A flow model is a generator that moves random noise toward a realistic image by following a learned velocity field. Recent work trains these fields so that the paths are nearly straight lines, and a straight path can be crossed in one or two large jumps instead of hundreds of small ones. That speed is what clinical MRI reconstruction needs, since current flow-based solvers still take minutes per scan. The catch is that straightness is learned on clean, curated training data. Real clinical scans are different: patients breathe, scanners come from different vendors, and noise floors vary. On those inputs we expect the learned paths to bend again, and a one-jump reconstruction then lands somewhere plausible but wrong. That is exactly the kind of error that turns into a hallucinated structure.
This proposal treats the re-bending itself as the signal. A model that predicts average velocity over an interval, as in Mean Flows for One-step Generative Modeling, can also report its instantaneous velocity at the same point in one parallel batch. On a straight path the two agree. Their disagreement is a nearly free measure of local curvature. We use it for two jobs at once. It decides, per scan and per step, whether to trust a single jump or to subdivide the interval. It also produces a pixel-wise reliability map that radiologists and downstream models can consume. Put simply, the reconstructor spends extra computation only on scans where its own geometry says it is off-distribution, and it tells you when that happened.
Core Claim
Let $u_\theta(x_t, r, s)$ be a network predicting the average velocity that carries a noisy image $x_t$ from time $r$ to time $s$. Its instantaneous velocity is then $u_\theta(x_t, t, t)$. We define the curvature gap at state $x_t$ as
$$g(x_t, t) = \frac{1}{d}\,\big\| u_\theta(x_t, t, 1) - u_\theta(x_t, t, t) \big\|^2 ,$$
where $d$ is the number of pixels. It is zero exactly when the predicted path from $t$ to the data endpoint at $t=1$ is straight, by the MeanFlow identity $u = v - (s-t)\,du/ds$ with $v$ the instantaneous velocity. We claim three things. First, under clinically realistic shifts (motion, cross-vendor acquisition, low signal-to-noise), $g$ rises and predicts reconstruction error, including error in the unsampled part of k-space, the raw frequency-domain measurement an MRI scanner records. Second, a controller that subdivides steps only when $g$ exceeds a threshold calibrated on clean data achieves a better fidelity-versus-latency trade-off than any fixed step count. Third, the gap flags hallucinations that the usual data-consistency residual cannot see.
The novelty is not "use MeanFlow for MRI." That combination is itself an open gap in the file. The novelty is using the model's own straightness violation as a free, calibrated test of when few-step inference is safe.
Relation to Prior Work
Two lines of work supply the machinery. The first is the flow-matching inverse-problem solvers. Flower alternates predicting the clean endpoint, a proximal data-consistency step, and moving forward in time. ICTM splits the maximum-a-posteriori reconstruction (the single most probable image given the measurements) into local objectives along the path and reports about 1.6 minutes per image. We reuse Flower's predict-refine-progress loop but replace its fixed multi-step schedule with a curvature-gated one. Neither paper gives any signal of when its reconstruction is unreliable, and ICTM explicitly produces a single deterministic output. The second line is average-velocity and rectified models: MeanFlow, InstaFlow and VoiceFlow. These show that straightened paths survive one or two steps where curved diffusion collapses, but only in unconstrained generation. We borrow the average-velocity parameterization precisely because it exposes both velocities from one network.
Several papers motivate the proposal by documenting the failure it targets. The 2020 fastMRI Challenge found that models trained on Siemens data degraded on GE data and fabricated vessel-like structures from surgical artifacts. Evaluating the Impact of Medical Image Reconstruction on Downstream AI Fairness found that generative reconstruction shifts sex and age disparities while aggregate AUROC stays within about 1%. Its models were slow multi-step ones, so nobody knows what one-step compression does. GTF2M brings flow matching to fastMRI but uses a multi-step forward-backward cycle and ignores motion and vendor shift.
Two papers constrain the design. Diffusion Bridge or Flow Matching? shows that flow matching with scarce data produces conflicting velocity fields. Curvature may therefore be elevated even in-distribution for rare protocols, which is why we calibrate the threshold per protocol rather than globally. DIAMOND steers flows away from artifacts, but it needs an external differentiable detector and adds per-step backward passes. Our signal needs no external model and no backpropagation.
Research Program
We train an average-velocity prior on fully sampled, coil-combined fastMRI brain and knee images from fastMRI. The measurement model is the standard linear undersampling operator. We deliberately do not model motion or vendor differences in the forward operator. They are the unmodeled clinical shift the gap must detect, which also sidesteps the linear-operator limitation that Flower and ICTM share.
At inference, each step makes one batched call that returns $u_\theta(x_t,t,1)$ and $u_\theta(x_t,t,t)$. It then jumps to the endpoint, applies a Flower-style proximal data-consistency step, and computes $g$. If $g$ falls below the threshold $\tau$, the reconstruction is returned. Otherwise the interval is halved and the loop continues, up to a cap of eight jumps. The threshold $\tau$ is set as a quantile of $g$ on clean held-out scans of the same protocol, so the expected fraction of clean scans that escalate is fixed in advance. The per-pixel version of $g$ at the final step is saved as the reliability map.
The analysis has three parts. The first measures how well $g$ tracks true error across a graded corruption suite. The second traces the latency-fidelity frontier the controller produces. The third tests whether the flags coincide with hallucinated content. Training cost stays modest, since MeanFlow reports about 16% overhead from its Jacobian-vector product.
Baselines
The naive reference is the same average-velocity prior run with a single fixed jump plus data consistency. This is the "just use MeanFlow" answer. It is the fastest possible point, so beating it means matching its latency on clean scans while reducing error and hallucination on shifted ones.
The strongest multi-step flow competitor is ICTM, because it is the one flow-prior solver in the file evaluated on medical compressed sensing. It reports PSNR 24.03 on HCP T2w at sampling ratio 1/10, at roughly 1.6 minutes per image (ICTM). We rerun it on HCP T2w to reproduce that number and on our fastMRI splits, using the same flow-prior backbone class and training data and its published step settings. Beating it means equal or better PSNR/SSIM at a small fraction of its wall-clock time on the same GPU. Flower is run the same way. Its reported CelebA deblurring PSNR of 34.98 dB is on natural images and is not comparable to ours. GTF2M is included as the unsupervised flow reference on fastMRI; no reported numeric result was found on Lacuna for GTF2M.
The strongest non-generative competitor is a supervised unrolled network of the kind that topped the 2020 fastMRI Challenge, where the winner reached SSIM around 0.964 at 4x multi-coil (2020 fastMRI Challenge). We train it on identical data with the same augmentation budget. We expect it to match or beat us on clean in-distribution PSNR. Our claim is about shifted data and reliability flagging, so beating it means lower error or fewer undetected hallucinations under shift, at comparable latency.
Three ablations isolate the key ingredient. The first is a fixed-step schedule matched to the controller's average number of network evaluations. It tests whether adaptivity, rather than extra compute, drives the gains. The second replaces $g$ with the measurement residual $\|y - Ax\|$ as the gating signal. The third replaces $g$ with a random gate at the same escalation rate. All methods get the same tuning budget on a shared validation split.
Why not just…?
Why not just always run four or eight steps? If shifted scans are rare, a fixed small step count might be good enough, and latency differences of a few passes are small. This is a real threat on clean data. The matched-compute fixed-step ablation settles it. The proposal wins only if, at equal mean evaluations, gating lowers error on the shifted subset, where fixed schedules spend compute uniformly. It also wins if the reliability map adds detection value that fixed schedules simply lack.
Why not just use the data-consistency residual as the reliability signal? It is free and classical. However, after a proximal consistency step the residual on sampled k-space lines is driven near zero by construction. Hallucinations live in the unsampled null space, where the residual is blind, as in the challenge's fabricated vessels. The residual-gated ablation tests this directly. We report the area under the error-detection curve for both signals, split by whether the error lies in sampled or unsampled frequencies. If the residual matches $g$ on null-space errors, the core claim fails.
Why not just draw several posterior samples and use their spread? Sample variance is the standard uncertainty estimate, but it costs one full reconstruction per sample. We compare at equal total evaluations, with $S$ one-jump samples against one gated reconstruction. Sample spread and $g$ may be complementary, and the experiment will show whether $g$ is the cheaper route to the same detection power.
Why not just train a supervised network with heavy augmentation? Reducing Textural Bias and AnyStar show that augmentation buys robustness to the corruptions it simulates. The same work notes that overly aggressive noise distorts anatomy and that geometric priors fail on branching structures. Augmentation also gives no per-scan signal of failure. We apply the same augmentation to both the supervised baseline and our prior, so the comparison isolates the flagging and adaptive-compute contribution.
Proposed Experiments
The main testbed is fastMRI multi-coil brain and single-coil knee at 4x and 8x. It includes three shifted conditions. The first is a vendor split, training on one scanner family and testing on the other, mirroring the challenge's transfer track. The second is retrospectively simulated rigid and respiratory-like motion, applied as per-line k-space phase errors at graded severities in the style of AF2R. The third is added noise at decreasing signal-to-noise. HCP T2w serves as the reproduction check against ICTM.
Metrics are PSNR and SSIM, wall-clock time per slice and number of network evaluations, and the Spearman correlation between $g$ and true per-scan error. We also report the area under the detection curve for error maps and a hallucination test. In that test, small synthetic structures are inserted into or removed from ground truth before undersampling, and we count the invented or erased structures that fall outside flagged regions.
A secondary experiment uses retrospectively simulated k-space from UCSF-PDGM. It chains each reconstructor to a fixed downstream glioma segmenter and classifier and reports Dice, AUROC and sex and age disparity in sensitivity and specificity. The question is whether one-jump reconstruction widens disparity and whether gating narrows it.
Three results decide the question. First, $g$ must correlate with error on shifted scans and must beat the residual on null-space errors. Second, at matched mean compute, the gated controller must dominate the fixed schedule on the shifted subset while staying near one-jump latency on clean scans. Third, $g$ must stay calibrated per protocol despite the low-data curvature predicted by the Diffusion Bridge analysis. If the first check fails, the proposal reduces to "MeanFlow for MRI" and should be reported as such.
References
- Mean Flows for One-step Generative Modeling
- Flower: A Flow-Matching Solver for Inverse Problems
- Flow Priors for Linear Inverse Problems via Iterative Corrupted Trajectory Matching
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation
- VoiceFlow: Efficient Text-To-Speech with Rectified Flow Matching
- Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction
- Evaluating the Impact of Medical Image Reconstruction on Downstream AI Fairness and Performance
- Unsupervised Accelerated MRI Reconstruction via Ground-Truth-Free Flow Matching
- Diffusion Bridge or Flow Matching? A Unifying Framework and Comparative Analysis
- DIAMOND: Directed Inference for Artifact Mitigation in Flow Matching Models
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI.
- Reducing Textural Bias Improves Robustness of Deep Segmentation Models
- AnyStar: Domain randomized universal star-convex 3D instance segmentation
- Realistic Restorer: artifact-free flow restorer(AF2R) for MRI motion artifact removal