List of Papers Browse by Subject Areas Author List
Abstract
The adoption of contrast agents in medical imaging is essential for accurate diagnosis. While highly effective and characterized by an excellent safety profile, the use of contrast agents has its limitation, including rare risk of allergic reactions, potential environmental impact and economic burdens on patients and healthcare systems.
This work addresses the contrast agent reduction (CAR) problem, aiming to minimize the administered dosage while preserving image quality. Unlike existing deep learning methods that simulate high-dose images from low-dose inputs via end-to-end models, we propose a learned inverse problem (LIP) approach. By learning an operator that maps high-dose to low-dose images, we reformulate CAR as an inverse problem, solved through regularized optimization to enhance data consistency.
Numerical experiments on pre-clinical images demonstrate improved accuracy compared to traditional methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3839_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/devangelista2/LIP-CAR/
Link to the Dataset(s)
N/A
BibTex
@InProceedings{EvaDav_LIPCAR_MICCAI2025,
author = { Evangelista, Davide and Morotti, Elena and Colombo Serra, Sonia and Luo, Pengpeng and Valbusa, Giovanni and Bianchi, Davide},
title = { { LIP-CAR: a learned inverse problem approach for medical imaging with contrast agent reduction } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15970},
month = {September},
page = {374 -- 384}
}
Reviews
Review #1
- Please describe the contribution of the paper
In this study, the Contrast Agent Reduction (CAR) problem is reformulated as an inverse problem. A neural network is first trained to approximate a forward operator from high-dose to low-dose images, and the reconstruction is subsequently performed via regularized iterative optimization.
- 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 study introduces a new approach to the CAR problem. Unlike conventional end-to-end methods that directly predict high-dose images from low-dose inputs, it reformulates the task as a learned inverse problem.
-
To address the ill-posedness of the inverse problem, the study introduces a gradient-based regularization term that guides the optimization process.
-
The proposed method, LIP-CAR, achieved better performance than end-to-end trained networks in SSIM, PSNR, and RMSE.
-
- 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 concern that the proposed method may be more sensitive to input image quality during inference. While end-to-end networks, given sufficient training diversity, can robustly reconstruct even noisy inputs, the proposed optimization-based approach may suffer from unstable convergence. For example, severe noise or artifacts in the input image could distort the loss landscape, leading the optimization to incorrect solutions. This dependence on input quality may limit its applicability in real-world settings. Could the authors provide their perspective on this issue?
- The learned forward operator is estimated purely from data, but whether it is mathematically suitable for inverse reconstruction remains unclear. Providing experimental results on more datasets and model architectures could help further validate the approach.
- For example, the learned forward operator may lack desirable properties such as continuity or differentiability. As a result, the inverse problem formulation may converge to local minima or become unstable under different scenarios.
- An ablation study on the GenTV regularizer should be provided, as it plays a crucial role in the reconstruction process. Evaluating this allows for a better understanding of how much the regularization term contributes to the overall performance.
- In addition, could the author provide how much difference there is between the result predicted by solving the inverse problem using the synthesized output as a prior and the actual ground truth?
- The paper does not provide any evaluation of the learned forward operator, which is a central component of the proposed approach. Including such results would be necessary to support the validity of the proposed approach.
-
- 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.
- 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
I agree that the proposed approach has the potential to improve the CAR problem. However, the current experimental setup lacks sufficient validation and leaves several aspects unaddressed.
- 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?
This study introduces a new approach to the CAR problem by reformulating it as a learned inverse problem through the use of a forward operator. However, some concerns remain regarding the insufficient justification and empirical validation of several key components in the proposed framework.
- 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.
While I still have concerns about the limited experimental validation and the appropriateness of the forward operator, I am ok with accepting the paper given its contributions.
Review #2
- Please describe the contribution of the paper
Proposing a framework for contrast agent reduction based on solving a learned inverse problem. They propose to learn an operator mapping high dose to low dose and solve the auxilary optimization problem of minimizing an energy integrating this operator for data fitting.
- 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.
Learned energies and their optimization, which is essentially proposed here have been heavily used in MRI reconstruction tasks. The paper hence proposes a novel formulation for contrast agent reduction that has not been used for this task, which is a relevant task in medicine. Additionally, they are the first to show this on a high relaxivity contrast agent.
Their experimental valuation includes a box plot to show statistical significance.
- 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 experimental validation is weak in the sense that only self-ablations are shown. None of the cited works, which do contrast reduction were compared to. Visually the differences are also barely visible and are more likely caused by the differences in how the intensity ranges are mapped or scaled in the plot, which explains why Image 39 in LIP-CAR the gray matter is brighter. In general the real challenge in this task is to retain very small lesions as has been shown in “Metastasis Detection Using True and Artificial T1-Weighted Postcontrast Images in Brain MRI” by Pinetz et al. Investigative Radilogy 2024 which is not addressed at all here. Additionally, the proposed framework uses a lot more compute than the comparison methods as there are 150 Neural Function Evaluations used to produce one image instead of 1.
-
It would also be nice to include how the low-dose operator compares to other works which generate low-dose images such as: “Simulation of arbitrary level contrast dose in MRI using an iterative global transformer model” by Pasumarthi et al. MICCAI 2023 or “Faithful synthesis of low-dose contrast-enhanced brain mri scans using noise-preserving conditional gans” by Pinetz et al. MICCAI 2023
-
The choice of Total Variation is not motivated. Total Variation assumes piece wise constant images, which is not the case in MRI and this is why the results look very smooth. Additionally, solving the optimization using ADAM is also not motivated. Settings like these lead to natural choices of optimizers as has been shown for example by Kobler et al in “Total Deep Variation for Linear Inverse Problems” CVPR 2020
-
- 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.
- 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
You should cite related work on inverse problems. At least an overview or survey paper should be cited for interested readers to get further information.
- 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?
I think the general approach is interesting to the MICCAI community and overshadows the missing comparison methods and the non-ideal choices made.
- 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.
While the paper itself is just a proof of concept work, the setup is unique in trying to adopt contrast reduction with rat brains and the overall approach is interesting enough (with open source code) that it merits publication.
However, there are still multiple issues, such as weak evaluation, clinical utility in adopting a 2D approach and the optimization framework itself could be significantly improved by adopting different optimizers and regularizers.
For me it is a truly borderline paper.
Review #3
- Please describe the contribution of the paper
This paper introduces an iterative scheme to solve the contrast agent reduction problem. This is done by using neural netorks to train both the forward and the backward operators.
- 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 solution is mathematically correct. This highly reduces the chances of artifacts and hallucinations.
The usage of known mathematical operations is interesting, and it seems to improve the quality of the results.
The paper is well written. The equations are correct and introduced in a way which makes them easy to understand
- 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.
Figure 1 is quite unclear and the text is very small
I am not sure splitting 3D volumes into 2D slices is a good idea. Essential volumetric information is lost.
Figure 3 is of poor quality. I can’t make out what it says.
“Indeed, Figure 3 firmly remarks the LIP-CAR success in terms of both SSIM, PSNR (the higher, the better) and RMSE metrics (the lower, the better). The boxplots referring to LIP-CAR (the green ones) consistently denote that we hit the best values, overcoming NN-L2H and NN-PL2H (respectively represented by the cyan and blue boxplots) and strongly improving the quality of the acquired xP and xL images (red and orange boxplots).” The confidence bars of the box plots seem to overlap. Can we claim significant improvements, especially with a small test dataset of 240 samples
Detectability of the tumours in improved images can be quantified by a reader study of some sort. I feel this will make the argument stronger.
- 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.
- 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?
While I feel the methodology is interesting, the results don’t seem to speak very well. The quantitative results in particular are only sparesely discussed, and the figure is of poor quality. There should be a table discussing the results and comparing with SOTA/ablations
- 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
Author Feedback
We thank the reviewers for their feedback. We appreciate that they agree the main strength of our work lies in its methodological contribution, applying well-established regularized optimization to the emerging CAR task. Below, we provide joint responses to the weaknesses (W) raised by reviewers R1, R2, and R3.
ABOUT THE MODEL
R3-W1: Concern that LIPCAR is sensitive to input image quality. Based on a well-consolidated optimization theory and model-based formulation, we argue that LIPCAR is not particularly prone to sensitivity issues. Since training/test data often share acquisition protocols, variations mainly stem from noise, handled by the regularizer.
R3-W2: Is the learned forward operator mathematically suitable? The NN-based forward operator is inherently continuous. While the obj function may be non-convex, potentially leading to local minima, our experiments show strong performance, as in other successful NN-based inversion methods like DIP or DGP.
R1-W3 and R3-W3: Choice of TV and ADAM not motivated. Need of ablation study on GenTV. As noted in the Conclusions, our implementation is a proof-of-concept. The regularized framework is modular, and literature offers many alternative priors and solvers that could replace our choices - still within the LIPCAR setting. We chose TV for its ability to promote sharp boundaries, supporting our goal of contrast enhancement over realism. While we used ADAM for simplicity, future work could explore solvers tailored to TV. A deeper investigation of GenTV, especially its behavior when based on the GT image, is a valuable future direction.
R2-W2: why splitting 3D to 2D The 2D slice-based analysis was adopted for simplicity and aligned with the industrial partner’s focus. LIPCAR is readily extendable to 3D volumetric data in all its modular components.
ABOUT THE WEAKNESS OF THE EXPERIMENTAL VALIDATION
R1-W1: Real challenge is on very small lesions. R2-W3: Can we claim significant improvements with a small test dataset? R2-W4: Tumor detectability by a reader study. This research is in collaboration with an industrial partner (with affiliated authors) focused on simulating higher contrast agent doses to enhance tissue contrast. As the study involves rat brains, small lesions were not the primary target. The data were acquired by the partner’s team, and we believe this pre-clinical dataset is sufficient to validate our methodology. Results were confirmed by domain experts and supported by metrics. We are also interested in future work with human data and reader studies, pending real image availability.
R1-W1: Only self-ablations are shown. We selected Bône et al. as the state-of-the-art competitor because it is a strong representative of pure DL-based CAR methods. While additional approaches exist (e.g., those suggested by R1), they mainly rely on end-to-end DL without incorporating any model-based formulation involving image-dependent parameters, which represents the distinctive novelty of our LIPCAR framework.
R1-W2 and R3-W4: Evaluation of the learned forward operator and comparisons. Due to space constraints, we omitted this clarification, but we will include the references in the revised manuscript. In future work, we plan to compare our ResUNet with the networks mentioned. As the choice of architecture is modular within LIPCAR, ResUNet can be replaced with any of the proposed networks eventually.
OTHER
R1-W1: Computational cost Correct observation. The computational demand of 150 NFEs reflects the current iterative approach that can be optimized; we see potential for future efficiency improvements.
R2-W1: Figs 1 and 3 not clear. To improve clarity, we will reprocess them at higher resolution.
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’
This paper is recommended for acceptance, as all reviewers have reached a unanimous positive consensus.
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