List of Papers Browse by Subject Areas Author List
Abstract
Multi-modal medical image synthesis is pivotal for alleviating
clinical data scarcity, yet existing methods fail to reconcile
global anatomical consistency with high-fidelity local detail.
We propose FermatSyn, which addresses three persistent
limitations:
(1) SAM2-based Prior Encoder that injects domain-aware
anatomical knowledge via LoRA$^{+}$ efficient fine-tuning of a
frozen SAM2 Vision Transformer;
(2) Hierarchical Residual Downsampling Module (HRDM)
coupled with a Cross-scale Integration Network (CIN) that
preserves high-frequency lesion details and adaptively fuses
global–local representations; and
(3) continuity constrained Fermat Spiral Scanning
strategy within a Bidirectional Fermat Scan Mamba (BFS-Mamba),
constructing an approximately isotropic receptive field that
substantially reduces the directional bias of raster or spiral serialization.
Experiments on SynthRAD2023, BraTS2019, BraTS-MEN, and BraTS-MET
show FermatSyn surpasses state-of-the-art methods in PSNR, SSIM,
FID, and 3D structural consistency.
Downstream segmentation on synthesized images yields no significant
difference from real-image training ($p{>}0.05$), confirming
clinical utility. Code is available at \url{https://github.com/gatina-yone/FermatSyn}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3777_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YuaFen_FermatSyn_MICCAI2026,
author = { Yuan, Feng AND Gao, Yifan AND Li, Haoyue AND Gao, Xin},
title = { { FermatSyn: SAM2-Enhanced Bidirectional Mamba with Isotropic Spiral Scanning for Multi-Modal Medical Image Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes several improvements to the LSGAN framework with a Mamba backbone. Firstly, the encoder is a finetuned SAM2-VTE model with a parallel Convolutional encoder. Secondly, a new scanning strategy for Mamba is introduced to reduce visual artifacts. The experiments show good synthesis results.
- 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 motivation for adopting a new scanning strategy is compelling. The authors clearly demonstrate how existing scanning methods can introduce visual artifacts, and how the Fermat scan addresses these issues. The three identified gaps are also well explained.
The experimental evaluation spans a range of tasks, and the proposed method achieves strong performance across them.
The downstream clinical validation adds an important dimension, illustrating both the practical utility of the synthesized images and their level of realism.
- 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.
Some parts of the methodology, such as configurations and hyperparameters, would be more appropriately placed in the implementation details section.
The statement that “even- and odd-indexed channels in stacked convolutional feature maps tend to encode complementary low- and high-frequency statistics, respectively” would benefit from supporting evidence or a relevant reference.
Additionally, certain elements should be clarified further, including acronyms (e. g. , DWConv) and statistical notation (e. g. , p<0.05).
- 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 introduces a novel method with a clear motivation and extensive experiments. Therefore, I recommend acceptance.
- 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.
I recommend accepting, as the authors have addressed my concerns.
Review #2
- Please describe the contribution of the paper
The manuscript describes a framework that generates missing modalities with anatomical consistency and high-fidelity local details with SAM2-enriched intergrated encoders. Fermat Spiral Scanning strategy is proposed to construct near-isotropic receptive field and improve the volumetric fidelity.
- 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 experiments are well-designed with impressive qualitative results. Inference time, 3D consistency and downstream performance are reported, which evaluate the model’s superiority from different aspects.
- 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) Some crucial technical details are confusing or missing without appropriate citations, making it hard to reproduce the results. For example, the authors implement a HRDM-based encoder to enhance conventional pooling. However, they do not mention what is conventional pooling and why it should be used in medical image translation. The definition and details for some abbreviations, such as HRDM, the inherent SE attention, and DWConv, are also missing without any citations. The operator ‘||’ in Equation 3 and 9 is never defined. 2) Inconsistent and confusing notations acorss the methodology. The authors define F_path.1 and F_path.2 in Section 2.1.2 but never use it, which should be the components of F_fusion. z_fwd and z_bwd are used in Equation 8 for bidirectional SSM but are later replaced by O_fwd and O_bwd in the following Equation 9.Fig. 1 displays f^a and f^b, but I cannot find any definition for them throughout the manuscript. 3) Over-simplified analysis of experimental results. In Section 3.2, the authors report the qualitative results but never mention it. The inference time compared with the GAN baseline is not reported. 4) The framework utilized pretrained SAM, which provides strong structural priors by training on large-scale annotated data, to enhance anatomical consistency. The authors are encouraged to compare with other related literatures [1,2,3] that use the same priors. In addition, it would be good to compare with different pretrained encoders (i.e., those from medical domain segmentors), so that we can learn whether the benefits come from the overall framework design or the specific pretrained priors. 5) As the framework contains multiple components on top of SAM’s encoder, the total parameters and GPU memory consumption during inference should be reported in addition to the inference time.
[1] Pang, Haowen, et al. “Cascaded Diffusion Model and Segment Anything Model for Medical Image Synthesis via Uncertainty-Guided Prompt Generation.” International Conference on Information Processing in Medical Imaging. Cham: Springer Nature Switzerland, 2025.[2] Huo, Jiayu, Sébastien Ourselin, and Rachel Sparks. “SAM-I2I: unleash the power of segment anything model for medical image translation.” 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI). IEEE, 2025.[3] Song, Zhiyun, et al. “Uni-COAL: A unified framework for cross-modality synthesis and super-resolution of MR images.” Expert Systems with Applications 270 (2025): 126241.
- 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 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 overall framework and experiments are well-designed. However, the methodology section is hard to follow and contains confusing/missing abbreviations and notations. Comparisons with other SAM-based methods are missing.
- 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.
Most of my concerns are expected to be addressed if the manuscript is updated accordingly.
Review #3
- Please describe the contribution of the paper
This paper proposes FermatSyn, a novel framework for multi-modal medical image synthesis that integrates anatomical priors, multi-scale feature modeling, and improved sequence modeling. Specifically, the method introduces a SAM2-based prior encoder to incorporate structural knowledge, a hybrid encoder to preserve both global context and local details, and a bidirectional Mamba architecture with a Fermat spiral scanning strategy to reduce directional bias in feature serialization. Experiments show FermatSyn surpasses state-of-the-art methods in PSNR, SSIM, FID, effectively improving both anatomical consistency and local fidelity in synthesized 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 introduces a Fermat spiral-based serialization scheme for Mamba, which is a meaningful technical contribution. By achieving more isotropic spatial coverage compared to raster or rectangular-spiral scans, the method effectively reduces directional bias in sequence modeling. This idea is both intuitively appealing and empirically validated, and represents a non-trivial improvement over existing Mamba-based approaches. In addition, the incorporation of a SAM2-based prior encoder provides useful anatomical guidance, which helps improve structural consistency across modalities. The overall framework is supported by relatively comprehensive experiments across multiple datasets and tasks, demonstrating consistent performance gains over strong baselines.
- 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.Limited novelty beyond the scanning strategy While the Fermat spiral scanning is novel and interesting, other components (SAM-based prior encoder, multi-scale feature fusion, GAN-based training) are largely incremental extensions of existing ideas. For example, prior works have already explored SAM-based guidance for medical image translation (e.g., SAM-I2I), and multi-scale feature fusion is a common design in medical image synthesis. 2.Overly complex architecture with multiple components The proposed framework combines several modules (SAM2+LoRA/HRDM/CIN/BFS-Mamba), making it difficult to isolate the true source of performance gains. Although ablation studies are provided, the overall design appears somewhat engineering-heavy and may reduce interpretability. 3.Computational complexity and efficiency not fully analyzed Although inference time is reported, the paper does not provide a detailed analysis of computational cost (e.g., FLOPs, memory usage, scalability). The model size (87M parameters) is relatively large, and it is unclear how it compares to baselines in terms of efficiency. 4.Insufficient theoretical analysis of the scanning mechanism The advantage of the Fermat spiral is mainly demonstrated empirically (e.g., reduced variance in operator footprint and improved PSNR). However, the paper lacks deeper theoretical analysis of how isotropic sampling affects state space modeling or feature representation in Mamba. 5.Reproducibility concerns due to lack of code availability The paper states that the code will be released upon publication, but no code or implementation details are currently available. This limits immediate reproducibility and makes it difficult to fully verify the proposed method.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a meaningful technical improvement with solid empirical validation. While the overall framework includes several incremental components, the proposed scanning strategy is interesting and contributes to performance gains. I find the work above the acceptance threshold, although not strongly so.
- 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 rebuttal addressed several of my concerns regarding computational analysis, contribution attribution, and the role of the proposed components. In particular, the authors clarified that the gains are not solely attributed to the SAM2-based prior encoder and provided additional discussion and ablation results supporting the contribution of the proposed Fermat spiral scanning strategy. The added efficiency-related discussion and planned profiling details also partially addressed my concerns regarding computational complexity. Some concerns still remain. Beyond the proposed scanning strategy, several components of the framework are still relatively incremental extensions of existing ideas, and the overall architecture remains somewhat engineering-heavy due to the combination of multiple modules. In addition, the theoretical analysis of the proposed scanning mechanism is still limited, with the advantages of Fermat scanning mainly supported empirically rather than theoretically. The code is also not publicly available during review. Nevertheless, I believe the proposed scanning strategy is technically meaningful, and the paper is supported by solid experimental validation and consistent improvements over strong baselines. Therefore, I maintain my original weak accept recommendation.
Author Feedback
We warmly thank the reviewers for their constructive comments and for noting the motivation, artifact-reduction rationale, and technical contribution of Fermat spiral scanning (R2,R4), as well as the broad experimental evaluation and downstream validation (R2,R3,R4). We will incorporate the following clarifications into the updated version: 1.Methodological and Notational Clarity (R2&R3): We will define HRDM, CIN, DWConv, SE attention, and “||” as channel-wise concatenation, move configurations/hyperparameters to the Implementation section, and unify notation: F_path.1/F_path.2 as components of F_fusion; z_fwd/z_bwd as serialized inputs; O_fwd/O_bwd as Mamba outputs; and f^a/f^b in Fig.1 as the two encoder streams before CIN fusion. 2.Even/Odd Channel Statement and Statistics (R2&R3): We will rephrase the even/odd channel statement as a channel-splitting heuristic for complementary feature grouping, inspired by channel-split/frequency-aware designs, rather than an inherent property of convolutional channels. We will also clarify that p<0.05 denotes Wilcoxon signed-rank testing. 3.Complexity and Efficiency (R3&R4): Parameters/backbone: FermatSyn has 87.3M trainable parameters; SAM2-VTE is frozen, and only LoRA+, HRDM, CIN, and BFS-Mamba are updated. Runtime: Inference is 31 ms per 256×256 slice on one RTX4090, compared with 24 ms for I2I-Mamba and 148 ms for SA-LDM under the same hardware. Additional profiling: We will add FLOPs and peak GPU memory, clarify the profiling protocol, and provide same-hardware inference-time comparisons for baselines with available implementations where feasible. 4.Contribution Attribution and Relation to SAM-based Translation Works (R3&R4): Responding to R3’s request for comparison with SAM-guided translation methods and R4’s concern about novelty attribution, we will better position FermatSyn relative to related works such as SAM-I2I and discuss the choice of SAM2-VTE relative to alternative pretrained encoders. We will clarify that SAM2-VTE is used as an anatomical-prior encoder rather than claimed as the sole novelty. The framework addresses three complementary bottlenecks: anatomical-prior injection, local-detail preservation, and directional-bias reduction. Submitted ablations disentangle them: SAM2-VTE improves PSNR from 23.67 to 27.35 on SynthRAD and from 24.35 to 27.18 on BraTS; HRDM+CIN further improves it to 27.74/28.45; Fermat scanning improves over rectangular spiral from 27.38 to 28.67 and from 28.13 to 29.53 under the same backbone. Thus, the gains are not attributed to a single component. 5.Fermat-Scanning Rationale (R3&R4): Fermat scanning provides a geometric rationale for more isotropic radial/angular coverage than raster or rectangular-spiral scans, reducing directional anisotropy in Mamba serialization. We will avoid overstating this as a complete theory of SSM dynamics and instead present it as geometric motivation supported by operator-footprint analysis and controlled ablations: Fig.2 shows lower footprint variation, and Table 5 shows consistent gains under the same backbone. 6.Result Discussion and Figure Clarity (R3): We will strengthen Sec.3.2 by discussing Fig.3, including tumor-core ROIs and ventricular-wall details, and connect qualitative observations to PSNR/SSIM/FID, 3D consistency, and downstream segmentation. 7.Reproducibility and Code (R3&R4): The revision will include the above definitions, notation fixes, hyperparameters, profiling details, and clearer result discussion. As stated in the submission, code will be released upon publication while preserving anonymity during review. We would like to express our gratitude to the reviewers for their valuable suggestions!
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a medical image synthesis framework integrating a SAM2-based prior encoder with a Mamba architecture using a novel Fermat spiral scanning strategy.
While reviewers appreciate the Fermat scanning approach for reducing directional bias, the architecture is highly complex (combining SAM2, LoRA, HRDM, CIN, and BFS-Mamba). The paper is currently not reproducible. Moreover, complexity obscures the true source of performance gains and raises concerns about computational overhead. Finally, the manuscript suffers from notation inconsistencies and missing definitions.
Points to address in the rebuttal: 1.Methodological & Notational Clarity (R2, R3): Resolve the numerous inconsistencies in the text pointed out by the reviewers. Provide a citation/justification for the claim that even/odd-indexed convolutional channels encode low/high-frequency statistics (R2).
2.Complexity and Efficiency (R3, R4): The framework is heavy (87M parameters). Provide a detailed computational analysis (FLOPs, GPU memory consumption, and inference time) compared against your baselines.
3.SAM Priors vs. Fermat Gains (R3, R4): Using SAM for image translation has been already explored. Using your existing ablations/data, clarify whether the performance gains are primarily driven by the pretrained SAM2 prior or the novel Fermat scanning strategy. Discuss how this compares conceptually to other recent SAM-based translation models (e.g., SAM-I2I).
4.Given the complexity of the approach, releasing the code would be greatly appreciated.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All reviewers recommended acceptance after the rebuttal.
The authors are however required to incorporate the following changes in the camera-ready version:
1) resolve all notational inconsistencies, define all undefined abbreviations (HRDM, CIN, DWConv, SE attention, “||”), and unify notation across figures and equations.
2) rephrase the even/odd channel statement as a channel-splitting heuristic and provide appropriate supporting references.
3) report a complete computational analysis including FLOPs, peak GPU memory, and per-slice inference time compared against baselines under the same hardware.
4) better position the framework relative to existing SAM-based translation methods such as SAM-I2I.
5) release their code to ensure reproducibility.
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 reviewers unanimously recommend acceptance after rebuttal, noting a meaningful technical contribution with demonstrations of consistent improvements across multiple medical image synthesis tasks. Although some concerns remain regarding presentation clarity, architectural complexity, and limited novelty beyond the scanning mechanism, these issues are not fundamental and can be addressed in the final version.
A major issue I notice is the use of the wrong font throughout the document. The formatting must be aligned with the MICCAI Springer-based template.
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 careful consideration of the reviews and the authors’ rebuttal, I recommend accepting this paper for presentation at MICCAI. The proposed Fermat spiral scanning strategy represents a significant advance in medical image synthesis by reducing directional bias in feature serialization, thereby improving volumetric fidelity and minimizing visual artifacts. The extensive experiments across multiple datasets demonstrate consistent performance gains over strong baselines, and the inclusion of clinical validation underscores its practical relevance. The authors effectively addressed all concerns and highlighted the strategy’s contribution to performance improvement. All reviewers agree that the paper meets the acceptance standards, and the authors will release the full source code upon publication to encourage further research.
