Fast cardiac MRI records a short movie of the beating heart from heavily undersampled measurements, which means far fewer frequency-space samples than a full image needs. To fill in the missing information, reconstruction methods borrow detail from neighbouring frames. That only works if the method knows how the heart moved between frames. When the motion estimate is wrong, the borrowed detail lands in the wrong place, and the result is ghosted double edges, jelly-like tearing, or blurred boundaries. We call these alignment-induced artifacts. They are often worse than the original noise. The trouble is circular: motion is estimated from the same degraded frames the method is trying to repair, so the motion estimator ends up reading aliasing streaks, scanner noise and contrast quirks as if they were anatomy.
We propose to break this circle by changing what the motion estimator looks at. Instead of raw intensities, it will see a learned structural descriptor, a per-pixel summary of how each location resembles its neighbours. This descriptor is trained to stay the same when the sampling mask, the noise level, the scanner, or the tissue contrast changes. It is then trained end-to-end inside the reconstruction. The idea borrows its representation from cross-modality registration, where such descriptors let CT be aligned with MRI. Here we repurpose it to make motion estimation ignore acquisition corruption. The payoff we test for is a single model that keeps motion estimates, and therefore reconstructions, free of ghosting as acceleration, noise and acquisition protocol shift beyond what it was trained on.
Core Claim
Alignment-induced artifacts in learned dynamic reconstruction come largely from the motion estimator's input carrying acquisition signatures. Examples are streak patterns that follow the sampling mask, noise floors, and contrast. Removing those signatures at the motion estimator's input should reduce the artifacts more than the obvious remedy of augmenting the same raw-intensity estimator with the same corruptions. The claim is falsifiable. If a raw-intensity estimator trained with identical corruption augmentation matches the descriptor-based one on out-of-distribution protocols, the claim is wrong.
Where This Sits
The reconstruction side comes from Motion-Compensated MR CINE Reconstruction With Reconstruction-Driven Motion Estimation. That work trains a motion network (GRAFT) jointly with a non-rigid reconstruction solver, using only the final image loss and no intermediate warping loss. It reports artifact suppression up to 20x acceleration at about 5.5 seconds per slice. We keep this end-to-end training and reuse the method. Its motion network, however, still reads intensity images of the partially reconstructed frames. The file also lists motion-estimator vulnerability to heavily degraded inputs as an open problem. The paper was not evaluated under non-Gaussian noise or protocol shift.
The representation comes from Modality-Agnostic Structural Image Representation Learning for Deformable Multi-Modality Medical Image Registration (MASR-Net). MASR-Net computes neighbourhood self-similarity on deep features and trains them with random intensity curves and anatomy-aware contrastive learning. This lets it register abdomen MR to CT at 85.61% Dice, up from 37.32% before alignment. It also beats hand-crafted descriptors, which the authors describe as noise-sensitive (mutual information) or insufficiently discriminative (MIND). MASR-Net handles only static volumes and never faces undersampling or clinical noise. Our change is to extend its contrastive "same point, different appearance" transformations from intensity curves to acquisition transformations: different undersampling masks, coil noise, k-space spikes as in Dey et al., and motion blur. We then use the resulting descriptor as the motion estimator's input inside a dynamic reconstruction.
Two results constrain the design. Reducing Textural Bias Improves Robustness of Deep Segmentation Models shows that clean-trained segmenters collapse to 0–1% Dice under heavy corruption, while mixed Gaussian and impulse-noise training restores Dice above 90%. It also warns that overly aggressive corruption erases low-contrast boundaries. Two consequences follow. Augmentation is a strong competitor and must be matched exactly. We must also measure clean-data cost, not just robustness. Structure-Driven Unsupervised Domain Adaptation for Cross-Modality Cardiac Segmentation uses Canny edges plus landmarks to reach 85.8% Dice (MRI→CT) and 74.8% Dice (CT→MRI). It relies on clean, sharp edges, and edges are exactly what undersampling streaks destroy. For that reason we use learned self-similarity rather than edge maps.
The motivation for making robustness come from diverse corruptions rather than scale is the diversity finding in On the Robustness of Medical Vision-Language Models: Are they Truly Generalizable?. That paper found small, diverse multi-modality fine-tuning sets beat large single-modality ones for corruption robustness. Motion-free alternatives avoid alignment altogether, and we treat them as competitors rather than components. Gabor Primitives for Accelerated Cardiac Cine MRI Reconstruction fits a per-scan parametric model and reports +2.34 dB PSNR over compressed-sensing reconstruction at radial R≈23, but it takes 2–4 minutes per slice. The bidirectional ConvLSTM in Myocardial Segmentation of Cardiac MRI Sequences with Temporal Consistency for Coronary Artery Disease Diagnosis smooths implicitly, and its authors acknowledge it fails under ghosting and brightness heterogeneity.
Research Program
We will build three pieces. The first is a descriptor network $\phi$. It maps a coil-combined, partially reconstructed cine frame to a per-pixel self-similarity descriptor over local and dilated neighbourhoods, following the MASR-Net construction. It is pre-trained contrastively with a loss that pulls together the descriptor of the same anatomical location under two independently sampled acquisition transformations and pushes apart different locations. Ground-truth correspondence comes for free because both views are generated from the same fully sampled frame. The transformation family covers random Cartesian and radial-like masks at 8x to 24x, complex Gaussian coil noise at varied signal-to-noise ratio, sparse k-space spikes, in-frame motion blur, and random monotone and inverting intensity curves. The intensity curves stand in for contrast and modality change.
The second piece replaces the input of the GRAFT-style motion estimator in the reconstruction-driven pipeline: it now receives $\phi$ of each frame instead of the frame's intensities. The third piece is joint fine-tuning of $\phi$, the motion estimator and the reconstruction network, using only the final reconstruction loss, exactly as in the source pipeline. That keeps the comparison to it clean.
The analysis asks three questions in order. Does the descriptor input reduce motion-field error on degraded frames? Does lower motion error translate into fewer ghosting and blurring artifacts? Do both hold on protocols never seen in training? A cheap diagnostic comes first. We will measure how much mask identity and noise level can be linearly decoded from the motion estimator's intermediate features, with and without $\phi$. If $\phi$ does not reduce this leakage, the mechanism is not what we claim, whatever the image metrics say.
Modality generalization serves here as a design constraint and a secondary test, not as the main application. The file contains no paired dynamic data across physical modalities, so we do not claim CT-to-cine transfer. We do check that $\phi$ keeps its cross-modality ability by running it zero-shot as a registration descriptor on the Learn2Reg abdomen MR-CT task.
Baselines
The naive reference is the same unrolled reconstruction network with no motion compensation, trained with the same data and loss. It sets the floor that any use of motion must beat.
The strongest existing method is the reconstruction-driven motion-compensated pipeline of Motion-Compensated MR CINE Reconstruction With Reconstruction-Driven Motion Estimation. It is the closest system and the one that already solved the circular dependency on in-distribution data, so it is the most direct competitor. The file reports it as superior in PSNR and SSIM at 8x, 12x, 16x and 20x on OCMR, at about 5.5 s per slice. No reported per-factor PSNR/SSIM value was found on Lacuna for this method on OCMR, so we will re-implement it and report our own numbers. It will be trained on the same fully sampled OCMR training slices, with the same number of gradient steps and the same hyperparameter search budget as our model.
The ablation that isolates our key ingredient is this same pipeline trained with exactly the acquisition-transformation augmentations used to pre-train $\phi$, applied to its raw-intensity input. It gets the same compute and step count as our model. Beating it is the result that matters. We will count it as beaten only if, on held-out protocols, our model reduces endpoint motion error and improves PSNR by a margin larger than the seed-to-seed spread across three seeds, at no more than a small clean-data cost (for example, under 0.3 dB in-distribution). A second ablation swaps the learned $\phi$ for the fixed MIND descriptor. It tests whether a hand-crafted structural input is already enough.
The strongest motion-free competitor is Gabor primitives Gabor Primitives for Accelerated Cardiac Cine MRI Reconstruction. It reports +2.34 dB PSNR over compressed-sensing reconstruction at radial R≈23, at 2–4 minutes per slice. It will be run per scan with its published optimization budget. Beating it means matching its PSNR within its reported margin at more than 20x lower latency, or exceeding it on corrupted inputs. A compressed-sensing reconstruction with temporal sparsity (PICS) serves as the classical reference.
Why not just…?
Why not just augment the existing motion-compensated pipeline with noise and undersampling? This is the real threat. Augmentation alone restored segmentation Dice from 0–1% to above 90% in Reducing Textural Bias Improves Robustness of Deep Segmentation Models. If it works as well here, our descriptor adds nothing. The matched-augmentation ablation settles this directly. Our bet rests on a structural argument. Augmentation asks the motion network to become invariant to corruptions as a side effect of the reconstruction loss. The contrastive pre-training instead supplies explicit correspondence supervision under each corruption, which the reconstruction loss never provides. If the ablation ties our model on held-out protocols, we will report that augmentation suffices and drop the descriptor claim.
Why not just avoid motion estimation, using Gabor primitives or recurrent temporal smoothing? Removing warping removes warping artifacts. But the Gabor approach costs 2–4 minutes per slice, and the ConvLSTM approach in Myocardial Segmentation of Cardiac MRI Sequences with Temporal Consistency for Coronary Artery Disease Diagnosis is reported to fail under ghosting and brightness heterogeneity. Both are included as baselines. If Gabor primitives beat us on corrupted inputs, the conclusion is that per-scan fitting is worth its latency. That is a useful answer too.
Why not just use a fixed structural descriptor like MIND, or mutual information? Descriptors like these are cheap and need no training. MASR-Net reports that mutual information is noise-sensitive and MIND lacks discriminative power for fine anatomy, and it outperforms both on MR-CT. Neither was tested on aliased frames. The MIND-input ablation measures this directly in our setting.
Why not translate or denoise frames first, then estimate motion? A two-stage denoise-then-flow pipeline is the modular route. It reintroduces the error propagation that Motion-Compensated MR CINE Reconstruction With Reconstruction-Driven Motion Estimation removed. Generative denoisers also risk hallucinated structure, a concern raised in Gabor Primitives for Accelerated Cardiac Cine MRI Reconstruction. We include a denoiser-plus-GRAFT variant at equal compute to quantify this rather than assume it.
Proposed Experiments
Training uses fully sampled OCMR (v1.0) slices, restricted to 1.5T scans and Cartesian masks at 8x–16x, following OCMR (v1.0)--Open-Access Multi-Coil k-Space Dataset for Cardiovascular Magnetic Resonance Imaging. Evaluation is organized as a ladder of shifts: in-distribution 1.5T Cartesian; unseen 20x and 24x acceleration; unseen 3T scans; unseen radial-like masks; and OCMR's prospectively undersampled free-breathing scans, which contain natural rather than simulated corruption. Each held-out setting is also tested under injected corruptions from the MediMeta-C types that apply to MRI (Gaussian and impulse noise, motion blur, contrast change) plus k-space spikes. ACDC magnitude sequences, retrospectively undersampled, provide a second-site test with segmentation labels.
Image quality is measured by PSNR and SSIM on the fully sampled references. Motion quality is measured by endpoint error against flow estimated on the fully sampled frames. We also measure alignment artifacts directly. A segmenter trained on clean ACDC is applied to the reconstructions, and we report Dice plus frame-to-frame boundary jitter. Ghosting is quantified as spurious duplicate edges in the residual map. For the prospective scans, which lack ground truth, we rely on this segmentation-based jitter together with blinded reader ranking.
The decisive check is the gap between our model and the matched-augmentation ablation, plotted against distance from the training protocol. Our claim predicts the gap is near zero in-distribution and grows along the shift ladder, and that the leakage diagnostic falls alongside it. A flat gap refutes the claim. A final sanity test runs $\phi$ zero-shot on Learn2Reg abdomen MR-CT registration. Staying near MASR-Net's 85.61% Dice would confirm the descriptor kept its modality-agnostic character. A large drop would show that acquisition invariance was bought at the cost of contrast invariance, which would be worth knowing before anyone extends this to true cross-modality dynamic imaging.
References
- Motion-Compensated MR CINE Reconstruction With Reconstruction-Driven Motion Estimation
- Modality-Agnostic Structural Image Representation Learning for Deformable Multi-Modality Medical Image Registration
- AnyStar: Domain randomized universal star-convex 3D instance segmentation
- Reducing Textural Bias Improves Robustness of Deep Segmentation Models
- Structure-Driven Unsupervised Domain Adaptation for Cross-Modality Cardiac Segmentation
- On the Robustness of Medical Vision-Language Models: Are They Truly Generalizable?
- Gabor Primitives for Accelerated Cardiac Cine MRI Reconstruction
- Myocardial Segmentation of Cardiac MRI Sequences With Temporal Consistency for Coronary Artery Disease Diagnosis
- OCMR (v1.0)--Open-Access Multi-Coil k-Space Dataset for Cardiovascular Magnetic Resonance Imaging