Abstract

Three-dimensional (3D) medical image enhancement, including denoising and super-resolution, is critical for clinical diagnosis in CT, PET, and MRI. Although diffusion models have shown remarkable success in 2D medical imaging, scaling them to high-resolution 3D volumes remains computationally prohibitive due to lengthy diffusion trajectories over high-dimensional volumetric data. We observe that in conditional enhancement, strong anatomical priors in the degraded input render dense noise schedules largely redundant. Leveraging this insight, we propose a sparse voxel-space diffusion framework that trains and samples on a compact set of uniformly subsampled timesteps. The network predicts clean data directly on the data manifold, supervised in velocity space for stable gradient scaling. A lightweight Structure-aware Trajectory Modulation (STM) module recalibrates time embeddings at each network block based on local anatomical content, enabling structure-adaptive denoising over the shared sparse schedule. Operating directly in voxel space, our framework preserves fine anatomical detail without lossy compression while achieving up to 10x training acceleration. Experiments on four datasets spanning CT, PET, and MRI demonstrate state-of-the-art performance on both denoising and super-resolution tasks. Our code is publicly available at: https://github.com/mirthAI/sparse-3d-diffusion.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5369_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/mirthAI/sparse-3d-diffusion

Link to the Dataset(s)

LDCT-and-Projection-data dataset: https://drive.google.com/drive/folders/17OGCnf__mke6U-Dp1Jv2SL7T5CRFoh9r?usp=drive_link FDG PET dataset: https://drive.google.com/drive/folders/1aKQ2BWK-JrahmK3Kz0cB4RfzPH388U5f?usp=drive_link AortaSeg24 dataset: https://drive.google.com/drive/folders/1q5HQBPMf8uex_0cL5Jp504zL1krMurJB?usp=drive_link UHB FCD Lesion Brain MRI dataset: https://drive.google.com/drive/folders/1spkcCS-7007F1R1_WoAhTFyZ3FNuH7CM?usp=drive_link

BibTex

@InProceedings{JiaHon_StructureAdaptive_MICCAI2026,
        author = { Jiang, Hongxu AND Li, Fei AND Yu, Boxiao AND Zhang, Ying AND Smith, Kaleb AND Gong, Kuang AND Shao, Wei},
        title = { { Structure-Adaptive Sparse Diffusion in Voxel Space for 3D Medical Image Enhancement } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16888},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    (a)Propose a sparse voxel-space diffusion framework for 3D medical image enhancement, combining x0​-prediction with velocity supervision and sparse timestep subsampling, which avoids information loss from latent compression and achieves up to 10× faster training. (b)Introduce Structure-aware Trajectory Modulation (STM), a lightweight module that uses tri-planar structural encoding and block-wise time embedding modulation to adapt the denoising trajectory according to local anatomical complexity. (c)Conduct extensive experiments on CT, PET, and MRI datasets for both denoising and super-resolution, achieving state-of-the-art performance with better fine anatomical detail preservation than existing 3D diffusion models.

  • 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.

    (a)The innovative sparse diffusion framework operates directly in voxel space (avoiding compressed-representation information loss) and adopts x_0-prediction with velocity supervision. It uses only 5 uniformly subsampled timesteps (vs. 1000 in traditional models) to achieve up to 10× training acceleration without sacrificing resolution or anatomical details, solving the high computational cost of existing 3D diffusion methods. (b)The lightweight Structure-aware Trajectory Modulation (STM) module adaptively adjusts denoising behavior via multi-structure encoding and block-wise time embedding modulation. It addresses anatomical heterogeneity of 3D medical volumes, enabling individualized denoising paths for different patches and balancing efficiency with fine detail preservation. (c)The method is evaluated on 4 public datasets (CT, PET, MRI) for denoising and super-resolution, compared with 4 representative 3D diffusion baselines. Quantitative (PSNR, SSIM, etc. ) and qualitative evaluations, plus ablation studies, confirm its state-of-the-art (SOTA) performance and clarify each component’s contribution.

  • 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.

    (a)The idea of using sparse timestep sampling + x0​-prediction for conditional diffusion has been explored in recent works such as Fast Diffusion (Zhen et al. , CVPR 2023), and Skip Diffusion. The paper does not sufficiently highlight the difference from these fast diffusion frameworks; the core sparsity strategy is largely similar. (b)The STM module uses multi-planar projection + affine modulation on time embedding, which is a standard design in medical vision:Multi-planar (axial/coronal/sagittal) encoding has been widely used in 3D medical CNNs and Transformers (e. g. , nnFormer, UNETR++). Modulating time embedding with spatial content is standard in adaptive diffusion and conditional diffusion models. Thus, STM is a practical combination rather than a fundamentally new mechanism. (c)The paper only compares with 3D diffusion models (DDPM, LDM, WDM), but does not compare with recent fast CNN/Transformer methods (e. g. , HFN, UCTransNet) or GAN-based methods optimized for medical enhancement. It is unclear whether diffusion is still necessary for speed vs. performance.

  • 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?

    This paper presents a practical voxel-space sparse diffusion framework for 3D medical image enhancement, with effective STM structure-adaptive modulation and strong experimental results on CT/PET/MRI datasets. The method achieves 10× faster training while preserving fine anatomical details. However, the core sparse diffusion and x0​-prediction strategy has been explored in existing works such as SkipDiff (AAAI 2024) and progressive distillation (ICLR 2022), showing limited novelty. The STM module is a reasonable combination of multi-planar encoding and time embedding modulation, which is standard in medical vision models (nnFormer, UNETR++). The paper lacks theoretical analysis and rigorous clinical validation. Overall, the work is technically solid and empirically effective but only incrementally novel.

  • 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 #2

  • Please describe the contribution of the paper

    This paper proposes a 3D medical image enhancement framework that performs sparse diffusion in voxel space. By integrating x₀ prediction, velocity supervision, sparse time-step scheduling, and structure-aware trajectory modulation (STM), it achieves approximately 10-fold training acceleration and favorable enhancement performance on denoising and super-resolution tasks for CT, PET, and MRI images.

  • 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 propose a sparse diffusion framework in 3D voxel space that combines x 0 -prediction with velocity supervision and timestep subsampling, achieving up to 10× faster training while improving enhancement quality.

  • 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 novelty overlaps with existing works. The conditional x₀ prediction using degraded input priors is basically the same idea as CoreDiff[1]; sparse timestep scheduling is essentially few-step diffusion, which DDIM[2] has already explored; velocity supervision is a weighted combination of noise and x₀ prediction, similar to what RDDM has done; STM resembles the adaLN-Zero modulation in DiT[3] combined with multi-scale convolutions. The paper presents these existing techniques as independent contributions but does not sufficiently discuss the relationship and differences with the above works. The novelty is questionable.

    2.The nature of sparse diffusion needs clarification. Both training and inference operate on only 5 timesteps, which is essentially no different from directly training a 5-step diffusion model. The paper describes it as “subsampling from a dense schedule”, but lacks theoretical analysis on why uniform sampling is optimal.

    3.The data scale is small and statistical reliability is questionable. All four datasets have small sample sizes (50, 88, 60, 120 cases), with only 10-24 cases for testing. For example, LDCT-and-Projection-data contains 299 patients but only 50 are used without explanation. The reported PSNR improvements (~0.5dB) on such small test sets lack statistical tests, and no standard deviations or confidence intervals are reported. I suggest using more available data or providing statistical significance analysis.

    4.The selection of comparison methods is not comprehensive enough. First, all four baselines are 3D versions of general diffusion frameworks, lacking comparison with non-diffusion methods (e.g., 3D U-Net, 3D SRGAN), making it hard to judge whether diffusion models are truly necessary for this task. Second, there is no comparison with recent methods specifically designed for 3D medical image restoration, or extending similar 2D models to 3D as baselines such as 3D CoreDiff.

    [1] Gao Q, et al. CoreDiff: Contextual error-modulated generalized diffusion model for low-dose CT denoising and generalization [2] Song J, Meng C, Ermon S. Denoising diffusion implicit models [3] Peebles W, Xie S. Scalable diffusion models with transformers

  • 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

    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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    This paper proposes a practical and efficient method for 3D medical image enhancement. The core observation is reasonable, experiments cover multiple modalities and tasks, and the training speedup is significant. The main weakness is that the methodological novelty needs further discussion, the data scale is small, and baselines are insufficient. If the rebuttal can provide relevant information, it would strengthen the contribution of this paper.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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’ response largely addressed my concerns. However, I still believe that methodologically, the main contribution of this paper is extending the idea of Fast-DDPM to 3D medical image acceleration; the remaining designs (e.g., STM) are drawn from prior work. My decision to revise toward acceptance is primarily based on the endorsements from the other two reviewers and the relatively comprehensive experimental validation.



Review #3

  • Please describe the contribution of the paper

    This paper describes an approach for enhancement of 3D medical images by employing adaptive sparse diffusion. What this does is an educated information of the time step at inference time, for a more efficient use of the denoiser. This enables faster inferences for volumetric data.

    3D is split into overlapping patches, preserving the dimensionality of the data (3D). The diffusion model denoises an image while taking as reference a degraded version of it. Both training and inference use only a subset of the diffusion steps. This is in contrast to conventional diffusion-based approaches where training employs the whole T diffusion steps, and it is inference where the reduction is perceived.

    More importantly, the adaptive time embedding parameter and the structurally aware module are an interesting solution that takes into account the noise content and the fine details in the image, having a dynamic adaptation of these parameters for better reconstructions/enhancements.

  • 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 major achievement of this paper is the speedy reconstruction of 3D medical images, whose processing tends to be prohibitely expensive in terms of computational time and resources.

    The method exploits the structural information present in the condition itself to determine the best timestep, critically this seems to be done for each block in the UNet architecture.

    The output of the diffusion model corresponds to an estimation of the clean image, instead of the noise content in the image to denoise. The authors claim this helps to preserve the structure

  • 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 datasets employed are of rather small sizes, which is more critical due to the dimensionality of their samples.
    • Figures are rather small and of low resolution, making it difficult to visually inspect the reconstructions of the methods.
    • The computational efficiency is in terms of training iterations, but I think it would be of equally or greater importance to show also the inference time required by each one of the methods, as it is this what would be crucial should the method be implemented in clinical settings.
  • 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.

  • 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 find the paper very interesting though I have a few minor questions, that perhaps the authors could clear:

    • Did the authors try to use encoding of axial view as well? Is one of the three views more important than the others? What drove them to use the encoded sagittal and coronal views but leave axial out?
    • Could the authors provide more details about the UNet architecture used for evaluations? Would a different architecture yield better results?
    • There is a typo (double “to”, page 2, paragraph 2, second line from the end of paragraph).
  • 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.

    (5) Accept — should be accepted, independent of rebuttal

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

    The paper is well written and clear in conveying the goal of the approach. Evaluations and ablations studies are good, however, I am not entierely sure the authors provide enough details for reproducibility of the paper, particularly regarding network architecture. Perhaps this element is not as important as the others, as the main contributions of the paper lie in the STM component.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    Authors provided clear answers to the concerns. Method is novel and of value to be shared with the community.



Author Feedback

1.Novelty and relation to prior work (R1&R2). We agree that sparse diffusion, x_0-prediction, and adaptive modulation individually relate to prior work. Our contribution lies in combining these components into an efficient framework for high-resolution 3D image enhancement. Compared with prior acceleration-oriented methods, our framework differs in three aspects: (1) direct operation in 3D voxel space rather than latent-space or 2D settings; (2) sparse training and inference under the same diffusion parameterization, yielding up to 10x training acceleration; and (3) anatomy-adaptive trajectory modulation within the shared sparse schedule. SkipDiff and PD mainly target inference acceleration via skipping/distillation, whereas our sparse schedule is applied directly during training. CoreDiff incorporates modality-specific degradation priors into the forward diffusion process, whereas we retain the standard Gaussian diffusion parameterization and inject anatomical priors through conditional voxel-space denoising without modality-specific re-derivation.

We agree that multi-planar encoding and modulation are established ideas, and STM is not intended as a fundamentally new conditioning mechanism. Instead, it adapts denoising trajectories to local anatomical complexity under extremely sparse schedules by modulating time embeddings rather than features alone. Together, these components enable efficient direct 3D voxel-space enhancement with only five diffusion steps.

2.Baselines and comparison scope (R1&R2). We focused on representative 3D diffusion baselines (latent, wavelet, DDPM, DDIM) because the primary contribution of this work is efficient sparse diffusion in voxel space. We agree broader comparisons are valuable and will add CNN/GAN baselines to the revision. Under identical LDCT settings, RED-CNN-3D and WGAN-VGG-3D achieved SSIM values of 0.72 and 0.73, respectively, below both the diffusion baselines (0.74-0.81) and our method (0.82). We will also expand discussion of restoration-specific baselines in the revision.

3.Clarification of the sparse diffusion formulation (R2). Our method performs sparse sampling from the original T=1000 diffusion schedule while retaining the same noise parameterization and timestep embeddings. Our goal is not to design a fundamentally new shallow diffusion process, but to exploit the empirical redundancy of neighboring denoising steps in conditional enhancement tasks with strong anatomical priors. Table 2 shows that sparse schedules consistently outperform dense schedules in our setting, with performance peaking at K=5 in our experiments. Uniform sampling was chosen as a simple prior-free strategy, while learned or non-uniform schedules remain interesting directions for future work.

4.Dataset scale and statistical reliability. (R2&R3). For LDCT-and-Projection-data, only 50 patients belong to the chest subset with consistent reconstruction kernels, while the remaining cases involve different anatomies or acquisition settings. The other datasets were used in full or near-full. Patch-based training with random cropping and augmentation yields approximately 10^4 effective samples, which increases training diversity despite the limited number of patients. We will additionally report mean ± std and paired Wilcoxon signed-rank tests (p<0.01) in the revision.

5.Additional clarifications and reproducibility (R3). We will improve figure resolution, fix the typo, add additional implementation details (UNet configuration, optimization settings, patch extraction, runtime, and GPU memory), and release code upon acceptance. Regarding axial projections, axial context is already captured by the volumetric 3D branch, while sagittal/coronal projections provide complementary through-plane structural cues for anisotropic medical volumes. Preliminary experiments showed that removing sagittal/coronal branches reduced PSNR by $0.2$ dB, whereas adding an additional axial branch provided only marginal gains.




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.
    • Relatively small data sizes.
    • Moderate technical novelty.
    • Require further clarification for the proposed method.
    • More comprehensive comparisons are required.
  • 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’ response partially addressed reviewers’ concerns. However, the paper’s main contribution requires further clarification.



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.

    Reviewers acknowledge that this work has merit, especially due to the experimental evaluation. While technical contribution is somewhat incremental, the rebuttal is convincing in claiming that combination of existing techniques to speed up reconstruction is meaningful. The work should be of interest to the community and overall reviewer assessment is very positive. I tend to weakly accept this work, but would expect authors to at least give an outlook of flow matching based methods in the manuscript, which have recently been successfully used for performing image restoration tasks in a computationally faster manner. Moreover, in case of acceptance a tidy up of the references is necessary, as there are several works cited in their arxiv version, although peer reviewed versions of the papers exist.



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.

    After reviewing the rebuttal and the reviewers’ updated comments, I believe the major concerns have been adequately addressed. While Reviewer 1 notes that the methodological novelty is somewhat incremental, the paper provides comprehensive experimental validation and demonstrates the effectiveness of the proposed approach. Reviewer 2 is satisfied with the authors’ responses and considers the work valuable to the community. Overall, I believe the paper meets the bar for acceptance and support acceptance.



back to top