Neural networks now reconstruct MRI images from heavily undersampled scanner data in a single forward pass. They do this far faster than classical optimization and with lower pixel error. They also sometimes invent anatomy that is not there or erase anatomy that is, and nothing at test time flags it. This proposal trains a separate "auditor" network to draw a map of where a reconstruction contains structure that the scanner never measured. The auditor learns from synthetic mistakes: fake lesions inserted into, and real lesions removed from, otherwise correct images. Each edit is mathematically constructed so that it agrees exactly with the raw scanner measurements. A simple consistency check against the data therefore cannot see it, and the auditor must learn what the measurements actually support.
Two observations motivate this. First, reconstruction hallucinations live almost entirely in the part of the image that the measurements do not determine. Second, work on text verification shows that detectors trained on deliberately corrupted examples can catch real mistakes, provided the corruptions are subtle enough. We combine the two: the synthetic errors are confined to the unmeasured part of the image, and the detector is judged mainly on whether it catches the organic errors of real reconstruction networks.
Background and the gap
A non-iterative reconstructor is a network $f_\theta$ that maps undersampled frequency-domain measurements $y$ (MRI "k-space") directly to an image $\hat{x}=f_\theta(y)$. On fastMRI, even a plain U-Net does this with a 40–50% relative reduction in normalized squared error over total-variation compressed sensing (fastMRI). The known failures are what make it risky. The 2020 fastMRI Challenge found that top entries turned surgical artifacts into structures resembling normal vessels, and that pixel metrics did not catch it. fastMRI+ notes that SSIM and PSNR can stay high while a ligament tear disappears. Risk Quantification in Deep MRI Reconstruction shows that adversarially trained (GAN-loss) reconstructors look sharper but carry more bias and variance. These papers establish that the gap exists.
Range–null decomposition explains where the errors hide. Let $\mathcal{A}$ be the undersampled acquisition operator. Then any image splits into a part fixed by the measurements and a part in the null space of $\mathcal{A}$, which is invisible to them. Null Space Matters enforces exact agreement on the measured part and lets the network fill only the null space. That eliminates data-inconsistent errors, but by construction it cannot rule out a fabricated lesion that happens to lie in the null space. We reuse this decomposition as our method of generating errors.
The synthetic-error literature supplies the training recipe. Rule-based entity and number swaps train token-level hallucination detectors for table-to-text generation (Text-in-Context). Mask-and-fill corruption produces plausible factual errors, and it works only within a "Goldilocks" range of corruption strength (MFMA). Detectors trained on synthetic deletions transfer to organic retrieval-augmented generation errors (Chang et al.). A critic trained on perturbed reasoning traces gives localized feedback (REFINER). All of these operate on discrete text, where "subtle but wrong" is defined by language. In imaging, the physics gives a sharper definition of a subtle error: one the measurements cannot refute. The research file found no paper that injects synthetic errors to train a reconstruction auditor. The closest imaging work uses injected perturbations only as attacks: Darestani et al. apply projected-gradient perturbations in measurement space to show brittleness, and train nothing on them.
Core Claim
A localized hallucination auditor trained only on synthetic, measurement-consistent pathology edits will localize real reconstruction errors better than three alternatives: uncertainty maps from stochastic reconstructors, data-consistency residuals, and auditors trained on unconstrained image-space edits. The advantage should be largest for the GAN-trained reconstructors that hallucinate most. The null-space constraint is the key ingredient. It removes the trivial cue that an edited image disagrees with $y$, and it matches where real hallucinations occur.
Research Program
Constructing the errors. Start from a fully sampled fastMRI knee or brain slice $x$ with fastMRI+ bounding boxes. Build an edit image $e$ of one of three kinds:
- Erasure: inpaint an annotated lesion region with surrounding tissue texture.
- Insertion: paste an annotated lesion patch, taken from another patient, into anatomically compatible tissue.
- Distortion: locally warp a boundary such as the meniscus.
Then keep only the part of the edit the scanner could not have seen:
$$x' = x + \alpha\,(I - \mathcal{A}^\dagger \mathcal{A})\,e,$$
Here $\mathcal{A}^\dagger$ is the pseudoinverse of the acquisition operator, $I-\mathcal{A}^\dagger\mathcal{A}$ projects onto its null space, and $\alpha$ sets edit strength.
For single-coil Cartesian sampling, the projection simply zeroes the edit's measured k-space lines. For multi-coil data we apply the projection through the coil-sensitivity operator, using estimated sensitivities. The edited image $x'$ produces exactly the same measurements $y$ as $x$.
The projection also strips the edit's well-sampled low frequencies. What remains looks like a high-frequency detail with no support in the data, which is what a hallucinated detail plausibly looks like.
We sweep $\alpha$ and lesion size to locate the Goldilocks zone that MFMA found for text masking ratios. Edits that are too weak are below the noise floor. Edits that are too strong are no longer plausible anatomy.
Avoiding the obvious shortcut. If edits were applied only to clean fully sampled images, the auditor could learn "network-output texture versus real-image texture" instead of "unsupported structure." We therefore apply edits on top of actual reconstructions $\hat{x}$ from several reconstructors and acceleration rates. The label is the support of the projected edit, thresholded at the noise level.
The auditor. The auditor is a U-Net-style segmenter. Its inputs are the candidate image, the zero-filled image $\mathcal{A}^\dagger y$, and the mask. Its output is a per-pixel probability that structure is unsupported, with a separate channel for "removed" versus "added" structure. It runs in a single pass, so pairing it with a one-pass reconstructor keeps the whole pipeline fast.
Evaluation on organic errors. This is the decisive part. On held-out fastMRI+ slices we reconstruct with each network and derive organic error labels from the fully sampled reference. We use two kinds of label:
- Pixel level: null-space error $|(I-\mathcal{A}^\dagger\mathcal{A})(\hat{x}-x)|$ above a noise-calibrated threshold.
- Lesion level: annotated boxes where lesion-to-background contrast changes by more than a fixed fraction, flagged as erased or attenuated.
These labels are automatic and imperfect. A radiologist read of a small sample would calibrate them, and we will report agreement.
Baselines
The naive reference is the trivial auditor that ranks every pixel equally. Next to it we report SSIM and PSNR, the image-level metrics fastMRI uses (fastMRI), to show how often error-laden slices score well. This sets the floor and quantifies the metric blindness that fastMRI+ describes.
The data-consistency residual re-projects $\hat{x}$ through $\mathcal{A}$ and maps the mismatch with $y$ back to image space. It costs nothing and is the first check a practitioner would run. For data-consistent unrolled reconstructors it is zero by design, which is exactly the regime we target. It is computed on the same slices.
The strongest existing competitor is the probabilistic reconstructor of Edupuganti et al.. It is a variational autoencoder with data-consistency layers whose Monte Carlo samples give pixel-wise variance maps, paired with a SURE estimate of overall error. SURE is the only blind error estimate in the file with reported accuracy. Its correlation with true squared error is $R^2=0.97$ at 2-fold and $0.84$ at 16-fold undersampling on a GE knee dataset. Because SURE gives one number per image, we compare against it on slice-level ranking. The variance maps serve as the localized competitor. We will also compute variance from deep ensembles of our own reconstructors, so that uncertainty-based baselines are not handicapped by architecture. For localized hallucination detection on fastMRI+ there is no reported result found on Lacuna, so all localized numbers will be produced here under one protocol. That protocol fixes the same slices and acceleration masks, and matches GPU hours for training the auditor against training the stochastic reconstructor and ensemble.
The key ablation trains an identical auditor on the same edits without null-space projection, so the edits are unconstrained image-space changes. A second ablation trains on Gaussian or projected-gradient perturbations in the style of Darestani et al. rather than anatomical edits. To count as beating the baselines, the auditor must achieve higher pixel-level area under the precision–recall curve and higher lesion-level recall at one false alarm per slice than the best uncertainty map, on organic errors, for at least the GAN-trained and U-Net reconstructors. It must also beat the image-space ablation by a margin larger than seed-to-seed variation.
Why not just…?
Why not just use uncertainty maps from a stochastic reconstructor? This is the real threat. Variance maps are already localized and need no synthetic data. However, Edupuganti et al. report that GAN-trained models gain both sharpness and bias. Bias, meaning a confident wrong answer, is precisely what variance misses: a model that always hallucinates the same vessel has low variance there. Our comparison stratifies organic errors into high-variance and low-variance groups. If variance maps match the auditor on the low-variance group, the synthetic-edit approach adds nothing, and we will report that.
Why not just check data consistency, or use a range–null reconstructor so errors cannot occur? Null Space Matters guarantees agreement with the measurements, and the 2020 challenge still saw hallucinations from data-consistent unrolled entries. A null-space error passes any consistency test. The residual baseline will show directly what fraction of organic error mass it catches. We predict near zero for data-consistent networks.
Why not just train on unconstrained synthetic edits, which are easier to make? Unconstrained edits disagree with $y$. A detector given $y$ can learn to spot the disagreement instead of learning anatomy, and the text literature warns that corruptions that are too easy produce detectors that fail on organic errors (MFMA). The image-space ablation tests this directly. If it transfers to organic errors as well as the null-space version, the central claim is wrong.
Why not just run a pathology detector on the reconstruction? A detector can only report what it sees. It cannot tell a real lesion from a fabricated one, and it says nothing about erased lesions. Our lesion-level erasure labels quantify that blind spot.
Proposed Experiments
Data and reconstructors. We use fastMRI multi-coil knee and brain data at 4× and 8× acceleration, restricted to fastMRI+-annotated studies (1,172 knee and 1,001 brain). Studies are split by patient, and edit sources are never shared across splits. We train three reconstructors on the same data: the fastMRI U-Net baseline, an end-to-end variational network, and a GAN-loss U-Net. The GAN-loss model provides a high-hallucination regime.
Experiment 1: transfer from synthetic to organic errors. This is the decisive check. We train the auditor on synthetic null-space edits applied to reconstructions from two of the three networks, then test on organic errors of the third. We report pixel-level precision–recall area and lesion-level recall for erasures and insertions separately.
Experiment 2: the Goldilocks sweep. We vary edit strength $\alpha$ and lesion size and plot organic-error performance against synthetic-training difficulty. We expect a peak at intermediate difficulty, as reported for text masking.
Experiment 3: distribution shift. We evaluate on the Stanford knee data used by Darestani et al. and on the fastMRI-A hardest-slice subset. The question is whether the auditor still detects errors when reconstruction quality drops, which is the setting where an auditor matters most. Darestani et al. found that reconstruction accuracy degrades roughly in parallel across scanners; we test whether auditor accuracy does too.
Experiment 4: slice-level ranking against SURE. Aggregating the auditor map per slice, we compare its ranking of slices by true null-space error with SURE's ranking.
The proposal fails if Experiment 1 shows no gain over ensemble variance on low-variance errors, or no gain over the image-space ablation.
References
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI.
- Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction
- fastMRI+: Clinical Pathology Annotations for Knee and Brain Fully Sampled Multi-Coil MRI Data
- Risk Quantification in Deep MRI Reconstruction
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction
- Text-in-Context: Token-Level Error Detection for Table-to-Text Generation
- Masked Summarization to Generate Factually Inconsistent Summaries for Improved Factual Consistency Checking
- Detecting Hallucination and Coverage Errors in Retrieval Augmented Generation for Controversial Topics
- REFINER: Reasoning Feedback on Intermediate Representations
- Measuring Robustness in Deep Learning Based Compressive Sensing