acceptodds
Under review as a conference paper at ICLR 2027

CAST: Conditional Acquisition-Structured Transport for Accelerated MRI

Abstract

Generative models are increasingly popular in accelerated MRI reconstruction. However, generative methods typically introduce acquisition operators and data consistency only at inference. We investigate how acquisition information can improve flow matching-based generative MRI reconstruction when incorporated during training. The result is Conditional Acquisition-Structured Transport (CAST), which uses acquisition operators to structure both the source distribution and the velocity field of the flow matching model. Sampling via CAST admits a corresponding k-space trajectory that preserves acquired measurements. We evaluate CAST on 2D, 3D, and cardiac cine MRI. CAST significantly improves reconstruction over network-input conditioning alone and enables the perception–distortion trade-off to be adjusted at inference. Averaging multiple samples matches or exceeds state-of-the-art unrolled networks in PSNR, closing the distortion gap that generative methods typically exhibit, while single-sample reconstructions achieve greater distributional similarity to reference images than generative baselines. On cardiac cine MRI, CAST can reconstruct fine details and reduce artifacts.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
accelerated mri, flow matching, inverse problems, image reconstruction, deep learning, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RinAOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3(+1)
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

The authors propose CAST, a conditional flow matching approach for MRI which builds the forward operator into the source distribution and the velocity fields instead of only feeding it to the network via its inputs. The authors evaluate on fastMRI brain, Stanford 3D knee, and OCMR cine against a number of diffusion inverse solver competitors, as well as trusted baselines (e.g., the E2E-VarNet or Wen's CNF). The authors also show that input-only conditioned approaches underperform even training-free inference methods which leverage a pre-trained prior.

Critical strengths

S1) The construction is principled and minimal.

S2) The approach has great inference efficiency for a flow matching-based method, which particularly matter for medical imaging tasks.

S3) The evaluation is broad, covering different anatomy across different datasets, as well as 2D vs. 3D MRI.

S4) The authors include an excellent (often absent) discussion of the perception-distortion trade-off.

Critical weaknesses

W1) The input-only baseline, which carries the paper's main claim, is under-specified. For instance, its step count is never stated. If it ran the same 5 Heun steps as CAST, the comparison may not be fair. Additionally, CAST's three main ingredients (source centered on the zero-filled image, noise confined to the range of BHB^H, and projected velocity) are not ablated properly.

W2) Closely related constructions are uncited: e.g., MC-DDPM [1] and null-space networks [2]. Additionally, rcGAN [3] is cited twice but is not compared against empirically although it has public code. This is a natural conditional baseline and should be added. Further, Bendel's mask-agnostic rcGAN [4] should be cited as well. There, the Gaussian code is projected onto the measured subspace and the output onto its complement. This seems highly related to CAST's approach of projecting the noise to the unmeasured subspace.

W3) Measurement preservation does not hold for the final output of the model (as acknowledged in Sec 4.3), however this claim leads the abstract. The claim should be toned down, or if the final output is indeed measurement consistent this should be clarified.

W4) For the cine experiments PDDR is 4.5dB below CineVN on breath-hold, yet it ties CAST on SER. This suggests a serious tuning mismatch.

W5) The perception-distortion framing is borrowed from [3], but presented as a CAST-specific feature. This should be further clarified. Also, why stop at 6 samples in the average? Additionally, under the PD tradeoff, LPIPS is a distortion metric. So, the LPIPS-SSIM panel of Figure 2 does not properly show the tradeoff.

[1] Y. Xie and Q. Li. Measurement-conditioned denoising diffusion probabilistic model for under-sampled medical image reconstruction. In Medical Image Computing and Computer Assisted Intervention (MICCAI), 2022. arXiv:2203.03623.

[2] J. Schwab, S. Antholzer, and M. Haltmeier. Deep null space learning for inverse problems: convergence analysis and rates. Inverse Problems, 35(2):025008, 2019.

[3] H. Chung, J. Kim, and J. C. Ye. Direct diffusion bridge using data consistency for inverse problems. In Advances in Neural Information Processing Systems (NeurIPS), 2023

[3] M. Bendel, R. Ahmad, and P. Schniter. A regularized conditional GAN for posterior sampling in image recovery problems. In Advances in Neural Information Processing Systems (NeurIPS), 2023.

[4] M. Bendel, R. Ahmad, and P. Schniter. Mask-agnostic posterior sampling MRI via conditional GANs with guided reconstruction. NeurIPS 2023 Workshop on Deep Learning and Inverse Problems.

Questions for authors

Q1) What NFE count and normalization did the input-only model use? What does it achieve at 50-100 steps? 1000 steps?

Q2) Can you ablate the three ingredients of CAST separately?

Q3) Why do single-sample numbers differ between Tables 2 and 3, and which sample count underlies each column of Tables 3-5?

Q4) Is per-pixel sample variance calibrated against error?

Q5) Why does PDDR underperform CineVN by 4.5dB on brath-hold?

Q6) Can you report PSNR versus PP (the number of samples averaged) up to 32 against the 2P/(P+1)2P/(P+1) for every sampler? The rcGAN paper establishes the PSNR scaling law of averaging posterior samples.

Q7) Why was rcGAN omitted from empirical evaluation?

5 comments

Part 1/3

Thank you for your positive score and appreciation of CAST's simplicity, efficiency, extensive evaluation, and for noting that our perception–distortion trade-off discussion is excellent.

We now address the reviewer's remaining concerns and questions.

W1 / Q1 / Q2: Fairness of the comparison between CAST and input-only conditioning, concerning the baseline's integration budget, and separate ablations of CAST's components.

We thank the reviewer for pointing this out. We originally used five Heun steps (ten network evaluations) for input-only conditioning, matching CAST, without tuning its integration budget. This was an oversight.

Hence, we added a solver budget analysis in Appendix B.6.1 and Figure 7. Increasing input-only conditioning from five to 20 steps raises PSNR from 34.40 to 35.78 dB. Increasing to 40 steps adds only another 0.06 dB; given this small additional gain, we did not evaluate the requested 50–100 or 1000 steps. Table 2 now reports the 40-step result (80 network evaluations). CAST achieves 36.27 dB with five Heun steps (ten network evaluations). We revised our conclusions to reflect the updated comparison: CAST improves all four metrics over input-only conditioning while requiring fewer integration steps.

We also added a component ablation and discussion of the results in Appendix B.6.1. Figure 7 shows that the zero-filled-centered isotropic baseline reaches similar PSNR to input-only conditioning with fewer integration steps. CAST retains this inference efficiency and achieves higher PSNR.

Moreover, at a common zero-filled center, neither CAST's source nor CAST's velocity structure alone improves PSNR over the isotropic-source baseline; combining them gives the best result. For the variant with CAST's velocity and an isotropic source, we kept the network fixed and replaced only the source with CAST's structured source. PSNR increased from 34.76 to 36.25 dB, close to CAST's 36.27 dB, without retraining. This is consistent with reducing initial noise along eigenvectors where BHBB^H B attenuates the velocity output.

Regarding normalization (Q1): Both models use the same intensity normalization. We divide the measurements by the 99th percentile of the zero-filled volume magnitude.

W2 / Q7: Missing discussion of MC-DDPM, null-space networks and mask-agnostic rcGAN, and why rcGAN was omitted from the experiments.

The related work section now discusses MC-DDPM, null-space networks, rcGAN and its mask-agnostic version. MC-DDPM (Xie and Li, 2022) uses diffusion to generate missing k-space in its single-coil MRI setting; CAST uses flow matching with multi-coil acquisition operators in coil-combined image space. Null-space networks (Schwab et al., 2019) constrain their image corrections to the forward operator's null space; CAST's image correction need not lie in that null space. Mask-agnostic rcGAN (Bendel et al., 2023b) projects its noise inputs onto the measured subspace and enforces output data consistency; CAST instead structures the source and velocity of a flow-matching process.

Our initial generative baselines focused on flow and diffusion models. We have now added rcGAN (Bendel et al., 2023a) on our homogeneous fastMRI setup (Table 3) and a new low-field MRI experiment using the M4Raw dataset (Table 6). On fastMRI (Table 3), rcGAN ranks second in PSNR, achieving 38.44 dB versus 38.87 dB for CAST. On M4Raw (Table 6), it ranks third among reconstruction methods, achieving 31.82 dB PSNR versus 32.15 dB for CAST.

W3: Concern that the abstract overstates measurement preservation when the returned image need not be data consistent.

We agree that the scope of measurement preservation should be clearer. Section 4.3 guarantees MYt=yobsMY_t=y_{\mathrm{obs}} for the associated k-space trajectory. Data consistency of the returned image requires its forward-encoded measurements to match the observations, MFSXt=yobsMFSX_t=y_{\mathrm{obs}}, which CAST does not guarantee. We revised the abstract to describe sampling as completing missing k-space data; Section 4.3 states the limitation on data consistency of the returned image.

Part 2/3

W4 / Q5: Why PDDR has much lower breath-hold PSNR than CineVN yet similar free-breathing SER to CAST, and whether this indicates a tuning mismatch.

We tuned PDDR (Fürnrohr and Heckel, 2026) separately on breath-hold and free-breathing validation data over regularization strength and stopping iteration (Table 7; Appendix B.4.1). We discussed the results with the original PDDR authors, who considered them reasonable. PDDR tends to produce smoother reconstructions, whereas CineVN (Vornehm et al., 2025) is trained end-to-end and recovers more fine detail.

Breath-hold PSNR compares images with fully sampled references, whereas free-breathing SER measures prediction of randomly held-out acquired k-space data. The GRO masks (Joshi et al., 2022) used by the OCMR free-breathing data sample central phase-encoding lines more densely; hence, SER is biased towards that sampling density. Smoother reconstructions can therefore achieve high SER despite losing peripheral k-space detail. SER also has no explicit temporal-coherence term.

Assessing distortion without high-quality references remains an open problem; we therefore provide SER alongside qualitative comparisons in Figure 4 and the supplementary videos.

The supplementary videos show that CAST substantially reduces unnatural flickering relative to PDDR and CineVN.

W5: Proper attribution of the perception–distortion framing, the correction that LPIPS is a distortion metric, and why averaging stopped at six samples.

We added a mention of Chung et al. (2023) in Section 5.2. They examine the trade-off by varying the number of sampling steps; our experiments vary the number of samples averaged. The section also attributes the averaging-based trade-off to Bendel et al. (2023b).

We agree that LPIPS is considered a distortion metric under the definition of Blau and Michaeli (2018). We now describe LPIPS–SSIM (Section 5.2 and Figure 2 caption) as showing similar averaging behavior for reference-based deep-feature metrics, while KID–PSNR shows the classical perception–distortion trade-off.

Alongside independent sampling, we use antithetic averaging, following Jia et al. (2026): two reconstruction runs for the same acquisition use opposite Gaussian noise throughout reconstruction, and we average their outputs. Separate pairs use independent noise streams. The motivation is to reduce sampling variation when averaging a fixed number of reconstructions. The new Figure 6 shows that antithetic pairing improves PSNR at matched reconstruction counts. We now describe this procedure in Section 5 and give its motivation and details in Appendix B.1; our initial submission did not distinguish it from independent sampling.

The six-reconstruction averages used three antithetic pairs rather than six independent samples. We capped the paired candidates at three pairs (except for the new rcGAN baseline) because validation showed diminishing PSNR gains. This benefits inference cost, especially for the compute-intensive generative prior baselines.

Q3: Why single-sample results differ between Tables 2 and 3, and which sample count was used for each metric in Tables 3–5.

The original Table 2 averaged scores across 10 individual draws, whereas the single-sample entries in Table 3 used one fixed draw. Variation across draws is marginal, so we now use the same fixed realizations in both tables to keep the reported results consistent.

Section 5.2 now describes validation-based selection of the sampling procedure. Selected independent and antithetic reconstruction counts are shown beside the scores in Tables 3–6 and 8; the appendix gives the selection grids and per-dataset details.

The tables identify the selected procedures as, for example, (2-i) for two independent reconstructions and (6-a) for three antithetic pairs. An average requires all the indicated reconstruction runs. Per-reconstruction solver budgets and further details are given in Section 5.1 and Appendices B.2.1, B.3.1, B.4.1 and B.5.1.

Q4: Is per-pixel sample variance calibrated against error?

We interpret this as asking whether variability across CAST samples provides a reliable estimate of reconstruction uncertainty. We have not established such calibration. If the reviewer has a concrete suggestion, we would be happy to carry out that experiment or analysis.

Part 3/3

Q6: Can the authors report PSNR curves through 32 samples against posterior-sampling theory (Bendel et al., 2023b)?

We added curves through 32 reconstructions in Figure 6 for CAST, rcGAN, Wen's CNF (Wen et al., 2023), FlowDPS (Kim et al., 2025) and FlowPnP (Martin et al., 2025).

For independent exact posterior samples, E1/EP=2P/(P+1)\mathcal E_1/\mathcal E_P=2P/(P+1) describes expected squared error (Bendel et al., 2023b). CAST and rcGAN follow its shape more closely than FlowDPS and FlowPnP.

We thank the reviewer for their positive assessment and hope our clarifications and new experiments strengthen the reviewer's support for acceptance.

References

Bendel et al. Mask-agnostic posterior sampling MRI via conditional GANs with guided reconstruction. In NeurIPS Workshop on Deep Learning and Inverse Problems, 2023a.

Bendel et al. A regularized conditional GAN for posterior sampling in image recovery problems. In Advances in Neural Information Processing Systems, 2023b.

Blau et al. The perception-distortion tradeoff. In IEEE Conference on Computer Vision and Pattern Recognition, 2018.

Chung et al. Direct diffusion bridge using data consistency for inverse problems. In Advances in Neural Information Processing Systems, 2023.

Fürnrohr et al. Piecewise dynamic diffusion regularization for reconstruction of cardiac cine MRI. In Medical Image Computing and Computer Assisted Intervention, 2026.

Jia et al. Antithetic noise in diffusion models. In International Conference on Learning Representations, 2026.

Joshi et al. Technical report (v1.0)–pseudo-random Cartesian sampling for dynamic MRI. arXiv:2206.03630, 2022.

Kim et al. FlowDPS: Flow-driven posterior sampling for inverse problems. In IEEE/CVF International Conference on Computer Vision, 2025.

Martin et al. PnP-Flow: Plug-and-play image restoration with flow matching. In International Conference on Learning Representations, 2025.

Schwab et al. Deep null space learning for inverse problems: convergence analysis and rates. Inverse Problems, 2019.

Vornehm et al. CineVN: Variational network reconstruction for rapid functional cardiac cine MRI. Magnetic Resonance in Medicine, 2025.

Wen et al. A conditional normalizing flow for accelerated multi-coil MR imaging. In International Conference on Machine Learning, 2023.

Xie et al. Measurement-conditioned denoising diffusion probabilistic model for under-sampled medical image reconstruction. In Medical Image Computing and Computer Assisted Intervention, 2022.

Reviewer_t6twOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, the authors focus on multi-coil MRI reconstruction, where they study how acquisition information can be used during training with flow matching. They show that providing this information only through network inputs does not improve over reconstruction with pretrained image priors, whereas structuring the source distribution and velocity field with acquisition operators substantially improves the conditional model (CAST).

Critical strengths
  • The authors point out an important fact on how acquisition information should enter conditional flow matching during training. The input-only vs CAST comparison (Table 2) uses the same backbone and inputs, so the effect of structuring the source and velocity is isolated reasonably well.
  • The authors evaluate on both 2D and 3D
  • More efficient : CAST uses few integration steps, has one tuned inference hyperparameter versus ~3 for prior-based methods, and runs much faster than FlowDPS, FlowPnP, TPDM-DDS, and PDDR.
Critical weaknesses
  • All tables don't report standard deviation, hence difficult to understand if improvement is stat-sig or not. The mean values are too close to each other (see Tables 2, 3, 4). Authors need to do t-test
  • CAST changes two things at once: the source (zero-filled image plus coloured noise) and the velocity parameterization (B^H B h_theta). The authors need to do ablation to decouple/isolate contributions.
  • In the multicoil setting, the “acquisition-structured” velocity may impose no subspace restriction, so the theory is questionable. B maps C^n to C·(n-m) coil k-space entries. For the setups used (4 to 38 coils, R = 4 to 13), B has many more rows than columns. B^H B = I - S^H F^H M F S (for normalized maps) is then generically full rank, though ill-conditioned. If so, range(B^H) is all of C^n. Additionally, Appx A’s argument that the marginal velocity lies in range(B^H) already holds trivially and restricts nothing. Also, “restricting learned updates to unmeasured entries” holds only in the auxiliary coil k-space vector Y_t, not in image space. In image space, B^H B acts as a spatially varying preconditioner and noise-colouring operator, not a projection. The restriction is thus a true null-space projection only in the single-coil case (S = I). The authors should report the rank or spectrum of B^H B on their data. They should also rewrite the method’s motivation if the operator is full rank. The paper’s main claims depends on this.
  • Conceptually, compare the work against MC-DDPM [Xie, Yutong, and Quanzheng Li. "Measurement-conditioned denoising diffusion probabilistic model for under-sampled medical image reconstruction." International conference on medical image computing and computer-assisted intervention. Cham: Springer Nature Switzerland, 2022.] It also performs diffusion in the unsampled k-space region, conditioned on measurements.
  • Most generative baselines (FlowDPS, FlowPnP, Variational, TPDM-DDS, PDDR) are unsupervised, mask-agnostic priors. CAST is supervised on paired data
Questions for authors

please see weakness

4 comments

Part 1/2

Thank you for recognizing the importance of the mechanism behind CAST, CAST's computational efficiency and our broad evaluation setup.

We now address the concerns of the reviewer.

Regarding statistical significance and small mean differences.

We assess uncertainty using paired 95% bootstrap confidence intervals for CAST-minus-baseline differences, rather than t-tests, with subjects or volumes as resampling units. Appendices B.2.4 and B.4.2 report these intervals and identify comparisons whose intervals include zero (Tables 10 and 12). With only four Stanford 3D test volumes, we report paired difference ranges (Appendix B.3.2, Table 11) and do not claim statistical significance for this dataset. We have highlighted the statistical analysis in the revised main body.

We also added paired 95% confidence intervals for the updated Table 2 comparison (Appendix B.2.4, Table 9). CAST improves PSNR over input-only conditioning by 0.44 dB, with an interval of [0.37, 0.50] dB. All four metric intervals exclude zero in favor of CAST.

Several improvements are substantial: for example, in Table 3, CAST improves PSNR over generative priors by over one dB and substantially reduces KID.

The qualitative examples further support these results, showing improvements in fine image structure (Figures 1 and 3) and reduced streaking artifacts in free-breathing cine (see videos in supplementary material and Figure 4).

Regarding separating the contributions of the source distribution and velocity parameterization.

We thank the reviewer for the suggestion. We added a component ablation and discussion in Appendix B.6.1 (Table 14), comparing source and velocity choices. Neither component alone improves PSNR over an isotropic Gaussian source centered on the zero-filled image; combining the two components gives the best result.

For the variant with CAST's velocity and an isotropic source, we kept the network fixed and replaced only the source with CAST's structured source. PSNR increased from 34.76 to 36.25 dB, close to CAST's 36.27 dB, without retraining. This is consistent with reducing initial noise along eigenvectors where BHBB^H B attenuates the velocity output.

Regarding whether BHBB^H B imposes an image-space subspace restriction, its spectrum, and the implications for the theory.

The reviewer is correct that BHBB^H B can be full rank in multi-coil MRI and need not impose an image-space subspace restriction. Here, B=(I−M)FSB=(I-M)FS maps an image to its unmeasured multi-coil k-space coefficients.

Along an eigenvector of BHBB^H B, CAST scales both the injected noise and the network's velocity output by the corresponding eigenvalue. Thus, it injects less noise into components where the velocity parameterization also attenuates updates. The construction therefore matches the source noise to the velocity weighting, even when the operator is full rank. Section 4.2 now makes this interpretation explicit, and the component ablations in Appendix B.6.1 provide empirical support.

Appendix A establishes representability of the marginal flow-matching velocity for all ranks, using range⁡(BHB)=range⁡(BH)\operatorname{range}(B^H B)=\operatorname{range}(B^H). In the full-rank case, this statement does not restrict the velocity class. The result establishes compatibility with flow matching, not improved conditioning or learning efficiency. The benefits of the parameterization are evaluated empirically in Table 2 and Appendix B.6.1.

As requested, we now report the spectra for all our test sets (Appendix B.6.1, Figure 8). These show acquisition-dependent weighting, with small-eigenvalue tails in fastMRI and M4Raw and spectra concentrated closer to one in cine MRI

Regarding updates being restricted to unmeasured entries only in the auxiliary k-space trajectory.

We agree that this guarantee concerns the auxiliary k-space trajectory. Since MB=0MB=0, it satisfies MYt=yobsMY_t=y_{\mathrm{obs}}, independently of the rank of BHBB^H B. The returned image need not satisfy MFSXt=yobsMFSX_t=y_{\mathrm{obs}}, as stated in Section 4.3. The revised introduction and abstract describe the construction as k-space completion.

Regarding the conceptual difference from MC-DDPM.

MC-DDPM (Xie and Li, 2022) uses diffusion to generate unmeasured coefficients in its single-coil MRI setting and maps completed measurements to images through the inverse Fourier transform. Multi-coil encoding is not a bijection onto arbitrary coil arrays, so this construction does not transfer unchanged. CAST uses flow matching and defines its source and velocity in coil-combined image space using sensitivity encoding and adjoints, with network dimensions independent of coil count. We added a discussion of MC-DDPM in our related work section.

Part 2/2

Regarding CAST using paired supervision while pretrained priors are unsupervised.

The reviewer is correct that the training setups differ. Both CAST and the evaluated image priors use training images obtained from fully sampled acquisitions. CAST additionally uses paired undersampled measurements and acquisition operators during training; image priors incorporate this acquisition information only during reconstruction.

We also compare with baselines trained with supervision: Table 3 and 6 provide direct comparisons with supervised conditional generators, Wen's CNF (Wen et al., 2023) and the newly added rcGAN (Bendel et al., 2023), as well as the supervised unrolled network E2E-VarNet (Sriram et al., 2020). We also compare with E2E-VarNet for 3D and low-field MRI (Tables 4 and 6), and with CineVN (Vornehm et al., 2025) for cardiac cine (Table 5). CAST achieves comparable or higher PSNR and lower KID than these supervised baselines.

We hope the new ablations, spectral analysis and clarifications address the reviewer's concerns. We would be grateful if the reviewer reconsidered their assessment.

References

Bendel et al. A regularized conditional GAN for posterior sampling in image recovery problems. In Advances in Neural Information Processing Systems, 2023.

Sriram et al. End-to-end variational networks for accelerated MRI reconstruction. In Medical Image Computing and Computer Assisted Intervention, 2020.

Vornehm et al. CineVN: Variational network reconstruction for rapid functional cardiac cine MRI. Magnetic Resonance in Medicine, 2025.

Wen et al. A conditional normalizing flow for accelerated multi-coil MR imaging. In International Conference on Machine Learning, 2023.

Xie et al. Measurement-conditioned denoising diffusion probabilistic model for under-sampled medical image reconstruction. In Medical Image Computing and Computer Assisted Intervention, 2022.

I thank the authors for the additional experiments and addressing my comments. I have revised my assessment and updated the score,

Thank you for raising your score; we're glad our response was helpful!

Reviewer_M5DUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

For accelerated MRI reconstruction, existing methods based on pretrained generative model normally introduce acquisition operator at inference. This paper proposes to also leverage the acquisition information during training. It compares two mechanisms to integrate acquisition operator into conditional flow matching models: using them only as network input v.s. constructing the source distribution and velocity field based on acquisition operator. Experiments on three datasets (2D fastMRI dataset, 3D Stanford Knee dataset, cardiac cine MRI dataset) have demonstrated superiority of the proposed method over simple input conditioning and pretrained generative model prior method, as well as unrolled networks.

Critical strengths

Acquistion-structured formulation: Rather than using the acquisition operator only during inference or just using them as additional input to the network, it constructs the source distribution and velocity field based on the acquisition operator.

Broad experimental validation. Evaluation across multiple MRI acquisition settings suggests that the method is generalisable to a wide range of anatomical sites and MRI sequences.

Critical weaknesses

The methodological novelty is limited. The proposed CAST formulation can be regarded as residual conditional flow matching, where the target is to predict the residual x∗−xzfx^*-x^{\mathrm{zf}} from noise constrained in the non-measured space. Similar idea has been proposed by papes like HFS-SDE (High-Frequency Space Diffusion Model for Accelerated MRI, IEEE TMI 2024), which constructs the source from the low-frequency k-space plus high-frequency Gaussian noise, and restricts both the drift and diffusion terms of the SDE to the high-frequency subspace. Although CAST decomposes the whole k-space into measured and unmeasured subspace rather than splitting it into low- and high-frequency region, the general idea is the same, i.e. using an MRI operator to structure both the source distribution and the generative dynamics.

Another weakness is related to the scope of the work. It is relatively narrow in its current form. The method is proposed and evaluated for accelerated MRI reconstruction. In addition, the method section is closely tied to the MRI acquisition operator and k-space structure. Although the authors conducted experiments with several MRI setting, it is not clear that if the method could also work for inverse problems beyond MRI.

Questions for authors
  • In the results section, different datasets used different methods as baselines. Could the authors explain a bit on this? Are the competing methods evaluated consistently across the three datasets (2D fastMRI, 3D knee, and cardiac cine)? Otherwise, it is not straightforward to access the advantage of CAST across settings.

  • That would be good if the authors could provide an ablation study, using the same pipeline proposed in the paper (proposed structured source distribution and velocity field), but without the additional acquisition-derived network inputs and the back projected measurement residual? This ablation study could help find out whether the improvement comes from the proposed flow construction or the engineered acquisition conditioning.

4 comments

Part 1/2

Thank you for the review and for highlighting our broad evaluation setup spanning 2D, 3D and dynamic MRI.

We now address the reviewer's concerns and questions.

Regarding novelty and the relationship to HFS-SDE (Cao et al., 2024) and residual flow matching.

We agree that CAST shares with HFS-SDE the idea of structuring generative reconstruction with MRI operators, and that CAST admits a residual-flow interpretation. Our contribution is the specific acquisition-conditioned source–velocity construction allowing a simple and effective reconstruction process, supported by a new direct HFS-SDE comparison, controlled ablations and evaluation across 2D, 3D and dynamic MRI.

HFS-SDE trains a generative diffusion prior with a pre-defined low-/high-frequency split to define its diffusion region and uses the actual undersampled measurements only during inference, similar to other generative prior based reconstruction methods. CAST is a supervised learning approach based on flow matching for learning the conditional reconstruction distribution directly. It uses the actual acquisition mask and measurements already during training to jointly structure its source and velocity. During inference, HFS-SDE uses measurement guidance rather than holding acquired high-frequency coefficients fixed, whereas CAST integrates its learned conditional velocity without separate guidance and preserves all acquired coil measurements, including high-frequency coefficients, in its associated k-space completion.

HFS-SDE's low-/high-frequency split construction assumes a fully sampled k-space center during inference (Cao et al., 2024, Algorithm 1), which may not be satisfied for real-time cine MRI reconstruction. CAST requires no fully sampled center and supports such settings as we show in our free-breathing cine experiments (Section 5.4).

We added HFS-SDE as a direct baseline. On homogeneous fastMRI, Table 3 reports 38.87 dB PSNR for CAST versus 36.71 dB for HFS-SDE. CAST also improves SSIM, LPIPS and KID. CAST requires 10 network evaluations versus 800 for HFS-SDE for one sample.

Regarding the method’s close connection to the MRI acquisition operator and its evaluation only on MRI reconstruction.

CAST may extend to other linear inverse problems with a comparable decomposition into acquired and unacquired measurements, but our experiments do not establish generality beyond MRI.

We focus on MRI to test the same source and velocity construction across distinct reconstruction challenges: 3D MRI is computationally demanding and requires consistency across slices; free-breathing cine requires resolving cardiac and respiratory motion from limited measurements per frame; and low-field MRI has low signal-to-noise ratios.

Regarding different baselines, consistent evaluation across datasets and CAST’s advantages.

We select baselines suited to each reconstruction setting. For 2D MRI, we compare conditional generators—Wen's CNF (Wen et al., 2023) and rcGAN (Bendel et al., 2023)—and generative prior based methods, FlowDPS (Kim et al., 2025), FlowPnP (Martin et al., 2025) and HFS-SDE (Cao et al., 2024). For 3D MRI, we use TPDM-DDS (Lee et al., 2023; Chung et al., 2024) and Krainovic et al.'s (2024) variational reconstruction, which use 2D diffusion priors for 3D reconstruction. E2E-VarNet (Sriram et al., 2020) provides a state-of-the-art unrolled reference in both settings. For cardiac cine, we compare PDDR (Fürnrohr and Heckel, 2026) and CineVN (Vornehm et al., 2025), which model temporal information. Across these settings, CAST retains the same source and velocity construction while using backbones suited to 2D, 3D or dynamic MRI reconstruction, demonstrating CAST’s versatility and modularity.

The same full baseline set is not evaluated across all datasets; evaluation is consistent within each dataset. Methods share the acquisition, test set and evaluation protocol. We perform extensive hyperparameter tuning for our baselines on each setup separately. Appendices B.2.1, B.3.1 and B.4.1 document validation selection and inference settings.

Part 2/2

Regarding an ablation that removes the acquisition-derived network inputs and whether the improvement comes from CAST’s flow construction or the acquisition-derived conditioning inputs.

We performed the requested ablation and observed a decrease in PSNR from 36.27 to 33.92 dB.

This is expected because CAST initializes from the zero-filled image perturbed by noise. Additionally providing acquisition information as input gives the network direct access to the measurement information. We added this result in Appendix B.6.1.

The results in Table 2 address this question more directly: CAST and input-only conditioning use the same backbone and receive the same acquisition-derived inputs. CAST’s improvements therefore support the benefit of its joint source and velocity construction, beyond input conditioning alone.

We hope our response has addressed the reviewer's concerns. Given the direct HFS-SDE comparison, new experiments and clarifications, we would be glad if the reviewer considered raising their score.

References

Bendel et al. A regularized conditional GAN for posterior sampling in image recovery problems. In Advances in Neural Information Processing Systems, 2023.

Cao et al. High-frequency space diffusion model for accelerated MRI. IEEE Transactions on Medical Imaging, 2024.

Chung et al. Decomposed diffusion sampler for accelerating large-scale inverse problems. In International Conference on Learning Representations, 2024.

Fürnrohr et al. Piecewise dynamic diffusion regularization for reconstruction of cardiac cine MRI. In Medical Image Computing and Computer Assisted Intervention, 2026.

Kim et al. FlowDPS: Flow-driven posterior sampling for inverse problems. In IEEE/CVF International Conference on Computer Vision, 2025.

Krainovic et al. Resolution-robust 3D MRI reconstruction with 2D diffusion priors: Diverse-resolution training outperforms interpolation. arXiv:2412.18584, 2024.

Lee et al. Improving 3D imaging with pre-trained perpendicular 2D diffusion models. In IEEE/CVF International Conference on Computer Vision, 2023.

Martin et al. PnP-Flow: Plug-and-play image restoration with flow matching. In International Conference on Learning Representations, 2025.

Sriram et al. End-to-end variational networks for accelerated MRI reconstruction. In Medical Image Computing and Computer Assisted Intervention, 2020.

Vornehm et al. CineVN: Variational network reconstruction for rapid functional cardiac cine MRI. Magnetic Resonance in Medicine, 2025.

Wen et al. A conditional normalizing flow for accelerated multi-coil MR imaging. In International Conference on Machine Learning, 2023.

The concern about the novelty remains unaddressed. HFS-SDE of course differ from CAST in terms of whether the model is supervised or unsupervised. However, my intention in mentioning it was to show one precedent example of incorporating projection operator into the training of diffusion/flow matching model, i.e. starting from one fixed component within a subspace and restricting the sampling dynamics to the complementary subspace. Actually, Wen et al. (ICML 2023) provides a more direct comparison, although it was already included as one of the baselines in the paper. It also trained a generative model to predict component in the null space of the acquisition operator and composite it with the measured component. Their difference lies more in the choice of generative models, i.e. normalizing flow v.s. flow matching, rather than in the underlying reconstruction principle. The paper did not discuss how CAST is methodologically novel over Wen’ et al.

Is the CAST backbone for the 3D MRI experiment still based on 2D or 2.5D slice-wise input? If so, it is unclear why the other 2D baselines cannot be adapted to the 3D evaluation as well. More importantly, the baseline categories are not covered consistently across experiments: the 2D setting includes conditional generative models, pretrained generative-prior methods, and an unrolled network, whereas the 3D and cine settings include only subsets of these categories.

Thank you for the follow-up. We further clarify how CAST differs from prior methods and report new 3D comparisons addressing the reviewer’s concern about baseline coverage. We will incorporate these clarifications and results into the final manuscript.

Regarding CAST's novelty relative to HFS-SDE and Wen's CNF.

We agree that using MRI operators to structure training of generative models has precedents.

Here is how CAST differs from HFS-SDE and what the practical significance for cine MRI is. HFS-SDE defines its diffusion region outside a chosen central k-space band, assuming that this band is fully acquired (Cao et al., 2024, Equation 14 and Algorithm 1). Real-time cardiac cine requires short acquisition windows to temporally resolve cardiac and respiratory motion, leaving only limited measurements per frame. The sampling patterns vary across frames and do not fully sample the k-space center. CAST instead uses each frame’s actual acquired and missing locations, allowing its k-space completion to include missing center as well as outer coefficients. This construction therefore accommodates our real-time free-breathing cine experiments (Section 5.4).

Our related work discusses Wen's CNF. We provide a more detailed comparison below and will incorporate it into the final manuscript.

Wen’s CNF uses two separate networks to reconstruct the individual coil images. A conditioning U-Net extracts features from zero-filled coil images. A separate invertible flow network uses these features to transform isotropic Gaussian latent noise into missing coil-image components. The model requires a fixed number of coil channels, and the conditioning U-Net is pretrained before both networks are trained jointly.

CAST uses a single network to predict updates to one coil-combined image, with sensitivity maps handling coil encoding and combination outside the network. Its network dimensions therefore remain independent of coil count.

Unlike Wen’s CNF, CAST shapes its input noise using the actual sampling pattern and coil sensitivities: Gaussian image noise is encoded into coil k-space via sensitivity maps, restricted to unmeasured frequencies and mapped back before being added to the zero-filled reconstruction. The same operation transforms the network’s output into the velocity used for numerical integration during reconstruction. This matched construction goes beyond changing the generative model family.

CAST’s backbone can be freely chosen for the reconstruction setting: it need not be invertible, and we train it directly through flow matching without a separate conditioning network or pretraining stage. We can thus use different architectures without changing CAST’s source and velocity construction. We use an image U-Net for 2D MRI, extend it with neighboring-slice inputs for 3D MRI, and use a temporal U-Net for cine MRI (Sections 5.3–5.5).

Regarding the 3D backbone and baseline coverage across experiments.

CAST uses a 2D U-net that receives the center slice and its two neighbors along the readout axis, as does E2E-VarNet (Appendix B.3.1).

We initially prioritized established reconstruction methods for 3D and cine over implementing adaptations of existing 2D generators. Following the reviewer’s suggestion, we have now evaluated both Wen’s CNF and rcGAN with the same neighboring-slice information as CAST and E2E-VarNet, using the same Stanford 3D evaluation setup as Table 4.

ModelPSNR ↑SSIM ↑LPIPS ↓KID ×1000 ↓
Wen's CNF38.350.9220.18723.40
rcGAN39.550.9340.1589.44
CAST39.890.9380.1463.29

CAST remains best across all four metrics.

We hope these clarifications and additional results address the reviewer’s remaining concerns.

Comments on the paper

1 comment

We thank the reviewers for their feedback and for appreciating CAST’s acquisition-structured construction and broad evaluation (all reviewers), its simplicity (RinA) and inference efficiency (t6tw, RinA). We summarize the three main concerns raised by the reviewers and how we addressed them:

  1. Is CAST sufficiently distinct from existing methods? The reviewers asked how CAST differs from HFS-SDE, MC-DDPM and Wen’s CNF. We clarified how the actual acquisition mask and sensitivity maps shape both CAST’s starting noise and image updates learned through flow matching. Unlike Wen’s CNF, CAST directly reconstructs one coil-combined image with a backbone that need not be invertible and whose dimensions remain independent of coil count. Unlike HFS-SDE, it requires no fully sampled center, enabling our real-time cine experiments (Section 5.4). MC-DDPM instead uses diffusion for single-coil reconstruction. The added HFS-SDE comparison favors CAST across all four metrics (Tables 2–3).

  2. Does the mathematical interpretation justify the construction? We agree with t6tw that full-rank BHBB^H B imposes no image-space subspace restriction. We revised Section 4.2 to explain how the operator matches source noise to velocity weighting, and added the requested spectra and component ablations to assess this interpretation empirically (Appendix B.6.1). Appendix A establishes representability, not improved learning efficiency.

  3. Do controlled comparisons establish CAST’s benefit? RinA questioned the input-only baseline’s solver budget. We increased its budget to 40 Heun steps, where performance had plateaued, and revised our conclusions accordingly. CAST retains better performance across all metrics using only five Heun steps (Table 2). To address t6tw’s concern about differing training data requirements, we added rcGAN as an additional supervised learning baseline (Tables 2, 3 and 6). In response to M5DU’s concern about baseline coverage, we also evaluated Wen’s CNF and rcGAN in 3D MRI with the same neighboring-slice information as CAST and E2E-VarNet. CAST remains best among these added baselines. We will incorporate the additional 3D results into the final manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.