List of Papers Browse by Subject Areas Author List
Abstract
Positron Emission Tomography (PET) is an advanced nuclear medicine imaging technique widely used in the diagnosis and treatment of oncology and neurological diseases. However, PET images suffer from high noise levels due to statistical fluctuations and physical degradation factors during image acquisition. Recently, deep learning-based denoising methods have shown great performance for PET image quality enhancement. Most of these methods attempt to incorporate high-quality anatomical images (such as CT or MR), as network input to provide prior information into the PET denoising process. However, directly using CT or MR images as network input has limited effectiveness and lacks interpretability due to the significant differences between two modalities. Exploring how to make better use of anatomical priors remains a valuable research direction. In this study, we proposed an unsupervised PET image denoising framework that leverages the Bowsher prior to achieving cross-modality fusion and anatomical information
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1122_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)
simulation dataset: https://brainweb.bic.mni.mcgill.ca/anatomic_normal_20.html
clinical dataset: https://openneuro.org/datasets/ds004513/versions/1.0.3
BibTex
@InProceedings{WuZho_Bowsher_MICCAI2025,
author = { Wu, Zhongxue and Wu, Jiankai and Cui, Jianan and Feng, Yuanjing and Chen, Zan},
title = { { Bowsher prior Enhanced Unsupervised PET image Denoising } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {86 -- 96}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposed a two deep image prior (DIP) based deep learning models combined with the Bowsher prior to improve PET denoising in an unsupervised way. The basic idea is to use the first DIP to get a denoised image, and then apply the Bowsher operator to highlight the boundary effect to get Bowsher prior, and finally use another DIP to get the final images. The experiment is based on the computer simulation and real data. The compared methods include several traditional methods and one deep image prior method. Overall, the paper is well written. But my concern is related to the method novelty and fairness of compared methods.
- 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 used the Bowsher prior to improve the boundary clarity of the PET image, which would be considered as a technical contribution for this work. Another strength of this work is that authors make a clear description for the paper, and the experiment part is well organized .
- 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.
- Because the proposed method is not an end-to-end method, I am wondering how to balance two separate deep learning models to get the stable result if applying on different datasets.
- Regarding the concept of method, is it important to explicitly include the Bowsher prior image as the bridge to connect two DIP methods. My thinking is you already have the MRI as the prior which can provide rich boundary information. Do you really need Bowsher instead of other deep learning layers?
- The compared methods are not solid. The traditional methods are too old, and we all know it may not work better than AI-based methods. The only CDIP should be one part of your whole work. It’s not surprising CDIP can not work better than yours. You should compare your methods with other unsupervised deep learning based denoising methods. If possible, it’s better to include an opened model with supervised learning for comparison. It doesn’t matter to show worse results than supervised methods. We just want to balance the training cost and performance.
- In addition to global index comparison, there is lack of quantitative accuracy comparison among different methods.
- 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
First, I suggest to highlight the importance on why needing an explicit Bowsher operator as the bridge. Second, at least adding one compared method which uses the totally different model with yours. Third, adding the quantification ability comparison to show other advantages.
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The weak reject is mainly due to the method novelty and comparison fairness.
- 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
This paper presents a new unsupervised PET image denoising method based on the Deep Image Prior (DIP) framework. Rather than relying solely on the noisy PET image or combining PET and MR images (as done in CDIP), the authors propose to first compute a Bowsher prior using the output of CDIP and an MR image. This prior is then used as an additional input for a second DIP step, alongside the noisy PET and MR images. The second DIP step uses SA-Net, an original architecture proposed by the authors.
- 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 idea of computing a Bowsher prior before passing it to the network appears to improve the performance of the subsequent DIP step, and is fairly original. The ablation study is thorough and addresses many of the questions that naturally arise while reading the paper.
- 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.
One part of the ablation study (Experiment 4) is unclear: “Use the Bowsher prior image computed from the noisy PET image to replace the MR image as prior image for the CDIP method.” While it makes sense to compute the Bowsher prior using the noisy PET image, the reasoning behind replacing the MR image with this prior is not obvious. It’s also unclear why this is tested in the CDIP method and not in SA-Net. This might be a misstatement, but I think it requires clarification from the authors.
Regarding the use of the CDIP output to compute the Bowsher prior: it is not obvious that this step brings a substantial benefit. The quantitative results do not indicate a significant difference when using overfitted / underfitted CDIP, or noisy PET image as input for computing the Bowsher prior. It seems plausible that this step could be removed with minimal impact on performance, while significantly accelerating the method.
The MR image is used at multiple points in the method (CDIP, Bowsher prior computation, and SA-Net), but the motivation for re-injecting this information multiple times is not fully clear. The authors mention that “Using only the Bowsher prior image as input results in decreased output image quality as MR image provides essential anatomical information that aids in detail recovery,” but it remains unclear why this anatomical information is not sufficiently preserved in the Bowsher prior alone.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
The writing is clear and well-structured. However, the visual differences in the qualitative results, especially in Figures 3 and 4, are difficult to perceive. Including zoomed-in regions or highlighting regions of interest could make these differences more apparent. For future work, it might be worth exploring whether the initial CDIP step can be omitted, as its benefit is not clearly demonstrated.
- 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 — could be 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 demonstrates clear improvements over CDIP for PET image denoising, with well-designed experiments and a novel use of a precomputed Bowsher prior as additional conditioning. However, I am not fully convinced of the necessity of the initial CDIP step. In particular, the results from the ablation study do not strongly support its contribution, and one experiment (Experiment 4) is somewhat confusing and might benefit from further clarification. Currently, different concepts are gathered in an empirical study and the added value seems incremental.
- 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 paper investigates a combination of methods aimed to improve PET images. Although preliminary, this study is relevant for the conference’s domain.
Review #3
- Please describe the contribution of the paper
This paper presents an unsupervised PET denoising framework that integrates anatomical prior information through a Bowsher prior formulation. The pipeline consists of a two-stage architecture: first, the CDIP method is used to generate a preliminary denoised image; then, this image, along with MR input, is used to create an anatomical prior via the Bowsher strategy. The resulting prior is incorporated into a denoising network (SA-Net) that utilizes attention mechanisms to fuse image and prior information. The proposed method is evaluated on both simulated and real PET data, and compared against several baseline approaches, including direct network-based denoising and traditional anatomical priors. Quantitative assessments (e.g., PSNR, SSIM) and qualitative visualizations indicate that the proposed method achieves superior noise suppression and structure preservation. The study highlights the potential of integrating MR-based anatomical priors in deep learning pipelines for improved PET denoising under low-count conditions.
- 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.
Methodological novelty through Bowsher prior integration: The key innovation lies in the explicit incorporation of a Bowsher prior—derived from MR and CDIP-based PET—as anatomical guidance into a deep unsupervised PET denoising network. This hybridization of classical priors with modern attention-based architectures is novel and technically elegant. Dual-stage unsupervised denoising strategy: The use of CDIP for initial denoising, followed by refinement with a structurally-aware SA-Net, demonstrates a thoughtful decomposition of the problem. This modular design reduces reliance on ground truth PET, making it more broadly applicable. Attention-based prior fusion mechanism: The paper leverages attention to balance raw PET and anatomical information, preserving important boundaries and suppressing noise adaptively. This provides interpretability and stronger localization performance. Solid empirical validation: Evaluation is carried out on both simulation and real patient data, including multiple organs and noise levels. Metrics like PSNR, SSIM, and LPIPS are used alongside visual inspection to support the claims. The comparisons cover a range of baselines, including traditional and learning-based methods. Clinical relevance: The method directly addresses the low-count PET scenario, a critical problem in clinical nuclear medicine. The framework allows improved image quality without requiring full-dose PET or labeled datasets, facilitating deployment in real-world imaging pipelines.
- 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.
Lack of statistical validation and uncertainty measures: The paper reports quantitative metrics (PSNR, SSIM, LPIPS) across datasets but omits any statistical significance testing (e.g., p-values or confidence intervals). Without reporting standard deviations or inter-patient variability, it is difficult to assess the robustness of the improvements. How confident can we be that these improvements generalize across subjects or scanners? No ablation or component analysis: The framework includes multiple modules—CDIP, MR-derived Bowsher prior, and an SA-Net architecture—yet no ablation study is conducted. It remains unclear how much each component contributes to final performance. Could simpler architectures (e.g., direct PET-to-PET learning or use of non-learned priors) yield comparable results? Conceptual ambiguity in unsupervised learning claim: Although the method is labeled “unsupervised,” it heavily relies on MR priors and CDIP outputs as pseudo-labels. This setup borders on weakly- or self-supervised learning. The conceptual framing of the method is unclear and should be more precisely defined. Overreliance on MR undermines CT/MR-free narrative: The paper positions itself as a CT-free solution, but in practice it requires MR scans. This weakens its claimed generalizability, especially in clinical workflows where MR is either unavailable or cost-prohibitive. No robustness testing against PET-MR misregistration: Anatomical priors are highly sensitive to spatial alignment. The method assumes perfect registration but does not simulate or evaluate performance under real-world misalignment. This is a major gap, especially given the reliance on MR-derived priors. Lack of clinical endpoint evaluation: While denoising improves image quality metrics, there is no evaluation on clinical impact—e.g., effect on lesion detectability, SUV accuracy, or diagnostic utility. It’s unclear whether the method merely denoises or also preserves subtle pathological features. Under-described network architecture and training setup: The SA-Net structure and training details are insufficiently reported. No information is given on training duration, optimizer configuration, loss weighting, or data augmentation. Without this, reproducibility and future extension become difficult. Failure modes and limitations not discussed: The paper lacks a critical reflection on when and how the method might fail. For example, what happens when the MR does not contain the pathology visible on PET? Is there a risk of anatomical bias that could suppress diagnostically important PET signals? No attention visualization or interpretability analysis: Given that the method uses attention mechanisms to fuse PET and MR inputs, visualizing attention maps could provide insights into what anatomical structures are being emphasized. The absence of such analysis limits the interpretability and trustworthiness of the model.
- 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
Dear authors, Your proposed method presents a thoughtful attempt to address denoising of ultra-low-dose PET images using a dual-stage learning framework that leverages anatomical priors and a self-attention mechanism. The motivation is strong, and the reported quantitative improvements are encouraging. However, for your work to be more compelling and suitable for clinical or translational application, several aspects merit deeper explanation or revision: You refer to your method as “unsupervised,” yet it relies on pseudo-labels from CDIP and anatomical MR priors. This creates ambiguity in the supervision paradigm. Is your approach better framed as self-supervised or weakly supervised? A precise clarification would greatly benefit the narrative and positioning of the work. Despite aiming for a CT-free workflow, your method depends on MR-derived priors, which are not always available in standard PET protocols. The “PET-only” narrative seems overstated. Could you discuss how the method would generalize in absence of MR, or explore a fully PET-based alternative? The proposed model is composed of multiple non-trivial components (CDIP stage, Bowsher prior, SA-Net with attention). Yet, no ablation study is provided. Without this, readers cannot assess the necessity or added value of each component. Similarly, while attention mechanisms are central to your design, no visualization or interpretability analysis is offered. This is a missed opportunity to validate your network’s behavior or guide potential clinical applications. Essential training details are missing, including learning rate, loss weighting, batch size, optimizer settings, and convergence behavior. These are necessary for replicability. You do not mention any code availability or plans for releasing it. This limits the community’s ability to reproduce or build upon your work. The quantitative evaluation lacks statistical significance measures (e.g., standard deviation, confidence intervals, or p-values). How do you justify that the observed improvements are consistent and generalizable? The anatomical robustness of the method is not validated. How would it perform in regions with subtle pathology, or under misalignment between PET and MR? While quantitative image quality improves, it is not clear whether this leads to better clinical decision-making (e.g., better lesion detection, preservation of small pathologies, improved SUV quantification). Even a limited clinical simulation or qualitative expert evaluation would significantly strengthen the work. In summary, this work proposes an innovative and timely approach, but it requires several technical, conceptual, and clinical clarifications to reach its full potential. We strongly encourage you to expand on these aspects either in the revision or future work.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents an important contribution to ultra-low-dose PET denoising using anatomical priors and a dual-stage attention-based framework. The proposed pipeline—leveraging CDIP outputs and MR-guided Bowsher prior—addresses a timely problem in PET imaging and demonstrates promising improvements across metrics such as PSNR, SSIM, and LPIPS. The idea of replacing anatomical images (e.g., CT) with MR-based priors to guide PET denoising is valuable and aligns with future directions in image-guided PET processing. However, several critical issues remain unresolved, which prevent a strong acceptance at this stage. The lack of statistical validation (e.g., uncertainty metrics, p-values), the absence of ablation studies for key components (CDIP, attention, prior), and insufficient discussion about reproducibility and robustness (e.g., misregistration, pathology-prior mismatch) limit the technical completeness of the study. The term “unsupervised” is also used somewhat loosely, given the heavy reliance on pseudo-labels and anatomical priors. If these concerns are addressed properly in the rebuttal, the paper can be considered a solid candidate for acceptance. Its core idea is relevant and has potential impact, but more technical rigor and clarity are needed to reach its full strength.
- 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.
After reviewing the authors’ rebuttal in detail, I believe that the key concerns raised during the initial review process have been adequately addressed. The authors provide thoughtful clarifications and new comparative results that help validate the originality and robustness of their proposed PET denoising framework. Specifically: The necessity and modular design of their dual-stage framework (CDIP + Bowsher Prior + SA-Net) are convincingly justified through ablation experiments and logical reasoning. The role of each component—particularly the Bowsher prior as a bridge and SA-Net’s attention mechanism—is clarified and substantiated with supportive results (e.g., Fig. 4 and Table 2). In response to methodological concerns, the authors offer clearer positioning of their approach as self-supervised rather than strictly unsupervised, improving conceptual accuracy. Quantitative performance is strengthened by comparisons against both supervised (3D U-Net) and unsupervised (DeepRED) baselines, along with significance testing (Wilcoxon rank sum) and standard deviations now included. They acknowledge limitations (e.g., PET/MR dependence, absence of clinical endpoint testing, and robustness under misregistration) and provide a realistic roadmap for future extensions and validation. While the paper still lacks some finer details (e.g., attention map visualization, full clinical interpretability), the authors commit to including these enhancements in the final version. Taken together, the method’s novelty, clinical relevance for ultra-low-dose PET, and the substantial effort to address reviewer feedback justify an accept recommendation. This work represents a meaningful methodological step forward in anatomically informed PET image denoising and provides a useful foundation for both academic and translational development.
Author Feedback
We sincerely thank the reviewers for their thoughtful feedback and constructive suggestions. We further address your concerns below.
Necessity of the Module and Ablation Experiments R2#2 R3#2 Bowsher Prior is the core component of our method which bridges dual DIP processes with its patch-based image fusion capability so that the network can extract structural information more efficiently. Replacing it with deep learning layers will lead to a framework similar to CDIP, which is proved less effective than the proposed method (Fig.2、3 and Table.1). R2#2 R1#2 CDIP acts as the critical initial denoiser, preventing the propagation of PET noise to Bowsher Prior and SA-Net, which lack inherent denoising capacity. Ablation study (Fig.4 Prior from noisy PET) confirms that omitting CDIP introduces residual noise/artifacts in results. R2#2 R3#1 SA-Net is used to enhance the uptake information in the Bowsher prior. The specially designed SA-Net can effectively limit noise overfitting during training and addressing the stop problem of the CDIP. Ablation study demonstrates SA-Net’s superiority over U-Net (Fig.4 Unet replace SA-Net) and its robustness to CDIP output variations (Fig.4 Prior from CDIP). Whether CDIP overfit (noisy) or underfit (blurry), our pipeline delivers consistent enhancements (cf.Fig.4 & Table.2). SA-Net’s strong generalization ensures stable performance across different datasets. R2#2 R1#1,3 The ablation study (Experiment 2 and 4) was designed to verify whether the Bowsher prior could replace the MR image for anatomical guidance to our method or CDIP. Thus, we used only Bowsher prior as input in SA-net (Fig.4 SA-Net input without MR) and CDIP (Fig.4 CDIP), demonstrating using Bowsher prior alone is insufficient for full anatomical guidance.
Comparative Experiments and Result Reliability R3#3 Thank you for pointing out that it is unfair to use CDIP as a comparison method, and we will move it to the ablation study. As suggested, we evaluated other unsupervised method (DeepRED) and supervised method (3D Unet) on the simulation dataset (clean PET labels unavailable for clinical dataset). Our method achieves superior PSNR/SSIM (31.147dB/0.912) over unsupervised DeepRED (29.704dB/0.875), though below supervised 3DUNet (34.263dB/0.966, 5-fold cross-validation). R2#1; R3#3,4 We’ve reported standard deviations across the quantitative metrics to assess inter-subject variability in Table 1 & 2. As suggested, we conducted the Wilcoxon rank sum test on the results, which proved that our method was significantly better than Gaussian filtering, NLM, BM4D, CDIP, and DeepRED in CNR improvement, PSNR, and SSIM (p<0.001 for all tests). R2#6 Clinical endpoint evaluation of our method will be performed in our future work. However, DIP-based PET denoising methods [1-2] have been proven to be good at preserving subtle pathological features while denoising.
Misregistration R2#5,8 Our method uses noisy PET itself as the training label, which ensures output distributions match original uptake patterns. Fig.2 shows that our approach can successfully reconstruct simulated lesions in PET image despite the absence of MR visible pathology, which proves the ability of our method to deal with misregistration.
Limitation and Future work (R2#4) Our method is currently tested exclusively on PET/MR data. Future studies will test on PET/CT dataset. (R2#7) Section 2.2 details the SA-Net architecture, and Section 2.3 reports training time and optimizer settings; loss weighting and data augmentation were not used. (R2#9) We will highlight ROIs in Figs.3-4 and show visualizing attention maps to improve visual comparisons. (R2#3) We will claim our method as self-supervised learning. Future studies will rigorously evaluate CDIP’s necessity, Bowsher prior’s bridging role, and SA-Net’s architectural contributions.
[1] K, Gong, et al. IEEE Trans. Med. Imaging 38.7 (2018): 1655-1665. [2] Q, Liu, et al. Med. Image Anal 95 (2024): 103180.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Upon review of the feedback and rebuttal, it appears that most of the concerns have been addressed, despite no change in the reviewers’ ratings.