List of Papers Browse by Subject Areas Author List
Abstract
Synthesizing PET from MRI offers a radiation-free route to functional biomarkers but remains fundamentally ill-posed due to the weak coupling between anatomy and metabolism. Existing medical generative models trained on limited molecular imaging data often overfit or generate artifacts, while natural-image foundation models cannot be directly transferred because their priors and semantics are misaligned with radiological domains. We introduce PETFlux, a task-aligned adaptation framework that repurposes a pretrained natural-image backbone for the specific demands of MRI-to-PET synthesis. The framework addresses the key barriers of domain mismatch, structural-functional ambiguity, and spatial over-smoothing through three synergistic strategies: (i) Low-Rank Adaptation (LoRA) via rectified flow matching to realign natural priors with PET distributions; (ii) LLM-augmented prefix tuning grounded in an evidence-based clinical knowledge base to inject patient-specific metabolic semantics; and (iii) a log-scale focal frequency loss to preserve clinically critical functional detail. Experiments on the ADNI dataset demonstrate that PETFlux achieves state-of-the-art synthesis fidelity and yields substantial gains in downstream diagnostic classification, highlighting a principled pathway for repurposing natural foundation models in molecular imaging by effectively integrating anatomical constraints and patient metadata.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0655_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
Link to the Dataset(s)
BibTex
@InProceedings{YinYua_PETFlux_MICCAI2026,
author = { Yin, Yuan AND Xiong, Honglin AND Huang, Haolin AND Li, Yonghao AND Kong, Yan AND Zhu, Yitao AND Jiang, Zhuoxin AND Sun, Kaicong AND Shen, Dinggang AND Wang, Qian},
title = { { PETFlux: Taming Natural Foundation Models for Accurate MRI-to-PET 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
The paper proposed a framework for MRI-to-PET synthesis that adapts a large pretrained natural-image foundation model to a clinically important but highly ill-posed medical imaging problem. Relative to prior MRI-to-PET work that mainly relies on GANs, Transformers, state-space models, or standard diffusion models trained on limited medical data, the paper proposes a more principled way to transfer foundation-model priors into this domain. It shows how to combine foundation models, parameter-efficient fine-tuning, and clinically grounded metadata to improve MRI-to-PET synthesis.
- 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.
This paper has several major strengths. First, it tackles a clinically important problem, since MRI-to-PET synthesis could reduce radiation exposure, cost, and limited access associated with PET while still providing useful functional information for diseases such as Alzheimer’s. Second, the paper presents a clever adaptation strategy that repurposes multiple pretrained foundation models for this downstream medical imaging task, instead of training a large model from scratch on limited paired data. Third, the use of LoRA-based fine-tuning is practically appealing, as it enables effective adaptation of a large backbone with much lower computational cost. Fourth, the proposed conditioning mechanism is particularly interesting: the paper transforms clinical metadata into semantic guidance that the generative model can better understand, which is a novel and well-motivated way to address the ambiguity of MRI-to-PET synthesis. Finally, the overall evaluation is strong, including multiple baselines, two tracers, ablation studies, and downstream diagnostic utility, which makes the empirical results convincing.
- 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.
1.2D Slice-based Limitation: The method operates on 2D slices, which fundamentally limits its ability to capture volumetric context. Given the paper’s own emphasis on the ill-posed nature of MRI-to-PET synthesis and its dependence on complex structural-functional relationships, this is a critical weakness. It is unclear how the model can faithfully synthesize global metabolic patterns (e. g. , distributed across the whole brain) when it only processes information slice-by-slice, potentially losing vital inter-slice dependencies.
2.Lack of Technical Detail and Validation for the LLM/Clinical Knowledge Base: The manuscript attributes the synthesis success to LLM-guided clinical descriptions derived from a knowledge base, yet provides insufficient detail on its curation, validation, and handling of conflicting guidelines. This lack of transparency severely hinders reproducibility. Furthermore, the integration of explicit, clinically informative metadata (e. g. , MMSE, CDR-SB) as narrative text makes it difficult to disentangle the contribution of true image translation from the injection of strong disease priors. A robustness analysis against noisy or missing metadata is required to substantiate the claims.
3.Insufficient Framework Generalization and Versatility Validation: While proposing a novel adaptation framework for foundation models, the evaluation is confined to a single dataset (ADNI). Validation on external datasets with different protocols is essential to demonstrate real-world generalizability. Additionally, the ablation studies do not fully isolate the contribution of the proposed adaptation strategy itself. Applying the complete pipeline to another foundation model of comparable scale would more fairly evaluate the framework’s core versatility, separating it from the benefits of the specific pre-trained priors used.
4.Incomplete Clinical Validation and Proof of Equivalence: Although downstream classification results are promising, a noticeable performance gap remains between synthetic and real PET, failing to establish clinical equivalence. The current quantitative metrics (SUVR, AUC) lack depth. A more comprehensive clinical validation is needed, including qualitative expert assessment (e. g. , blinded scoring of image quality and diagnostic confidence) to rigorously verify the claimed resolution of “clinically misleading textures” and prove the clinical utility of the generated images.
- 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.
(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 overall score reflects several major concerns that collectively impact the paper’s technical rigor, generalizability, and demonstrated clinical utility. The primary factors leading to this assessment are as follows: 1.Lack of Technical Detail and Validation for the LLM/Clinical Knowledge Base: The manuscript attributes the synthesis success to LLM-guided clinical descriptions derived from a curated knowledge base. However, insufficient detail is provided regarding the construction, validation, and handling of potentially conflicting guidelines within this base. This lack of transparency severely hinders reproducibility. More critically, the integration of explicit, clinically informative metadata (e.g., MMSE, CDR-SB) as narrative text makes it difficult to disentangle the contribution of genuine image translation from the model’s reliance on injected, strong disease priors. A robustness analysis against noisy or missing metadata is required to substantiate the core claim that the model learns the true MRI-to-PET mapping. 2.Insufficient Framework Generalization and Versatility Validation: While proposing a novel adaptation framework for foundation models, the evaluation is confined to a single dataset (ADNI). Validation on external datasets with different imaging protocols is essential to demonstrate real-world generalizability. Furthermore, the presented ablation studies do not fully isolate the contribution of the proposed adaptation strategy itself. Applying the complete pipeline to another foundation model of comparable scale would provide a fairer evaluation of the framework’s core versatility, separating its benefits from those inherent to the specific pre-trained priors of the model used in this study. 3.Incomplete Clinical Validation and Proof of Equivalence: Although downstream classification results are promising, a noticeable performance gap persists between synthetic and real PET images, failing to establish clinical equivalence. The current reliance on quantitative metrics like SUVR and AUC lacks depth. A more comprehensive clinical validation is needed, incorporating qualitative expert assessment (e.g., blinded scoring of image quality and diagnostic confidence). This is necessary to rigorously verify the claimed resolution of “clinically misleading textures” and to ultimately prove the clinical utility of the generated images.
- 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 #2
- Please describe the contribution of the paper
The paper proposes PETFlux, a framework for MRI-to-PET synthesis that adapts a pretrained natural-image foundation model to medical imaging. The method combines LoRA-based rectified flow adaptation, LLM-guided conditioning from clinical metadata, and a spectral loss to improve fine metabolic detail. Experiments on ADNI with both AV45 and FDG PET show better synthesis metrics and stronger downstream classification performance than the compared baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper addresses MRI-to-PET synthesis problem with a timely foundation-model adaptation approach.
- Method is technically interesting and combines the main components in a coherent way.
- Evaluation includes both image synthesis metrics and downstream diagnostic utility.
- The method is tested on two tracers, and the gains appear consistent across both AV45 and FDG.
- 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.
- My main concern is that the method combines several ideas at once, so the true source of improvement is still a bit hard to isolate.
- The ablation showed only trend plots, and it does not fully separate how much comes from LoRA adaptation, the LLM-based conditioning, and the spectral loss.
- I also think the paper needs stronger statistical support. The results in Table 1 are promising, but only single values are reported. There are no confidence intervals, repeated runs, or significance tests, so it is hard to know whether the improvements are stable and clearly meaningful.
- Only one dataset ADNI was evaluated.
- It suggests to compare a metadata-only conditioning baseline without LLM-generated narrative text to show that this component is necessary beyond simpler metadata conditioning.
- 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
The paper is interesting and generally well written. To make it stronger, I suggest few improvements:
- Make the novelty claim more precise in the introduction and discussion, and explain clearly what is new beyond combining existing components.
- Improve the result section by adding uncertainty information such as standard deviation, confidence intervals, or significance testing.
- Increase fontsize for numbers in Figure 3.
- 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?
I think this paper is slightly above the acceptance threshold. The problem is important, the method is technically solid, and the empirical results are consistently strong across two PET tracers. My main concerns are that the novelty is more in the overall combination than in a single clearly new method, and the paper does not provide statistical testing or uncertainty estimates to show how reliable the improvements are.
- 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 most of my main concerns. I think the paper is acceptable and would be suitable as a highlighted poster.
Review #3
- Please describe the contribution of the paper
The main contribution of this paper is to perform MRI-PET tasks using a tuned Flux backbone. Two other innovations are the introduction of LLM as a condition and a frequency domain loss.
- 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 charts are clear and easy to understand.
2.I greatly appreciate that the authors provided ACC/AUC metrics in addition to PSNR/SSIM, which enhances the persuasiveness of the experiments.
3.The ablation experiments are excellent.
- 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.I would like to see a comparison of training and inference costs (GPU memory and time). I completely understand that the authors don’t need to worry that their current method is inferior to pix2pix, etc., and I would appreciate it if they could report some realistic metrics. It’s normal for methods based on Flux-like generative models to have such issues.
2.In fact, if PETFlux works in the latent space, I think it would be more reasonable for the comparison methods to also be trained in the latent space. If the authors could provide training results for 1-2 comparison methods in the latent space (if time is limited, providing only the pix2pix method would suffice, and the comparison results could be named pix2pix-latent), I would appreciate it. In my experience, VAE compression will result in some loss of accuracy, and PSNR/SSIM may decrease. However, this is not absolute, and I would prefer to see the authors’ practical application on this task.
3.In Figure 1, the encoder and decoder are generally VAEs, right? I understand that FLUX has strong VAE capabilities, but if the authors fine-tune the VAE on in-domain images, is it possible to further improve the metrics? Two additional experiments can be added here: (1) Changes in SSIM/PSNR after fine-tuning VAE (input x, decoded x after VAE) (2) Changes in SSIM/PSNR of generation task (x->y) after fine-tuning VAE.
4.Regarding the introduction of LLM. I feel it is very similar to some recent works [1][2]. Could you please add a discussion about them in the final version? Also, how would the effect change if the LLM embedding were replaced with the embedding extracted by sigLIP/DINO as the condition? This is the experimental addition I most want to see (sigLIP may be better aligned with the language; I am more concerned about the performance of DINO. If time permits, I suggest providing the effects of both DINOV2 and DINOV3.If resources are limited, only DINOV3 can be provided).
5.Other minor issues: (1) Line break issues in 3.1.Dataset. has no line break, while LLM Prompting, etc., have line breaks. (2) Many paragraphs have orphaned lines, which is not aesthetically pleasing. I suggest correcting this in the final version.
Ref: [1] Shi J, Li C, Gong T, et al. CoD-MIL: chain-of-diagnosis prompting multiple instance learning for whole slide image classification[J]. IEEE Transactions on Medical Imaging, 2024, 44(3): 1218-1229.[2] Yang S, Zhou H, Yang Y, et al. LCM-Net: LLM-driven Cross-Modality MoE Feature Fusion Network for Cancer Survival Analysis[J]. IEEE Transactions on Medical Imaging, 2026.
- 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
Considering the rebuttal limitations of MICCAI, the authors could prioritize answering question 4; further answers to questions 1-3 would be even better. Based on the rebuttal, I would consider further increasing the score from 4 to 5.
- 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 motivation is quite good, and the charts are clear and easy to understand. It would be even better if my questions were answered.
- 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.
After reading the rebuttals, I recommend accepting the paper. The authors have reasonably addressed my main concerns, resolving most issues, including computational cost, potential spatial baseline settings, and VAE tuning. The reported efficiency data clarifies the trade-offs of PETFlux, and the explanations regarding pix2pix-latent and Flux VAE tuning are acceptable. I believe the discussion of computational and time costs is sincere; FLUX-like networks have faced similar issues in my own research, and efficiency optimization could be an important direction for future research, especially considering potential clinical applications.
In conclusion, this paper demonstrates a meaningful application, clear experimental results, robust ablation studies, and practical clinical indicators. I encourage the authors to further refine the final version, discuss the relevant LLM conditionalization work I mentioned in the original comment, and add the corresponding references. Therefore, I support the acceptance of this paper.
Author Feedback
We thank all reviewers for their constructive feedback and for recognizing the clinical importance (R1/R2/R3), technical soundness (R1/R2), downstream diagnostic utility (R1/R2/R3), and clear presentation(R2/R3) of our work. We address the main concerns below.
LLM/Clinical Knowledge Base Details (R1Q2) We clarify that the LLM is not used to infer subject-specific diagnoses or PET findings. Our knowledge base follows fixed, auditable rules: MMSE/CDR-SB are mapped to coarse severity levels for each tracer. The LLM only rewrites these rule-selected profiles into standardized visual descriptions. The prompts contain no subject labels or PET data, ensuring no information leakage. We will release the mapping rules, dictionaries, prompt templates, construction scripts, and LLM query scripts. To address shortcut concerns, we evaluate missing/noisy metadata at inference: generic instruction with zero prefix (missing) and mismatched prompts/metadata (noisy). PSNR drops from 26.43/27.37 (full) to 25.01/26.99 and 25.48/27.08 (AV45/FDG), showing degradation but no collapse. Metadata remains useful as auxiliary guidance, and metadata-prefix-only is worse than full PETFlux.
2D Limitation (R1Q1) We acknowledge this limitation. PETFlux uses slice-wise generation because Flux is native 2D, while full-resolution 3D generation remains memory-demanding and often relies on memory-saving approximations. Nevertheless, our evaluation is not limited to slice-level metrics: we reconstruct PET volumes via regional SUVR error and downstream 3D diagnostic AUC, which assess whether clinically relevant metabolic information is preserved at the volume level. 3D adaptation will be discussed as future work.
Generalization/clinical validation (R1Q3/R2Q4) We agree that external cohorts and blinded expert assessment are needed before clinical translation. PETFlux improves synthetic PET quality and diagnostic utility, but remains a research-stage framework, not a replacement for real PET. ADNI is a large, longitudinal, multi-center benchmark (e.g., ResDM IJCAI 2024; DIReCT IPMI 2025, no clinical validation). Evaluation on 2,600+ AV45/FDG pairs supports first-stage validation. External validation, protocol-shift analysis, and blinded reader studies are important next steps.
Ablation Study (R2Q1/Q2/Q5/R1Q2) Our ablation study is component-wise, not incremental. Each module—baseline, Log-FFT, LLM text, metadata prefix, and full PETFlux—is evaluated independently to isolate its contribution to performance. The caption of Fig. 3 will be revised to clarify this and avoid confusion.
Statistical reliability (R2Q3) Table 1 reports mean values due to space constraints, but we agree that uncertainty is important. Representative PSNR improvements are statistically significant: AV45 PSNR 26.43 with 95% CI [25.82, 27.03], p = 5.60 × 10^-13; FDG PSNR 27.37 with 95% CI [26.87, 27.68], p = 2.51 × 10^-10.
References and formatting (R3Q4/Q5) We thank the reviewer for the helpful references and will include cited works and fix formatting. Alternative visual embeddings (DINO/SigLIP) will be discussed as future comparison; performance with DINOV2/3 is a valuable extension.
Computational cost (R3Q1) PETFlux is more expensive than lightweight GANs, but LoRA updates only 2.45% of parameters. At batch size 1, inference requires 36.86GB/4.87 s/slice; Pix2pix is 0.31GB/0.0052s/slice; Pix2pix-turbo is 6.42GB/0.3751s/slice.
Latent-space baseline (R3Q2) Pix2pix is a lightweight image-space translation model; forcing it into the same VAE latent space mainly tests VAE compression rather than standard pix2pix. Our table already includes latent generative baselines, SD v1.5 and pix2pix-turbo, while pix2pix-latent is a useful future control.
VAE tuning (R3Q3) Our preliminary VAE-tuning trials did not improve synthesis quality, likely because modifying the pretrained Flux VAE shifts the latent manifold expected by the DiT backbone and weakens LoRA adaptation.
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 reviewers appreciate the timely and principled adaptation of natural-image foundation models to the clinically vital task of MRI-to-PET synthesis. Strengths include the clever use of LoRA for parameter-efficient fine-tuning, the consistent performance gains across multiple tracers (AV45 and FDG), and the inclusion of downstream diagnostic metrics (ACC/AUC) to validate clinical utility. However, the reviewers raised several concerns that must be addressed in the rebuttal. Primary among these is the lack of transparency and validation for the LLM-guided conditioning; authors need to clarify how clinical metadata is curated. Reviewers also highlighted the limitations of the 2D slice-based approach in capturing 3D metabolic patterns and requested a robustness analysis against missing or noisy metadata to ensure the model isn’t simply memorizing disease priors. Furthermore, the rebuttal could include a comparison of computational costs (GPU memory/time), and a discussion on how the framework generalizes to external datasets beyond ADNI.
- 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 were satistfied with the rebuttal. The authors should address the remaining concerns in the camera-ready version.
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.
R1 is the only reviewer to not respond, and had the most concerns about the manuscript. The authors make it clear that there was significant minunderstanding on the contribution of the LLM module in the proposed method. R1 major points 2 and 3 are similar critiques on the limitations on dataset validation and reliance on only quantitative measures without the inclusion of clinically relevant metrics. Both of these are not addressed, however there are similar methods in the literature that rely on the quantitative metrics with only ADNI as a dataset. R2 and R3 are happy to accept and state their questions have mostly been addressed.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors did sufficiently address the concerns of the reviewers in the rebuttal. The work represents a clinically important contribution that is technically sound and clearly presented with robust ablation studies.