List of Papers Browse by Subject Areas Author List
Abstract
NA
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1196_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/nature0012300/SAFE-Diff.git
Link to the Dataset(s)
LiTS dataset: https://competitions.codalab.org/competitions/17094
KiTS19 dataset: https://github.com/neheller/kits19
BibTex
@InProceedings{Sneha_SAFEDiff_MICCAI2026,
author = { Sneha AND Gupta, Nivedita AND Paul, Angshuman},
title = { { SAFE-Diff: Structurally Anchored Diffusion for Anatomically Faithful CT Image Super-Resolution } },
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
This paper presents SAFE-Diff, a hybrid framework for CT image super-resolution that aims to improve perceptual quality while preserving anatomical fidelity. The method combines two complementary stages. In the first stage, a deterministic residual reconstruction network predicts a coarse high-resolution image focused on recovering global structure and stable anatomical content. In the second stage, a conditional diffusion refinement module enhances local textures and fine details using truncated DDIM sampling with only a small number of denoising steps, improving efficiency compared with standard diffusion approaches.
A further contribution is the use of stationary wavelet transform SWT fusion, where low-frequency structural information from the first stage is combined with high-frequency detail generated in the second stage. This design explicitly separates anatomical consistency from texture enhancement and aims to reduce the risk of unrealistic hallucinated details.
The paper evaluates the method on public abdominal CT datasets LiTS and KiTS, using image quality metrics such as PSNR, SSIM, LPIPS, and FID. Additional experiments include cross-dataset generalization and downstream segmentation performance, demonstrating that the generated super-resolved images may also benefit subsequent analysis tasks. The work contributes both a practical methodological framework and an application study in clinically relevant CT enhancement.
- 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) CT super-resolution is important for improving visualization and potentially enabling downstream tasks when only low-resolution acquisitions are available. 2) The paper addresses a known tradeoff in image super-resolution: deterministic CNN-based methods often optimize fidelity but may produce overly smooth outputs, whereas diffusion models can generate sharper details but may be slower and less reliable. Combining these two paradigms is a sensible and practically useful direction. 3) The use of truncated DDIM sampling with only a few denoising steps is attractive and improves practicality over conventional diffusion pipelines with many iterative steps. 4) The wavelet-based fusion strategy is intuitive and interpretable. Preserving low-frequency anatomical structure while using diffusion for high-frequency detail is a reasonable way to improve realism while maintaining consistency. 5) The experiments cover multiple datasets, several full-reference and perceptual metrics, cross-dataset testing, ablation studies, and downstream segmentation. This breadth strengthens the empirical contribution. 6) Improvements in LPIPS and FID suggest that the method generates visually more realistic textures than several baselines. 7) The focus on anatomically faithful enhancement rather than purely perceptual sharpness is especially relevant in medical imaging settings.
- 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.(MAJOR) While the framework is well engineered, many individual components are established ideas, residual reconstruction, diffusion refinement, DDIM acceleration, wavelet fusion, transformer blocks. The novelty lies more in integration than in a fundamentally new algorithmic principle. 2.(MAJOR) CT is inherently volumetric, yet experiments appear based on 2D slices. This may ignore inter-slice continuity and volumetric consistency, which are important in clinical workflows and 3D analysis tasks. 3.(MAJOR) A key motivation is reducing anatomically implausible details, but the paper does not provide a dedicated hallucination analysis, reader study, uncertainty analysis, or region-level error assessment to directly support this claim. 4.(MAJOR) Some ablation results suggest that removing the wavelet fusion module may improve certain metrics. This weakens the claim that the fusion strategy is consistently beneficial and deserves deeper discussion. 5.(MINOR) Although downstream segmentation is useful, no radiologist assessment, lesion measurement study, or task-specific diagnostic evaluation is included. 6.(MINOR) The paper reports quantitative metrics, but statistical significance testing or confidence intervals would strengthen claims of superiority. 7.(MINOR) It would help to explicitly state whether all baselines were retrained under identical preprocessing, splits, and optimization 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 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.
(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 relevant problem and presents a practical hybrid framework that combines deterministic reconstruction with efficient diffusion refinement. The experimental section is reasonably comprehensive and shows competitive or strong performance, particularly on perceptual metrics and downstream tasks. These are meaningful strengths.
However, the methodological novelty is moderate, since the approach primarily combines known components in a sensible architecture rather than introducing a fundamentally new technique. In addition, important claims regarding anatomical faithfulness and hallucination reduction are not directly validated, and the 2D slice-based setting limits clinical realism for CT data.
Overall, I view this as a useful contribution that is slightly above the acceptance threshold, especially if the rebuttal satisfactorily addresses questions regarding ablation behavior.
- 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
The paper proposes a lightweight, two-stage framework for CT super-resolution. It addresses the “hallucination” and “structural drift” issues of pure diffusion models by using a deterministic Stage 1 (Residual Prediction Net) to anchor anatomy and a truncated Stage 2 (Diffusion Refiner) for perceptual sharpness. A key novelty is the integration of Stationary Wavelet Transform (SWT) to fuse low-frequency anatomical stability from Stage 1 with high-frequency details from Stage 2.
- 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.Structural Fidelity: The “Structurally Anchored” design effectively mitigates the common risk of hallucinations in generative models, which is critical for medical imaging. 2.Efficiency: By utilizing a truncated diffusion trajectory (starting from t=98 and using only 2 DDIM steps), the model achieves an inference time of 236ms, making it practical. 3.Methodological Innovation: The use of 2-level SWT for frequency-domain fusion is a clever way to bypass the pixel-level trade-off between blurriness and noise, ensuring that anatomical sub-bands are preserved. 4.Evaluation: The 46.9% improvement in LPIPS over SOTA is impressive, and the downstream liver segmentation validation (Table 5) provides strong evidence of “anatomical faithfulness.”
- 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.Ablation of t_{inf}: While the authors state t_{inf} = 98 is used, a more detailed ablation study on the sensitivity of structural fidelity vs. perceptual quality across different truncation points would strengthen the paper. 2.Dataset Diversity: While LiTS and KiTS are standard, they are both abdominal CT. Testing on different modalities or body parts (e.g., Lung or Neuro) would better demonstrate generalization. 3.Anonymity Breach: The inclusion of a direct GitHub link in the Abstract violates the double-blind review guidelines of MICCAI 2026.Authors are expected to use anonymized links (e.g., via Anonymous GitHub) to preserve their identity during the review process.
- 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.
(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?
The paper addresses a core problem in medical SR: the trade-off between perception and distortion. The solution is technically sound, extremely efficient for a diffusion-based method, and validated with both low-level metrics and high-level clinical tasks (segmentation). It aligns with MICCAI’s interest in “anatomically faithful” synthesis.
- 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 presents SAFE-Diff, a two-stage framework for CT image super-resolution that aims to resolve the trade-off between structural fidelity and perceptual sharpness. The primary contribution is a novel architecture that uses a deterministic Residual Prediction Net in Stage 1 to recover anatomical foundations, followed by a truncated diffusion refiner in Stage 2 for texture enhancement. A final Stationary Wavelet Transform (SWT/ISWT) fusion step is used to integrate the stable low-frequency structures from the first stage with the high-frequency details from the second, ensuring anatomical consistency.
- 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.
- Unlike standard diffusion models that start from pure noise, this work innovatively anchors the generative process using a structural residual map. This effectively mitigates common hallucination risks in medical imaging by keeping the diffusion process constrained to a truncated trajectory.
- The use of a 2-level SWT to fuse specific sub-bands from different stages is a clever way to preserve anatomical ground truth (low frequencies) while allowing the diffusion model to do what it does best: generating sharp textures (high frequencies).
- The model is efficient, utilizing only ~22.65M parameters and achieving an inference time of 236ms per slice, which is significantly faster than most of the other baselines.
- A particularly strong aspect of the evaluation is the use of nnU-Net v2 for liver and kidney segmentation. Showing that the super-resolved images yield near-identical Dice scores to the ground truth images (0.96 vs 0.97) provides strong evidence of clinical reliability.
- 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.
- While the paper compares against several diffusion models, it would benefit from citing and comparing against very recent sampling-efficient diffusion models that used similar truncation or distillation strategies to ensure the claimed 46.9% LPIPS improvement is benchmarked against the latest “fast” diffusion architectures. There are many sampling-efficient diffusion models published in recent years.
- While the paper mentions t_max=300 for training and t_inf=98 for inference, more intuitive explanation is needed regarding how these specific values were determined. Is there a sensitivity analysis showing how t_inf impacts the hallucination-vs-sharpness trade-off?
- In Table 3, the jump in FID when adding Stage 2 and SWT is clear, but the PSNR actually drops slightly compared to Stage 1 alone. The authors should clarify if this slight loss in pixel-wise accuracy could impact diagnostic measurements, even if the segmentation masks remain stable.
- The paper mentions the use of “db2” for the SWT. It would be helpful to know if other wavelet bases were tested and why db2 was selected as the optimal choice for CT anatomy.
- 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.
(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?
Challenging task and innovative solution, but missing some comparisons and clarifications.
- 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 thank all reviewers (R1-R3, Meta) for their feedback. We highly appreciate their valuable comments. We address the concerns below.
1.On Novelty (Meta-Q1, R1-Q1): Our core contribution is a principled “Structurally Anchored” design, which balances the perception-distortion trade-off. We use a deterministic Stage 1 network to capture the underlying anatomy and guide the generative process. A diffusion module then acts purely as a refiner to add sharp details, ensuring the final image is perceptually sharp. Finally, the SWT fusion strategy explicitly grounds and preserves low-frequency anatomical truth in the final image. We will further clarify these in Sections 1 and 2 of the revised manuscript.
2.On SWT fusion (Meta-Q2, R1-Q4): SWT fusion provides the most balanced LPIPS-PSNR results with almost the same LPIPS and a marginal improvement in PSNR. However, we agree that FID degrades slightly and that performance gains are not consistent across all metrics in every experiment. We will add this discussion to the revised Section 3.
3.Use of 2D CT data (R1-Q2): We acknowledge that inter-slice continuity and volumetric consistency is vital for clinical workflows. This study focused on exploring the diffusion as a refiner module and mitigating the perception-distortion trade-off. We will extend this work for 3D CT in the future, and it is mentioned in the conclusion.
4.On experimental settings (R1-Q7): All baseline models were trained by us using the exact same settings, including identical preprocessing pipelines and data splits. This will be mentioned in the revised Section 3.
5.On hallucination/error analysis, radiologist or diagnostic assessment, and statistical test (R1-Q3, Q5, Q6), dataset diversity (R2-Q2): We appreciate these valuable suggestions. As per the rebuttal guidelines, we do not include these results in this manuscript. However, in our future work, we will look into each suggestion.
6.Explanation and sensitivity analysis for t_max and t_inf (R2-Q1, R3-Q2): We determined t_max=300 and t_inf=98 through hyperparameter tuning using validation performance. We will clarify this in the revised Section 3.We do not include new results on sensitivity analysis as per the rebuttal guidelines.
7.Anonymity Breach (R2-Q3): We wish to clarify that the GitHub repository is fully anonymized, resulting in no breach of anonymity.
8.On comparisons (R3-Q1): We will compare with newer diffusion models in our future work. We refrain from reporting any new results in accordance with the rebuttal guidelines.
9.PSNR drop vs. diagnostic impact clarification (R3-Q3): The slight PSNR drop reflects the perception-distortion trade-off. Evaluating its direct diagnostic impact is an important suggestion that we will take up in our future work. We will mention this in the revised Section 3. 10.Justification for selecting the “db2” wavelet (R3-Q4): Standard DWT downsamples images, which causes blocky artifacts. We use the shift-invariant SWT with the db2 filter, a well-known baseline for decomposition. We will explore modality-specific wavelets for CT and MRI in our future work. This will be mentioned in the revised Section 4.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 novelty is based on integrating individual components instead of fundamentally new algorithms.
- The ablation study could be improved and requires further discussion.