List of Papers Browse by Subject Areas Author List
Abstract
Positron Emission Tomography (PET) is a powerful imaging technique but involves radiation exposure due to the use of radioactive tracers. A promising solution to mitigate this risk is reconstructing standard-dose PET (SPET) from low-dose PET (LPET). Previous studies have primarily focused on attenuation-corrected PET data; however, the attenuation correction process can amplify noise and artifacts, especially in low-dose scenarios. Additionally, PET scans are often paired with CT scans for attenuation correction, further contributing to radiation exposure. To address these challenges, we propose a new paradigm that reconstructs Attenuation-Corrected SPET (AC SPET) and standard-dose CT (SCT) images from the original Non-Attenuation-Corrected LPET (NAC LPET)) and low-dose CT (LCT) data through a collaborative reconstruction framework.
Key components of our proposed method include: (1) a coarse-to-fine learning strategy, wherein specialized reconstruction basis is initially built by processing each modality individually, followed by Domain Adapters to facilitate cross-modal feature correlation; (2) a hybrid Mamba-powered Expert Network that effectively captures long-range dependencies between different regions of whole-body PET/CT images; and (3) a Physics-informed Mutual Loss function to enforce consistency between the PET and CT domains, ensuring robust and reliable reconstruction results. Extensive experiments on the collected dataset demonstrate that our model achieves diagnostic-quality reconstruction while significantly reducing radiation exposure.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2987_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)
TCIA dataset: https://www.cancerimagingarchive.net/collections/
BibTex
@InProceedings{TanZix_ANew_MICCAI2025,
author = { Tang, Zixin AND Jiang, Caiwen AND Cui, Zhiming AND Shen, Dinggang},
title = { { A New Paradigm for Low-dose PET/CT Reconstruction with Mamba-powered Progressive Network and Physics-informed Consistency } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15970},
month = {September},
page = {3 -- 12}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel coarse-to-fine PET-CT dual-modal reconstruction framework, which combines a Mamba-based expert network and a physical prior mutual information loss (LPM). The method first uses modality-specific networks to reconstruct PET and CT images separately, and then conducts refined fusion through a cross-modal feature injection module. This paper proposes a new coarse-to-fine PET-CT bimodal reconstruction framework, which integrates a Mamba-based expert network and a physical prior mutual information loss (LPM). The method first uses modality-specific networks to reconstruct PET and CT images separately, and then conducts refined fusion through a cross-modal feature injection module. The proposed mutual information loss explicitly guides the PET and CT images to be consistent at the statistical level during the fine PET reconstruction stage, thereby improving the consistency between the anatomical structures and metabolic information of the 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 application of Mamba blocks to bimodal image reconstruction is a new trial, and the design of coarse-to-fine strategy is ingenious for the bimodal reconstruction task. Additionally, mutual information is used to measure the statistical consistency between PET and CT, guiding the reconstruction process from a physical perspective with a clear theoretical motivation, and improving the consistency between anatomy and metabolism.
- 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.The proposed method exhibits insufficient innovation and lacks robust theoretical foundation. 2.There is a dearth of experimental data to substantiate its efficacy,for instance in table 2, the experiment results are incomplete. 3.No comparison of time efficiency is provided. The Mamba module, potentially imposing a higher computational burden compared to CNNs, lacks essential efficiency metrics such as running time or parameter numbers in the paper, which would have been crucial for a comprehensive evaluation. 4.The strategy of cross-modal injection remains unexplained and unoptimized. The injection sequences of PET→CT and CT→PET are set as fixed configurations without any discussions on whether they represent the optimal choices or exploration of alternative combinations. This oversight limits the comprehensiveness and thoroughness of the research, hindering a more in-depth understanding of model design, performance and potential improvements.
- Please rate the clarity and organization of this paper
Poor
- 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 authors are suggested to complete the experimental results and incorporate relevant theoretical formulas for the proposed model. Additionally, a comparison of inference time and the number of model parameters should be included to validate the efficiency and advantages of employing Mamba. If feasible, making the code and pre-trained model publicly accessible is recommended, which will contribute to expanding the reproducibility and influrence of the proposed model.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents a new pipeline for the task of bimodal low-dose PET-CT image reconstruction. The model exhibits insufficient innovation in methodology and lacks robust theoretical foundation and motivation in model design. Furthermore, the experimental results in table 2 are incomplete and could not sufficiently validate its efficacy. These key concerns make me provide a reject recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Due to concerns about methodological novelty, theoretical foundation, and time efficiency evaluation, the authors did not adequately address these issues. Therefore, I am inclined to recommend a reject decision.
Review #2
- Please describe the contribution of the paper
The authors proposed a novel deep learning framework to reconstruct high-quality PET and CT images from low-dose PET (LPET) and low-dose CT (LCT) scans, which can significantly reducing radiation exposure during PET/CT imaging procedures. It uses coarse-to-fine learning strategy that begins with modality-specific networks and adds cross-modal enhancement via domain adapters, and also leverages Mamba blocks in the architecture. This approach has promising clinical implications for safer PET/CT imaging.
- 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 method jointly reconstructs attenuation-corrected PET and standard-dose CT, which is quite novel. 2) The integration of Mamba blocks enables the network to capture long-range dependencies in 3D volumes. This is a notable advance over conventional CNN.
- 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) Synthetic dataset only. The low-dose PET and CT inputs are generated via synthetic downsampling and noise modeling, especially downsampling will not only lead to noisier image but also introduce artifacts in the image.
2) Current manuscript lacks clinical evaluation, such as lesion detectability or radiologist assessment. Without this, it’s hard to assess true diagnostic value in clinical workflows.
- 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
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 — 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 manuscript is novel but the evaluation is not clinically relevant enough.
- 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.
The study lacks real data evaluation, which might be one of the weakness. Other than that, it should be of potential interest to the readers/attendees.
Review #3
- Please describe the contribution of the paper
The authors propose a novel Mamba-powered expert network for low-dose PET/CT reconstruction. It is an interesting exploration to reconstruct standard-dose PET images (SPET) directly from the non-attenuation-corrected low-dose PET images (NAC LPET). Both the quantitative and qualitative evaluations demonstrate the advantages of the proposed model over other state-of-the-art 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.
- The proposed alternating enhancement scheme takes into account both CT and PET information when performing reconstruction.
- The authors integrate the tri-orientation memba into the network to adaptively capture spatial features in 3D, making the network more task-appropriate.
- The authors provide a comprehensive comparison between the proposed model and the other competing models, making the conclusion more confident.
- 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.
- There is a typo in Figure 1. One of the symbols in the CT expert network should be D_CT instead of D_PET, if I understand correctly based on the following descriptions of the article.
- The technical descriptions can be further refined. In section 2.1, at the end of page 3, the authors mention that “the trained E_CT is frozen in this step, while the D_CT network remains trainable”. Why was part of the expert network frozen during training? Is there a pre-training step for each modality? Is the proposed network shown in Figure 1 trained in an end-to-end manner?
- 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
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?
I recommend that this paper be accepted. The authors propose a novel network for low-dose PET/CT reconstruction and the method achieves good results.
- 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 the paper. The author mentioned in their feedback that they will provide a more detailed description of the pipeline.
Author Feedback
Q1: Innovation and injection strategy (R1) Our work aims to reduce radiation exposure from both PET and CT scans in combined PET/CT imaging, where paired acquisitions are standard in clinical practice. To the best of our knowledge, existing methods address dose reduction in PET or CT separately, without leveraging their intrinsic metabolic-anatomical correlation. Since the two modalities are co-registered and come from the same patient, we argue that they can mutually enhance each other’s reconstruction — an idea well established in multi-sequence MR imaging. Supported by such theoretical foundation, we propose the first joint low-dose PET/CT reconstruction framework. Motivated by clinical workflow — where CT is acquired before PET and used for attenuation correction — we design a two-stage pipeline: first enhancing low-dose CT, then using it to guide PET generation. This sequence aligns with real-world practice and preserves the physical dependency between modalities, improving interpretability and reliability. Building on this foundation, we further enhance the backbone network by incorporating Tri-oriented Mamba Blocks for long-sequence modeling and introducing a Physics-Informed Mutual Loss for robust cross-modal alignment. Q2: Technical description (R3) Our training strategy follows a similar approach to that of [23]. The two-stage pipeline is not end-to-end. In the first stage, the PET and CT Expert Network are pre-trained separately, allowing each to learn domain-specific features. In the second stage, the pre-trained CT encoder is fixed to preserve the modality-specific knowledge it has acquired, preventing it from being overwritten by cross-modal signals introduced during PET reconstruction. Due to space limitations, we have only briefly referred to this strategy in the current version with a citation to [23]. We will include a more detailed description in the final version if given the opportunity. Q3: Synthetic dataset and clinical evaluation (R2) Due to the high cost of PET imaging, low-dose PET is rarely performed in clinical practice except for research purposes, especially with matched low-dose CT. Consequently, no public dataset currently supports our joint reconstruction study. We follow established practices in the field (e.g., [8], [15]) by using synthetic data. The simulation approaches have demonstrated strong correlation with real low-dose outcomes, enabling our experiments to closely approximate real-world scenarios. Nonetheless, we are actively collaborating with clinical partners to collect real data and obtain SUV assessments for future validation. Q4: Incomplete results in Table 2 (R1) The “N/A” in our ablation study (Table 2) is not due to missing data or incomplete experiments. Rather, the M4 aims to evaluate the effectiveness of the Physics-informed Mutual Loss in improving PET reconstruction. Since the CT reconstruction had already been addressed in the earlier configurations (M1–M3), we omitted the CT results for M4 as “N/A”. Q5: Inference time and number of parameters (R1) Our two-stage Mamba-based design is motivated by both the clinical imaging workflow — where CT is acquired before PET and used for attenuation correction — and the need for reliable reconstruction outcome. Although a single-stage CNN-based framework may seem more compact, our staged approach offers better interpretability, cross-modal alignment, and overall image quality. To mitigate potential increased computational cost, we employ parameter sharing and network reuse strategies. Specifically, the second stage builds on the fine-tuning of the first-stage model, avoiding the introduction of a large number of new parameters and maintaining high efficiency. Q6: Open-source commitment (R1) We plan to release our code, implementation details, and dataset after the anonymous review period. Q7: Typo in Figure 1 (R3) We thank Reviewer #3 for careful review. We will thoroughly proofread our manuscript in the next revision.
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’
The authors proposed an interesting approach and have address the majority of the reviewers concerns
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’
N/A