Do Fast Flow Scans Make Sick Vessels Look Healthy? Auditing and Bounding Z-Score Shrinkage in Accelerated 4D-Flow MRI

Doctors judge a child's blood vessels with Z-scores, not millimetres. A Z-score says how many standard deviations a measurement sits from what is normal for a body of that size and age, so −2.84 means "severely undersized for this child." Fast MRI scans gather only a fraction of the raw data and rely on a neural network to fill in the rest. That network learned what vessels and blood flow usually look like, so it may quietly pull abnormal measurements toward normal. We propose to measure this effect directly as error in the Z-score rather than error in pixels. We also propose a calibrated stopping rule that tells the scanner when it has collected enough data to trust the Z-score.

The stakes are clearest when doctors monitor operations that restore the body's natural blood flow. In infants with Tetralogy of Fallot, stenting the right ventricular outflow tract restores the heart's own pulsing ejection into the lungs. In Right ventricular outflow tract stenting promotes pulmonary artery development in tetralogy of fallot, treatment success was read as left pulmonary artery Z-scores rising from −2.84 to −0.78 and right pulmonary artery Z-scores from −2.84 to −0.47 on CT angiography. A learned reconstruction that pulls Z-scores toward zero would mimic exactly this treatment effect. The failure would therefore not be random noise. It would be biased in the direction of a false "the intervention worked."

Core Claim

Our hypothesis is that learned 4D-flow reconstructions shrink Z-scores toward zero. Here 4D-flow MRI means MRI that measures three-directional blood velocity throughout a 3D volume over the heartbeat. We expect the shrinkage to be measurable as a slope below one between reconstructed and true Z-scores, to grow with acceleration, and to be largest for the most abnormal vessels. Pixel-level metrics will hide it.

We further claim that it can be bounded. We will calibrate a prediction interval directly on the Z-score using paired full and undersampled acquisitions. The scanner can then keep collecting data until that interval clears the clinical decision threshold, typically |Z| = 2. Acquisition time thus becomes a per-patient quantity set by diagnostic certainty rather than a fixed acceleration factor.

Cross-species data serve as the stress test. Large animals give hemodynamics close to human scale but with different heart rates and calibers, which is exactly the shift under which the calibration should break if it is going to.

Why this is not already done

The closest work is learned super-resolution of 4D flow. LoFNO is a localized Fourier neural operator, a network that learns mappings between whole fields. On 95 Aneurisk carotid aneurysm geometries simulated with large-eddy CFD and downsampled 4x at a signal-to-noise ratio of 10, it reports a velocity test error of 0.0452, compared with 0.0684 for linear interpolation, and about 30% lower wall shear stress error. SRflow is a residual CNN with a direction-aware loss. It reports 10% (cardiovascular) and 30% (cerebrovascular) gains in peak-velocity-to-noise ratio over cubic B-spline at 4x.

Both papers score reconstructions on aggregate field error, and both downsample in image space rather than undersampling raw k-space (MRI's frequency-domain data). Neither asks whether a downstream normative score is biased, and neither covers abnormal tails or non-human scales.

The shape-prior work in Fighting MRI Anisotropy learns a continuous heart-shape model from 153 CT angiography scans. It reaches a right-ventricle Dice score of 0.91 on 140 cardiomyopathy MRI scans. That is evidence that strong priors produce plausible smooth anatomy, which is what motivates our concern, but the paper does not test how those priors behave on hypoplastic or stenosed anatomy.

The normative-modelling literature defines the target quantity but never touches reconstruction. Kawasaki guidance and MIS-C follow-up show why size-adjusted Z-scores beat millimetre cutoffs in growing children. Franken et al. shows that reference curves themselves fail across populations, which warns us that normative surfaces are population-specific.

Research Program

We define the Z-score of a quantity $X$, such as a branch diameter, peak systolic velocity, or mean wall shear stress, as $Z = (X - \mu(\theta))/\sigma(\theta)$. Here $\mu$ and $\sigma$ are the mean and spread of healthy anatomy given covariates $\theta$ such as body surface area and heart rate.

For a reconstruction method at acceleration $R$, we fit the relation between the Z-score from the accelerated reconstruction, $\hat Z_R$, and the Z-score from the full acquisition, $Z^\ast$:

$$\hat Z_R = \alpha_R + \beta_R Z^\ast + \varepsilon,$$

The slope $\beta_R < 1$ is our measure of shrinkage. We also report how often a vessel is classified differently at the |Z| = 2 cutoff.

The first step is a k-space simulator for phase-contrast MRI. Starting from Aneurisk CFD velocity fields, it produces raw-like complex k-space with realistic noise, undersampled at R = 2 to 12. This matters because the acceleration being audited is the physical one, not the image-space downsampling used by LoFNO and SRflow.

The second step builds a pathology spectrum. We deform the 95 Aneurisk geometries to create graded narrowing, down to roughly −4 in diameter Z, and graded dilation, then re-run the CFD. The unperturbed geometries define synthetic normative surfaces for the flow quantities. We state plainly that these are in-silico norms, a stand-in for the dynamic normative curves that clinical practice lacks.

The third step creates species-scaled variants. We rescale caliber, heart rate and viscosity so that the Womersley number (the ratio of pulsatile to viscous forces) and the Reynolds number (inertial to viscous forces) match porcine and lapine ranges. We take Wang et al. and Jiang et al. as the reason large animals are the right intermediate step for high-pressure arterial flow.

The method itself is split conformal calibration, a distribution-free procedure that turns held-out errors into prediction intervals with guaranteed coverage. We apply it to the Z-score itself. On held-out pairs we compute residuals $r = |\hat Z_R - Z^\ast|$ and take their $(1-\delta)$ quantile $q_R$. The interval is

$$\big[\hat Z_R - q_R(\hat Z_R),\ \hat Z_R + q_R(\hat Z_R)\big],$$

where $q_R$ is allowed to depend on $\hat Z_R$, so that the interval widens in the abnormal tail.

Acquisition then becomes sequential. The scanner adds k-space segments, lowering $R$, until the interval excludes the decision threshold or a time budget runs out.

The guarantee holds only when calibration and test data are exchangeable, meaning drawn from the same distribution. The human-to-pig shift breaks this by design, so it is the central test. We will measure how much coverage degrades, and how many animal scans are needed to recalibrate.

The final stage, contingent on facilities, is prospective porcine 4D-flow imaging with surgically created pulmonary or aortic narrowing. The full acquisition is the reference, and invasive pressure gradients serve as an independent anchor, following the catheter comparison in Ralovich et al..

Baselines

The naive reference point is to compute Z-scores directly from the undersampled data with zero-filled or linear reconstruction. The comparison uses the same simulated k-space and no training. On the Aneurisk benchmark, linear interpolation has a reported test error of 0.0684 (LoFNO). This baseline matters because classical reconstruction is noisy but carries no learned pull toward normal. It may therefore show higher variance but slope $\beta_R \approx 1$, and that is the trade-off we must expose. Cubic B-spline is run alongside it, since SRflow uses it as its reference.

The strongest existing method is LoFNO, with a reported test error of 0.0452 and about 30% lower wall shear stress error on Aneurisk at 4x. SRflow is the second learned competitor. No Aneurisk-specific SRflow number appears in the file; its reported result is the PVNR gain over cubic B-spline.

Both are retrained by us on the same healthy-geometry training split, inputs, and tuning budget, with the same GPU-hour cap per method. Beating them means two things at equal or shorter mean scan time. First, our sequentially stopped protocol must show a smaller absolute bias $|1-\beta|$ for vessels with |Z| > 2. Second, it must show fewer misclassifications at the cutoff while holding empirical coverage at or above $1-\delta$. Matching their field error is not required.

Two ablations isolate the key ingredient. The first applies the same conformal interval but uses a fixed acceleration instead of sequential stopping. This tests whether adaptivity buys scan time. The second uses an interval calibrated on pixel error and propagated to Z, instead of one calibrated on Z directly. This tests whether calibrating the clinical quantity itself matters.

The full acquisition is the upper reference.

Why not just…?

Why not just scan at low acceleration? Choosing R = 2 everywhere probably removes most shrinkage. It wastes time on clearly normal or clearly abnormal patients, though, and long free-breathing exams add the motion corruption that motivates acceleration in the first place (Learning Optical Flow for Fast MRI Reconstruction). Our fixed-R ablation directly reports scan time against misclassification. If a fixed low R matches the sequential rule at equal mean time, we will say so, and the contribution shrinks to the audit.

Why not just train on pathological examples? Adding stenosed and hypoplastic geometries to training is a real threat and may remove much of the bias in-distribution. We test it explicitly by retraining LoFNO with the deformed geometries included. Retraining cannot certify individual patients, however. It also does not address the species shift, where the pathology may look different again. If augmentation brings $\beta_R$ to about 1 with good coverage across species, the calibration layer becomes a safeguard rather than a necessity, and we would report it that way.

Why not just enforce data consistency? Unrolled reconstruction networks re-impose agreement with the measured k-space at every step, which limits how far the prior can override the data. That should reduce hallucination at low R. At high R, though, the unmeasured frequencies are exactly where fine stenoses live, so the prior still decides them. We include a data-consistent variant among the learned baselines, under the same budget, so that this question is answered empirically.

Decisive Experiments

The audit succeeds if, on the in-silico pathology spectrum, LoFNO and SRflow show a slope $\beta_R$ significantly below 1 that worsens with R and with |Z|, while their pixel error stays favourable. That pattern would show pixel metrics hiding the clinically relevant bias. The audit fails if both methods track Z-scores as faithfully as classical reconstruction at every R.

The cross-species check succeeds or fails on coverage. If intervals calibrated on human-scale flows hold coverage on porcine and lapine Womersley regimes, the method transfers. If coverage collapses, the result to report is how many animal scans are needed to restore it. That number is a practical answer to the gap that preclinical imaging protocols are not standardized.

The intervention-mimicry check reproduces the TOF scenario in silico. Paired geometries move from Z ≈ −2.8 to Z ≈ −0.6, and we ask whether accelerated reconstruction overstates the apparent improvement. If learned reconstruction inflates the measured improvement of a stenosed vessel after "treatment," the case for Z-calibrated acquisition is made.