Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Magnetic resonance (MR) imaging provides excellent soft-tissue contrast without ionizing radiation, but lacks electron density information essential for radiotherapy dose calculation. Existing MR-to-CT synthesis methods mostly only depend on spatial-domain learning, leading to texture overfitting and compromised bone delineation and dosimetric accuracy. Critically, most latest deep learning approaches lack rigorous clinical validation on dose computation and treatment planning, limiting their translational utility. To address these gaps, we propose DPMix-GAN, a generative adversarial framework integrating dual-domain perception and patch-wise feature dynamic mixing for robust CT synthesis. DPMix-GAN fuses spatial and frequency features to capture complementary anatomical representations, and employs dynamic patch-wise mixing of high-order statistics to preserve anatomical structure. Experimental results on a clinical dataset across brain and pelvic regions shows that DPMix-GAN generated anatomically accurate synthesis CT with well-preserved boundaries. Moreover, clinical validation on radiotherapy treatment plans confirmed dosimetric equivalence between synthetic and real CT, with gamma passing rates exceeding clinical acceptance thresholds. The code is available at https://github.com/AprilT0621/DPMix-GAN.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0524_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/AprilT0621/DPMix-GAN

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiJia_DPMixGAN_MICCAI2026,
        author = { Li, Jiapeng AND Li, Mingzhu AND Zhang, Xiao AND Zhang, Shuyuan AND Yuan, Cheng AND Zhang, Lian AND Feng, Hongying},
        title = { { DPMix-GAN: Dual-Domain Perception and Dynamic Patch-Wise Mixing GAN for 3D CT Synthesis in MR-only Radiotherapy } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The manuscript presents DPMix-GAN, a generative adversarial network (GAN) designed for 3D synthesis of CT images from MRI to enable MR-only radiotherapy planning. The primary goal of this framework is to provide accurate electron density maps for radiation dose calculation while overcoming the soft-tissue contrast limitations inherent in traditional CT scans. A key technical novelty is the Dual-Domain Perception Module (DPM), an architectural component embedded at the U-Net bottleneck that employs Fast Fourier Transform (FFT) to jointly model spatial features and frequency-domain responses. Additionally, the model employs Patch-wise Feature Dynamic Mixing (PFDM), a discriminator strategy that perturbs localized high-order statistics (mean, variance, skewness, and kurtosis) to prevent the model from overfitting to modality-specific textures. Beyond standard image metrics, the authors ensure clinical equivalence through dosimetric validation, performing treatment planning and dose recalculation to quantify the accuracy of the synthesized images for actual patient care.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    • The integration of the Fast Fourier Transform directly into the generator’s bottleneck is a sophisticated approach to capturing high-frequency anatomical boundaries, such as bone-tissue interfaces, which are often over-smoothed in spatial-only GANs.

    • The inclusion of Gamma index analysis and Dose-Volume Histogram congruence, featuring a 96.23% passing rate, provides a much higher level of clinical confidence than intensity-based metrics alone.

    • Testing the framework on both brain and pelvis datasets demonstrates a significant degree of anatomical robustness across different parts of the body.

    • The use of higher-order statistics, specifically skewness and kurtosis, in the discriminator regularizer is a clever and effective extension of standard feature-mixing techniques.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The test set is limited to only 16 brain and 15 pelvic cases, a small sample size that makes it difficult to definitively claim state-of-the-art performance or long-term clinical reliability.

    2.Reliance on a private dataset hinders reproducibility, and the authors should benchmark their method on public datasets like SynthRad2023/2025 (that provides MR/CT pairs for hundred of patients, in abdominal, pelvic and brain regions) to compare against established standards like 3D nnU-Net and SwinUNETR.

    3.Using rigid registration for pelvic anatomy is insufficient, as deformable registration is required to account for variations in bladder and bowel states to ensure valid ground-truth alignment for training. Please at least discuss that a deformable step was qualitatively judged unncessary. 4.The intensity normalization process is underspecified, leaving it unclear whether global min-max, per-patient, or z-score approaches were used, which is critical given the different intensity distributions of MR and CT. 5.The chosen patch size of 128x128x8 is extremely thin for a 3D architecture, effectively resulting in a 2.5D “thick-slice” implementation rather than a true volumetric 3D model. Please provide justifications of such approach rather than a standard 3D patch-based approach. The authors should explicitly state if VRAM constraints limited the implementation and justify why a more isotropic patch size, such as 32x32x32, was not evaluated. Also, the use of a RTX 4070 was mentioned, please explicit the VRAM of such hardware (I assume 12GB?). 6.The 3D-UNet baseline scores are significantly lower than established standards in public challenges, raising concerns about whether the baseline was properly optimized or hampered by the thin-patch configuration. Also, a more modern baseline, such as a Residual UNet, should have been considered given its widespread success in current medical imaging synthesis and segmentation fields.

    1. There is a lack of explanation regarding how the 2D VGG network processes 128x128x8 medical inputs, making the reported perceptual supervision difficult to evaluate or replicate. Also, applying the VGG-19 perceptual loss, a 2D network trained on natural RGB images, to 3D medical patches creates a conceptual mismatch that is not addressed in the text. Recent approaches tried to account for this conceptual mismatch using 3D medical features. This should at least be discussed. (e. g. Longuefosse et al. : Anatomical feature-prioritized loss for enhanced MR to CT translation, Phys Med Biol. 2025) 8.The true value of the DPM module remains unquantified because the paper lacks a comparison between implementing FFT within the generator architecture versus simply using FFT as a loss function (e. g. spectral loss or frequency loss, as in [Morell-Ortega et al. : Robust deep MRI contrast synthesis using a prior-based and task-oriented 3D network, Imaging Neurosci. 2025]) 9.The study lacks segmentation-based validation, such as using TotalSegmentator to confirm if the model correctly reconstructs the geometry of critical structures rather than just generating realistic textures.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    I highly recommend the authors to participate in SynthRAD2023 and 2025 challenges (post-challenge task or simply validation set). I am not affiliated with the organizers; however, I suggest this to ensure the proposed model’s generalizability is rigorously evaluated against multi-center, high-quality public benchmarks that represent the current state-of-the-art in synthetic CT generation.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The major factor leading to this recommendation is the weak baseline problem combined with an unjustified anisotropic implementation. By using a patch depth of only 8 slices, the authors have created a environment that likely favors their specific architectural additions (DPM/MSwin-T) while preventing standard 3D baselines (3D-UNet) from reaching their full potential. Furthermore, the lack of deformable registration for the pelvis and the very small test cohort (N ~= 15) may harm the generalizability of the results. While the clinical dosimetric validation seems well-executed, it cannot fully compensate for these methodological shortcuts, along with a lack of segmentation results. Benchmark results on a public dataset (e.g. SynthRAD) would much benefit the statements made.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I thank the authors for their detailed response and for providing the anonymized code repository to support reproducibility. The paper offers a valuable contribution to MRI-only radiotherapy. While the dataset is small and single-center, this is acceptable for a conference paper presenting preliminary clinical feasibility. The authors have promised to integrate relevant discussions regarding preprocessing, 3D medical feature-based losses, and future works on public benchmarks into the final manuscript, which is appreciated. However, a few justifications in the rebuttal remain unconvincing: the claim that 3D-UNet intrinsically causes over-smoothing is incorrect, as this artifact usually comes from intensity-based (L1/L2) losses rather than the architecture itself, leaving the poor UNet baseline performance a likely implementation issue. Furthermore, the anisotropic 1281288 patch size essentially reduces the model to a 2.5D approach rather than a true 3D patch-based approach, a choice likely driven by the 12GB GPU memory constraint rather than purely methodological advantages. Some discussions or mention for future works around this issue would be appreciated. Nevertheless, given the overall novelty of DPMix-GAN and the clinical relevance of the validation, the paper’s contributions outweigh these limitations, and I maintain an overall positive evaluation.



Review #2

  • Please describe the contribution of the paper

    The paper proposes DPMix-GAN, a paired MR-to-CT synthesis framework aimed at MRI-only radiotherapy planning. Two components are introduced: (1) a Dual-domain Perception Module (DPM) placed at the generator bottleneck that processes spatial features together with FFT-derived frequency features through multi-shape window self-attention and fuses them back to the spatial domain; and (2) a 3D Patch-wise Feature Dynamic Mixing (PFDM) regularizer applied to the PatchGAN discriminator, which perturbs local feature statistics (mean, variance, skewness, kurtosis) with those of a randomly sampled real CT patch to discourage reliance on modality-specific textures. Evaluation is performed on an in-house dataset of 193 paired brain and pelvic MR/CT scans, with both image-level metrics and clinical dosimetric validation using MC-square dose recalculation, DVH curves and Gamma analysis.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Clinically meaningful evaluation. Unlike most MR-to-CT synthesis papers, the authors go beyond intensity-level metrics and report dosimetric equivalence on recomputed treatment plans, with a Gamma passing rate of 96.23% and a DVH comparison on CTV and OARs. This is the right level of validation for the MRI-only RTP claim and directly addresses a gap the authors highlight in the introduction.
    • The dual-domain design is well motivated. Placing spatial/frequency MSwin-T branches at the bottleneck is a reasonable way to recover high-frequency bone interfaces that pure spatial GANs tend to over-smooth, and the ablation (Table 2, Fig. 5) shows a consistent MAE reduction from the DPM module with visibly cleaner skull boundaries.
    • PFDM is a sensible discriminator regularizer for the small-data regime. Mixing the content patch with high-order style statistics of a real CT patch pushes the discriminator away from texture shortcuts; the ablation supports this with reduced internal hallucinations (Fig. 5 error maps) and a further ~2 HU gain when combined with DPM.
    • Consistent quantitative improvements across two anatomical regions (brain and pelvis) against three reasonable baselines (3D-UNet, SwinUNETR, MTT-Net), with tighter standard deviations than competitors on most metrics.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Single-center, small-scale evaluation. The entire validation relies on 193 paired scans from one hospital with test sets of only 16 brain and 15 pelvis cases. A central claim of the introduction is that standard GANs “generalize poorly across centers and patient cohorts,” yet no cross-center or cross-scanner evaluation is provided. Limit the generalization narrative to the single-center setting.
    • Baseline coverage is narrow. The discussion explicitly positions DPMix-GAN against diffusion-based MR-to-CT approaches and against structure-aware GANs , but none of these are actually compared.
    • Anonymized code link is provided but not populated.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses a clinically relevant problem and is one of the few MR-to-CT synthesis submissions I have seen that actually recomputes treatment plans and reports Gamma/DVH results rather than stopping at MAE/SSIM. The method is reasonable, the two components are supported by a coherent ablation, and quantitative gains over the chosen baselines are consistent across two anatomical regions. However, the methodological novelty is incremental relative to prior spatial/frequency translation and localized style-mixing work, the evaluation is single-center with a small test set, the dosimetric analysis is under-powered.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper proposes DPMix-GAN, a GAN-based framework for MR-to-CT synthesis in MRI-only radiotherapy, combining a dual-domain perception module (DPM) and a patch-wise feature dynamic mixing module (PFDM). It evaluates the method not only with standard image-level metrics, but also with dosimetric validation for radiotherapy planning.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper addresses a clinically meaningful problem, namely MR-to-CT synthesis for MRI-only radiotherapy, rather than treating image synthesis as an isolated technical task. A strength is the use of real paired clinical CT–MR data across two anatomical regions, together with dosimetric validation beyond standard image-level metrics. This makes the evaluation more practically relevant to radiotherapy planning.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Limited novelty of DPM. DPM reuses the same MSwin-T block for both spatial and frequency-domain features, rather than introducing a frequency-specific modeling design. Its novelty therefore appears limited.

    Insufficient justification for bottleneck-only design. DPM is applied only at the bottleneck, but the paper does not justify why this is the most effective stage for preserving high-frequency anatomical information, especially since the features at this stage are already highly compressed.

    Baseline set is limited. The comparison includes only one clearly recent task-specific baseline (MTT-Net, 2024), while the other baselines are more generic architectures. This makes the benchmark less convincing.

    Ablations show usefulness, not mechanism. The ablations support that DPM and PFDM are empirically helpful, but they do not directly validate the stronger interpretation of why these modules work.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I assign this paper a 4/6 rather than 3/6 mainly because it goes beyond standard image-level evaluation: the study is based on real clinical data and includes dosimetric validation, which is particularly important for MRI-only radiotherapy. These aspects materially strengthen the practical relevance of the work. That said, I still have notable concerns regarding the limited strength of the baseline set and the fact that the mechanism-level interpretations of DPM/PFDM are stronger than what the current evidence directly establishes.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    After considering the rebuttal, I support acceptance mainly due to the paper’s clinical relevance, not because my technical concerns were fully resolved. The rebuttal does not substantially alleviate my concerns about the limited baseline set, and I remain unconvinced by the justification for excluding broader recent comparisons. I am also not fully persuaded by the claimed novelty or design rationale of DPM/PFDM, including the bottleneck-only placement. Nevertheless, the use of real paired clinical MR–CT data and dosimetric validation makes the work valuable for MRI radiotherapy, which justifies an accept decision despite these remaining concerns.



Author Feedback

We sincerely thank the AC and reviewers for appreciating the novelty of our DPMix-GAN that fuses spatial-frequency features and employs the patch-statistics-perturbed discrimination for CT synthesis in MRI-only radiotherapy, as well as the dosimetric validation. Major concerns are categorized and addressed below. 1.Inadequate baselines (R1Q6, R2Q2, R3Q3). Our baselines cover representative types: 3D-UNet as a CNN baseline, SwinUNETR as a Transformer baseline, and MTT-Net as a recent task-specific model, across which our method consistently demonstrated superior performance. Diffusion-based methods were excluded in baselines due to the accompanying clinical concerns such as stochastic generation and inference efficiency. All baselines used the same preprocessing and training configuration on a 12GB RTX 4070S. 3D-UNet would intrinsically lead to over-smooth synthesis and poorer sharp-boundary recovery, confirmed in Fig. 3.Its performance may also be sensitive to different data and patch size. We definitely agree that broader comparisons against more recent models would improve the evaluation and such follow-up benchmarks are in progress. 2.Small/Single-center data (R1Q1Q2, R2Q1). The data was limited to single center with small validation sets due to the strict selection criteria, i.e., paired MR-CT data with shortest acquisition intervals, consistent MR protocols, and planning information, to conduct treatment planning-orientated evaluation as a clinical feasibility validation. Besides, the model was consistently tested using independent test sets across two anatomical regions and MR protocols and achieved satisfactory performance, indicating it did not simply overfit validation-specific features. We fully agree with reviewer and cross-institute benchmarks and benchmarks against public data like SynthRAD2023 with the same metrics are ongoing for fairer comparison. 3.DPM/PFDM design and ablation (R1Q8, R3Q124). Sharing the same architecture of the two MSwin-T branch is a deliberate symmetric design to keep spatial and frequency features aligned for effective attention fusion. Current bottleneck placement enables global frequency modeling with the largest receptive field, while shallower-stage FFT features are more likely dominated by local texture patterns. Compared to Morell-Ortega et al., DPM injects frequency cues into feature transformation and decoding, promoting sharper boundaries and consistent structural recovery, proved by the ablation results in Fig. 5 that Baseline+DPM reduced errors near sharp skull boundaries. Meanwhile, PFDM perturbs high-order discriminator statistics and suppresses internal artifacts, consistent with reduced reliance on modality-specific texture cues. We will revise the relevant descriptions and add discussion to clarify this distinction. 4.Concerns on preprocessing, VGG loss, and segmentation validation (R1Q34579). In preprocessing, all rigidly registered pairs were carefully inspected with marginal cases of localized mismatches observed and manually adjusted. Both CT and MR were min-max normalized to [-1,1] per-volume with a prior clip to [-1000, 2000] for CT. Volumes of interest were usually thin in z after resample, like pelvic 300×300×(42–96). Therefore, the thin 128×128×8 patch was intentionally selected to preserve large in-plane context while retaining local cross-plane information. VGG loss was computed slice-wise and averaged as an auxiliary term, while 3D consistency is enforced by the 3D generator, adversarial loss, and voxel-wise supervision. We will revise to add discussions on the direct use of 3D medical feature-based losses including Longuefosse et al.. Our evaluation included DVH analysis, which necessitates a prior accurate segmentation. We agree that systematic direct segmentation-based validation is valuable future work. 5.Code unavailable (R2Q3). The code has now been released in the anonymized repository to support reproducibility.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper addresses a clinically relevant problem of MR only planning by generating synthetic CT. All reviewers appreciated the novelty of the DPM approach that combines Fast Fourier Transform to extract features in the frequency domain together with the spatial domain and the patch-based discriminator that perturbs the textural features to improve robustness to imaging variations. The rigor of experimental evaluation was seen as high especially given the treatment replanning and assessment of dosimetric accuracy in addition to standard image-based metrics. However, lack of comparison to current state-of-the-art methods was seen as weakness and rationale should be clarified together with a clear explanation of how the current method overcomes the gaps in prior methods. Please also clarify some of the experiment details and address all of the reviewers’ concerns as best as possible.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The authors addressed the reviewers concerns – the paper has technical merits and should be presented at the meeting. It would however be good to acknowledge the limitation of small dataset in the discussion of the final paper.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    All reviewers are positive about this paper. Given its clinical relevance, I lean toward acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Unfortunately, one of the reviewers missed the post-rebuttal evaluation - despite my reminders, but the other two agree that the paper has sufficient merit to be part of MICCAI 2026.Myself, I believe the topic is relevant and the contribution sound and properly evaluated for the scope of a conference paper. Yet, I agree with reviewer 1 that it would be highly recommendable to include the promised discussions (see post-rebuttal decision justification) and in particular attack the raised issue with respect with the fairness of the statements regarding 3D U-Net and the rather 2.5D nature of the approach. I also encourage the authors to take the criticisms of reviewer 3 in the decision for post-rebuttal justification seriously discussing the limitations raised. This should be part of the camera-ready version.



back to top