A Label-Free Test of Message-Passing Theory on Real MRI Operators

Message-passing algorithms for inverse problems, such as recovering an image from undersampled scanner measurements, come with an unusual promise. At every iteration the algorithm claims to know exactly how wrong its current guess is. Specifically, it claims that the input to its denoising step equals the true image plus white Gaussian noise of a variance it reports. When that claim holds, the error is predictable in advance. On real measurement operators, such as Cartesian MRI sampling with multiple receive coils, the claim silently fails, and the algorithm can diverge or quietly mis-tune itself. This proposal turns the promise into a test that can be run on real data without a ground-truth image. The test is built from a difference signal that vector approximate message passing already computes. We ask whether it predicts divergence, reconstruction error, and scanner-model mismatch better than the simple checks practitioners already use.

The opportunity is that two problems usually treated separately have the same symptom. The first is that these algorithms are sensitive to the structure of the measurement matrix. The second is the gap between simulated and real evaluation. When the operator is too structured, the algorithm's self-reported noise stops matching its actual error. The same happens when the operator the software assumes differs from the one the scanner applied, which is what happened across vendors in the fastMRI challenge. A measurable version of "does the theory hold on this scan?" would serve as a stability diagnostic, a deployment monitor, and a way to validate theory on data where no clean reference exists.

Background and the gap

Approximate message passing (AMP) is an iterative method whose per-iteration error is tracked exactly by a scalar recursion called state evolution. That exactness depends on a correction term, the Onsager term, which cancels the echo of past errors. A Simple Derivation of AMP and its State Evolution via First-Order Cancellation derives this cancellation. On the convergence of approximate message passing with arbitrary matrices shows that stability is governed by the peak-to-average ratio of squared singular values. It also shows that damping restores convergence only at a cost of roughly $\sqrt{\kappa(A)}$ in speed, and that damping requires spectral knowledge rarely available in practice.

Vector AMP (VAMP) splits each iteration into a linear minimum-mean-squared-error step and a denoising step. It extends the guarantees to matrices whose right singular vectors are uniformly random. Denoising based Vector Approximate Message Passing shows it holds 20–24 dB PSNR up to condition number $10^4$, where the AMP-based denoising method collapses to 7–8 dB. That result comes from synthetic transform matrices built to satisfy the assumption.

These papers motivate the proposal: they identify the matrix property that breaks the theory. The work that constrains it is MRI Image Recovery using Damped Denoising Vector AMP. It found that VAMP diverges on real single-coil Cartesian fastMRI knees and restored stability with damping of variance amplitudes plus an ADMM warm start (DD-VAMP++). As the file notes, those fixes break the state-evolution guarantee, so the algorithm's reported variances are no longer known to mean anything.

Learned variants face the same issue. AMP-Inspired Deep Networks for Sparse Linear Inverse Problems verified error Gaussianity inside LAMP with Q-Q plots, but only because ground truth was available on synthetic data. Time-Embedded Algorithm Unrolling for Computational MRI learns iteration-dependent Onsager weights on retrospective fastMRI, with no check that the learned terms still whiten anything.

On the theory side, Macroscopic Analysis of Vector Approximate Message Passing in a Model Mismatch Setting predicts that mismatch drives VAMP's fixed points unstable, but it tests this only synthetically. The benchmark papers, fastMRI: An Open Dataset and Benchmarks for Accelerated MRI and Results of the 2020 fastMRI Challenge, document the retrospective-only evaluation and the cross-vendor drop, which the challenge traced to frequency-oversampling defaults.

Nobody in this set measures, on real scans, whether the Gaussian-white-error assumption holds. Nobody uses its failure as a signal.

Core claim

VAMP maintains two estimates of the image at each iteration: $r_1^t$, fed to the denoiser, and $r_2^t$, fed to the linear step. Theory says their errors are independent, white, and Gaussian, with precisions $\gamma_1^t$ and $\gamma_2^t$ that the algorithm reports. The difference $r_1^t - r_2^t$ equals the difference of those errors, because the unknown true image cancels. It can therefore be computed on any real scan.

Under the theory, its power spectrum in k-space should be flat at level $1/\gamma_1^t + 1/\gamma_2^t$. We define a band-wise calibration ratio

$$\rho_b^t = \frac{\| P_b\, F (r_1^t - r_2^t)\|^2}{n_b\,(1/\gamma_1^t + 1/\gamma_2^t)},$$

where $F$ is the Fourier transform, $P_b$ selects k-space band $b$ (for Cartesian MRI, groups of phase-encode lines ordered by distance from the center and by sampled versus skipped), and $n_b$ is the number of entries in the band. State evolution predicts $\rho_b^t \approx 1$ for every band. The audit score is the spread of $\log \rho_b^t$ across bands, together with the lag-one correlation between successive differences.

Unlike residual checks on measured data, this statistic covers the unsampled part of k-space. That is where aliasing and hallucinated structure live.

The claim has three parts. The score rises monotonically as the operator moves away from the rotation-invariant ideal. It rises before divergence or error blow-up, giving early warning. It detects forward-model mismatch of the kind behind the fastMRI Transfer Wall, and it ranks per-scan reconstruction error better than simpler label-free proxies.

The file does not include variable-density or colored-noise AMP variants, so we cannot rule out that related spectral corrections exist elsewhere. The novelty claimed here is the use of the difference signal as a ground-truth-free certificate, and its evaluation as such.

Research program

The first component is instrumentation. We add the audit to VAMP, to DD-VAMP++ (reimplemented following the paper), and to a learned unrolled VAMP whose Onsager weights vary by iteration in the TE-Unrolling style. Where DD-VAMP++'s damping changes the meaning of $\gamma$, we compute the ratio with the damped precisions the algorithm actually uses. Each iteration then yields a band-wise calibration profile at negligible cost, since both $r_1$ and $r_2$ already exist.

The second component is a controlled ladder of operator structure. The rungs are synthetic operators with Haar-random right singular vectors, where the theory holds; the DCT and Hartley matrices with swept condition number from the D-VAMP study; single-coil random Cartesian masks; equispaced masks; and multi-coil operators $A = M_{\text{mask}} F S$ built from fastMRI coil maps. The theory holds on the first rung, so a ratio near one there validates the statistic before we trust it where it matters.

The third component is a controlled mismatch program on real fastMRI multi-coil raw data. We perturb the assumed operator in ways matching documented physical causes: mishandled readout oversampling (the Transfer Wall cause named in the challenge paper), coil-map errors from estimating maps on fewer calibration lines, and sub-pixel k-space shifts mimicking gradient delay. This is still retrospective. It is not prospective validation, and we will say so. What it does is test the physics the file identifies as hidden by retrospective masking.

The fourth component is a causal check. If the audit localizes the violation to specific bands, replacing VAMP's scalar precision with a per-band precision should close the gap and improve error. If the gap closes but error does not improve, the audit measures something that does not matter.

Baselines

The naive reference is to trust the algorithm's own report. The predicted per-iteration error is $1/\gamma_1^t$, and divergence is flagged only when the image blows up. This is what every deployed VAMP variant implicitly does, and it is the thing the audit must beat. It is run on the same scans at no extra cost.

The strongest label-free competitor is held-out k-space cross-validation. It withholds a small fraction of acquired lines, reconstructs without them, and measures prediction error on the withheld lines. It is widely used, needs no theory, and works with any reconstructor, which makes it the real threat. It is given the same compute as the audit, and we tune its hold-out fraction on the 30-image tuning split used by DD-VAMP++.

The second competitor is a Monte Carlo Stein-type risk estimate. It combines the data residual with a randomized estimate of the denoiser's divergence, the same Jacobian probe DD-VAMP++ already uses, and matches the audit's number of probe evaluations.

The audit beats a baseline if its rank correlation with true per-scan NMSE, and its area under the ROC curve for detecting injected mismatch and impending divergence, both exceed the baseline's with bootstrap confidence intervals that do not overlap. The early-warning criterion is flagging at least several iterations before the error rises.

For reconstruction quality, which matters only in the causal-check stage, the reference numbers from the file are as follows. DD-VAMP++ reports −25.475 dB NMSE and 0.939 SSIM, and the best PnP-ADMM reports −24.694 dB NMSE, both on 128×128 single-coil fastMRI knees at 4× (MRI Image Recovery using Damped Denoising Vector AMP). The 2020 challenge winner reports SSIM around 0.964 at 4× on multi-coil brain (Results of the 2020 fastMRI Challenge); that is a different setting and 200M parameters, so it is context, not a target. For TE-Unrolling, no reported NMSE or SSIM value was found on Lacuna.

The key ablation compares per-band precision VAMP against scalar VAMP with identical damping, denoiser, initialization, and iteration budget. It isolates whether the diagnosed coloring is the operative failure.

Why not just…?

Why not just use held-out k-space cross-validation? It is simpler, and on the measured region it may match the audit. However, it only scores predictions of acquired lines, so it is structurally blind to errors in skipped phase-encode lines. That is where the challenge paper found hallucinations and where Cartesian aliasing concentrates. The experiment settles this directly. We report correlation separately for sampled-band and skipped-band error. If cross-validation equals the audit on both, we will say the theory-derived statistic adds nothing and reduce the claim to the early-warning result.

Why not just damp heavily, or switch to PnP-ADMM or an unrolled network and ignore state evolution? Damping buys stability at a $\sqrt{\kappa(A)}$ slowdown (On the convergence of approximate message passing with arbitrary matrices), and it still gives no statement about whether the reported noise level is right. PnP-ADMM and unrolled networks avoid the question rather than answer it. The audit also applies to learned unrolled VAMP. In the multi-coil mismatch experiment we measure whether the learned models' error degrades without any internal signal, which is the failure the audit is meant to catch.

Why not just use a Stein-type risk estimate? It estimates total mean squared error but assumes the noise model is correct. Under operator mismatch the effective noise is not the nominal one, so the estimate can be confidently wrong. We include it at matched compute. If it detects the injected oversampling and coil-map errors as well as the audit, the band-wise structure of the audit is unnecessary.

Decisive experiments

The first check is calibration. On Haar-random operators, $\rho_b^t$ must sit near one in every band. On the DCT condition-number sweep it must rise with $\kappa$. If it fails on the ideal rung, the statistic is invalid and the project stops there.

The second check is per-scan ranking and detection on fastMRI multi-coil knees and brains, at 4× and 8× with random and equispaced masks, under the three injected mismatch families at graded severity. We measure how well the audit ranks per-scan NMSE against ground truth. We also measure detection AUROC and warning lead time against the three baselines, with DD-VAMP++, plain VAMP, and learned unrolled VAMP as the instrumented reconstructors.

The third check is causal. Per-band precision VAMP should close the audit gap and improve NMSE over scalar VAMP on the same scans. On the single-coil 128×128 4× protocol it should at least match the −25.475 dB DD-VAMP++ result with less or no damping. If the audit correctly flags mismatch but the band-wise fix does not help, we report the audit as a diagnostic only.