Telling Corrupted Measurements from Hidden Fibers in Diffusion MRI

Diffusion MRI estimates the direction of nerve fiber bundles in each small brain cube (a voxel) by measuring how water diffuses along 30 to 100 gradient directions. Most white matter voxels contain several bundles that cross. Resolving them requires spherical deconvolution, a sharpening inversion that is so sensitive to noise that a single corrupted measurement can produce a fake fiber peak or erase a real small one. Robust statistics would normally handle corrupted measurements. Here it runs into a basic ambiguity. Inside one voxel, a large misfit at one gradient direction could be an artifact, or it could be a small fiber the model has not yet captured. Per-voxel outlier rejection cannot tell the two apart.

Our proposal resolves this ambiguity by pooling information across voxels. Acquisition artifacts such as motion dropouts, slice signal loss and vibration are shared: they hit the same gradient direction in many voxels of a slice at once. Missed fibers are local and depend on tissue geometry. We therefore estimate, over a neighborhood of voxels, a robust covariance of the fit errors across gradient directions. We use Tyler's M-estimator, a covariance estimator that ignores each sample's overall size and so tolerates voxels with very different noise levels. We then refit every voxel's fiber orientation distribution (the function on the sphere whose peaks are fiber directions) by weighted least squares under this covariance. The design keeps the extra cost to roughly one additional deconvolution pass, which matters because iterative robust estimators are usually what makes such pipelines too slow for whole-brain use.

Core Claim

Model the vector of measurement errors at voxel $i$ across the $p$ gradient directions as $e_i = \tau_i z_i$. Here $\tau_i > 0$ is an unknown voxel-specific noise scale, set by coil sensitivity and tissue, and $z_i$ has a direction-by-direction scatter matrix $\Sigma$ shared within a neighborhood. Under this model, artifacts appear as inflated or correlated entries of $\Sigma$, and the nuisance scales $\tau_i$ are exactly what Tyler's estimator is built to cancel. This is the "compound-Gaussian with unknown texture" setting that motivates Tyler's estimator in radar.

Our claim is that a constrained deconvolution weighted by a regularized Tyler estimate of $\Sigma$ will do three things relative to per-voxel robust deconvolution. It will suppress artifact-induced false peaks. It will preserve small fibers crossing at acute angles that per-voxel reweighting tends to discard. It will do both at a cost within a small constant factor of standard deconvolution. The new ingredient is not "robust deconvolution", which exists. It is using cross-voxel scatter structure in direction space as the signal that separates artifact from anatomy.

Relation to Prior Work

The work that motivates the gap comes from robust diffusion estimation. Robust estimator framework in diffusion tensor imaging fits single diffusion tensors with a modified least-trimmed-squares estimator that tolerates up to 16 corrupted directions out of 30. On a Tourette syndrome cohort, it recovered a corpus callosum effect at $p < 0.005$ that constrained least squares missed. Its single-tensor model collapses in crossing regions, which is where most white matter lives (Validation and regularization in diffusion MRI tractography; Connectivity-Based Parcellation of the Cortical Mantle Using q-Ball Diffusion Imaging). Its trimming fraction is also hand-set.

Trimming works for a six-parameter tensor fit with spare degrees of freedom. A 45-coefficient deconvolution has far fewer spare degrees of freedom, so per-voxel trimming loses its footing. That is why we move the robustness from the voxel to the neighborhood.

Two crossing-fiber methods constrain the design. Spherical Deconvolution of Multichannel Diffusion MRI Data with Non-Gaussian Noise Models and Spatial Regularization (RUMBA-SD) builds Rician and noncentral-chi likelihoods into Richardson-Lucy iterations with a total-variation prior. It resolves crossings 5 to 10 degrees narrower than its Gaussian counterpart, but it models noise as independent per-measurement intensity noise. It has no notion of a gradient direction that is bad across a whole slice.

SR²-CSD (see also the companion version) pairs high-order deconvolution with a self-tuned total-variation prior. It reports up to 30% lower angular error and 40% lower peak-number error at low SNR on the HARDI 2013 phantom, but it names its two-step quadratic program as the main cost and assumes additive Gaussian noise. We treat both as the strongest classical competitors. Our weighting is complementary to their spatial priors: SR²-CSD smooths the fiber estimates across space, while we pool the error structure across space and leave the fiber estimates voxel-wise.

The method itself comes from the robust covariance literature. Robust Shrinkage Estimation of High-Dimensional Covariance Matrices supplies the regularized Tyler estimator, which shrinks the estimate toward the identity with weight $\rho$. It also proves that a unique fixed point exists for any $\rho \in (0,1]$, which we need because neighborhoods may contain only a few times more voxels than directions. Tyler's and Maronna's M-estimators: Non-asymptotic concentration results and Marčenko-Pastur law for Tyler's M-estimator give finite-sample and spectral behavior that we use to size neighborhoods. Convergence and Fluctuations of Regularized Tyler Estimators characterizes the estimator's fluctuations for fixed dimension, which is our regime since $p$ is fixed by the protocol.

On overhead, Frank-Wolfe-based Algorithms for Approximating Tyler's M-estimator speeds up Tyler's estimator for large dimensions. We note candidly that with $p \le 100$ the cubic per-iteration cost cited in the shrinkage literature is negligible. The real overhead in this problem is the nested outer loop of refitting and re-estimating over a million voxels, and we attack that structurally.

The learned alternatives, Robust Fiber ODF Estimation Using Deep Constrained Spherical Deconvolution for Diffusion MRI and Equivariant spatio-hemispherical networks for diffusion MRI deconvolution, gain robustness and speed through training. They give no explicit account of which measurements were distrusted. Our estimator produces that account, the matrix $\Sigma$, which can be inspected and checked against known artifacts.

Research Program

A technical obstacle. If $\Sigma$ were estimated from ordinary fit errors of a high-order deconvolution, it would see little of the corruption. At order 8, the 45 coefficients against 60 directions give each measurement an average leverage of 0.75, meaning the fit can largely bend to absorb a corrupted value. That absorption is precisely how spurious crossing peaks arise. At order 12 the problem is underdetermined.

We therefore estimate $\Sigma$ from leave-one-out prediction errors of an unconstrained low-order fit at order 4. These are available in closed form as $\tilde e_{ij} = e_{ij} / (1 - h_{jj})$, where $h_{jj}$ is the leverage of direction $j$. Since every voxel shares the same gradient table, the leverages are computed once per protocol. Low order keeps leverage small, so artifacts stay visible in the errors rather than being fitted.

The estimator. In each neighborhood we solve the regularized Tyler fixed point on the leave-one-out errors $\tilde e_i$. A neighborhood is an in-plane tile of a few hundred white-matter voxels within one slice, chosen because dropouts are slice-locked:

$$\Sigma = (1-\rho)\frac{p}{n}\sum_{i=1}^{n}\frac{\tilde e_i \tilde e_i^{T}}{\tilde e_i^{T}\Sigma^{-1}\tilde e_i} + \rho I.$$

Here $n$ is the number of voxels in the tile and $\rho$ is the shrinkage weight. We set $\rho$ by the subspace-style rule of Hoarau et al., adapted to the fixed-dimension setting, and compare it against held-out-direction cross-validation.

We then whiten the measurements and the design matrix by the Cholesky factor of $\Sigma$, once per tile. Each voxel is solved by the usual non-negativity-constrained deconvolution, single- or multi-tissue, so existing solvers in MRtrix3 or DIPY are reused unchanged.

Overhead control. We test whether a one-step estimator suffices: a single scatter estimate followed by a single weighted refit. One-step M-estimators often retain most of the robustness of the fully iterated solution, and this approach removes the outer loop entirely. Tyler fixed points in neighboring tiles are warm-started from each other. The target is whole-brain runtime within 2 to 3 times standard deconvolution. We report wall-clock time against SR²-CSD and RUMBA-SD on identical hardware.

Analysis. We will characterize when cross-voxel pooling identifies corruption. An artifact affecting direction $j$ in a fraction $f$ of tile voxels shows up in $\Sigma_{jj}$ once $f$ is large relative to the estimator's fluctuation scale, which the concentration results above make explicit as a function of $n$. A missed fiber in a single voxel contributes only $O(1/n)$ to $\Sigma$, so it cannot be mistaken for an artifact. This gives a concrete, testable rule for choosing tile size. Larger tiles separate artifact from anatomy more reliably but miss spatially localized corruption.

Baselines

The naive reference is standard constrained spherical deconvolution at order 8 with no outlier handling, as in MRtrix3/DIPY, together with its high-order variant Super-CSD. Everything must beat these on corrupted data and match them on clean data.

The strongest classical competitor on artifact-corrupted crossing estimation is SR²-CSD (Taskin et al.). It reports up to 30% lower angular error and 40% lower peak-number error than CSD and Super-CSD at low SNR on the HARDI 2013 phantom. RUMBA-SD (Canales-Rodríguez et al.) is the strongest competitor that models non-Gaussian noise. Both are run with their authors' tuning procedures on the same corrupted inputs and the same response functions, and we report their runtime alongside ours. Beating them means lower angular error and peak-number error under injected shared corruption at matched or lower runtime, with no loss on clean data. We also test our weighting stacked under the SR²-CSD prior. If the gains add, that settles that the two mechanisms are complementary.

The strongest learned competitor is Deep CSD (Yao et al.). It reports angular correlation coefficient 0.891 versus 0.826 for CSD on MASiVar, and about 0.902 when HCP directions are subsampled from 96 to 50. We compare on the same MASiVar scan-rescan pairs. Our method needs no training, so matching 0.891 would be the bar on clean multi-site data. On corruption patterns absent from Deep CSD's training data, we expect to exceed it.

A per-voxel robust deconvolution is the direct test of our key claim. It uses iteratively reweighted constrained deconvolution with Huber weights on per-voxel standardized errors, the natural extension of the trimming in Maximov et al. to crossing models. It gets the same iteration budget as our method.

Three ablations isolate the ingredients. The first replaces Tyler with the sample covariance of the same leave-one-out errors, which tests whether scale-invariance matters. The second replaces the full $\Sigma$ with a diagonal robust variance per direction, a median absolute deviation across tile voxels, which tests whether off-diagonal structure matters. The third uses ordinary rather than leave-one-out errors, which tests the leverage argument.

Why not just…?

Why not just reweight each voxel's measurements robustly, the way robust tensor fitting does? This is the most natural alternative. In a voxel with a small acute-angle fiber, the measurements most aligned with that fiber carry the largest errors under a fit that misses it. Per-voxel reweighting will downweight exactly the evidence for the fiber. The corpus offers no direct measurement of this effect, so our experiments settle it. On phantoms with crossings below 45 degrees and no corruption, per-voxel reweighting should raise the false-negative rate for the minor fiber while our method should not. If per-voxel reweighting shows no such loss, our main argument weakens to a speed and interpretability argument, and we will report that.

Why not just detect and replace bad slices during preprocessing, then run ordinary deconvolution? This is a real threat for whole-slice dropouts, which hard detection handles well and cheaply. Our method subsumes it in that case, since a dropped slice inflates $\Sigma_{jj}$ to near infinity and is excluded. The difference appears for partial attenuation, spatially localized corruption, and correlated artifacts across directions, which a binary keep-or-drop rule cannot weight. The corruption experiments sweep severity from total dropout to 20% attenuation. If detect-and-replace matches us across the sweep, we will say the covariance machinery is unnecessary. As Maximov et al. note, choosing the rejection level is itself a hard tuning problem.

Why not just use a diagonal per-direction weight instead of a full covariance? It is cheaper and might suffice if artifacts are independent across directions. The diagonal ablation tests this directly. Interleaved acquisitions and vibration plausibly produce correlated errors, but we do not assume it. If the diagonal version matches the full one, we adopt it and the contribution becomes the cross-voxel Tyler pooling and the leave-one-out step.

Why not just train a network on corrupted data? Deep CSD and equivariant networks are fast and strong on the distributions they see. A network trained on simulated dropouts would likely handle simulated dropouts. Our check is generalization to corruption types held out of training, compared against Deep CSD retrained with our corruption generator. If the network generalizes as well, our method's remaining advantage is that it needs no training and returns an inspectable $\Sigma$, and we will state it at that size.

Proposed Experiments

The main synthetic tests use the HARDI 2013 numerical phantom at SNR 10, 20 and 30, and the DiSCo phantom. To each we add corruption representative of motion dropouts, partial attenuation, and correlated vibration patterns. Each corruption is injected slice-wise into a chosen fraction of gradient directions, from 1 to 25%. On these we measure angular error, peak-number error, and false-positive peak rate, stratified by crossing angle. Where obtainable, the 132-phantom library of Canales-Rodríguez et al. extends the crossing-angle sweep down to 1 degree.

The first decisive check is the clean-data control. Without corruption, our method must match standard deconvolution within noise. If it does not, the weighting is costing anatomy. The second decisive check is the acute-crossing minor-fiber recall comparison against per-voxel reweighting described above. We also report tractography valid and invalid connections with the Tractometer setup used by Elaldi et al. on DiSCo.

In vivo, we use HCP subjects with injected corruption, because the uncorrupted acquisition serves as a reference. We also use MASiVar scan-rescan and traveling-subject pairs (Cai et al.). There we report angular correlation against Deep CSD's 0.891 and the coefficient of variation of fiber-derived measures across scanners, against the reported 6.49% interscanner median. We do not claim harmonization of systematic scanner bias. Tyler's invariance cancels voxel scale, not vendor-specific angular bias, and we will report whether interscanner variation changes at all.

For a sanity check on real artifacts, we visualize the estimated $\Sigma_{jj}$ maps on raw HCP and MASiVar data and compare them against slice-dropout flags from standard preprocessing quality control. Agreement shows that the estimator finds real artifacts rather than anatomy. Finally, whole-brain runtime for one-step versus fully iterated versions, against SR²-CSD and RUMBA-SD on identical hardware, determines whether the overhead claim holds.