About the MRE Inverse Problem

Magnetic resonance elastography does not measure stiffness. It measures motion. Everything Sentinel does follows from that gap, and from the fact that closing it is a genuinely hard problem rather than a calculation.

This page is about why the problem is hard and why Sentinel answers it the way it does. For the equations, see the Mathematical Reference; to actually run a reconstruction, start with Your First Inversion.

What the scanner gives you

In an MRE exam, a driver vibrates the tissue at a fixed frequency and a phase-contrast MRI sequence encodes the resulting displacement into image phase. What comes out is a complex-valued displacement field sampled on the imaging grid: at every voxel, how far and with what phase the tissue moved.

Stiffness is nowhere in that data. It is a property of the tissue that governs how the wave propagates, and it has to be inferred from the wave you observed.

Forward and inverse

The forward problem is well behaved. Given a distribution of material properties, the equation of motion determines the displacement field those properties would produce. There is one answer, it depends continuously on the inputs, and a finite element discretization computes it to whatever accuracy the mesh supports.

The inverse problem runs the other way: given the displacement field, find the properties. It is the question clinicians actually want answered, and it is ill-posed.

Why it is ill-posed

"Ill-posed" is not a synonym for "difficult". It means the problem fails one of Hadamard's conditions — a solution exists, is unique, and depends continuously on the data — and in MRE it fails the last two in ways that matter.

The data does not determine the answer. Different property distributions can produce displacement fields that differ by less than the measurement noise. This is worst exactly where the wave is weak: where amplitude is low, the displacement carries almost no information about local stiffness, and the reconstruction there is being decided by something other than the data.

Small errors in the data become large errors in the answer. The map from displacement back to properties amplifies noise. MRE data is not especially clean — phase wraps, partial-volume effects at boundaries, low SNR far from the driver — so this is not a hypothetical concern.

The data is incomplete. Displacement is sampled on a finite voxel grid, in a finite field of view, at one frequency. The tissue's behaviour is continuous and its property field has structure below the voxel scale.

None of this makes reconstruction impossible. It means the data alone is not enough, and any method that appears to invert the data without adding information has added it implicitly somewhere.

Two ways to respond

One response is direct inversion: assume local homogeneity, then solve the governing equation algebraically at each point for the modulus that satisfies it. Algebraic Helmholtz inversion is the familiar example. It is fast, it is simple, and where its assumptions hold it works well.

Its weakness is structural rather than incidental. Solving the wave equation pointwise requires spatial derivatives of the measured displacement, usually second derivatives — and differentiation amplifies exactly the high-frequency noise the data already has. Smoothing first controls the noise but blurs the boundaries you were trying to resolve, and the local-homogeneity assumption is least true precisely at those boundaries.

Sentinel takes the other response: model-based iterative inversion. Rather than differentiating the data, it builds a finite element forward model, computes the displacement a candidate property field would produce, compares that against what was measured, and adjusts the properties to reduce the difference. The comparison is what closes the loop.

The advantage is that the measured displacement is only ever used as data to match, never as something to differentiate. The derivatives are taken of the model, which is smooth and known exactly. The cost is that every step requires a forward solve, which is why this is the expensive approach — and why so much of Sentinel is concerned with making that solve fast.

What the model has to capture

Two properties of soft tissue shape the discretization.

It is viscoelastic: it stores and dissipates energy, so the shear modulus is complex, ${\mu }^{\ast }={\mu }^{\prime }+i{\mu }^{″}$, with the storage modulus ${\mu }^{\prime }$ and the loss modulus ${\mu }^{″}$. That makes the stiffness matrices complex symmetric rather than Hermitian — a detail with real consequences for which linear solvers apply.

It is nearly incompressible: the bulk modulus exceeds the shear modulus by about six orders of magnitude. Discretizing that naively produces volumetric locking, so Sentinel uses a mixed displacement–pressure formulation with pressure degrees of freedom alongside displacement. That choice is why the element matrices have a saddle-point structure, and in turn why About Static Pressure Condensation is worth so much — it removes those pressure unknowns exactly, before the solve.

Restoring well-posedness

Since the data does not determine the answer, something else must. That something is regularization: an explicit statement of what a plausible property field looks like — smooth, or piecewise smooth with edges, or close to a known anatomical prior, or simply within physical bounds.

This deserves to be stated plainly rather than buried: regularization is not a numerical convenience, it is part of the answer. Change the regularizer and you change the reconstruction, because you have changed what the method believes in the absence of data. Reporting a stiffness map without reporting how it was regularized leaves out something the reader needs.

The practical consequences are in How to Choose and Configure Regularization.

Why it is also a large problem

Everything above would be true of a small problem. A brain reconstruction is not small: hundreds of thousands of nodes, a complex-valued solve at every iteration, and a property field with many thousands of unknowns.

Worse, the natural formulation — one global inverse problem over the whole brain — needs boundary conditions on the brain surface, which is where the displacement data is least trustworthy. Sentinel's answer is to decompose the domain into overlapping zones whose boundaries lie in the interior, where the measurements are reliable, and to consolidate their results. That argument is made properly in Zone-Based Domain Decomposition, and the loop built around it in Inverse Solver Pipeline.

Further reading

• Zone-Based Domain Decomposition — why the domain is split, and how.

• Inverse Solver Pipeline — the iteration this problem statement implies.

• About Static Pressure Condensation — the near-incompressibility consequence.

• Mathematical Reference — the governing equations and element kernels.