acceptodds
Under review as a conference paper at ICLR 2027

Learning Shared Bidirectional Latent Dynamics for Translation Between CT and MRI

Abstract

Bidirectional image translation is an important problem in computer vision because many related visual domains require mappings in both directions. A representative case is translation between CT and MRI, which provide complementary anatomical and tissue information but may not always be simultaneously available. Existing methods commonly train separate models for translation from CT to MRI and from MRI to CT, treating the two directions as independent tasks and leaving their shared structure insufficiently explored. To address this limitation, we formulate the two translation directions as latent state evolution processes governed by shared dynamics. Modality specific VAEs encode CT and MRI into latent representations, and a shared velocity model transports the source representation toward the target. Since the same dynamics model must represent two different translation directions, we introduce latent anchors that combine the source state with direction information and guide velocity prediction. The anchor is used as a reference during latent evolution rather than as the predicted target itself. An image space refiner further corrects residual errors after decoding. Experiments on two public paired CT and MRI datasets demonstrate consistent performance across both translation directions and MRI contrasts, with the proposed framework achieving the highest mean PSNR and SSIM among the compared methods. Ablation studies further show that direct parameter sharing introduces a performance gap, while explicit direction modeling, particularly the proposed anchor formulation, largely recovers this loss. Source code is provided in the supplementary material.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Bidirectional Image Translation, Medical Image Translation, Shared Latent Dynamics, Flow Matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fq6dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified framework for bidirectional CT-to-MRI and MRI-to-CT translation using shared latent dynamics. The framework employs modality-specific VAEs and a shared velocity field to model both translation directions within a single parameterization. To alleviate the performance degradation associated with parameter sharing, the authors introduce Anchor BiFlow, which constructs a source- and direction-conditioned latent anchor to guide velocity prediction. An image-space refiner is further employed to correct residual reconstruction errors. Experiments on SynthRAD2023 and PCM2025 demonstrate competitive quantitative performance and favorable computational efficiency, with only 4.15M parameters and approximately 6.26 ms inference latency.

Critical strengths
  1. Modeling both CT–MRI translation directions through shared latent dynamics is well motivated and practically relevant. Evaluation on two datasets, both translation directions, and multiple MRI contrasts supports the potential utility of the framework.

  2. The complete model achieves competitive translation quality with 4.15M parameters and a reported average inference latency of 6.26 ms (Appendix C.2). This represents a favorable balance between translation quality and computational cost relative to the iterative diffusion and flow baselines at the evaluated inference settings.

Critical weaknesses
  1. Equation (14) directly supervises the anchor toward the target latent. With α=1, the integration endpoint is this anchor plus an accumulated residual. Without direct anchor decoding or a capacity-matched single-pass residual baseline, the incremental benefit of iterative dynamics remains unclear.

  2. Table 2 supports recovery of the sharing gap before refinement, but the final advantage over Uni Flow remains unresolved. The MRI→CT PSNR gain is 0.09 dB before refinement, while adding refinement yields a further 0.43 dB. A Uni Flow baseline with the same refiner is needed to assess the advantage under comparable refinement.

  3. Tables 2–3 do not define the reported ± values or specify the repetition protocol. Clarifying the independent runs, random seeds, and uncertainty would help assess the stability of the small pre-refinement gains.

Questions for authors
  1. How does Anchor BiFlow compare with direct anchor decoding and a capacity-matched single-pass residual predictor under comparable supervision? These comparisons would clarify the incremental benefit of iterative dynamics.

  2. Does the advantage over Uni Flow persist when both use the same VAE representations and image refiners under comparable training settings? Please report parameter counts, training budgets, and supervision settings.

  3. How many independent training runs were conducted for Tables 2–3, and what do the reported ± values represent? Are the small pre-refinement gains consistent across training seeds?

3 comments

We thank the reviewer for their careful review. Our point-by-point responses are provided below:

Reply to Q1: For SynthRAD2023, we introduced two ablation variants to compare against our full model, Anchor BiFlow (5-step). The Anchor-only variant decodes directly from the predicted anchors. Anchor BiFlow (1-step) retains the same anchor mapper and velocity network, performing a single application of the predicted residual onto the anchors before decoding. Both the one-step and five-step models are trained using a joint bidirectional approach. Including the two VAEs, the shared anchor mapper and velocity network, and the refinement modules for both directions, both models have a total parameter count of 4.15M. During the generation of endpoints, the one-step and five-step models invoke the shared velocity network once and five times, respectively. As shown in the table below, both before and after refinement, the complete model achieves the highest average SSIM in both directions.

MethodTrainingRefinerCT→MRI PSNR ↑CT→MRI SSIM ↑MRI→CT PSNR ↑MRI→CT SSIM ↑
Anchor-onlyJoint—19.28 ± 0.040.7447 ± 0.00319.21 ± 0.080.7771 ± 0.001
Anchor BiFlow (1-step)Joint—19.32 ± 0.070.7495 ± 0.00219.22 ± 0.050.7771 ± 0.002
Anchor BiFlow (5-step)Joint—19.44 ± 0.020.7634 ± 0.00219.26 ± 0.020.7831 ± 0.002
Anchor-onlyJoint✓19.44 ± 0.040.7651 ± 0.00219.26 ± 0.090.7883 ± 0.001
Anchor BiFlow (1-step)Joint✓19.47 ± 0.140.7695 ± 0.00219.31 ± 0.050.7897 ± 0.001
Anchor BiFlow (5-step)Joint✓19.47 ± 0.050.7794 ± 0.00419.69 ± 0.080.8068 ± 0.006

Reply to Q2: On SynthRAD2023, Uni Flow + Refiner and Anchor BiFlow + Refiner use the same frozen VAE and refiner architecture. Both use losses for flow matching, latent endpoints, image reconstruction, and target VAE projection consistency. Anchor BiFlow also uses an anchor loss. Uni Flow trains one model per direction for 40 epochs each, while Anchor BiFlow trains one joint model for 40 epochs. Both use five training seeds and 16 CT–MRI pairs per batch. Anchor BiFlow achieves higher mean SSIM in both directions.

MethodTrainingParams (M)CT→MRI PSNR ↑CT→MRI SSIM ↑MRI→CT PSNR ↑MRI→CT SSIM ↑
Uni Flow + RefinerSeparate2.47 × 219.51 ± 0.060.7696 ± 0.00219.32 ± 0.090.7909 ± 0.002
Anchor BiFlow + RefinerJoint4.1519.47 ± 0.050.7794 ± 0.00419.69 ± 0.080.8068 ± 0.006

Reply to Q3: For each configuration in Tables 2 and 3, we report mean ± SD over five independent runs using the same random data split and training seeds 1234, 111, 222, 333, and 444. The table below reports two-tailed paired tests comparing five-step Anchor BiFlow with each variant before refinement.

MethodTrainingCT→MRI PSNRCT→MRI SSIMMRI→CT PSNRMRI→CT SSIM
Uni FlowSeparate————
Source-Cond BiFlowJoint—p<0.05p < 0.05—p<0.05p < 0.05
Dir-Cond BiFlowJointp<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Decomp BiFlowJoint————
Anchor-onlyJointp<0.05p < 0.05p<0.05p < 0.05—p<0.05p < 0.05
Anchor BiFlow (1-step)Joint—p<0.05p < 0.05—p<0.05p < 0.05
Anchor BiFlow (5-step)(contrast)Joint————
“—” indicates that the difference did not reach the p<0.05p<0.05 significance level.
Compared to independently trained Uni Flow, the differences across all four metrics failed to reach statistical significance. This is expected, as training a model for a single direction is inherently simpler than training one for both directions simultaneously.

When the two directions require different velocity predictions and the source state lacks sufficient information to distinguish between them, a shared model struggles to fit both mappings effectively. Anchors conditioned on the translation direction mitigate this issue, allowing the bidirectional model to achieve average performance comparable to that of Uni Flow prior to refinement.

Other: We would like to thank the reviewer once again for their careful review; we will incorporate the necessary revisions and additions to address the points raised in the updated version of the manuscript.

Furthermore, we have added comparative experiments with SynDiff [1] and SelfRDB [2], as shown in the table below:

MethodTrainingParams (M)FLOPs (G)SynthRAD2023 CT→MRISynthRAD2023 MRI→CTPCM2025 T1 CT→MRIPCM2025 T1 MRI→CTPCM2025 T2 CT→MRIPCM2025 T2 MRI→CT
SelfRDB [2]Separate39.73 × 2141.83 × (10–20)18.85 / 0.774019.43 / 0.802315.92 / 0.370910.67 / 0.325113.55 / 0.314810.54 / 0.3306
SynDiff [1]Joint79.46141.66 × 416.65 / 0.679017.68 / 0.756615.15 / 0.343110.87 / 0.318513.05 / 0.308210.05 / 0.2978
ProposedJoint4.1514.75 + 1.61 × 519.47 / 0.779419.69 / 0.806819.43 / 0.509214.94 / 0.402317.60 / 0.490715.16 / 0.4105

[1] M. Özbey et al., “Unsupervised Medical Image Translation With Adversarial Diffusion Models,” IEEE Transactions on Medical Imaging, 2023.

[2] F. Arslan et al., “Self-consistent recursive diffusion bridge for medical image translation,” Medical Image Analysis, 2025.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than eight hours remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Reviewer_HgwQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a shared bidirectional latent dynamics framework for CT–MRI translation. Modality-specific VAEs are first trained and frozen, after which both translation directions are modeled as conditional latent state evolution processes driven by a single shared velocity field. The central methodological proposal is Anchor BiFlow, which constructs a source-dependent, direction-aware latent anchor and uses it as an explicit reference for shared velocity prediction, rather than directly predicting the target latent or relying only on a scalar direction code. The framework also includes an image-space refiner and latent-space robustness training for the VAE decoders. The claimed contributions are the unified shared-dynamics formulation, the anchor-based direction representation, and validation on paired CT–MRI datasets showing improved mean PSNR/SSIM and reduced boundary artifacts relative to separate and jointly trained baselines.

Critical strengths

Strength 1: The problem framing is clean and addresses a real inefficiency in bidirectional medical image translation. Most CT→MRI and MRI→CT methods are trained as two separate models, which doubles training cost and parameter count and provides no mechanism for sharing structure across directions. Formulating both directions as two trajectories under one shared dynamics model is a natural and practically relevant idea, and it gives the paper a clear conceptual contribution beyond another conditional GAN or diffusion variant.

Strength 2: The Anchor BiFlow design is a sensible way to inject direction information without collapsing the shared model. A scalar direction code alone is a weak signal, and directly predicting the target latent risks degenerating into two separate predictors in disguise. Using a source-dependent, direction-aware anchor as an intermediate reference for the velocity field is a reasonable middle ground, and the ablation comparing Source-Cond, Dir-Cond, Decomp, and Anchor variants is informative for understanding where the benefit comes from.

Critical weaknesses

Weakness 1: The manuscript as provided is materially incomplete, and the central quantitative table is unreadable, which prevents evaluation of the paper’s main empirical claim. Table 1, which is supposed to report the quantitative comparison on SynthRAD2023 and PCM2025 across all 12 metric–task combinations, contains only the header row in the provided file; the actual numbers are missing. Several pages contain long runs of repeated indices, Figure 1’s caption begins with “igure 1,” and parts of the method and appendix appear truncated or corrupted. This is not a cosmetic issue: the paper’s headline claim is that the proposed method achieves the highest mean PSNR and SSIM across all evaluated settings, and without a readable Table 1 that claim cannot be verified. For a methods paper whose contribution is primarily empirical, this is a serious evaluability failure and supports rejection in the current form.

Weakness 2: The empirical attribution of the proposed Anchor BiFlow mechanism is not sufficiently isolated, and the theoretical analysis does not provide meaningful guidance about why this specific anchor design should outperform the alternatives. From Table 2, Anchor BiFlow improves over Source-Cond BiFlow by a relatively small margin before refinement, and the image-space refiner and VAE robustness training contribute additional gains. The paper does not separate the contribution of the anchor from the contribution of the refiner, the VAE robustness strategies in Table 3, or the choice of five-step Euler integration. Moreover, the analysis in Appendix A.1 essentially shows that explicit direction modeling expands the hypothesis class and therefore cannot increase the optimal risk; this inclusion relation applies equally to Dir-Cond, Decomp, and Anchor, so it does not explain why Anchor BiFlow should be better. Since the paper’s core methodological claim is specifically about direction representation in shared dynamics, this weak attribution is consequential.

Questions for authors
  1. The provided manuscript is missing the body of Table 1 and has corrupted pages. Can you provide a complete, readable version? Without the actual numbers for all 12 metric–task combinations, the paper’s central quantitative claim cannot be assessed.

  2. In Appendix A.1, Equation (17) defines vq∗=zsrcq−zsrcqv_q^* = z_{\mathrm{src}}^q - z_{\mathrm{src}}^q, which is identically zero. Is this a typographical error, and if so, what is the intended definition? This matters for verifying the flow-matching derivation.

  3. In Algorithm 1 and the surrounding text, the notation for latent states appears inconsistent (e.g., zqq(k)z_q^q(k), z~qq\tilde{z}_q^q, zq,(k)z^{q,(k)}). Can you clarify the exact update rule and the role of the anchor correction term α(zancq−zsrcq)\alpha(z_{\mathrm{anc}}^q - z_{\mathrm{src}}^q)?

  4. Can you provide a component-wise ablation that separates the anchor, the image-space refiner, and the VAE robustness training? This would directly clarify how much of the final gain is attributable to Anchor BiFlow itself rather than to the surrounding pipeline.

  5. In Table 4, are the reported latency and FLOPs for the proposed method measured with the image-space refiner included? Also, is the comparison against separate methods using the combined parameter count of two models? A clearer accounting of joint versus separate cost would make the efficiency claim more convincing.

  6. The method relies on paired CT–MRI data for flow matching, endpoint supervision, and anchor supervision. How sensitive is the framework to pairing quality, and can it be extended to unpaired or weakly paired settings?

  7. How many random seeds or runs are used for the PSNR/SSIM results? Can you report variance or paired statistical tests to establish that the reported improvements are robust, especially given that some margins in Table 2 are small?

3 comments

We thank the reviewer for providing a substantial number of comments. We understand that AI tools may be used to assist with the review process, given the high volume of manuscripts; if you have utilized AI for this purpose, we kindly ask that you carefully verify the feedback provided. Our point-by-point responses are as follows:

Reply to Q1: We submitted a fully compiled PDF. Therefore, I find it difficult to understand the basis for the reviewer's question here. Including what you mentioned:“Several pages contain long runs of repeated indices, Figure 1’s caption begins with “igure 1,” and parts of the method and appendix appear truncated or corrupted. ”

Reply to Q2: The formula in the appendix has always appeared as follows: Equation (17) in Appendix A.1 is vq∗=ztgtq−zsrcq,v_q^\ast=z_{\mathrm{tgt}}^q-z_{\mathrm{src}}^q, I am also struggling to understand the basis for your question.

Reply to Q3: In Algorithm 1, qq denotes the translation direction, kk represents the current step, and KK is the total number of integration steps. At each step, the velocity is computed based on the current state zˉq=zq,(k)\bar{z}^q = z^{q,(k)}, and the state is updated as follows:

\left[u_\theta^{q,(k)} + \alpha\left(z_{\mathrm{anc}}^q - z_{\mathrm{src}}^q\right)\right].$$ The anchor point is calculated once before the integration begins. While the learned term $u_\theta^{q,(k)}$ varies according to the current state and time, the anchor correction term $\alpha(z_{\mathrm{anc}}^q - z_{\mathrm{src}}^q)$ remains constant across all steps, pointing from the source state toward the anchor point. Given a step size of $1/K$, the cumulative displacement contributed by this correction term after $K$ steps is $\alpha(z_{\mathrm{anc}}^q - z_{\mathrm{src}}^q)$. **Reply to Q4:** Tables 2 and 3 in the original manuscript examine the contributions of three components. The following SSIM gains correspond to CT→MRI and MRI→CT, respectively. - **Anchor (Table 2).** Without a refiner, Anchor BiFlow yields improvements of 0.012 and 0.006 compared to Source-Cond BiFlow. - **Refinement in image space (Table 2).** Incorporating a refiner into Anchor BiFlow results in SSIM gains of 0.016 and 0.024. - **VAE robustness training (Table 3).** With Anchor BiFlow fixed and the refiner disabled, applying “Perturb” and “Consistency” during VAE pretraining improves SSIM by approximately 0.005 in both directions. We have also included an “Anchor-only” baseline trained jointly in both directions. It directly decodes the predicted anchors and uses the same decoder and refiner architecture. As shown in the table below, Anchor BiFlow with five inference steps achieves higher average PSNR and SSIM than Anchor-only in both directions on SynthRAD2023, with and without refinement. | Method | Training | Refiner | CT→MRI PSNR ↑ | CT→MRI SSIM ↑ | MRI→CT PSNR ↑ | MRI→CT SSIM ↑ | |---|---|:---:|---:|---:|---:|---:| | Anchor-only | Joint | — | 19.28 ± 0.04 | 0.7447 ± 0.003 | 19.21 ± 0.08 | 0.7771 ± 0.001 | | Anchor BiFlow | Joint | — | 19.44 ± 0.02 | 0.7634 ± 0.002 | 19.26 ± 0.02 | 0.7831 ± 0.002 | | Anchor-only | Joint | ✓ | 19.44 ± 0.04 | 0.7651 ± 0.002 | 19.26 ± 0.09 | 0.7883 ± 0.001 | | **Anchor BiFlow** | Joint | ✓ | **19.47 ± 0.05** | **0.7794 ± 0.004** | **19.69 ± 0.08** | **0.8068 ± 0.006** | **Reply to Q5:** Table 4 in the appendix reports the total parameter count required for bidirectional translation. For methods trained separately for each direction, the parameter counts of the two models are summed. Anchor BiFlow has 4.15M parameters in total, including two VAEs and two refinement modules. Latency and FLOPs are reported for a single translation in one direction, including refinement in image space. The original entry “16.36 × 5” counted source encoding, anchor prediction, target decoding, and refinement five times each. These modules are executed once per translation. Only the velocity network is executed five times. We have recalculated the FLOPs for the 11 methods listed in the original table using the actual number of calls to each module. With a batch size of 1 and an input resolution of 256×256, Anchor BiFlow has a latency of 6.26 ± 0.16 ms and a computational cost of 22.79G FLOPs (counting each multiplication and accumulation as two operations in total). **Reply to Q6:** This paper focuses on paired CT–MRI translation; weakly paired or unpaired scenarios have not yet been addressed. We will clarify the scope of our current method in the "Limitations" section of the revised manuscript and identify these two scenarios as directions for future research.

Reply to Q7: For all methods presented in Table 2, models were trained independently using five different random seeds (1234, 111, 222, 333, and 444) based on a fixed data split. The reported PSNR and SSIM values represent the mean ± standard deviation across these five training runs. In the table below, "Anchor BiFlow + Refiner" is used as the reference. For the MRI→CT direction, both PSNR and SSIM are significantly higher than those of all 11 comparison groups. For the CT→MRI direction, the results show significant differences in 4 groups for PSNR and in 10 groups for SSIM.

MethodTrainingRefinerCT→MRI PSNRCT→MRI SSIMMRI→CT PSNRMRI→CT SSIM
Uni FlowSeparate——p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Uni FlowSeparate✓—p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Source-Cond BiFlowJoint——p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Dir-Cond BiFlowJoint—p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Decomp BiFlowJoint—p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Anchor-onlyJoint—p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Anchor BiFlow (1-step)Joint—p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Anchor BiFlow (5-step)Joint——p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Decomp BiFlowJoint✓——p<0.05p < 0.05p<0.05p < 0.05
Anchor-onlyJoint✓—p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Anchor BiFlow (1-step)Joint✓—p<0.05p < 0.05p<0.05p < 0.05p<0.05p < 0.05
Anchor BiFlow (5-step)(contrast)Joint✓————

Other: We would like to thank the reviewer once again for their careful review; we will incorporate the necessary revisions and additions to address the points raised in the updated version of the manuscript.

Furthermore, we have added comparative experiments with SynDiff [1] and SelfRDB [2], as shown in the table below:

MethodTrainingParams (M)FLOPs (G)SynthRAD2023 CT→MRISynthRAD2023 MRI→CTPCM2025 T1 CT→MRIPCM2025 T1 MRI→CTPCM2025 T2 CT→MRIPCM2025 T2 MRI→CT
SelfRDB [2]Separate39.73 × 2141.83 × (10–20)18.85 / 0.774019.43 / 0.802315.92 / 0.370910.67 / 0.325113.55 / 0.314810.54 / 0.3306
SynDiff [1]Joint79.46141.66 × 416.65 / 0.679017.68 / 0.756615.15 / 0.343110.87 / 0.318513.05 / 0.308210.05 / 0.2978
ProposedJoint4.1514.75 + 1.61 × 519.47 / 0.779419.69 / 0.806819.43 / 0.509214.94 / 0.402317.60 / 0.490715.16 / 0.4105

[1] M. Özbey et al., “Unsupervised Medical Image Translation With Adversarial Diffusion Models,” IEEE Transactions on Medical Imaging, 2023.

[2] F. Arslan et al., “Self-consistent recursive diffusion bridge for medical image translation,” Medical Image Analysis, 2025.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than eight hours remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Reviewer_h2eFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Anchor BiFlow for CT-to-MRI and MRI-to-CT translation using one shared flow model. It uses separate VAEs, a learned anchor to guide latent updates, and an image refiner. Experiments on SynthRAD2023 and PCM2025 report the highest mean PSNR and SSIM among the tested methods. The compact shared model is interesting, but important comparisons are missing. The unclear evaluation details and unexplained poor baseline results also reduce confidence in the reported gains. I do not think the current evidence is strong enough to support acceptance.

Critical strengths
  1. The full system supports both directions with a total of 4.15M parameters. Its five-step inference is also much faster than the tested diffusion methods.

  2. The full model achieves the highest mean PSNR and SSIM in all six settings in Table 1. The ablations provide some evidence that anchor guidance and image refinement improve the outputs.

Critical weaknesses
  1. The anchor is directly trained to match the target latent code in Equation 14. The paper does not compare against an anchor-only model using the same decoder and image refiner. Without this test, the benefit of five flow steps over a simpler direct mapping remains unclear.

  2. Metric calculation and baseline fairness need clarification. Table 1 reports very low scores, including PSNR of 7.29 dB and SSIM of 0.0103 for DDIM on PCM2025 T1 MRI→CT. The authors should clarify intensity normalization and the PSNR data range, and confirm that all methods use the same evaluation code. Baseline training and tuning details are also needed to explain these results and establish a fair comparison.

  3. The PCM2025 dataset paper reports only 18 patients, so an 8:2 subject split would leave roughly four patients for testing. This paper does not report exact patient counts, results across different splits, or patient-level confidence intervals. The reported gains may therefore depend strongly on which patients were selected for testing.

  4. Figure 8 is difficult to read and understand. The panels, points, and axis labels in Figure 8 are too small to read clearly. It is difficult to follow how the latent points change between steps or judge whether they move closer to the target. Larger panels, clearer labels, and arrows showing the movement would make the figure more useful.

Questions for authors

See Weaknesses.

3 comments

We thank the reviewer for their careful review. Our point-by-point responses are provided below:

Reply to W1: We have added an “Anchor-only” baseline. This model directly decodes the predicted anchors, utilizing the same decoder and refiner architecture.

As shown in the table below, on the SynthRAD2023 dataset, Anchor BiFlow uses five flow inference steps and achieves higher average PSNR and SSIM values in both translation directions than the Anchor-only baseline, regardless of whether a refiner is employed.

MethodTrainingRefinerCT→MRI PSNR ↑CT→MRI SSIM ↑MRI→CT PSNR ↑MRI→CT SSIM ↑
Anchor-onlyJoint—19.28 ± 0.040.7447 ± 0.00319.21 ± 0.080.7771 ± 0.001
Anchor BiFlowJoint—19.44 ± 0.020.7634 ± 0.00219.26 ± 0.020.7831 ± 0.002
Anchor-onlyJoint✓19.44 ± 0.040.7651 ± 0.00219.26 ± 0.090.7883 ± 0.001
Anchor BiFlowJoint✓19.47 ± 0.050.7794 ± 0.00419.69 ± 0.080.8068 ± 0.006

Reply to W2: On PCM2025 T1 MRI→CT, DDIM achieved a PSNR of 7.29 dB and an SSIM of 0.0103. Following the original DDIM implementation, we trained the model for 100 epochs using AdamW with a learning rate of 2×10⁻⁴ and a batch size of 16. We used 50 DDIM sampling steps at inference.

For Table 1, all methods used the same code to compute PSNR and SSIM on full 256×256 grayscale slices with data_range=1.0 and average the scores across slices. For SynthRAD2023, we used a CT window level of 40 and width of 80 and normalized MRI intensities to [0, 1]. For PCM2025, we scaled PNG intensities directly to [0, 1]. We also converted DDIM outputs from [−1, 1] to [0, 1] before evaluation.

Reply to W3: To assess sensitivity to patient selection, we conducted an additional experiment on PCM2025 using five random seeds to generate five patient splits. Each split included 14 of the 18 patients for training and 4 for testing. Each patient’s CT, T1, and T2 slices were kept together. Across the five splits, 13 distinct patients appeared in the test sets. On each split, we trained Anchor BiFlow + Refiner and BiBBDM for 40 epochs using the same training seed throughout.

For each task, we averaged PSNR and SSIM first across slices within each patient and then across test patients. The table reports gains of Anchor BiFlow + Refiner over BiBBDM. For one of the five splits, we estimated 95% confidence intervals by resampling the four test patients. The last two columns show the range of gains across all five splits.

Translation taskPSNR gain (dB) [95% CI], one splitSSIM gain [95% CI], one splitPSNR gain range across 5 splits (dB)SSIM gain range across 5 splits
T1 CT→MRI2.58 [1.71, 3.40]0.083 [0.047, 0.122]1.94–3.070.057–0.104
T1 MRI→CT2.35 [1.12, 3.58]0.059 [0.006, 0.117]1.68–3.160.028–0.083
T2 CT→MRI3.00 [1.99, 4.00]0.112 [0.061, 0.168]2.43–3.780.084–0.143
T2 MRI→CT2.77 [1.42, 4.19]0.065 [0.016, 0.118]1.83–3.420.033–0.094

The magnitude of the gains varied across splits, but Anchor BiFlow + Refiner achieved higher mean PSNR and SSIM than BiBBDM on all four tasks in every split.

Reply to W4: We have enlarged the subplots and scatter plots in Figure 8 to make the underlying distributions at each step clearer.

Other: We would like to thank the reviewer once again for their careful review; we will incorporate the necessary revisions and additions to address the points raised in the updated version of the manuscript.

Furthermore, we have added comparative experiments with SynDiff [1] and SelfRDB [2], as shown in the table below:

MethodTrainingParams (M)FLOPs (G)SynthRAD2023 CT→MRISynthRAD2023 MRI→CTPCM2025 T1 CT→MRIPCM2025 T1 MRI→CTPCM2025 T2 CT→MRIPCM2025 T2 MRI→CT
SelfRDB [2]Separate39.73 × 2141.83 × (10–20)18.85 / 0.774019.43 / 0.802315.92 / 0.370910.67 / 0.325113.55 / 0.314810.54 / 0.3306
SynDiff [1]Joint79.46141.66 × 416.65 / 0.679017.68 / 0.756615.15 / 0.343110.87 / 0.318513.05 / 0.308210.05 / 0.2978
ProposedJoint4.1514.75 + 1.61 × 519.47 / 0.779419.69 / 0.806819.43 / 0.509214.94 / 0.402317.60 / 0.490715.16 / 0.4105

[1] M. Özbey et al., “Unsupervised Medical Image Translation With Adversarial Diffusion Models,” IEEE Transactions on Medical Imaging, 2023.

[2] F. Arslan et al., “Self-consistent recursive diffusion bridge for medical image translation,” Medical Image Analysis, 2025.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than eight hours remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Thank you for the response and the new experiments. The five patient splits, the clearer evaluation details, the new baselines, and the larger Figure 8 address several of my concerns. However, in some cases the five flow steps add only a small gain over Anchor-only (for example, 19.44 vs. 19.47 dB for CT→MRI with the refiner). The very low DDIM scores (7.29 dB PSNR and 0.0103 SSIM) are still not explained, and they look like a failed run. The gap to SelfRDB also grows from about 0.3–0.6 dB on SynthRAD2023 to about 3.5–4.6 dB on PCM2025, so I am still not sure the baselines were well tuned. I have raised my score to 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.