector table leaves MD unchanged. b-values above the fitting limit (1500 by default) do not enter the fit.
References: Basser, Mattiello and LeBihan 1994, Biophys J

Voxels with a negative eigenvalue

The share of brain voxels where the fitted tensor has a negative eigenvalue, which is not physical. Lower is better.

In detail

What it measures: How often the tensor fit gives a negative diffusivity along some direction.
How it is computed: Before negative eigenvalues are set to 0, the percentage of brain voxels whose fitted tensor has at least one below zero.
How to read it: Lower is better. A small share is usual (noise in voxels with low radial diffusivity, voxels at the edge of the mask); a clear rise against the dataset means noise, motion or artefacts larger than the diffusion contrast, or b=0 volumes that do not match the diffusion-weighted ones.
What raises or lowers it: Low SNR, motion between volumes, signal dropout, a mask that includes voxels outside the brain, or a table that does not match the volumes (for example in the wrong order).
Caveats: FA computed after the clipping looks normal in these voxels; this measure is how they are seen.
Compared with MRIQC: Shown beside MRIQC's fa_degenerate, which targets the same failure (an unphysical fit) with a different definition; the two values are not interchangeable.
References: Basser, Mattiello and LeBihan 1994, Biophys J

Neighbouring direction correlation

The mean correlation of each diffusion-weighted volume with its nearest neighbour in q-space (Yeh 2019). Higher is better.

In detail

What it measures: Neighbouring directions should give similar images; a volume corrupted by motion, dropout or artefacts correlates less with its neighbour.
How it is computed: Each diffusion-weighted volume's neighbour is the volume nearest it in q-space (the square root of the b-value times the direction, either sign). The correlation is Pearson's over the brain voxels (every k-th voxel in a large brain); the value is the mean over the diffusion-weighted volumes.
How to read it: Higher is better, against the same protocol: it falls at higher b-values (lower SNR) and with fewer directions (neighbours further apart). The 'Nearest-direction correlation' plot shows which volumes fall below the rest.
What raises or lowers it: Motion, signal dropout, spikes and low SNR lower it, and so does misregistration between volumes.
Caveats: A mean over all volumes, so one bad volume barely moves it: look at the plot. Not comparable between protocols with different b-values or numbers of directions.
Compared with MRIQC: Same idea as MRIQC's ndc; values differ with the brain mask and the sampling.
References: Yeh et al. 2019, NeuroImage

Plots

Directions on the sphere

Each diffusion volume's gradient direction, one colour per shell, seen along one axis. An even spread covers the sphere; gaps and clusters do not.

In detail

What it measures: Where the gradient directions of each shell point. A direction and its opposite measure the same thing, so every direction is folded onto the half of the sphere facing the viewer.
How it is computed: Each b-vector is normalised and reversed when it points away from the viewing axis, then drawn with Lambert's equal-area projection, so an even spread looks even: the centre is the viewing axis, the outer circle the plane across it, and the faint rings lie 30 and 60 degrees from the axis. The b=0 volumes have no direction and are not drawn.
How to read it: Look for empty regions and for points on top of one another, and view along each axis: a scheme can look even from one side and leave a gap seen from another. The largest gap of each shell, beside its name, puts a number on it.
What raises or lowers it: A poorly designed scheme, a scheme meant for a full sphere acquired as half of one, and volumes dropped at conversion or by a scan stopped early.
Caveats: A property of the table, not of the image: an image whose table is wrong or flipped looks the same here. The axes are the image's, as the .bvec stores them.
References: Jones, Horsfield and Simmons 1999, Magn Reson Med; Jeurissen, Leemans and Sijbers 2014, Med Image Anal

b-value of each volume

Every volume's b-value in the order acquired, one colour per shell and the b=0 volumes in grey. Shows how the b=0 volumes are spread through the scan.

In detail

What it measures: The order in which the shells and the b=0 volumes were acquired.
How it is computed: Each volume's b-value from the .bval, at its position in the series. Values up to the b=0 limit count as b=0 and the others are grouped into shells rounded to the nearest 50, as the quality check groups them (Settings > Quality control > Diffusion).
How to read it: b=0 volumes spread through the series let drift and motion be followed over the whole scan; all of them at the start cannot. Interleaved shells share heating and motion evenly. A point off its shell's line is a b-value unlike its neighbours'.
What raises or lowers it: The acquisition protocol; series reordered or joined after the scan.
Caveats: It shows the table, not whether the images were acquired in that order.
References: Vos et al. 2017, Magn Reson Med

Displacement

How far each volume sits from the first b=0, in mm, with volumes far out of place marked. Lower is better.

In detail

What it measures: The head's position over the series, as one distance per volume.
How it is computed: The three absolute translations plus the three absolute rotations as arcs on a 50 mm sphere by default, each relative to the first b=0, with eddy-current shifts removed when the gradients explain them. The dashed line is 2 mm by default (Settings > Quality control > Diffusion); a marked volume is beyond it and a robust outlier of its shell.
How to read it: Lower is better. The b=0 volumes are the head's reliable track; diffusion-weighted volumes scatter around it by the registration's noise, more at higher b. Follow the b=0 points for slow motion and look for single volumes far above their neighbours.
What raises or lowers it: Head motion; registration noise on low-signal volumes.
Caveats: A displacement from the first b=0, not a frame-to-frame change: a head that moved once early and stayed there reads high for the rest of the scan.
References: Power et al. 2012, NeuroImage

Translation (x, y, z)

Where each volume sits against the first b=0, in mm, along the image's three axes. Flat near zero is best.

In detail

What it measures: The head's position over the series along three axes.
How it is computed: The three translations of each volume relative to the first b=0, along the image's voxel axes, with the eddy-current part removed when the gradients explain at least 30 % of a shell's translations by default (Settings > Quality control > Diffusion).
How to read it: Neither direction: read the shape. A slow drift is the head settling, a step a movement; scatter that follows the diffusion-weighted volumes and not the b=0 ones is registration noise or eddy current left in.
What raises or lowers it: Head motion, eddy currents, and registration noise.
Caveats: A screening registration, not the correction an analysis uses.

Rotation (pitch, roll, yaw)

How each volume is turned against the first b=0, in degrees. Flat near zero is best.

In detail

What it measures: How the head is turned over the series, and so how the gradients turned with it.
How it is computed: The components of each volume's rotation vector relative to the first b=0, about the image's first, second and third voxel axes (pitch, roll and yaw for the usual axial acquisition), in degrees.
How to read it: Neither direction: read the shape. Rotations above a few degrees mean the b-vectors should be rotated with the motion correction.
What raises or lowers it: Head rotation, most often nodding.
Caveats: A screening registration: small values are within its noise.
References: Leemans and Jones 2009, Magn Reson Med

Slice signal

Every slice of every volume against what the tensor predicts, as a robust z within its shell. Blue is signal lost (a dropout), red is signal added.

In detail

What it measures: Where in the series, slice by slice, the signal departs from the tensor's prediction.
How it is computed: Rows are slices (the top slice at the top), columns volumes. Each cell is the slice's mean log residual (observed against predicted) as a robust z among the volumes of its shell, clipped at 10 either way and coloured from -8 to 8; slices with too little brain are blank. Dropouts found by the check are flagged, and a click goes to that volume and slice.
How to read it: A single blue cell is a dropout in one slice of one volume. A blue column is a whole volume that lost signal, a red one a volume that gained it. A pattern repeating every few rows in one column points at multiband slices acquired together.
What raises or lowers it: Motion during the diffusion encoding, cardiac pulsation, table vibration, and interleave artefacts.
Caveats: The z is relative to the shell: a slice that is bad in every volume of a shell does not stand out.
References: Andersson et al. 2016, NeuroImage

b=0 signal

The brain's median signal in each b=0 volume against the first, in percent. A slope is drift.

In detail

What it measures: How the scanner's signal changes over the series, as the b=0 volumes show it.
How it is computed: The median over the brain mask of each b=0 volume, over the first b=0's, times 100, drawn at each b=0's position in the series. The drift measure is a fit through these points.
How to read it: Flat at 100 is best. A steady slope is drift; a single point off the line is motion or an artefact in that volume.
What raises or lowers it: Gradient and shim heating; motion.
Caveats: Only as many points as b=0 volumes; with all of them at the start, it says nothing about the end of the scan.
References: Vos et al. 2017, Magn Reson Med

Spiking voxels

Voxels far above the tensor's prediction, per million brain voxels, per volume. Lower is better.

In detail

What it measures: Which volumes hold single voxels much brighter than they should be.
How it is computed: Each voxel's residual against the tensor's prediction is compared with its own median across the shell; a spike is more than 8 robust spreads above by default (Settings > Quality control > Diffusion) and more than about 65 % above. The count per volume is scaled per million brain voxels.
How to read it: Lower is better; peaks are volumes to inspect.
What raises or lowers it: RF spikes; motion that moves bright tissue into usually dark voxels.
Caveats: Volumes of a shell with fewer than four volumes are not judged.

Odd against even slices

Per volume, odd slices against even slices relative to the tensor's prediction, as a robust z within the shell. Near zero is best.

In detail

What it measures: Motion between the two passes of an interleaved slice acquisition.
How it is computed: Per volume, the median residual over odd slices minus that over even slices, as a robust z within the shell. The dashed line is 5 by default (Settings > Quality control > Diffusion); a volume is marked when |z| is above it and the difference is above about 5 % in signal.
How to read it: Near zero is best; the mark uses the size of z, so a volume far below zero is marked as well as one far above. A marked volume shows a venetian-blind pattern in a sagittal or coronal view.
What raises or lowers it: Head motion between the passes of an interleaved acquisition.
Caveats: Meaningful for interleaved acquisitions; with other slice orders it points at another slice-wise effect.

Nearest-direction correlation

Each diffusion-weighted volume's correlation with the volume nearest it in q-space. Higher is better.

In detail

What it measures: Whether each volume looks like its nearest neighbour in direction and b-value.
How it is computed: Pearson's correlation over sampled brain voxels between each diffusion-weighted volume and its nearest neighbour in q-space (the square root of the b-value times the direction, either sign); b=0 volumes have none. Volumes with a robust z below -5 among all diffusion-weighted volumes (not per shell) are marked.
How to read it: Higher is better. Points far below the rest are volumes unlike their neighbours: check them in the 'Slice signal' plot and the image. With several shells, each sits at its own level, lower at higher b.
What raises or lowers it: Motion, signal dropout, spikes, low SNR, misregistration between volumes.
Caveats: With several shells the marks are judged over all diffusion-weighted volumes together, so a high shell that is a minority of the volumes can be marked as a whole for its naturally lower correlation.
References: Yeh et al. 2019, NeuroImage

Eddy-current shift

The shift of each diffusion-weighted volume its gradient direction explains, in mm, along the image's three axes.

In detail

What it measures: The part of each volume's displacement that eddy currents, not the head, caused.
How it is computed: In each shell of at least 10 volumes, the translations are regressed on the gradient's components; the explained part is shown when it accounts for at least 30 % of the shell's translations by default (Settings > Quality control > Diffusion). The plot appears only when some shell passes.
How to read it: Lower is better. The shift is usually largest along the phase-encoding axis and follows the gradient direction from volume to volume; a voxel or more calls for distortion correction, which is routine.
What raises or lowers it: Strong, fast-switching diffusion gradients and high b-values.
Caveats: Only the shift; the stretch and shear eddy currents cause are not shown.
References: Andersson and Sotiropoulos 2016, NeuroImage

Functional time series

The plots under the time course of a 4-D image, and the maps of the whole run in the viewer's QC menu.

Plots

Framewise displacement

How far the head moved from one volume to the next, in mm (Power 2012). Lower is better; volumes above the dashed line are marked.

In detail

What it measures: Frame-to-frame head movement as one number per volume: the absolute changes of the three translations plus those of the three rotations, the rotations as arcs on a 50 mm sphere by default (Settings > Quality control > Functional time series (BOLD)).
How it is computed: From fMRIPrep's confounds table when the run has one of the right length, else estimated here: each volume registered rigidly to the first volume after any non-steady-state ones, on a grid of about 3.5 mm by default (Settings > Quality control > Functional time series (BOLD)), using the reference's 15,000 strongest edges. Settings > Quality control > Methods chooses whether the confounds are read. With the confounds, their framewise_displacement column is shown, recomputed from their parameters when the radius is set to other than 50 mm. The first volume has no FD.
How to read it: Lower is better. The dashed line is 0.5 mm by default (Settings > Quality control > Functional time series (BOLD)), Power 2012's and fMRIPrep's usual threshold; stricter analyses use lower ones. Isolated peaks are sudden movements, a raised baseline is restlessness throughout. The note says whether the numbers come from the confounds or from the estimate.
What raises or lowers it: Head motion, swallowing, coughing and talking; with short repetition times also the apparent motion breathing causes, which shifts the image along the phase-encoding direction.
Caveats: The estimate is a screening registration: against fMRIPrep's it agreed closely for a participant who moved (r 0.94 to 0.98) but read about 1.4 times higher for a still one, whose motion is near the registration's noise. FD per volume depends on the repetition time: the same movement spread over more, shorter volumes gives smaller values, so a threshold set for one TR does not carry to another.
Compared with MRIQC: MRIQC summarises the same FD as fd_mean and counts volumes above its own threshold (fd_num, fd_perc).
References: Power et al. 2012, NeuroImage

Translation (x, y, z)

Where the head is along three axes, in mm, relative to the reference volume. Flat near zero is best.

In detail

What it measures: The head's position over the run along three axes.
How it is computed: The three translations of the same motion as FD (fMRIPrep's trans_x, trans_y, trans_z, or the estimate). For the estimate the reference is the first steady volume and the axes are the image's voxel axes, which are left-right, back-front and down-up for the usual axial acquisition.
How to read it: Neither direction: read the shape. A slow drift over the run is common (the head settling); a step is a sudden movement; a total range larger than a voxel means the same voxel holds different tissue at different times.
What raises or lowers it: Head motion and slow settling; with short repetition times, breathing.
Caveats: Relative to a reference: the values depend on which volume that is, not only on how the head moved.

Rotation (pitch, roll, yaw)

How the head is turned, in degrees, relative to the reference volume. Flat near zero is best.

In detail

What it measures: Pitch about the left-right axis (nodding), roll about the back-front axis (tilting to a shoulder), yaw about the vertical axis (shaking the head).
How it is computed: The three rotations of the same motion as FD, in degrees (fMRIPrep's rot_x, rot_y, rot_z, or the components of the estimate's rotation vector). For the estimate the axes are the image's voxel axes, which match those names for the usual axial acquisition.
How to read it: Neither direction: read the shape. One degree moves a point 50 mm from the centre of rotation by about 0.87 mm, so small angles matter at the edge of the brain. Nodding (pitch) is the commonest.
What raises or lowers it: Head motion, most often nodding; swallowing and breathing.
Caveats: Relative to a reference volume, as the translations are.

DVARS

How much the whole image changes from one volume to the next, in percent of the mean signal. Lower is better; volumes above the dashed fence are marked.

In detail

What it measures: The size of the change between consecutive volumes, over the head.
How it is computed: Over the head mask of a reference volume (voxels brighter than a fifth of its 98th percentile): the root mean square of each volume minus the previous one, over the run's median global signal, times 100. The dashed fence is the 75th percentile plus 1.5 interquartile ranges of the run's own DVARS by default (Settings > Quality control > Functional time series (BOLD)), FSL's rule. Non-steady-state volumes at the start, left out by default (same section), do not enter the fence or the marks; volume 0 has no value.
How to read it: Lower is better. Peaks that coincide with FD peaks are motion; peaks without FD point at spikes, slice dropouts or sudden physiological changes. Read it with FD and the carpet plot.
What raises or lowers it: Motion, RF spikes, slice dropouts, deep breaths, and non-steady-state volumes at the start.
Caveats: The fence comes from the run itself, so a run full of motion has a high fence, and a clean run still marks its largest values. It is not standardised, so its scale depends on the run's noise and does not compare between protocols.
Compared with MRIQC: Not on the scale of MRIQC's dvars_std, dvars_nstd or dvars_vstd, which are standardised.
References: Power et al. 2012, NeuroImage; Afyouni and Nichols 2018, NeuroImage

Outlier voxels

The percentage of the head's voxels that are outliers in each volume (AFNI's 3dToutcount rule). Lower is better.

In detail

What it measures: How many voxels are far from their own usual values at the same moment.
How it is computed: 8000 voxels by default (Settings > Quality control > Functional time series (BOLD)) are sampled from the head mask and each voxel's series has a quadratic trend removed. A value is an outlier when it lies further from its voxel's median than qginv(0.001/N) x sqrt(pi/2) x MAD, N the number of volumes. The dashed line is 5 % by default (same section), afni_proc.py's censoring default.
How to read it: Lower is better. A volume above the line has many voxels far from their usual values at once: motion, a spike or a dropout.
What raises or lowers it: Motion, RF spikes, slice dropouts, and non-steady-state volumes.
Caveats: From a sample of voxels, so small, local events can be missed. Voxels at the head's edge become outliers with small movements, so motion dominates the plot.
Compared with MRIQC: The same rule as MRIQC's aor (AFNI's outlier ratio), shown per volume where MRIQC reports the mean.
References: Cox 1996, Comput Biomed Res

Slice spikes

The largest robust z of any slice against its own course, per volume. Lower is better; volumes above the dashed line are marked.

In detail

What it measures: Single slices that depart from their own course in one volume.
How it is computed: Each slice's mean over the head, along the slice axis (the sidecar's SliceEncodingDirection, else the third axis). What the whole volume did (an offset and a scaling of the run's typical slice profile) is removed, then a quadratic trend over time; each slice becomes a robust z across volumes, and the plot shows the largest |z| per volume. The dashed line is 6 robust SD by default (Settings > Quality control > Functional time series (BOLD)).
How to read it: Lower is better. A marked volume has one slice that departed from its own course; the summary names the slice, so check that volume and slice in the image.
What raises or lowers it: RF spikes, a slice spoiled by motion during the volume (interleaved or multiband acquisitions), and artefacts confined to single slices.
Caveats: Whole-volume changes are removed on purpose, so global signal changes and slow drift are not spikes.

Global signal

The mean signal over the head, per volume. Neither direction: read the shape.

In detail

What it measures: How the overall signal level changes through the run.
How it is computed: The mean over the head mask (voxels brighter than a fifth of a reference volume's 98th percentile) in each volume, in the scanner's units.
How to read it: Look for steps (a scanner adjustment or a large movement), a bright start (non-steady-state volumes the scanner did not discard), slow drift, and slow oscillations. Read it with FD and DVARS: a change seen in all three is motion.
What raises or lowers it: Scanner drift and heating, non-steady-state volumes, motion, and physiology: breathing, heart rate and arousal all move it.
Caveats: Not a noise measure: much of its slow variation is physiological and real.

Carpet plot

Every sampled voxel's signal over time, one row each, from the edge of the head inwards (Power 2017). Vertical bands are events across the head.

In detail

What it measures: The whole run at a glance: what happened to every part of the head at every volume.
How it is computed: 8000 sampled head voxels by default (Settings > Quality control > Functional time series (BOLD)), each with a quadratic trend removed and scaled to its own SD (clipped at 3), ordered by distance from the head's edge (the edge at the top) and averaged into 240 rows. Colours span z from -2 to 2; non-steady-state volumes, left out by default, are blank.
How to read it: A clean run looks like even texture. Vertical bands across many rows are events that touched the whole head at once: motion, a deep breath, a spike. Bands strongest at the top (the edge) are typical of motion.
What raises or lowers it: Motion, breathing and arousal changes, RF spikes, and slice dropouts.
Caveats: Rows are ordered by depth, not by tissue as in Power's original, so grey matter, white matter and CSF mix in each row. Each voxel is scaled to its own SD, so differences in amplitude between voxels do not show.
Compared with MRIQC: MRIQC's functional report draws a carpet plot too, ordered by tissue.
References: Power 2017, NeuroImage

Maps

Mean

The mean of each voxel through the run: a picture of the anatomy and of where signal is lost.

In detail

What it measures: Where the signal is, and where it is missing.
How it is computed: The mean over the run's volumes, without the non-steady-state volumes at the start when Settings > Quality control > Functional time series (BOLD) leaves them out (the default). Shown inside the head: voxels above a fifth of the map's 98th percentile.
How to read it: Look for dark regions inside the head: susceptibility dropout near the sinuses and ear canals (orbitofrontal cortex, temporal poles), and brightness that falls from the edge to the centre (receive-coil sensitivity). The temporal SNR is low where the mean is dark, for that reason rather than noise.
What raises or lowers it: Susceptibility differences at air-tissue boundaries, coil sensitivity, slice coverage, and wrap-around.
Caveats: In scanner units: compare patterns, not values. Voxels below a fifth of the 98th percentile are hidden, so faint ghosts outside the head do not show.

Standard deviation

The standard deviation of each voxel through the run, around a slow trend. Bright means the signal moves.

In detail

What it measures: Where the signal varies over time.
How it is computed: Around a quadratic trend by default, after the non-steady-state volumes (Settings > Quality control > Functional time series (BOLD)), in the scanner's units; shown where the mean map is shown.
How to read it: Bright in vessels, ventricles and the brainstem is expected (pulsation). A bright rim at the brain's edge means motion; bright copies of the head along the phase-encoding direction mean ghosting.
What raises or lowers it: Motion, cardiac and respiratory pulsation, ghosting, and thermal noise.
Caveats: In scanner units: compare patterns within a run, not values between scanners. It is hidden wherever the mean is below a fifth of its 98th percentile, so a ghost shows only where it overlaps the displayed head.

Temporal SNR

Each voxel's mean over its standard deviation through the run, after slow drift is removed. Higher is better.

In detail

What it measures: How steady each voxel's signal is: how much of it is signal rather than fluctuation.
How it is computed: The mean map over the standard deviation map (both after the non-steady-state volumes, the SD around a quadratic trend by default). The summary gives the median over the head and the share of the head below 20 by default (both Settings > Quality control > Functional time series (BOLD)).
How to read it: Higher is steadier. Compare runs of the same protocol: it depends on field strength, voxel size, acceleration, the coil and the repetition time, so no single number is a pass mark. Grey matter usually reads lower than white matter, because physiological fluctuations are larger there. Patterns: a low ring at the brain's edge (motion), low values in ventricles and along vessels (pulsation), low where the mean is dark (dropout), alternating slices (slice-wise motion or spikes).
What raises or lowers it: Thermal noise (small voxels, high acceleration), physiological noise (breathing, heartbeat), motion, and signal dropout.
Caveats: It does not grow in step with image SNR: as voxels get larger or the field stronger, physiological fluctuations take over and it levels off. It is computed on the data as stored, before motion correction, so motion lowers it.
Compared with MRIQC: MRIQC reports tsnr, the median temporal SNR in the brain; BIDS Manager's median is over the displayed head region, so the values differ.
References: Krueger and Glover 2001, Magn Reson Med; Triantafyllou et al. 2005, NeuroImage

MEG and EEG

The signal viewer's QC checks a recording channel by channel and segment by segment, and compares each channel only with the others of its type (magnetometers with magnetometers, EEG with EEG). It marks flat channels, channels louder than the others of their type that also do not follow them, and channels at a normal level that do not follow their type (by the standard deviation, the peak-to-peak amplitude and the correlation with the others), the power at the line frequency, segments where many channels jump at once, and muscle activity. Each finding can be turned into a bad channel or segment with Mark as bad. The thresholds are under Settings > Quality control > MEG and EEG QC (Settings).