List of Papers Browse by Subject Areas Author List
Abstract
Low-dose computed tomography and low-dose positron emission tomography reduce examination time and radiation exposure, crucial for radiation-sensitive patients and repeated exams. However, dose reduction yields photon-starved measurements that amplify noise and structured artifacts, undermining reliable lesion segmentation.
Although deep learning advances low-dose reconstruction, most models lack explicit dose modeling. Some rely on dose metadata, often missing or poorly aligned with actual degradation, leaving residual dose-dependent distortions.
To address these limitations, we define an observation-driven dose (OD-dose) as a latent coordinate of task-relevant corruption severity and estimate it by fusing periodicity degradation cues and mutual information cues extracted from sinograms. The estimated OD-dose modulates intermediate features via dose-aware feature-wise polynomial modulation, enabling nonlinear, task-aligned dose-adaptive restoration. The proposed module is a lightweight, plug-and-play component that can be attached to diverse reconstruction backbones. Experimental results confirm that DosePlug improves reconstruction and lesion segmentation on LDCT and LDPET benchmarks, while maintaining stable gains on real-world datasets. The code is available at https://github.com/kikidohyun/DosePlug.git.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1517_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/kikidohyun/DosePlug.git
Link to the Dataset(s)
CT-ICH Dataset: https://www.physionet.org/content/ct-ich/1.3.1/
AutoPET Dataset: https://www.autopet.org/fdgpetct.html
UDPET Dataset: https://udpet-challenge.github.io/
Mayo Clinic LDCT Dataset: https://www.cancerimagingarchive.net/collection/ldct-and-projection-data/
BibTex
@InProceedings{KiDo_DosePlug_MICCAI2026,
author = { Ki, Do Hyun AND Lee, Yeong Jong AND Yoo, Seok Bong},
title = { { DosePlug: Observation-Driven Dose Modeling for Robust Low-Dose Reconstruction and Segmentation } },
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
This paper presents an origial approach to low-dose computed tomography (LDCT) and positron emission tomography (LDPET) reconstruction by introducing “DosePlug,” a plug-and-play module that estimates an observation-driven dose (OD-dose) directly computed from sinogram (ie: doesn’t require physical and acquisition metadata). The core contribution lies in defining OD-dose as a latent corruption severity metric derived from Mutual Information (MI) and periodicity degradation, which is then used for nonlinear feature modulation. This work addresses challenges in reducing radiation exposure while maintaining diagnostic fidelity for lesion segmentation.
- 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- OD-Dose: The concept of “Observation-Driven Dose” as a task-relevant latent coordinate rather than relying on physical metadata (as mAs) is an interesting contribution. From the authors, it allows better alignment with the image/sinogram degradation observed in low-dose acquisitions. 2- Plug-and-Play Architecture: The ability to insert the module into diverse backbones (Diffusion, CNN, Transformer) without retraining the entire backbone is highly practical for assessment. Here, 2 backbones are tested (VisNET and RED). 3- Evaluations: 2 real-world datasets (UDPET, Mayo LDCT) beyond synthetic simulations (based on AutoPET and CTICH dataset) for image quality measurements; AutoPET and CTICH for diagnostic (lesions segmentation quality); coherent ablation study.
- 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.
- Is the approach work only on 2D images or also on 3D ?
- Some claims are difficult to understand : a- why using squares in the sinogram periodicity library, no circles? (closest to reality?) b- What “%” really means? ( x-ray/injected dose or exposure time), c- “MI from an input sinogram can provide an internal cue for OD-dose estimation because. . . “ why MI? no better feature for that? d- Why use a second-order polynomial for modulation? (exponential law?)
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
- The authors can highlight a few results in tables
- (curiosity) Is there a link with prototypes and using both the sinogram periodicity library and the attention alignment?
- The optimal performances (DSC/HD95) @100% can be reminded. It is not clear if the scores obtained are good enough for clinical usage, nor if the % used corresponds to a clinical reality.
- 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?
This paper introduced novel concepts that are well tested and convincing. The authors may wish to take my previous comments (in weakness) into account in order to clarify certain aspects of the text.
- 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 low-dose CT reconstruction method. Based on sinograms clues (distortion, information loss), they estimate the corruption to reconstruct the image without relying on dose metadata.
- 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.
To my knowledge, the corruption estimation from the sinogram-level cues is novel, with its integration in the pipeline using task-driven latent dose estimation. The results consistently improve standard reconstruction 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.
1) Results are not compared with the best results (leaderboard) in AUTOPET and CTICH. 2) Segmentation is only evaluated with simulated low-dose. Real low-dose is evaluated with reconstruction. 3) How often is dose metadata non-available ? In clinical practice, would it really be missing, i.e. recent scanner not producing them? Which ones ? Or is the problem only in PACS extraction, processing pipelines discarding the tags/reports info ? 4) Foundation models trained on multiple doses such as [1] were shown to be quite robust to dose variation [2], could you evaluate how they perform in these segmentstion tasks ? [1] Pai, Suraj, et al. “Vision foundation models for computed tomography.” arXiv preprint arXiv:2501.09001 (2025). [2] Martín Asiain, María, et al. “Impact of CT dose on AI performance: A comparison of radiomics, deep, and foundation models in a multicentric anthropomorphic phantom study.” Medical Physics 53.3 (2026): e70374.5) Fig 1, it is a bit confusing that the DSCs are average DSCs. It seems like DSC is 0 for some LDPETs, but 15/20% are reported. However I agree that it is important to report the average. I don’t have a solution, but maybe the authors can think of something. 6) Fig 1, It would be good to include your results in Figure 1 (your reconsteuction, segmentations and DSCs). Like that we really get a full picture of the problem and the provided solution. 7) Fig 1, It is RED backbone for PET and VisNet for CT ? Clarify.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall well designed method, with good evaluation including ablation studies, and convincing results. Some missing clinical motivation and comparisons.
- 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 #3
- Please describe the contribution of the paper
Authors propose DosePlug, a lightweight and plug-and-play module designed to improve image reconstruction and lesion segmentation in low-dose CT(LDCT) and PET(LDPET).
- observation-driven dose(OD-dose): a latent coordinate of task-relevant corruption severity inferred directly from sinograms, eliminating the need for physical dose metadata at inference
- dose-aware feature-wise polynomial modulation(FwPM): it is applying an OD-dose-conditioned second-order polynomial reparameterization for stable nonlinear modulation, mitigating noise amplificationo restore corrupted sinogram structures using a learnable library
- periodicity-aware attention: it restores corrupted sinogram periodicity via a sinogram periodicity library(SPL)and provides periodicity degradation cues for OD-dose estimation.
- 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.
- Novel Metadata-Free Formulation: By defining OD-dose based on sinogram to directly quantify image degradation.
- Exceptional Efficiency and Plug-and-Play Design: DosePlug is highly versatile, supporting diverse backbones(CNN, Transformer, and Diffusion model).
- Superior Performance at Ultra-Low Doses: The model demonstrates remarkable robustness under extreme conditions; for instance, at a 1% dose level, it improved DSC and PSNR compared to a baseline SOTA model VisNet.
- Clinical Feasibility and Validation: The evaluation metrics, such as rPSNR, were selected in collaboration with nuclear medicine clinicians to ensure clinical relevance.
- Real-World Generalization: Beyond synthetic simulations, the module maintained stable gains on real-world datasets like UDPET and Mayo Clinic LDCT, proving its practical utility in diverse acquisition environments.
- 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.
- Limitation to 2D Domains: The current implementation focuses on 2D reconstruction, whereas medical imaging often requires processing 3D volumetric data. The authors acknowledge that extending DosePlug to 3D domains remains future work.
- Sensitivity to Non-Dose Artifacts: The stability of OD-dose estimation may decrease when encountering extreme domain shifts that are unrelated to dose, such as specific scanner-dependent artifacts.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 recommendation for this paper is primarily justified by its high practical utility and the improvement of reconstruction and lesion segmentation on LDCT and LDPET benchmarks at a minimal computational cost.
- Novel Metadata-Free Formulation: By defining OD-dose based on sinogram to directly quantify image degradation.
- Exceptional Efficiency and Plug-and-Play Design: DosePlug is highly versatile, supporting divere backbones(CNN, Transformer, and Diffusion model).
- Superior Performance at Ultra-Low Doses: The model demonstrates remarkable robustness under extreme conditions; for instance, at a 1% dose level, it improved DSC and PSNR compared to a baseline SOTA model VisNet.
- Clinical Feasibility and Validation: The evaluation metrics, such as rPSNR, were selected in collaboration with nuclear medicine clinicians to ensure clinical relevance.
- Real-World Generalization: Beyond synthetic simulations, the module maintained stable gains on real-world datasets like UDPET and Mayo Clinic LDCT, proving its practical utility in diverse acquisition environments.
- 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 sincerely thank the reviewers for their positive assessment and constructive comments.
(Percentage) The percentages used in this paper to indicate low-dose levels refer to the dose ratios relative to the full-dose reference. In CT, this corresponds to a reduction in the tube current-time product, i.e., mAs, while in PET, it corresponds to reduced injected activity or acquisition time.
(Second-order polynomial) As described in the manuscript, conventional FiLM-style first-order modulation has limited capacity to model image distortions that vary nonlinearly with the dose level. Therefore, we adopt a second-order polynomial to introduce stable nonlinear feature modulation.
(Real low-dose segmentation) Due to the lack of publicly available real low-dose datasets with lesion masks, real low-dose segmentation experiments are currently limited. As noted in the Conclusion section, future work will focus on this evaluation.
(Dose metadata availability) We agree that dose metadata is rarely missing during routine clinical acquisition. Our point is that, from a dataset and post-processing perspective, dose metadata may become unavailable due to preprocessing pipelines. To avoid the misleading impression that dose metadata is frequently missing in clinical practice, we will revise the manuscript to clarify this point.
(Scanner-dependent artifacts) Compared with models that directly use dose metadata as input, OD-dose-based modulation inferred from input observations is expected to be more robust to scanner and protocol shifts. However, as noted in the Conclusion section, we acknowledge that OD-dose estimation may become less stable under severe domain shifts unrelated to dose, such as scanner-specific artifacts. We will also clearly state this point as a future research direction.
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.
All reviewers are positive about this paper. The authors should address reviewers’ concerns in the final version.