Abstract

SPECT imaging faces persistent challenges from soft-tissue attenuation artifacts in clinical practice. While CT-based correction remains the clinical reference standard, associated radiation risks and infrastructure requirements limit its widespread adoption. To address this, we propose a Meta-Learning-Driven CT Morphology Disentangled Diffusion Model (MetaMorph-Diff), which achieves CT-independent attenuation correction. First, we design a Morphological Structure-Attentive Fusion module that explicitly guides the diffusion process using CT-derived anatomical priors. During training, its Morpho-Attentive Alignment submodule establishes voxel-level physical constraints between SPECT features and attenuation distributions by leveraging CT anatomical priors. During inference, its Morpho-Disentangling Gate achieves complete disentangling from CT dependencies through learned morphological embeddings. Crucially, the model uses only SPECT images during inference to achieve accurate attenuation correction without relying on CT data. Second, we propose a multi-region adaptive meta-learning strategy, which enhances cross-anatomical generalization capability by optimizing model initialization parameters, enabling a single model to achieve consistent and accurate correction across diverse anatomical regions. Our method surpasses existing approaches with higher-precision attenuation distribution prediction and stronger multi-region correction adaptability. The code is available at https://github.com/yhr1020/MetaMorph-Diff.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2259_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yhr1020/MetaMorph-Diff

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YanHao_MetaLearningDriven_MICCAI2025,
        author = { Yang, Haoran and Fan, Jiansong and Li, Lihua and Pan, Xiang},
        title = { { Meta-Learning-Driven CT Morphology Disentangled Diffusion Model for Multi-Region SPECT Attenuation Correction } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15972},
        month = {September},
        page = {349 -- 358}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces MetaMorph-Diff, a novel framework for attenuation correction (AC) in SPECT imaging without the use of CT during inference. The framework employs a meta-learning-driven diffusion model to adapt across multiple anatomical regions and features a Morphological Structure-Attentive Fusion (MSAF) module. During training, CT images are used to inform the model through Morpho-Attentive Alignment (MA), while at inference, Morpho-Disentangling Gate (MDG) allows the model to function CT-free. The work demonstrates state-of-the-art performance in attenuation map prediction for brain, thyroid, and myocardial SPECT across multiple k-shot settings.

  • 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. The model effectively merges a conditional diffusion process (based on SR3) with MAML (Model-Agnostic Meta-Learning) to achieve generalizable CT-free AC across multiple anatomical regions.
    2. The paper includes comprehensive ablation and comparison studies against baselines like CycleGAN, VQ-I2I, and Pix2Pix, where MetaMorph-Diff outperforms in terms of SSIM, PSNR, MAE and ED across 3 anatomies.
    3. By combining meta-learning (MAML) with diffusion modeling, the paper effectively enables few-shot generalization across anatomies.
    4. The MSAF module with MA and MDG cleverly uses CT features during training and then discards CT at inference. The transition from dual-modal (CT+SPECT) to single-modal (SPECT-only) inference is technically sound and well-motivated.
  • 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 model is trained and tested on a private in-house dataset from a single center. This undermines generalizability.
    2. Although Fig. 1 illustrates the training and inference pipeline, there’s minimal discussion on how anatomical consistency is preserved without CT during inference.
    3. There is no analysis of where or when the model fails, which is crucial in medical imaging tasks.
    4. The combination of MS-SSIM and L2 is empirically validated (Table 2), but lacks theoretical or medical grounding. Why not use perceptual losses or region-aware penalties (e.g., Dice loss for ROI)? Justifying and exploring other loss designs could enhance robustness.
    5. Statistical significance tests are missing.
  • 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.

  • 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/2025/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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    Please refer to the weaknesses.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 all of my concerns, specifically mentioned why a private in-house dataset was used for this paper.



Review #2

  • Please describe the contribution of the paper

    The authors propose a method to perform attenuation correction in SPECT without the need for CT, which mitigates radiation risks. They utilize morphological structure attentive fusion in a dual stage model; during training the CT and SPECT features are aligned, and inference uses only SPECT with learned embeddings. Multi-region adaptive meta-learning is used to improve generalization across diverse anatomical regions using k-shot learning, and a hybrid loss function gives better structural fidelity and voxel-level accuracy.

  • 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.

    They show an innovative use of diffusion models that tailor the model with CT-guided learning but CT-independent inference. The meta-learning in SPECT improves performance in a multi-region context, and voxel-level alignment employs attention-based voxel-wise alignment between the CT and SPECT features. The authors compare their model performance to several typical GAN, which their model outperforms in all 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.

    The dataset is limited coming from only Chinese hospitals to which external access is restricted. The data homogeneity also induces bias in the results, which would be accounted for by testing on a more diverse population. Attenuation corrections are the most problematic when dealing with populations with high BMI or significant chest mass, which is unlikely to be well-represented in the chosen cohort. A comparison of the results for small-breasted vs large-breasted women for example is very important. While the focus on brain, thyroid, and heart provides a good first step, there are many SPECT use cases that are left out, such as bone, renal, and gastrointestinal scans. There is also no discussion of the computing requirements to implement this into a routine workflow. The Phillips Precedence SPECT-CT is a very old system, so it would be interesting to see how the algorithm performs on a modern SPECT.

  • 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 provide sufficient information for reproducibility.

  • 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/2025/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 — could be accepted, dependent on rebuttal

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

    The routine looks sound from a theoretical standpoint and is well formulated. I have concerns about the uniformity of the dataset and lack of accountability for bias.

  • 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



Review #3

  • Please describe the contribution of the paper

    The paper introduces a Meta-Learning-Driven CT Morphology Disentangled Diffusion Model that enables CT-independent attenuation correction using a multi-region adaptive meta-learning strategy and a novel hybrid loss function.

  • 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) The paper utilizes meta-learning and diffusion to achieve consistent and accurate correction using only SPECT across various anatomical regions. 2) It introduces novel components such as Morphological Structure-Attentive Fusion and an Adaptive Meta-learning Strategy with inner- and outer-loop tasks, demonstrating strong technical innovation. 3) It provides a generalized approach that applies to multi-region SPECT attenuation, setting it apart from existing methods. 4) The paper includes robust state-of-the-art method comparisons and comprehensive ablation studies to validate its performance.

  • 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) Figure 1 is overly complex and lacks detailed explanations for the individual modules. 2) The methodology section is insufficiently detailed, particularly in describing the Multi-Region Adaptive Meta-Learning module, including unclear weight passing and the specifics of backpropagation updates. 3) Although a hybrid loss function is proposed by combining various loss components, there is no comparative table or ablation study that details the characteristics of each loss and their contributions when combined.

  • 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

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

    1) The paper demonstrates strong technical novelty by integrating meta-learning and diffusion approaches with a Multi-Region Adaptive Meta-Learning strategy. 2) The evaluation is thorough, including comprehensive comparisons with state-of-the-art methods and detailed ablation studies, which substantiate the method’s effectiveness. 3) However, the overall readability is diminished by insufficient detail in figure descriptions and manuscript clarity, which impacts the presentation of the research.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    The problem and limitations of existing methods were clearly identified, and the idea and proposed approach are excellent. However, the explanation mainly focuses on the roles of modules and simple equations, resulting in insufficient explanation in the figures. The authors claim to have designed a hybrid loss function, explaining the advantages of each component and stating that the balancing parameter alpha was tuned experimentally. However, the exact value of alpha is not reported, and detailed analysis is missing in the ablation study. They should provide a comparison table among different loss functions to prove and analyze the claimed advantages of their proposed loss. Currently, only a simple comparison with MSE is presented. Although reference [22] provides theoretical support for the MS-SSIM+L1 combination, this work uses L2 instead of L1. Therefore, if comparisons are made, it would be more appropriate to compare with the L1 combination rather than only MSE. The manuscript mainly lists advantages but fails to provide essential proof and analysis due to space limitations, which undermines the basic structure of the paper. Most of the results focus on comparisons with other methods, but a detailed analysis of the model structure is necessary to make the claimed advantages truly convincing.




Author Feedback

Thank you to all reviewers (R1&R2&R3) for your careful review and valuable comments. We take your suggestions seriously and respond as follows: R1: Lack of statistical significance testing. A:Thanks for pointing this out. We performed statistical tests using IBM SPSS 27.0. One-way ANOVA showed that all metrics had P-values < 0.01, indicating significant differences between models (P < 0.05 considered significant). We will include these results in the revision. R1&R2: Limitations of the dataset. A:Thanks for your review.The dataset used in this study is constrained by the absence of publicly available SPECT AC datasets and ongoing challenges related to privacy and ethics in cross-institutional data sharing in China. To mitigate single-center and data homogeneity issues, we used MAML.During data collection, we also ensured diversity in sex, age, body/disease types. Currently, we focus on three commonly imaged anatomical regions to establish an initial algorithmic framework. As stated in Section 4, we plan to expand to additional tasks and actively promote multi-center collaboration. R1&R3: Hybrid loss function. A:Thanks for your interest in our loss function design. The combination of MS-SSIM+L1 used in Ref.22 provides the theoretical basis for our loss function. In transferring this concept to medical imaging, we found that L2 performed better than L1 for our task. L2 emphasizes global pixel intensity consistency, which is critical in medical images where even subtle grayscale variations can impact both clinical interpretation and quantitative analysis.Regarding perceptual loss (R1), We experimented with a wide range of combinations of mainstream loss functions. Perceptual loss stresses visual similarity but overlooks numerical details,making it less suitable for tasks requiring voxel-level precision.Our method achieves high voxel consistency, as demonstrated by Fig.2. For the concern about missing ablation studies and comparison tables (R3), we noted in Section 2.3 that we did explore various loss function combinations and weighting strategies. Due to space limitations, we could not include all intermediate results. R1&R3: Anatomical consistency without CT, lack of failure case analysis, and insufficient method details. A:Thank you for your helpful comments. Due to space constraints, we were unable to elaborate on these aspects. Regarding anatomical consistency during inference without CT (R1), we briefly explained in Section 2.1 that the MDG leverages SPECT-derived features and learns anatomical priors through meta-learning.Concerning model failures (R1), Section 3.2 mentions that our model shows superior accuracy in both ROIs and overall regions compared to other methods. However, we will expand the analysis of failure cases in the revised manuscript.Regarding the lack of detail in the Multi-Region Adaptive Meta-Learning module (R3), we will provide a more thorough description of its update mechanism in the revision. R2: Lack of discussion on computational requirements for clinical workflow integration. A:Our model can process multiple slices in seconds on modern GPUs (e.g., NVIDIA RTX 3090), with resource demands within clinically acceptable limits. We will add performance evaluation details in the revision. R2: Use of the outdated Philips system. A:Thanks for the concern regarding older systems. However, the Philips Precedence system remains widely used in China. While newer SPECT systems may include deep learning enhancements, Philips has exited the nuclear medicine market,leaving older models without support.Our method provides a “virtual upgrade” path to enhance image quality on such systems, which is particularly valuable for extending their clinical utility in resource-limited regions. R3: Fig.1 is too complex. A:Due to space limitations, we attempted to present the complete framework-including the core model and its MAML strategy—in a single figure. we will improve the layout to enhance readability in the revision.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Though one “positive” reviewer changed his/her mind about the final decision as he/she think the ablation study is needed. I believe the study itself is novel and should be interested to the community.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top