acceptodds
Under review as a conference paper at ICLR 2027

Free-Breathing Ungated Cine Cardiac MRI in Seconds via a Cross-Slice Transformer

Abstract

Cine cardiac MRI is the clinical gold standard for measuring heart function, but a full exam typically takes an hour or more, since it depends on repeated breath-holds and electrocardiogram (ECG) gating that are difficult or unreliable for patients with dyspnea or arrhythmia. Real-time MRI captures each 2D frame within a fraction of a heartbeat, needing neither, but slices are acquired sequentially, each capturing the heart at a different respiratory position and cardiac state. Reassembling them into a coherent 3D volume, known as slice-to-volume reconstruction (SVR), still relies on breath-holds, ECG, or self-gating in existing methods, which also take minutes per subject. We present , to our knowledge the first cardiac SVR method that is simultaneously , , and . Rather than synchronizing slices, we on a reference slice that defines the target cardiac state: a cross-slice transformer reasons jointly over the entire stack and, in a single forward pass, predicts dense 2D-to-3D deformations that correct respiratory and cardiac motion together, and varying the reference yields a 3D cine. Trained on 628 subjects from five public datasets, CardioStitch reconstructs a full cine in 2 s, over faster than optimization-based methods. It outperforms baselines given the full real-time series of every slice plus ECG or self-gating on all image metrics, cutting respiratory error from 4.2 to 0.9 mm, and on all functional metrics under simulated atrial fibrillation. CardioStitch also transfers without retraining to real free-breathing, ungated scans, including from a patient with atrial fibrillation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cardiac MRI, slice-to-volume reconstruction, motion correction, dense correspondence, medical image reconstruction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gcsVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CardioStitch, a feed-forward method that builds a 3D cardiac cine from free-breathing, ungated real-time short-axis slices. It fine-tunes VGGT with a slice-location token and a reference-slice token, predicts a dense 2D-to-3D displacement per pixel, and splats the slices into a volume. Training uses a splatting reconstruction loss with a heart-region term plus an observation-consistency loss that samples the GT volume at the predicted positions. All quantitative results come from simulated breathing (rigid translation) and simulated HRV/AF applied to gated breath-hold cine from five public datasets. The method beats static SVR and dense-temporal baselines on most metrics and runs in about 2 s. There is also a qualitative test on real scans from two subjects.

Critical strengths
  1. The problem is real and well framed. Breath-holds and ECG gating are a true pain point for heart failure and AF patients, and conditioning on a reference slice instead of binning cardiac phases is a clean way around the synchronization problem. Predicting geometry rather than generating pixels also means every voxel comes from an observed pixel, which matters a lot for clinical trust.
  2. The analysis goes beyond a results table. The loss ablation shows clearly that L_obs handles breathing and L_h handles cardiac motion, and the attention-masking experiment in the appendix (block 3 reads the reference, blocks 12-17 fix breathing) is a nice piece of evidence that the model works the way the authors claim.
Critical weaknesses
  1. Almost all of the numbers come from a fairly simple simulation. Breathing is modeled as a rigid 3D translation (a sin^6 curve with a random tilt), and the same motion model is used for both training and testing. Real breathing is nonrigid and has rotation, hysteresis, drift and slice-to-slice variation, and none of that is tested. The simulated AF only re-times the same 12-phase gated cycle, so real beat-to-beat changes in shape and filling are not really there. On top of that, EPE is computed after removing the per-subject mean error, and all image metrics are computed after a 6-DOF rigid registration. So I think the 0.9 mm number very likely overstates what will happen on real ungated data.
  2. The real-data evidence is too thin to support the clinical story. There are only two subjects (one healthy, one AF), and there is no quantitative comparison against anything, only LAX images and LV volume curves. Also, the comparison is hard to read: CardioStitch is the only method trained on this exact simulator, while the baselines are either used outside their design assumptions (static SVR on a beating heart), run with components turned off (SVRTK robust statistics), or re-implemented without official code (CiNeVol). FC-SVR and Stolt-Ansó et al. are listed in Table 1 but not compared.
Questions for authors
  1. How does performance change when the test-time motion differs from the training model, e.g. nonrigid or slice-specific breathing, rotation, or amplitudes outside U(11.45, 26.15) mm? Even one such stress test would help a lot.
  2. The location token only encodes a scalar through-plane position z_i. How would the method deal with oblique or non-parallel slices, or with different slice thickness/gap than in training?
  3. For the 3D cine, all non-reference slices are fixed to one frame and only the central slice changes. So under AF, base and apex motion is never observed and has to be inferred. How sensitive are the results to the choice and quality of the reference slice (e.g. a basal reference, or a reference frame with artifacts)?
  4. For the two prospective subjects, can you report EF/EDV/ESV against a standard breath-hold cine of the same subject?
  5. CiNeVol gets the exact R-peak times and the true breathing signal and still loses on most metrics. Did you check that your re-implementation reproduces the numbers in the original paper on any comparable setting?
4 comments

(Part 1/3)

We sincerely thank the reviewer for the thoughtful feedback and for recognizing the clinical motivation, the geometry-based reconstruction formulation, and the analysis of the learned cardiac and respiratory motion. Your suggestions motivated a substantially extended prospective evaluation and several additional experiments to assess the robustness and fairness of our evaluation.

Q1: Does the reported accuracy depend on rigid registration or per-subject mean-error removal during evaluation?

A1: We agree that reporting both corrected and uncorrected results is important. Following the reviewer's suggestion, we evaluated all methods without rigid registration for image metrics and without per-subject mean-error removal for respiratory EPE.

methodEPE, mean removed (mm)EPE, raw (mm)PSNR, registered (dB)PSNR, raw (dB)
SVRTK4.795.6721.8219.52
NiftyMIC4.695.0120.0320.37
NeSVoR6.386.9820.9119.60
Fetal CMR 4D4.805.8522.2320.31
CiNeVol4.235.2323.2722.46
CardioStitch (Ours)0.920.9024.5024.25

For CardioStitch, these evaluation corrections change EPE by only 0.02 mm and PSNR by 0.25 dB. CardioStitch maintains its advantage without either correction, indicating that the reported performance does not depend materially on these evaluation procedures.

We have included these results in Appendix A.11 (Table A9).

We acknowledge that the respiratory EPE is measured on simulated data with known motion, rather than on prospective acquisitions. For real scans, we have instead added independent clinical-reference comparisons, as described in Q3.

Q2: How robust is CardioStitch to respiratory motion outside the training distribution and to realistic cardiac rhythm variations?

A2: We agree that our original respiratory simulation represents a simplified model of real breathing. As suggested, we evaluated the unchanged model on all 180 test subjects under six out-of-distribution respiratory settings.

Test-time respirationEPE, ours (mm)EPE, naive stacking (mm)PSNR (dB)
as in training0.964.724.23
amplitude ×1.51.756.923.55
amplitude ×22.968.823.04
waveform sin² (less time resting at end-expiration)1.154.923.80
waveform sin¹⁰ (more time resting at end-expiration)0.894.224.52
different AP/SI ratio (0.7)0.994.724.18
per-slice amplitude varies ±30%0.934.624.27

Even at twice the training respiratory amplitude, CardioStitch maintains substantially lower EPE than naive stacking (2.96 vs. 8.83 mm), with a PSNR reduction of only 1.2 dB relative to the in-distribution setting. Changes in respiratory waveform, AP/SI ratio, and slice-dependent amplitude have relatively small effects.

Regarding cardiac motion, our AF simulation varies beat duration and contraction amplitude within each subject's gated-cine trajectory. It isolates irregular cardiac timing but does not simulate arbitrary regional changes beyond that trajectory. We acknowledge that rotation, hysteresis, drift, non-rigid respiration, and more complex beat-to-beat cardiac motion remain important directions for further investigation.

The additional respiratory experiments are included in Appendix A.5 (Table A5), and the limitations of the respiratory simulation are acknowledged in Sec. 6.

(Part 2/3)

Q3: Can the prospective reconstruction be quantitatively validated against standard breath-hold cine MRI?

A3: Following the reviewer's suggestion, we expanded the prospective evaluation from 2 to 13 subjects (10 healthy volunteers and 3 patients, including one AF patient), each with a separately acquired breath-hold ECG-gated cine.

Without retraining, CardioStitch recovers 89% of the reference stroke volume (95% CI: 82–96%) and achieves an EF error of 9.6 pp. Its EF is underestimated in 12 of 13 subjects, primarily due to an overestimation of ESV, consistent with the negative EF bias observed in the retrospective evaluation. The mean EDV bias is +9 mL (6%), while the mean ESV bias is +19 mL, explaining the systematic underestimation of EF.

Fetal CMR 4D, using 40 consecutive frames per slice and cardiac timing information, achieves a comparable EF error (p = 0.84) and lower EDV error (p = 0.01). However, it requires 22–49 minutes of optimization per subject, compared with seconds for CardioStitch.

These results provide quantitative evidence of CardioStitch's applicability to real free-breathing, ungated acquisitions, while also revealing its remaining functional estimation bias. We emphasize that the separately acquired gated cine is a clinical reference rather than beat-specific ground truth.

The extended evaluation is included in Sec. 5.2 (Table 4) and Appendix A.9.

Q4: Are the baseline comparisons fair given their different design assumptions, input requirements, and implementation settings?

A4: We clarify that Dangi et al. is retrained using the same data and simulator as CardioStitch, whereas the five optimization-based methods perform subject-specific reconstruction and require no population training.

The two synchronization-based baselines receive additional information unavailable to CardioStitch: CiNeVol receives the exact simulated R-peak times and respiratory signals, and Fetal CMR 4D receives the full temporal series of every slice and the nominal R–R interval.

Regarding individual configurations:

  • Static SVR methods: SVRTK, NiftyMIC, and NeSVoR assume static anatomy. They are included as representative SVR methods applicable to a single slice stack, rather than methods specifically designed for asynchronous cardiac motion.
  • SVRTK: Robust statistics are disabled because, in our single-stack setting, they reject every slice.
  • FC-SVR: Its original training requires ground-truth slice correspondences to a static anatomy. Extending its supervision to asynchronous cardiac states would require registration-derived pseudo ground truth.
  • Stolt-Ansó et al.: Its alignment objective relies on slice intersections, which are absent in a single parallel short-axis stack.

To provide an additional same-data learning-based comparison, we retrained CardioStitch without cross-slice attention using identical training settings. Its EF error increases from 7.5 to 37.2 pp under AF, demonstrating the importance of cross-slice reasoning.

We have clarified the baseline assumptions and configurations in Sec. 4 and Appendix A.10.1 and included the retrained ablation in Sec. 5.3 (Table 5).

Q5: How does CardioStitch handle different slice spacings, thicknesses, and non-parallel acquisition geometries?

A5: CardioStitch encodes physical slice positions in millimeters rather than discrete slice indices, allowing different spacings and gaps to be represented directly. Our training data cover slice spacings from 5 to 12 mm.

slice spacingnEF error (pp)Dice
12 mm957.00.857
10 mm748.00.843
9.5–9.6 mm32.30.850
8 mm47.60.849
5–6 mm412.60.66

Performance is relatively consistent across commonly represented spacings. The four subjects with 5–6 mm spacing exhibit lower accuracy, potentially due to the limited number of thin-slice stacks in training (only 8 of 628 training subjects have spacing ≤6.5 mm). The prospective scans have a slice spacing of 10 mm and are reconstructed without any fine-tuning on the prospective subjects.

The current method moves observed pixels without explicitly deconvolving the slice profile. Consequently, slice thickness is not explicitly modeled, and the reconstructed through-plane resolution remains limited by the acquired slices and partial-volume effects.

For non-parallel slices, a scalar through-plane position is insufficient to represent the full acquisition geometry. Extending the location encoding to incorporate full spatial slice poses would be a natural direction for generalizing the framework to multi-view acquisitions.

The subgroup analysis is included in Appendix A.11 (Table A10), and the limitations regarding acquisition geometry are acknowledged in Sec. 6.

(Part 3/3)

Q6: How sensitive is CardioStitch to the position and image quality of the reference slice?

A6: We conducted two additional sensitivity analyses on 60 test subjects.

First, the model was trained using only the central slice as the reference. We moved the reference slice at test time away from the central slice toward either the apex or the base, computing EF from the ED and ES frames.

Distance from center (slices)01234
EF error (pp), toward the apex8.610.412.515.823.6
EF error (pp), toward the base8.610.516.125.943.5

The results degrade gradually as the reference moves away from the center, with greater sensitivity toward the base. A one-slice deviation increases EF error by approximately 1.9 pp (p = 0.005). However, the automatically selected central slice lies on average only 0.7 slices from the middle of the LV in the test set and contains the LV in all 13 prospective stacks.

Second, we evaluated the effect of image degradation in the reference slice.

reference imagePSNR change (dB)EF error change (pp)
half resolution−0.14+1.0
Gibbs ringing−0.14+0.2
ghosting−0.17+0.6

CardioStitch exhibits limited sensitivity to these tested image degradations.

We acknowledge that a single reference frame cannot fully characterize arbitrary regional cardiac dynamics, particularly under arrhythmia. Extending the framework to exploit multiple temporal observations per slice is therefore an important direction for future work.

These sensitivity analyses are now included in Appendix A.3.

Q7: Has the CiNeVol reimplementation been validated, and why does it exhibit relatively large respiratory motion errors despite receiving oracle signals?

A7: CiNeVol's original evaluation uses in-house patient data and licensed XCAT phantoms that are unavailable to us. Therefore, we cannot directly reproduce its original quantitative results. Nevertheless, our implementation follows the architecture, loss functions and weights, PSF sampling, optimizer, and iteration budget described in the original paper and supplementary material.

Its performance is consistent with the expected behavior of its synchronization-based formulation. CiNeVol uses ECG-derived R-peak times to assign cardiac phases, assuming that the interval between two consecutive R-peaks corresponds to a complete cardiac cycle. Under HRV, CiNeVol achieves the lowest EF error among all evaluated methods (5.58 pp), with PSNR close to CardioStitch (24.26 vs. 24.59 dB). Under AF, truncated cardiac cycles violate the complete-cycle assumption of phase assignment, increasing its EF error to 10.50 pp.

As its authors report, CiNeVol also underestimates cardiac contraction. Consistent with this observation, our reimplementation exhibits a negative EF bias, which increases in magnitude from −2.7 pp under HRV to −8.1 pp under AF.

Regarding respiratory motion, even with oracle respiratory signals, CiNeVol achieves a through-plane EPE of 4.23 mm, with a correlation of 0.32 between estimated and true displacements. For comparison, predicting no respiratory displacement gives an EPE of approximately 4.7 mm. We hypothesize that this is related to an ambiguity in per-subject optimization: a through-plane displacement changes the observed cross-section, which can be explained by modifying the reconstructed anatomy rather than correcting the slice location.

In contrast, CardioStitch leverages a population-trained prior learned from 628 subjects to help resolve this ambiguity. Incorporating a population-level prior is also identified as a future direction in the original CiNeVol paper.

We have clarified the implementation details in Appendix A.10.1 and the functional accuracy comparison in Sec. 5.1.

We thank the reviewer again for the constructive feedback and hope that these additional experiments provide a more comprehensive assessment of CardioStitch's accuracy, robustness, and practical limitations.

Thank you for providing the additional experiments. I appreciate the author's efforts, especially the in-depth assessment of the real data and the results of the unregistered cases. This topic indeed addresses some of my past concerns.

However, I am still not completely convinced by the significant improvement brought by the baseline model. CardioStitch was specifically trained for this simulation scenario, based on the real anatomical volume. Multiple baseline models were tested under conditions where the initial design was not applicable. In particular, the performance of CiNeVol was only slightly better than without motion correction, which makes me doubt how much we can draw conclusions from this comparison. The results of the real data are useful, but EF still has a systematic underestimation problem, and the new robustness test does not truly cover more complex real-world movement scenarios. Overall, I think the test paper has improved, but my main concerns still remain, so I will keep the score at 2 points.

Reviewer_FCZjOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CardioStitch reconstructs a coherent 3D cardiac cine from free-breathing, ECG-free short-axis slices. Instead of first determining which slices share the same cardiac phase, it selects one reference slice (the central slice’s frame) to define the target cardiac state. A VGGT-based transformer with location and reference-state tokens then predicts, in one forward pass, dense 2D→3D displacements that align every slice to that cardiac state and to a common end-expiratory respiratory configuration. Aligned slices are splatted into a volume, and sweeping the reference frame yields a 3D cine.

It is trained on 628 subjects from five public gated cine datasets with simulated free-breathing (rigid translational respiration and time-warped cardiac phases). It is evaluated under simulated AF and HRV against SVRTK, NiftyMIC, NeSVoR, Dangi et al., Fetal CMR 4D, and a reimplemented CiNeVol. The paper reports the best image metrics (EPE ~0.9 mm vs ~4.2 mm), the best functional metrics under AF, and a runtime of ~2 s. Two prospective real-time scans (one healthy, one AF) are shown qualitatively.

Critical strengths
  1. Clear, practical formulation. Conditioning on a reference slice instead of phase synchronization is a clean way to pose cardiac SVR. One forward pass, no gating, no per-subject optimization, and ~2 s runtime are a real advantage over methods that take minutes.

  2. Careful analysis of what the model learns. Loss ablations separate respiratory vs cardiac correction (Table 4, Fig. A1). Test-time attention masking localizes reference reading to block 3 and respiratory correction to blocks 12–17 (Appendix A.2), which is more informative than typical baseline tables alone.

  3. Honest limitations and transparent curation. The paper states single-stack geometry limits, thick-slice resolution, and the small prospective set (n=2). Dataset filtering (898/1337 retained) is documented, and real free-breathing scans including an AF patient are included without retraining.

Critical weaknesses
  1. Simulation does not support strong AF / free-breathing claims. All quantitative results use the same gated cine as ground truth, and the respiratory model at test time is the same rigid translational model used in training. Every simulated cardiac state lies on the subject’s gated trajectory; beat-to-beat variation is limited to truncating or prolonging that trajectory (A.4), so no beat contains morphology or contraction patterns absent from the gated cine. CardioStitch’s AF and HRV results are also nearly identical on several metrics (e.g., EPE 0.917/0.916, Dice 0.847/0.847). So the experiment mainly shows that the method is insensitive to this timing perturbation. It does not show robustness to real arrhythmia.

  2. Baseline comparison is weak. Static SVR methods (SVRTK, NiftyMIC, NeSVoR) get a single-frame stack they were not designed for, and their respiratory EPE (4.7-6.4 mm, Table 2) is essentially no better than naive stacking (4.77 mm, Table A2). CiNeVol is a reimplementation with oracle signals, yet still reports ~4.2 mm EPE. Dangi et al. is retrained, but there is no same-data baseline without cross-slice attention, so the “cross-slice transformer” contribution is not isolated.

  3. No architectural ablations. Only the loss terms (Table 4) are ablated. The location token, reference-state token, VGGT initialization, frozen DINOv2, and choice of central slice as reference are not tested. The test-time attention masking in A.2 is informative but is not a retrained frame-only model. Table 4 supports that LobsL_{obs} helps, but there is not an isolating experiment to support the claim that coverage normalization is responsible for the problem (Fig. 3). Only a toy illustration is given.

  4. Reporting and clinical framing are thin. Results are single aggregated numbers across the 180 test subjects (no SDs, CIs or paired tests) and do not provide a breakdown by dataset, pathology or the held out Canon/Philips domains described in A.3. Under HRV, Fetal CMR 4D and CiNeVol have lower EF error than CardioStitch (5.98 / 5.58 vs 7.88). Units for EF, LVM and VTC are not given in Table 3.

Questions for authors
  1. The AF simulation (A.4) only truncates or prolongs the subject’s gated cardiac trajectory, so every beat reuses the same morphology and contraction pattern. Can you evaluate on a simulation where these also change (e.g., scaling the amplitude or mixing phases from different cycles)? If any real arrhythmia data with even a coarse independent reference are available, that would strengthen the AF claim.

  2. Please add a control that the output depends on non-reference slice content (e.g., shuffle/replace non-reference slices, reference-only baseline, or regional wall-motion cases).

  3. How are slices extracted at respiratory-shifted positions (interpolation scheme)? Do results hold under OOD respiration (different waveform, amplitude outside training range, rotation/deformation)?

  4. Can you validate the CiNeVol reimplementation against the behavior reported in the original paper? Please also add FC-SVR (retrained) and a no-cross-slice-attention baseline on the same simulator. Also, CardioStitch keeps the acquired slice spacing while baselines are reconstructed at 1.4 mm isotropic (A.6.1). Appendix A.6 re-blurs PSF-deconvolved outputs with a matched kernel and scores all methods on the same slices; how does this account for the resolution mismatch when computing LV volume, EF and VTC?

  5. For the two prospective subjects, which frames are used for non-reference slices, and is there a gated breath-hold reference so EF/volume agreement can be reported?

6 comments

(Part 1/5)

We sincerely thank the reviewer for the detailed and constructive feedback, and for recognizing the practical formulation, the analysis of learned motion, and the transparent discussion of data curation and limitations. Following your suggestions, we have conducted additional architectural ablations, controlled experiments, robustness evaluations, and statistical analyses, as well as a substantially extended prospective evaluation.

Q1: How does CardioStitch perform under realistic arrhythmia and respiratory motion beyond the training simulation?

A1: We agree that our simulated AF remains within each subject's gated-cine trajectory. Although the simulation varies both beat duration and contraction amplitude, it does not introduce new regional deformation patterns. Specifically, when a short beat is interrupted, the subsequent beat starts from a partially filled ventricle and ejects less. Mixing phases from different cardiac cycles is not possible with our retrospective datasets, since each subject's gated cine contains only one cardiac cycle.

The purpose of this experiment is to isolate the impact of irregular cardiac timing on synchronization-based reconstruction: from HRV to AF, CiNeVol's EF error increases from 5.58 to 10.50 pp, while ours remains relatively stable (7.88 vs. 7.49 pp).

For respiratory motion, we evaluated the unchanged model on all 180 test subjects under six out-of-distribution settings. The first row repeats the training-range breathing simulation using the same random draws as the other settings.

breathing at test timeEPE, ours (mm)EPE, naive stacking (mm)PSNR (dB)
as in training0.964.724.23
amplitude ×1.51.756.923.55
amplitude ×22.968.823.04
waveform sin² (less time resting at end-expiration)1.154.923.80
waveform sin¹⁰ (more time resting at end-expiration)0.894.224.52
different AP/SI ratio (0.7)0.994.724.18
per-slice amplitude varies ±30%0.934.624.27

The performance degrades gradually with increasing respiratory amplitude. Changes in waveform, AP/SI ratio, and slice-dependent amplitude have relatively small effects.

Regarding the simulation procedure, respiratory-shifted slices are extracted using trilinear interpolation, and intermediate cardiac states are obtained by linear interpolation between adjacent cine phases.

We have also extended the prospective evaluation to 13 real free-breathing, ungated subjects, including one AF patient, each with a separate breath-hold gated cine as a clinical reference. CardioStitch recovers 89% of the reference stroke volume across all 13 subjects, compared with 19% for naive stacking. For the AF patient, who exhibits the greatest beat-length variation among the 13 subjects, this ratio is also 89% (see Q9).

The new respiratory experiments are included in Appendix A.5 (Table A5). Rotation and non-rigid respiration remain untested, as acknowledged in Sec. 6 and Appendix A.5.

We emphasize that the simulated AF experiments isolate irregular cardiac timing, while the prospective experiments provide complementary evidence of performance on real acquisitions. Neither establishes the accuracy of arbitrary beat-to-beat regional deformation recovery, which remains an important direction for further investigation.

Q2: What is the contribution of cross-slice attention beyond the training data and loss functions?

A2: We agree that a retrained model without cross-slice attention is an important control. As suggested, we retrained CardioStitch under identical settings, except that all attention blocks operate within individual slices. The parameter count, training data, losses, and training recipe remain unchanged.

PSNR (dB)EPE (mm)EF error (pp)
w/o cross-slice attention (AF)23.422.2837.2
Ours (AF)24.500.927.5
w/o cross-slice attention (HRV)23.542.2538.2
Ours (HRV)24.590.927.9

Removing cross-slice attention increases the EF error from 7.5 to 37.2 pp under AF (paired Wilcoxon p < 10⁻³⁰, 180 subjects). This confirms the importance of cross-slice information exchange for cardiac-state alignment, complementing our original test-time attention-masking analysis.

These results are now included in Sec. 5.3, Table 5 (AF), and Table A1 (HRV).

(Part 2/5)

Q3: What are the contributions of the conditioning tokens and pretrained components, and how sensitive is the model to reference-slice selection?

A3: Following the reviewer's suggestions, we conducted the following additional ablations and sensitivity analyses.

Physical-location token. We retrained the model without the physical-location token under identical training settings. Removing the token has little effect on respiratory EPE or EF error but reduces PSNR by approximately 0.3 dB (lower in 76–77% of subjects, paired Wilcoxon p < 10⁻¹⁵). This is consistent with the fact that physical slice coordinates also enter the reconstruction through Eq. (5).

Reference-state token. Removing the reference-state token at test time, while keeping all images and positions unchanged, increases the EF error from 7.7 to 48.5 pp (60 test subjects, p < 10⁻¹⁰). The reconstructed EF decreases from 55% to 12%, confirming the importance of explicit reference conditioning.

Reference-slice selection. The model was trained using only the central slice as the reference. We moved the reference slice at test time from the center toward either the apex or the base on 60 subjects, computing EF from the ED and ES frames.

reference position (slices from centre)01234
EF error (pp), toward the apex8.610.412.515.823.6
EF error (pp), toward the base8.610.516.125.943.5

A one-slice deviation increases EF error by approximately 1.9 pp (p = 0.005), with greater sensitivity toward the base. In practice, the central slice is on average only 0.7 slices away from the middle of the LV in our test set and contains the LV in all 13 prospective subjects.

Reference image quality. Half-resolution images, Gibbs ringing, and ghosting reduce PSNR by only 0.14–0.17 dB and increase EF error by at most 1.0 pp under the tested perturbations.

Pretrained components. We initialize Cardiac VGGT from VGGT and freeze the DINOv2 encoder during training. Unfreezing DINOv2 did not provide consistent improvements during development. We acknowledge that a systematic comparison of alternative initialization and pretraining strategies, including training from scratch, remains outside our current evaluation.

These new architectural and reference-sensitivity experiments are included in Sec. 5.3, Tables 5 and A1, and Appendix A.3.

Q4: Does the reconstruction genuinely depend on the anatomical content of the non-reference slices?

A4: Yes. As suggested, we conducted three additional input controls on 60 test subjects, keeping the trained model unchanged. All changes are statistically significant compared with the unmodified model (p ≤ 0.012).

input to the modelPSNR change (dB)EF error (pp)VTC error (%)
Original input—7.76.8
Shuffled non-reference slices−5.511.617.7
Non-reference slices from another subject−9.213.031.3
Reference slice only−9.737.564.1

All controls substantially degrade reconstruction quality. Moreover, the reference-only input covers only approximately one-quarter of the heart ROI. EF is computed only for cases with a recognizable LV in the latter two settings (52 and 34 of 60 subjects, respectively), so the reported EF errors underestimate their overall degradation.

These results demonstrate that CardioStitch relies on the observed anatomical content and spatial arrangement of the non-reference slices rather than reconstructing the heart from the reference slice alone.

Additionally, we measured the stroke volume of the non-reference slices alone, obtaining 84% of the ground-truth value, comparable to 85% for the whole heart. This ratio decreases to 5% when all cross-slice attention is masked at test time. These results further support the recovery of cardiac contraction beyond the reference slice through cross-slice information exchange.

These experiments are included in Appendix A.3 (Table A3).

(Part 3/5)

Q5: Is there direct experimental evidence that coverage normalization weakens through-plane motion supervision?

A5: We agree that a controlled experiment independent of the network would help isolate the effect of coverage normalization. We conducted such an experiment during development and have now included it in Appendix A.4.

Specifically, we bypass the network and treat each slice's through-plane displacement as a free parameter, optimized directly by gradient descent on 16 subjects. Respiratory shifts are simulated as in training, so the true displacement is known. We measure the slope of the recovered displacement against the true displacement, where a slope of 1 indicates full recovery.

Optimization objectiveRecovered/true shift slopeHeart PSNR (dB)
Coverage-normalized splatting0.3620.8
Stop-gradient on normalization denominator0.6616.5
Normalized splatting + observation consistency0.9020.9

With the coverage-normalized splatting loss, the optimized shifts recover only 36% of the true respiratory displacement in terms of the fitted slope. Removing the normalization from the gradient improves the slope to 0.66, but introduces more coverage holes and substantially reduces reconstruction quality. In contrast, adding the observation-space term (L_obs) improves the slope to 0.90 while largely preserving reconstruction quality.

This controlled experiment supports our motivation for combining the two complementary operators. Consistent results are observed in the trained model: without smoothness regularization, adding the observation consistency loss reduces respiratory EPE from 2.417 to 0.867 mm (Table 5).

We note that this controlled optimization represents an optimistic setting, since the input slices are exact resamples of the target volume. Nevertheless, it isolates the effect of coverage normalization independently of network architecture and training.

The experiment is now described in Appendix A.4.

Q6: Are the baseline comparisons fair, and why are FC-SVR and other learning-based methods not included?

A6: SVRTK, NiftyMIC, and NeSVoR are designed for static anatomy. We include them as representative SVR methods applicable to a single slice stack and have clarified their design assumptions in Sec. 4. Their respiratory EPE is close to or worse than that of naive stacking, indicating that these subject-specific registration methods have difficulty resolving respiratory motion in this setting. In contrast, CardioStitch, which learns a prior across subjects, achieves approximately 0.9 mm EPE, compared with 4.7–6.4 mm for these baselines.

Dangi et al. is retrained using the same training data and simulation as CardioStitch. We have additionally included the requested retrained model without cross-slice attention (Q2).

Regarding FC-SVR, its training requires ground-truth slice motion relative to a static anatomy. While our simulator provides respiratory displacements, the cardiac correspondences between different cardiac states would require registration-derived pseudo ground truth, changing its original supervision protocol. We therefore do not include it as a quantitative baseline.

Regarding CiNeVol, we could not directly reproduce the original reported numbers because the in-house patient data and licensed XCAT phantoms used in the original evaluation are unavailable to us. Nevertheless, our implementation follows its original paper and supplementary material and receives the ground-truth cardiac timing and respiratory signals. It is the strongest baseline under HRV, achieving the lowest EF error (5.58 pp).

Even with the true respiratory signal, CiNeVol achieves a through-plane EPE of approximately 4.2 mm, with a correlation of 0.32 between the estimated and true displacements. Predicting no displacement gives an EPE of approximately 4.7 mm. We hypothesize that its limited respiratory correction is related to the through-plane ambiguity of per-subject optimization, which can explain displaced observations through changes in reconstructed anatomy rather than slice locations. A population-trained prior helps CardioStitch address this ambiguity. Incorporating a population-trained prior is also identified as a future direction by the authors of CiNeVol. Further implementation details are provided in Appendix A.10.1 and our response to Reviewer gcsV.

Resolution matching. Before evaluation, baseline reconstructions that deconvolve the imaging PSF are reconvolved with a subject-specific slice PSF and resampled to the acquired slice locations. All methods and the ground truth are therefore evaluated on the same spatial grid using the same segmentation model. We have clarified this protocol in Appendix A.10.

Reviewer_aaSpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CardioStitch turns asynchronously acquired cardiac MRI slices into a 3D cine without explicit gating or synchronization. A reference-conditioned transformer predicts how to move the slices into alignment. Most evaluation uses simulated acquisitions and with additional examples from real scans.

Critical strengths
  1. The reference-slice formulation is interesting, and reconstruction is fast. The simulated results across several datasets are promising.

  2. The loss ablations and attention interventions help explain how the model uses the reference and corrects respiratory motion.

Critical weaknesses
  1. I am not yet convinced that the method recovers the correct whole-heart dynamics under AF. Only the reference input changes over time; the other input frames stay fixed. The AF simulation changes the timing of a gated cine while keeping its motion trajectory. It does not test beat-to-beat regional motion changes beyond each subject’s gated-cine trajectory.

The real-scan results in Fig6 show smoother reconstructions and LV curves. There is no independent check that the contraction away from the reference slice is correct.

Questions for authors

Is there an independent check of regional motion or contraction amplitude outside the reference slice in the prospective scans? A controlled test with motion varying beyond the fixed gated-cine trajectory would also help.

2 comments

(Part 1/2)

Thank you for the careful review and for highlighting the reference-slice formulation, the reconstruction speed and the attention analysis. Your concerns regarding cardiac motion beyond the reference slice motivated additional experiments and a substantially extended prospective evaluation. Below, we provide point-by-point clarifications and present the corresponding additional experimental results.

Q1: Can CardioStitch recover whole-heart cardiac dynamics when only the reference slice changes over time?

A1: The reference frame specifies the target cardiac state, while each non-reference slice provides an observation of the subject's anatomy. CardioStitch predicts dense correspondences to bring these asynchronous observations into the target state.

To specifically evaluate cardiac motion beyond the reference slice, we conducted an additional experiment on 60 test subjects. We measured the stroke volume of the non-reference slices alone, i.e., the change in their combined LV volume over the cardiac cycle.

Evaluationstroke volume / GT
Whole heart0.85
Non-reference slices only0.84
Non-reference slices (cross-slice attention masked at test time)0.05

The non-reference slices recover a similar proportion of their ground-truth contraction amplitude as the entire heart, despite their input frames remaining fixed throughout the reconstructed cardiac cycle.

Furthermore, masking attention from the non-reference slices to the reference slice in block 3 increases their EF error from 8.8 to 48.7 percentage points (pp), approaching naive stacking (57.2 pp). This provides complementary evidence that the target cardiac state is communicated across slices through attention.

The same behavior is observed on prospective scans: masking reference attention in block 3 reduces the LV volume swing of the non-reference slices from 91.8 to 2.4 mL in the AF patient and from 52.0 to 1.7 mL in the healthy volunteer.

The stroke-volume analysis is included in Appendix A.3, complementing the attention-masking analysis in Appendix A.2.

While these results support the effectiveness of our reference-conditioned formulation in recovering global cardiac contraction, we acknowledge that incorporating multiple temporal observations per slice could further improve the recovery of spatially heterogeneous cardiac dynamics. Due to computational resource constraints, we have only evaluated the single-frame setting thus far. However, our framework can be easily extended to incorporate multiple frames per slice while remaining synchronization-free, which we consider an important direction for future work.

Q2: Does the AF simulation capture realistic beat-to-beat cardiac motion beyond the gated-cine trajectory?

A2: We agree that the current AF simulation does not introduce regional deformation patterns absent from the subject's gated cine. However, it varies both the duration and the contraction amplitude of individual beats: when a short beat is interrupted, the subsequent beat starts from a partially filled ventricle and ejects less.

The purpose of this experiment is to isolate the impact of irregular cardiac timing on synchronization-based reconstruction. From HRV to AF, the EF error increases from 5.98 to 7.79 pp for Fetal CMR 4D and from 5.58 to 10.50 pp for CiNeVol, while CardioStitch remains relatively stable (7.88 vs. 7.49 pp).

We acknowledge that the current simulation does not establish robustness to regional cardiac motion beyond the gated-cine trajectory. As discussed in A1, incorporating multiple temporal observations per slice is an important direction for addressing this limitation.

(Part 2/2)

Q3: Is cardiac contraction beyond the reference slice independently validated on prospective real-time scans?

A3: Following the reviewer's suggestion, we expanded the prospective evaluation from 2 to 13 subjects (10 healthy volunteers and 3 patients, including one with AF). Each subject also underwent a separate breath-hold ECG-gated cine scan, providing an independent clinical reference.

vs. breath-hold gated cinestroke volume, method / gatedEF error (pp)
Naive stacking0.1947.9
SVRTK0.2143.5
NiftyMIC0.1747.1
NeSVoR0.1847.1
Dangi et al.0.2344.3
Fetal CMR 4D1.1612.1
Ours0.89 (95% CI 0.82–0.96)9.6

Without retraining, CardioStitch recovers 89% of the reference stroke volume (95% CI: 82–96%), substantially outperforming the single-frame baselines in EF accuracy (paired Wilcoxon p < 0.001). Fetal CMR 4D, given 40 consecutive frames per slice and their cardiac timing, achieves a comparable EF error (p = 0.84) but requires 22–49 minutes of optimization per subject, compared with seconds for CardioStitch. CiNeVol is not included in the prospective comparison because it requires ECG-derived cardiac phases, whereas no ECG signal was recorded during these real-time acquisitions.

These results provide quantitative evidence of recovered whole-LV contraction on real acquisitions. Nevertheless, the separately acquired gated cine is a clinical reference rather than beat-specific ground truth, particularly for the AF patient. Independent validation of regional, beat-to-beat deformation remains an important direction for future work.

The extended evaluation is included in Sec. 5.2 (Table 4) and Appendix A.9.

We thank the reviewer again for the valuable suggestions and hope that these additional experiments clarify both the capabilities and limitations of our reference-conditioned formulation.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response

(Part 1/2)

We sincerely thank all three reviewers for their thoughtful and constructive feedback, and for recognizing the practical significance of synchronization-free cardiac SVR, the reconstruction efficiency of CardioStitch, and the analysis of its learned motion.

The reviewers' main concerns center on: (1) the recovery of whole-heart dynamics from a single asynchronous slice stack; (2) the validity of our simulated AF and respiratory motion and the generalization to real scans; (3) the contributions of individual architectural and methodological components; (4) the fairness and statistical rigor of the evaluation.

To address these concerns, we have conducted extensive additional experiments and revised the manuscript. We summarize the major updates below.

1. Extended prospective evaluation with independent clinical references (aaSp, FCZj, gcsV). We expanded the prospective evaluation from 2 to 13 subjects (10 healthy volunteers and 3 patients, including one with AF), each with a separately acquired breath-hold ECG-gated cine as a clinical reference. Without retraining, CardioStitch recovers 89% of the reference stroke volume (95% CI: 82–96%), compared with 17–23% for naive stacking and single-frame SVR baselines. It achieves an EF MAE of 9.6 percentage points, compared with 12.1 for Fetal CMR 4D, with no statistically significant difference (p = 0.84). Notably, Fetal CMR 4D receives 40 frames per slice and cardiac timing information and requires 22–49 minutes of optimization per subject, whereas CardioStitch requires only one frame per slice and reconstructs a 3D cine in seconds. These results provide additional quantitative evidence of the potential of synchronization-free reconstruction on real acquisitions. (Sec. 5.2, Table 4, Appendix A.9)

2. Additional analysis of whole-heart dynamics and reference conditioning (aaSp, FCZj, gcsV). To assess cardiac motion beyond the reference slice, we measured the stroke volume of the non-reference slices alone, obtaining 84% of the ground-truth value, comparable to 85% for the entire heart. We further conducted input controls by shuffling, replacing, and removing non-reference slices, demonstrating that the reconstruction depends on their observed anatomical content rather than the reference slice alone. Additional experiments evaluate the sensitivity to reference position, reference image quality, and the reference-state token. (Appendix A.3, Tables A3–A4)

3. Additional architectural ablations and validation of the training objectives (FCZj). We separately retrained CardioStitch without cross-slice attention and without the physical-location token under identical training settings. Removing cross-slice attention increases the EF error from 7.5 to 37.2 percentage points under simulated AF, confirming its importance for recovering cardiac dynamics. We also conducted a controlled optimization experiment to isolate the effect of coverage normalization on through-plane motion supervision and demonstrate how observation consistency improves motion estimation. (Sec. 5.3, Tables 5 and A1, Appendix A.4)

(Part 2/2)

4. Robustness to out-of-distribution respiratory motion (FCZj, gcsV). We evaluated the unchanged model under six respiratory motion settings outside the training distribution, including larger motion amplitudes, different respiratory waveforms, different AP/SI motion ratios, and slice-dependent amplitudes. Even at twice the training motion amplitude, CardioStitch maintains substantially lower respiratory EPE than naive stacking (2.96 vs. 8.83 mm). We acknowledge that rotational and non-rigid respiratory motion remain untested. (Appendix A.5, Sec. 6)

5. More comprehensive and transparent evaluation (FCZj, gcsV). We added standard deviations, bootstrap confidence intervals of paired differences, paired statistical tests, and subgroup analyses across datasets, vendors, held-out acquisition domains, pathologies, and slice spacings. We additionally report results without rigid registration or per-subject mean-error removal, clarify the resolution-matching protocol and baseline configurations, and discuss the advantages of dense-temporal baselines under relatively regular cardiac rhythms. (Sec. 5.1, Appendices A.10–A.11, Tables A7–A10)

6. Clarification of the AF simulation and current limitations (aaSp, FCZj, gcsV). Our AF simulation is designed to isolate the impact of irregular cardiac timing on synchronization-based reconstruction, rather than to reproduce all forms of pathological cardiac motion. We acknowledge that it does not capture regional deformation patterns beyond each subject's gated-cine trajectory. Moreover, the current single-frame formulation limits the temporal information available for recovering spatially heterogeneous cardiac dynamics. Extending CardioStitch to jointly exploit multiple temporal observations per slice while preserving its synchronization-free formulation is an important direction for future work.

Together, these additional experiments strengthen the empirical support for our central contribution: a population-trained, feed-forward framework that reconstructs a coherent 3D cardiac cine from asynchronous slices without explicit cardiac synchronization or patient-specific optimization.

The major revisions are highlighted in blue in the revised manuscript. Detailed responses to individual reviewers are provided separately.

We again thank all three reviewers for their valuable suggestions, which have helped us improve both the technical validation and clarity of the paper.

AuthorsOpenReview ↗

Follow-up on Author Responses

Dear Reviewers,

Thank you again for your constructive feedback. We have provided detailed responses to your concerns and updated the manuscript with additional experiments and analyses.

As the discussion period draws to a close, we would greatly appreciate your feedback on whether our responses have addressed your main concerns. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.