- 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 , 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 (the number of samples averaged) up to 32 against the 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 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, 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.
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.
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 BHB 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.
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.
We agree that the scope of measurement preservation should be clearer. Section 4.3 guarantees MYt=yobs for the associated k-space trajectory. Data consistency of the returned image requires its forward-encoded measurements to match the observations, MFSXt=yobs, 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.