Fast MRI scans skip most of the raw frequency measurements (called k-space) and let a neural network fill in the rest. This makes scans shorter, but it creates a safety problem. When a radiologist sees a small ligament tear in an accelerated scan, nothing tells them whether the scanner actually measured the tear or the network drew it from what it learned about typical knees. The problem gets worse as reconstruction networks start using neighboring slices to fill gaps. A tear visible on slice 12 may then be partly copied from slice 13. We propose to jointly learn the undersampling pattern and a reconstruction network that sees several adjacent slices. Every reconstructed voxel then comes with an exact, ground-truth-free provenance map: how much of the local detail came from this slice's own measurements, how much was borrowed from neighboring slices, and how much was invented by the network's learned prior. Training penalizes invented detail inside pathology regions.
The key observation is that, with hard data consistency, the decomposition is not a heuristic saliency map. Hard data consistency means the network must keep every measured frequency exactly as acquired. The reconstruction of each slice then splits linearly into a part built from measured frequencies and a part the network filled in. Running the same network with its neighbor inputs withheld further splits the filled part into "borrowed" and "invented." This gives an audit signal that is tied to scanner physics. It also turns the sampling pattern from a pure quality knob into a lever on how much of the diagnostically relevant signal is actually measured.
Core Claim
We claim three things. First, per-slice sampling patterns that are staggered across slices, learned together with a multi-slice decoder, preserve annotated pathology better than 2D-learned masks at the same acceleration. Second, the provenance map predicts where pathology will be erased or hallucinated better than uncertainty maps do. Third, adding a provenance penalty inside pathology regions shifts the learned sampling toward frequencies that carry pathology, at small cost in image-wide PSNR. The second claim matters most and stands even if the first turns out weak.
Where This Sits
Decoder-aware sampling is the mechanism we reuse. Several methods learn masks jointly with a reconstruction network: LOUPE relaxes binary masks into learnable probabilities, PUERT does the same with an unrolled ISTA decoder, and End-to-End Sequential Sampling adapts the mask within a scan. All three treat each slice in isolation. PUERT's "explicability" is architectural, meaning its layers mirror optimization steps. It does not say whether a given structure in the output was measured. We keep static, population-level masks so the method avoids the online-latency problem of active acquisition. The change is that masks may differ by slice position, and the decoder reads neighboring slices.
Volumetric context loss is the application regime. Roth et al. showed that slice-independent 2D processing produces jagged out-of-plane results: 3D smoothing lifted pancreas Dice from 56% to 68%. That result motivates using cross-slice context. It also constrains the design, because post-hoc smoothing cannot restore information that was never measured. Our twist is that borrowing across slices is itself an interpretability hazard, and it needs to be accounted for rather than simply added.
Clinical interpretability supplies the requirement and the evaluation lens. Two uncertainty lines are the closest audit competitors. Risk Quantification in Deep MRI Reconstruction estimates error without ground truth using SURE (Stein's Unbiased Risk Estimator), but it produces a single global score per 2D image under a fixed mask. Reducing Uncertainty with Active Acquisition produces pixel-wise variance and uses it to choose the next k-space line. Neither separates measured signal from signal generated by the prior. A confident hallucination can have low variance.
Active Sampling for MRI-based Sequential Decision Making goes the other way. It samples k-space directly for ACL-tear classification and drops the image. Deliberative work with clinicians (Upshaw et al.) indicates that this loss of auditability is what blocks adoption. Task-adapted CT reconstruction shows that task-driven pipelines deliberately distort images (its "PSNR paradox") and offer no account of which distortions are real. Our provenance map is designed to provide that missing account.
Research Program
The data are fastMRI multi-coil knee volumes, which are acquired as stacks of 2D Cartesian slices undersampled along one phase-encode direction. Each slice therefore gets its own 1D line mask. Differing masks per slice are realizable on standard hardware, although our experiments are retrospective. In-plane resolution is fine but slices are thick. This anisotropy caps how much neighbor borrowing can help, which makes the dataset an honest test.
The sampler learns a probability over phase-encode lines for each position in a repeating slice period (for example period 2 or 3), relaxed as in LOUPE. The decoder is an unrolled network with hard data consistency that takes slices $z-1$, $z$, and $z+1$ as input. During training we randomly withhold the neighbor inputs. This keeps the neighbor-blind version of the decoder in-distribution, so it can be used for attribution.
For slice $z$, let $M_z$ be the binary mask, $y_z$ the measured k-space, $F$ the (coil-combined) Fourier operator, and $\hat k_z$ the decoder's k-space estimate. Hard data consistency gives an exact split:
$$\hat x_z = F^{-1}(M_z y_z) + F^{-1}\big((1-M_z)\hat k_z\big) = m_z + f_z .$$
Here $m_z$ is the measured component and $f_z$ is the filled component. We also compute $f_z^{\text{blind}}$, the filled component when the neighbor inputs are withheld. We then define borrowed $= f_z - f_z^{\text{blind}}$ and invented $= f_z^{\text{blind}}$.
For a region $R$, we compute each component's share of the band-passed detail energy, meaning the fine-scale structure that carries small lesions. The invented share, $\rho_R$, is the audit quantity shown to the reader as a map. It needs only $y$ and the network, not a fully sampled reference.
Training minimizes reconstruction loss plus $\lambda$ times the mean of $\rho_R$ over fastMRI+ pathology boxes (fastMRI+). Because the mask is population-level, this penalty cannot cheat by looking at a particular patient's lesion location. It can only shift sampling toward frequency bands that pathology tends to occupy. One caveat: the measured component is aliased, so "measured share" is a statement about how much of the lesion's fine-scale contrast lies in acquired frequencies, not about how visible the lesion is in the zero-filled image. We will report the measure under that definition.
Baselines
The naive reference point is equispaced Cartesian undersampling with a fully sampled center, reconstructed slice by slice with the fastMRI U-Net (fastMRI). The file reports this U-Net only as a 40–50% relative NMSE improvement over total variation.
The strongest reconstruction-only competitor is i-RIM (i-RIM), which reports multi-coil 4x test-set NMSE 0.0052, PSNR 39.52, and SSIM 0.928. We will retrain it on our splits and masks rather than compare against the leaderboard figure, since the test sets differ. Beating it means matching its PSNR within about 0.3 dB while clearly improving pathology preservation inside fastMRI+ boxes.
For learned sampling, we run LOUPE-style per-slice masks and the T-Step sequential sampler (End-to-End Sequential Sampling), which reports SSIM 92.91 against LOUPE's 92.44 at 4x. Both get the same decoder backbone, data, and hyperparameter-search budget. The sequential sampler is an upper reference that uses adaptivity our static masks deliberately give up.
For diagnosis, the reference is the MORL image-less sampler of Du et al., which reports ACL AUC 0.886 at 13 sampling steps. We run a fixed ACL classifier on our reconstructions at matched line budgets. Coming within a few AUC points while also delivering an auditable image would count as success; exceeding it is not required.
For auditing, the competitors are pixel-wise variance maps obtained by Monte Carlo sampling (Zhang et al.) and a localized SURE-style estimate following Edupuganti et al., each computed on the same reconstructions.
Two ablations isolate the key ingredients. The first compares staggered masks with the same mask on every slice. The second compares the provenance penalty on versus off ($\lambda=0$). The first isolates volumetric sampling; the second isolates the audit-driven objective.
Why not just…?
Why not just use fixed, hand-designed staggered masks with a multi-slice decoder? This is the main threat. Offsetting a regular pattern by one line per slice already gives neighbors complementary frequencies with no learning at all. It is run as a baseline at matched acceleration. If it matches the learned masks on pathology preservation, we will drop claim one and present the provenance audit and penalty as the contribution. The penalty can be applied to fixed-pattern families too, for example by learning only the density profile.
Why not just show uncertainty maps? Variance measures disagreement within the prior. A prior that consistently fills in a meniscus edge produces low variance whether the edge exists or not. The counterfactual experiment below tests directly which signal predicts erasure and hallucination. If variance wins, that is a publishable negative result for provenance.
Why not just smooth 2D reconstructions across slices, or put Grad-CAM on the reconstructed image? Roth et al. show smoothing plateaus because it cannot recover unmeasured structure. A saliency map computed on $\hat x$ cannot tell $m_z$ from $f_z$, because it attributes to the image rather than to the measurements. Haddada et al. further show that a model can reach acceptable accuracy while its explanations invert clinical priorities: k-NN reached Kendall Tau −0.39 against expert consensus. Both options appear as slice-wise baselines in the pathology-preservation comparison.
Decisive Experiments
The first experiment is counterfactual lesion insertion and removal. Small synthetic lesions spanning one to three slices are added to fully sampled raw k-space, and some real annotated regions are altered by local contrast flattening. The data are then undersampled and reconstructed. From this we measure lesion-contrast recovery (erasure) and spurious lesion signal on adjacent slices where nothing was inserted (cross-slice hallucination).
The central test asks whether the per-lesion invented share $\rho_R$ predicts these failures, scored as per-lesion AUROC, better than variance or localized SURE on the same reconstructions. We expect the borrowed share to specifically flag the neighbor-leakage cases that 2D uncertainty methods cannot see.
The second experiment measures pathology preservation on real fastMRI+ boxes at 4x and 8x. Metrics are in-box SSIM and NMSE, and the detection rate of a fixed pathology detector relative to fully sampled images, reported alongside whole-image PSNR and SSIM. This lets any PSNR-paradox trade-off be seen explicitly.
The third experiment is a small reader study. Musculoskeletal radiologists grade suspected tears with and without the provenance overlay. We record whether flagged "invented" regions lower confidence on cases that are actually wrong, following the concordance-with-experts style of evaluation in Haddada et al. rather than raw accuracy alone.
References
- Learning-based Optimization of the Under-sampling Pattern in MRI.
- PUERT: Probabilistic Under-Sampling and Explicable Reconstruction Network for CS-MRI
- End-to-End Sequential Sampling and Reconstruction for MRI
- Deep convolutional networks for pancreas segmentation in CT imaging
- Risk Quantification in Deep MRI Reconstruction
- Reducing Uncertainty in Undersampled MRI Reconstruction With Active Acquisition
- Active Sampling for MRI-based Sequential Decision Making
- Priorities for Artificial Intelligence Applications in Primary Care: A Canadian Deliberative Dialogue with Patients, Providers, and Health System Leaders
- Empirical evidence of the task-adapted reconstruction framework for joint CT reconstruction and segmentation
- fastMRI+: Clinical Pathology Annotations for Knee and Brain Fully Sampled Multi-Coil MRI Data
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI.
- i-RIM applied to the fastMRI challenge
- Assessing the Interpretability of Machine Learning Models in Early Detection of Alzheimer's Disease