Abstract

While whole-body multimodal medical imaging scanners have been increasingly recognized for more effective medical applications, the excessive long acquisition time in PET–MR scanning is a major obstacle in more efficient clinical practice. Deep learning–based MRI translation provides a potential solution to reduce scan duration. However, current models often focus on specific anatomical regions and face challenges for whole-body scans that consists of highly heterogeneous feature distributions mainly due to (1) different anatomical regions across whole-body, and (2) lesions or pathological tissues. This paper tackles the challenges through a novel Heterogeneity-Adaptive Diffusion Schrödinger Bridge (HA-DSB) framework. By explicitly modeling translation as stochastic transport between source and target distributions, HA-DSB incorporates region context embeddings derived from a vision-language model (VLM) to enable region-specific modeling. To enhance fidelity of the pathological tissue, lesion-aware metabolic prior from PET is integrated directly into the bridge dynamics through a dual-stage guidance mechanism. Specifically, a PET-guided noise modulation module adaptively scales spatial diffusion perturbations during the forward process, while PET features are leveraged during the reverse process to selectively amplify lesion-relevant structures via an attention mechanism. Experiments demonstrate the superiority of our method across all body regions in whole- body MRI translation and show improved translation quality in lesion areas under PET guidance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2603_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/xyw-medical-research/HADSB

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanChe_HeterogeneityAdaptive_MICCAI2026,
        author = { Wang, Chengbo AND Yu, Jiacheng AND Bian, Linjie AND Qi, Ming AND Liu, Xiaosheng AND Che, Tongtong AND Zhang, Jichang AND Li, Shuyu AND Song, Shaoli AND Wang, Xiuying},
        title = { { Heterogeneity-Adaptive Diffusion Schrödinger Bridge for PET-Guided Whole-Body MRI Translation } },
        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 authors propose a Heterogeneity-Adaptive Diffusion Schrödinger Bridge (HA-DSB) framework for PET-guided whole-body MR translation. To handle highly heterogeneous feature distributions across the whole body, the method incorporates region context embeddings derived from a vision-language model (VLM). Furthermore, it integrates a dual-stage PET guidance mechanism, comprising forward process noise modulation and reverse process multi-scale attention, to enhance the structural fidelity of pathological tissues.

  • 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.

    1.Methodological Formulation: Explicitly modeling MR translation as a stochastic transport problem between source and target distributions via DSB is theoretically sound and effectively addresses the over-smoothing issues inherent in standard implicit diffusion alignment for whole-body datasets. 2.Targeted Adaptive Mechanisms: The dual-stage PET guidance is logically designed. Using PET uptake to spatially modulate the forward diffusion perturbations correctly redistributes the learning difficulty, preventing the model from heavily biasing toward dominant healthy tissues. 3.Good Intra-Domain and Lesion-Specific Evaluation: Compiling a dataset of 246 whole-body PET/MR subjects is a commendable effort given the prolonged acquisition time and high cost of this modality. More importantly, the explicit isolation and evaluation of a clinically confirmed lesion cohort (21 patients) demonstrates a rigorous experimental design, providing convincing evidence of the method’s superiority in synthesizing pathological regions within the source distribution.

  • 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.Clinical Assumption Limitations: The noise modulation mechanism assumes that lesions typically manifest as focal high-uptake signals in PET. This design fundamentally neglects hypometabolic or photopenic lesions. The framework’s performance on low-uptake pathological structures remains unverified and theoretically questionable. 2.Empirical Parameter Sensitivity: The spatial scaling map bounds for the noise modulator \mathcal{M}\phi are set to s{min}=0.2 and s_{max}=2.0 without theoretical justification. The paper lacks an ablation study analyzing the model’s sensitivity to these hyperparameters. 3.VLM Dependency and Reliability: The generation of region context embeddings relies entirely on Google Gemini 3 Pro. While the authors state its reliability has been empirically verified, the paper lacks transparency regarding how hallucinated labels or misalignments in slice-level organ classification were systematically handled or filtered during training. 4.The risk of distribution shift in single-center data: The dataset, while large, is collected from a single institution using a single scanner model. The framework’s robustness against multi-center distribution shifts and varying acquisition protocols is unverified, which limits its immediate clinical translation potential.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    No.

  • 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

    No.

  • 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?

    This submission addresses a highly significant clinical bottleneck (prolonged whole-body PET/MR acquisition) with a highly tailored and mathematically rigorous approach (HA-DSB). The quantitative and qualitative results, especially the improvements in lesion-containing areas, demonstrate clear clinical utility. However, the paper operates on a simplified assumption regarding PET lesion signatures and lacks sensitivity analyses for key hyper-parameters. I am recommending a Weak Accept, contingent upon a satisfactory rebuttal clarifying the robustness of the noise modulator and the framework’s applicability to non-hypermetabolic lesions.

  • 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.

    The authors have addressed and responded to the questions I raised. I have no further questions.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a heterogeneity-adaptive diffusion Schrödinger bridge framework for PET-guided whole-body MRI translation. Its main contribution is to explicitly model both anatomical and pathological heterogeneity in whole-body imaging through region-aware conditioning, PET-guided modulation, and PET-aware attention, leading to improved synthesis quality across body regions, especially on lesion-containing slices.

  • 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 and well-motivated problem. In whole-body PET/MR examinations, multiple MR sequences are often required, resulting in long acquisition time. In this context, PET-assisted whole-body MR-to-MR translation is clinically relevant because it may reduce the need for acquiring additional MR sequences and improve scanning efficiency. The method is technically sound and reasonably novel in adapting a Schrödinger bridge framework to heterogeneous whole-body PET/MR translation, and the experiments show consistent improvements over competitive baselines, with particularly encouraging results on lesion-containing slices. The paper is also clearly written and well presented.

  • 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.

    The ablation study could be stronger. Since the method includes several key components, more detailed ablations would be helpful to isolate the contribution of each module. And it seems that the author does not intend to make the project code open source.

  • 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?

    Because this paper has clear clinical significance, is written clearly and logically, and has a high degree of innovation in methods. However, since the authors did not express any intention to release the source code, it can only be given a score of four.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper proposed a heterogeneity-adaptive diffusion Schrödinger bridge framework for whole-body MRI translation in PET/MR imaging. The method incorporates region-context embeddings to model anatomical heterogeneity across body regions and introduces PET-guided lesion-aware mechanisms to improve translation quality in pathological areas.

  • 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 practically relevant problem, namely whole-body MRI translation in PET/MR, which is more challenging than region-specific translation due to large anatomical variability across the body. The method is designed to account for both regional heterogeneity and lesion-related information, and the overall framework is clearly motivated from the clinical and technical challenges described in the paper.

  • 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.
    • The reliability of the automatically generated body-location and organ labels is not sufficiently validated. The paper states that Google Gemini 3 Pro is used to generate these labels and that their reliability has been empirically verified, but no supporting experiments or detailed analysis are provided. This is a fundamental issue because the usefulness of the proposed conditioning depends directly on label quality.
    • It is unclear whether the frozen PubMedBERT encoder is suitable for encoding fine-grained anatomical labels in MRI. The paper does not provide evidence that PubMedBERT can adequately represent subtle organ-level distinctions in this imaging setting.
    • The rationale for fusing coarse-grained body-location information with fine-grained organ information through multi-head cross-attention is insufficiently justified. The paper does not explain why both are needed or whether this design is more effective than using either source alone. No ablation is provided to support the necessity of this component.
    • Table 1 reports per-region quantitative results on the test set, but statistical significance is not provided, making it difficult to judge whether the reported improvements are robust.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    I recommend rejection because several key design choices are not sufficiently validated or justified. In particular,

    • the paper does not provide convincing evidence for the reliability of the automatically generated anatomical labels, even though these labels are central to the proposed conditioning strategy.
    • The use of a frozen PubMedBERT encoder for fine-grained MRI-related anatomical semantics is also not adequately supported.
    • the motivation and effectiveness of fusing body-location and organ information are not clearly demonstrated.
    • the lack of statistical significance analysis weakens confidence in the reported quantitative improvements
  • Reviewer confidence

    Very confident (4)

  • [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



Author Feedback

We appreciate the Reviewers’ recognition of the novelty and importance of our work. We fully address the valuable comments from Meta-reviewer and the three Reviewers as below. Please note that the source code link was provided in Implementation Details of our paper [R2].

Response to Meta-Reviewer / [R1-R3] C1.Low-uptake lesions [R1] We clarify that our network does not assume high-uptake lesions in PET. Rather, through harnessing both PET uptake characteristics and MR anatomical context, our design accounts for uptake variability in clinical PET/MR data. As described in Section 2.3, our noise modulation is jointly conditioned on PET uptake information and the anatomical context c_reg, enabling an adaptive perturbation process that better incorporates heterogeneous uptake patterns within specific anatomical contexts.

C2.VLM labels reliability & model label dependency [R1,R3] We used the strategies including professional validation, VLM output constraints and technical design to measure and ensure the label reliability and robustness of model label dependency. Reliability: —- We clarify that ‘the empirical verification’ on label reliability was performed by professional validations on the random 110 slice samples. The reliability of the body-location was justified by about 96% accuracy. For organ labels, to ensure reliability, we constrained the VLM outputs to a predefined set of 52 major MRI-visible organs.

Dependency & Robustness: —- Technically, our model explicitly accounts for label uncertainty to reduce its dependence on organ-label quality. As shown in Eq. 5, organ labels are incorporated via cross-attention as a residual correction to the reliable body-location embedding, rather than as a dominant conditioning signal. The framework is therefore more robust to variations in organ-label quality. —- The region context embedding serves as a soft anatomical anchor, allowing the bridge to learn more fine-grained transport paths without altering the direct correspondence between the two distributions modeled by the Schrödinger Bridge. Therefore, the model is robust to variations in label quality.

C3.Distribution shifts [R1] Clinical whole-body PET/MR differs from fixed-region protocols and inherently involves distribution shifts caused by multi-bed acquisition and patient-level protocol variability. Our data thereby reflect realistic clinical variability. We agree the multi-center data will strengthen validation, and we strive to collect multi-center data for public access since to our best knowledge, no public whole-body PET/MR dataset is currently available.

C4.Hyperparameter bounds [R1] The bounds were introduced to avoid degenerate near-zero or excessively amplified noise strength. In our experiments, the network was not sensitive to the bound values

Response to remaining comments: Source Code [R2]: We actually put the code link in the implementation details. PubMedBERT encoding [R3]: The pretrained PubMedBERT is widely used in fine-grained biomedical NLP encoding. Our framework also has learnable adaptors after the frozen PubMedBERT (Fig. 1(d)) to form c_reg. We would like to emphasize that fundamentally different from image-text alignment that requires the accurate understanding of anatomical semantics, in our model, the organ labels mainly serve as discrete anatomical anchors to distinguish intra-region contexts, allowing the translation bridge to learn different region-specific bridge paths. Statistical Significance [R3]: We appreciate the suggestion. Our rationale was that the experimental comparisons consistently demonstrated the outperformance achieved by our method with approx. 5%–10% gains in SSIM and 3–4 dB gains in PSNR. Since these substantial improvements of our method were well experimentally validated, the statistical significance test that is often most informative for the observed improvement being marginal, was not reported. We will add statistical analysis if permitted.




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.

    Mixed reviews 2 weak accept and 1 reject. All reviewers suggested that the paper is well motivated and addresses an important question.

    Key issues to address in rebuttal include: (1) assumptions and clinical cases, particularly the reliance on hypermetabolic PET signals and lack of validation on low-uptake lesions (R1); (2) validation of key components, including the reliability of VLM-generated labels and justification of the conditioning design (R1, R3); (3) limited robustness and generalization, due to single-center data and lack of evaluation under distribution shifts (R1); (4) clarity on key modules and hyperparameters (R1, R2, R3).

    The authors should carefully address the reviewers’ comments and provide clarification on important methods and results questions.

  • 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.

    Reviewers 1 and 2 recommend accept. The work presents a principled, mathematically sound methodological contribution by leveraging a Diffusion Schrödinger Bridge for PET-guided whole-body MRI translation.

    While Reviewer 3 suggested reject initially. They have not provided post-rebuttal comments. Based on their initial reviewers, major concerns on design choices whereas it seems to be resovled from other reviewers comments.



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.

    The proposed HA-DSB turns PET into whole-body MRI with a Schrodinger-bridge diffusion model, using PET uptake and region labels to keep lesions sharp instead of washed out, which is innovative and reasonable.

    Of the four points for reject, two ask for work that MICCAI rules forbid at rebuttal (a label-fusion ablation and a significance test on Table 1, they are important though!), and two are fair (the Gemini-generated anatomy labels are unvalidated, and a frozen PubMedBERT may miss fine organ detail) but are already answered by the design and rebuttal. Specifically, organ labels enter Eq 5 only as a small correction to the more reliable body-location embedding, so label noise does not break the model, and R1 raised the same worry and judged it resolved.

    Accept as a plain poster, not a highlight, since it is single-center on one scanner, carries no significance test. PLEASE note leans on a camera-ready promise for the label evidence, otherwise I don’t see it is acceptable to MICCAI.



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.

    The paper presents a Heterogeneity-Adaptive Diffusion Schrödinger Bridge (HA-DSB) framework, incorporating region context embeddings through VLM and lesion-aware metabolic prior from PET through a dual-stage guidance mechanism. The paper is technically solid and solves a clinically relevant problem. Through rebuttal, the authors further clarified key points regarding the label dependency, reliability, encoder usage, hyperparameter, etc. I would like to recommend for acceptance



back to top