List of Papers Browse by Subject Areas Author List
Abstract
Positron Emission Tomography (PET) image reconstruction is inherently challenged by Poisson noise and physical degradation factors, which are further exacerbated in limited-angle acquisitions. While deep learning methods demonstrate promising performance, their generalization to unseen clinical data distributions remains limited without extensive retraining. We propose PET-Adapter, a test-time domain adaptation framework for generative PET reconstruction models pretrained solely on phantom data. Our method enables adaptation to clinical datasets with varying anatomies, tracers, and scanner configurations without requiring paired ground truth. PET-Adapter introduces layer-wise low-rank anatomical conditioning during adaptation and Ordered Subset Expectation Maximization-based warm-starting that initializes the generation from physics-informed reconstructions, reducing diffusion steps from 50 to 2 without compromising quality. Experiments across multiple clinical datasets demonstrate superior 3D reconstruction performance in both full-angle and limited-angle settings, highlighting the clinical feasibility and computational efficiency of the proposed approach. The code base for PET-Adapter will be made publicly available under https://github.com/ruveydayilmaz0/PetAdapter.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1581_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ruveydayilmaz0/PetAdapter
Link to the Dataset(s)
BrainWeb: https://zenodo.org/records/4897350
CERMEP: Data link available upon request. Paper link: https://pmc.ncbi.nlm.nih.gov/articles/PMC8446124/
AGIEF: https://openneuro.org/datasets/ds004054/versions/1.0.0
NX: https://openneuro.org/datasets/ds006917/versions/1.0.2
Huntington: https://www.repository.cam.ac.uk/collections/b2883f29-45f5-4574-9198-c0af38636fb4
BibTex
@InProceedings{YilRüv_PETAdapter_MICCAI2026,
author = { Yilmaz, Rüveyda AND Wu, Yuli AND Stegmaier, Johannes AND Schulz, Volkmar},
title = { { PET-Adapter: Test-Time Domain Adaptation for Full and Limited-Angle PET Image Reconstruction } },
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
The paper proposes a test-time domain adaptation framework for diffusion-based PET reconstruction, enabling models trained on phantom data to generalize to clinical datasets. It incorporates MR-guided conditioning through layer-wise modulation to introduce anatomical priors, and employs an OSEM-based warm-start strategy to significantly reduce diffusion sampling steps while preserving reconstruction quality.
- 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 key strength of this work is the comprehensive comparison with multiple state-of-the-art methods, including both classical reconstruction approaches (e. g. , OSEM, Bowsher) and recent diffusion-based techniques. The experimental evaluation spans several datasets with varying domain shifts, which strengthens the empirical validation. In addition, the proposed framework is relatively simple to train, as it relies on a pretrained model on phantom data and performs lightweight adaptation at test time without requiring paired ground truth data. This makes the approach practically appealing.
- 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 overall methodological novelty is limited, as the proposed framework mainly combines existing components (diffusion models, low-rank adaptation, and MR-guided conditioning) in an incremental manner rather than introducing fundamentally new techniques. Moreover, the effectiveness of the warm-starting (WS) strategy is not fully convincing. As shown in the ablation study, the configuration without WS but with more sampling steps (T = 50) achieves comparable or even better performance than the proposed WS with T = 2 . This raises questions about whether WS truly improves reconstruction quality or mainly serves as a computational trade-off.
- 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
1.Clarification of the benefit of warm-starting (WS): The ablation results suggest that WS does not outperform the configuration without WS when sufficient diffusion steps are used (e. g. , T = 50). The authors should clarify whether WS improves reconstruction quality or only reduces computational cost, and provide a more balanced discussion. 2.Fairness of comparison with baseline methods: The proposed method leverages both test-time adaptation and warm-start initialization, while baseline methods appear to use standard settings. This may lead to an unfair comparison. It would be helpful to include stronger baselines (e. g. , with similar initialization or adaptation) for a more rigorous evaluation. 3.Evaluation protocol and ground truth definition: For some clinical datasets, the ground truth is derived from MRI-based regional averaging rather than true PET measurements. This may bias the evaluation toward anatomically smooth reconstructions. The authors should discuss the limitations of this evaluation setup and its impact on the reported metrics. 4.Lack of qualitative evaluation in limited-angle (LA) setting: The paper does not provide qualitative reconstruction results for the LA experiments. Without visual inspection, it is difficult to assess whether the method introduces artifacts, such as MR-induced over-smoothing or sensitivity-related artifacts due to missing angular coverage. Including qualitative comparisons for LA cases would significantly strengthen the evaluation.
- 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 addresses a relevant and practical problem in PET image reconstruction and provides a well-engineered framework with strong empirical results across multiple datasets. Although the methodological contribution is somewhat incremental and certain design choices (e.g., warm-starting) are not fully justified, the work demonstrates reasonable technical quality and potential practical impact.
- 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 authors propose a new test-time fine-tuning method for diffusion model-based PET image reconstruction on out-of-distribution data. The core is low-rank MR-guided adaptation to enable fine-tuning and injection of MR prior. The authors also propose OSEM warm-starting to substantially shorten the sampling path of a DDPM. The idea of test-time domain adaptation for PET reconstruction, especially from phantom-only pretraining to diverse clinical datasets, is novel and practically relevant.
- 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) Test-time fine-tuning based on LoRA and the injection of MR image prior. Since PET data have significant variation across different datasets, test-time fine-tuning for diffusion samplers is a potential solution for OOD data. This topic remains novel in PET. The authors also propose the LoRA module that injects MR image as the prior, which is a flexible way of incorporating prior information. 2) Warm-start is a feasible solution for acceleration. Typical diffusion samplers usually suffer from high computation cost. Another benefit, as the authors claim, is the feasibility of applying a pure 2D score function on the warm-started image without losing slice consistency across axial slices. This opens a new path for real 3D reconstruction under the memory limit. 3) The experiment design is bold. Although this is a general framework for OOD reconstruction, the authors train the score function on phantom images and apply it to simulated datasets, where there is a large domain mismatch between the training set and the testing data. The results demonstrate the strong adaptation capability of the proposed method.
- 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) Unclear fine-tuning objective. I am not clear about \hat{x}^MAP at the beginning of p5.Is it Tweedie’s estimate \hat{x}_0? I suggest the fine-tuning objective be formulated more clearly.
2) This paper assumes that MR image is available; however, MR image prior is not guaranteed for many PET datasets. Could the authors comment on the method’s performance using plain LoRA without MR as the prior?
3) Only reconstruction on simulated data has been reported, which could be sufficient for a conference paper. Information on the simulation is not complete, such as the simulated scanner.
4) It is not very clear how the slice consistency is achieved. The author essentially claim a new way of circumventing direct 3D reconstruction. Is it true that the slice consistency is realized solely by a warm start and without further constraint in the sampling?
5) Potential unfairness when compared with baselines. Are the baselines implemented with the warm start? Could authors comment on the portion of the performance gain that is contributed to the warm start?
6) I suggest this work be referred to as well: Hashimoto et al., PET Image Reconstruction Using Deep Diffusion Image Prior, TMI, 2026
- 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.
(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 topic and the proposed method (LoRA fine-tuning) are relatively new in PET. The preliminary results are promising.
- 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 proposes PET-Adapter, an unsupervised test-time domain adaptation framework intended to reduce the generalization gap between generative PET reconstruction models trained on synthetic phantom data and unseen clinical data. The main contribution is a multi-component adaptation pipeline that combines low-rank anatomical conditioning based on MRI, an unsupervised adaptation objective that includes both measurement fidelity and anatomical regularization, and a physics-informed warm-start strategy based on OSEM reconstruction to reduce the computational cost of diffusion sampling.
- 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 major strength of the work is the warm-starting strategy. Instead of initializing diffusion sampling from pure Gaussian noise, the method starts from a perturbed OSEM reconstruction. This is a sensible design choice, since OSEM can already recover the low-frequency image structure efficiently, allowing the diffusion model to focus on refinement. In practice, this appears to reduce the number of sampling steps substantially, from 50 to 2, without an obvious loss in reconstruction quality.
Another strength is the use of layer-wise low-rank adaptation blocks to condition a pretrained unconditional score model on MR information at test time. This provides a practical way to inject anatomical guidance without retraining the full model. The experimental evaluation is also fairly broad, with comparisons across multiple datasets and methods.
- 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 aspects of the experimental setup are unclear. For example, it is not clear where the NX data come from, and the definition of the anatomical loss La should be stated more explicitly.
Although the experiments use clinical image volumes, the PET measurements y are still simulated using the Parallelproj framework. A stronger clinical validation would require raw scanner-acquired sinogram data. Even with realistic physics and noise modeling, simulated data may not fully capture scanner-specific effects such as detector-dependent artifacts, calibration imperfections, or more complex scatter behavior, which are precisely the kinds of domain shifts that test-time adaptation is meant to address.
In addition, while the overall integration is effective, the individual components are not fundamentally new. The conditioning mechanism based on low-rank adaptation and the broader idea of steerable conditional diffusion for test-time adaptation in medical imaging have both been explored in closely related settings.
Finally, while the paper includes ablations on warm-starting and TTDA, it does not separately ablate the anatomical guidance term La from the measurement fidelity term Lp. That makes it difficult to assess the specific contribution of the anatomical prior to the final performance.
- 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.
(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?
My overall score was mainly driven by the paper’s practical contribution and clear presentation. The paper is well written, and the proposed framework is easy to follow. Together with the fairly broad experimental evaluation, these strengths made the paper a positive submission overall.
- 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
Author Feedback
We thank all reviewers for their valuable and constructive feedback. We provide clarifications below regarding the main points raised in the reviews. We do not include additional experimental results, as new experiments cannot be added at this stage.
Effectiveness of warm-starting (WS): The main benefit of WS is computational efficiency rather than a direct improvement in reconstruction quality. As shown in Table 2, PET-Adapter with WS and T=2 diffusion steps achieves reconstruction quality comparable to the no-WS setting with T=50, while substantially reducing the number of sampling steps.
Fairness of comparison with baseline methods: To clarify the contribution of WS and TTDA, we included ablations of our method without WS and without TTDA in Table 2.The “w/o WS, T=50” setting corresponds most closely to the sampling regime used by the other methods, which also use T=50 diffusion steps in our experiments. We did not modify the baseline non-adaptive methods to include our adaptation mechanism, since doing so would change the original methods. Our goal was to compare against these methods as established non-adaptive baselines. That said, PET-Adapter could potentially be combined with other diffusion-based PET reconstruction methods to make them adaptable.
MRI-based regional averaging: For CERMEP and AGIEF, the provided PET images are relatively low-quality and blurry; therefore, we generated high-SNR reference volumes by averaging PET values within MRI-defined brain regions, as described in the paper. We agree that this evaluation protocol has limitations, we also tested the use of the original PET measurements as reference images and observed consistent results with the regional-averaging. We reported the regional-averaging results because they make anatomical and reconstruction details easier to inspect. For the NX dataset, we use the provided PET images directly as reference images, since they are already high-quality.
Qualitative LA reconstructions: We agree that qualitative LA examples would help readers better assess possible artifacts. Due to space limitations, the paper only includes qualitative results for the FA setting, while the LA results are reported quantitatively in the tables.
Availability of MR images: PET-Adapter still improves reconstruction quality when MR images are not available. However, performance is lower than in the MRI-guided setting, which is expected since MR images provide useful anatomical guidance during adaptation.
The simulated scanner: The CERMEP and AGIEF datasets were originally acquired using a Siemens Biograph mMR scanner, which is the scanner geometry we modeled in Parallelproj. For the NX dataset, the exact scanner acquisition model was not provided, so we could not reproduce the exact scanner geometry in our simulations. We acknowledge that this is a limitation.
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 paper presents a test-time adaptation method for brain PET reconstruction to adapt to different clinical domain data with a phantom pre-trained diffusion model. An MR-guided conditioning via layer-wise adaptation blocks and a warm start through OSEM reconstruction prior are introduced. The experiments cover multiple centers and multiple vendors, including both FA and LA conditions, and the results reached an optimal performance compared with SOTA methods.
The authors have reached a consensus on the acceptance of the paper. While the technical novelty is incremental, the reviewers find the integration of LoRA and warm-starting for PET reconstruction to be a sound and well-engineered contribution. Its reduced computation load (T=2) could further enhance its potential for practical application.
Though an early acceptance is recommended, the authors should still provide feedback regarding (1) the benefit of warm starting beyond reduced computation cost, given that similar performance was achieved with more inference steps; (2) the requirement of MR conditioning since the MR prior is not always available; (3) results are on simulated data only; (4) the pre-training phantom dataset is relatively small containing only 20 scans; and (5) further implementation details as asked by R2 and R3.