Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Dynamic contrast-enhanced Magnetic Resonance Imaging (DCE-MRI) is a critical tool for breast cancer detection and diagnosis. Yet, its reliance on contrast agents presents risks, increases cost, and is contraindicated for certain patients, necessitating a contrast-agent-free alternative. In this paper, we use a conditional Generative Adversarial Network (GAN) with a multi-scale enhancement consistency loss for contrast agent-free breast DCE-MRI synthesis from a multi-parametric MRI input (T1w, multi-b-value DWI, and ADC maps). The proposed loss explicitly enforces consistency of the predicted enhancement maps across multiple scales. The model generates multiple DCE post-contrast phases simultaneously, reducing inference time and computational overhead. The proposed framework achieved the best performance among previous methods on public and in-house datasets. The code is available at https://github.com/minnelab/DCE-MRI-Syn.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1983_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/minnelab/DCE-MRI-Syn

Link to the Dataset(s)

https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=89096426

BibTex

@InProceedings{YanZhi_Multiparametric_MICCAI2026,
        author = { Yang, Zhikai AND Zhang, Cristina AND Chen, Huijian AND He, Muzhen AND Moreno, Rodrigo},
        title = { { Multi-parametric MRI for Contrast Agent Free Breast DCE-MRI Synthesis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This is an interesting study in which a new way loss function element is suggested in the form of the Multi-Scale Enhancement Consistency (MSEC) which can be used for generation of breast DCE MRI series. The authors validate this method both using a validation set and a external in-house test set.

  • 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 major strength of this article is it’s contribution in for of the somewhat novel formulation of the he Multi-Scale Enhancement Consistency. Similar approaches were suggested in previous works by Fonnegra et al. (Computer in Biology and Medicine 196, 2025 110660) however I would still consider the proposed reformulation of the problem through the inclusion of the Multi-scale effect to be clearly novel in this space.

    I also consider the use of the BMMR-2 Dataset as the training-base to be novel as to my knowledge all previous studies which used multi-parametric MRI input in breast MRI used exclusively in-house datasets for training. An important aspect in this regard is the availability of the segmentation data which allows for the evaluation of the enhancement curve inside of lesion.

  • 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.

    Even tough the authors methods show a good level of originality the work suffers the most from a complete lack of a discussion of the results against recent works presented in the last years. Especially lacking is a comparison of the dynamic aspect results against the works of Schreiter et al (mentioned by the authors in reference 9), but also against works by Fonnegra et al (Computer in Biology and Medicine 196, 2025 110660), Osuala etl al (“Simulating dynamic tumor contrast enhancement in breast MRI using conditional generative adversarial networks,” J. Med. Imag. 12(S2) S22014 https://doi. org/10.1117/1.JMI. 12.S2.S22014) and by Lang et al. (Temporal Neural Cellular Automata: Application to Modeling of Contrast Enhancement in Breast MRI, MICCAI 2025).

    The evaluation of the input-data impact also lacks a discussion against the previous works which evaluated the use of multi-parametric input data for example: Chung et al (Deep learning to simulate contrast-enhanced breast MRI of invasive breast cancer. Radiology, 306(3), p. e213199.), Zhang et al (Synthesis of Contrast-Enhanced Breast MRI Using T1- and Multi-b-Value DWI-Based Hierarchical Fusion Network with Attention Mechanism, MICCAI 2023) and Liebert et al. (Impact of non-contrast-enhanced imaging input sequences on the generation of virtual contrast-enhanced breast MRI scans using neural network. European radiology, 2025 35(5), pp. 2603-2616.)

    Another issue is with some of the experiments. The authors use as one of the methods for comparison an attention based method for which they use just the T1w and ADC maps as input data. Using just such combination for the comparison against the proposed method is not fair as the network lacks the information from the DWI acquisitions. The authors also don’t state at any point of the materials and methods section a reason why just such combination of input data is used for the AAD method.

    From some of the experiments I’m also missing results on the in-house dataset for the baseline model and for the SSEC.

    In regards to the evaluation of the contrast-phases the authors show just the normalized intensities of at the respective contrast-phases as a proof of a better performance. However in clinical routine the aspect which is investigated is the increase/decrease of the signal intensity relative to the contrast-phase 1 or relative to the previous contrast-phase. I would like to invite the authors to perform such an evaluation of their results instead of focusing on just the signal-intensity.

    Another issue is that the authors perform no statistical evaluation of the results (repeated measurement ANOVA with most likely a Friedman test followed by a Nemenyi/Dunn post-hoc evaluation).

    There are several points with small writing errors: Page 4 - Evaluation Protocol Pagraph the is typo in method and in the word generation Page 5 - last paragraph missing space before AAD

  • 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 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?

    I believe that this work shows an overall good level in regards of quality but it doesn’t perform any sort of a discussion against previous work which in my view is an necessary aspect of any good scientific work. However I believe these issues can be addressed in the rebuttal phase.

  • 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 proposes a deep learning framework based on conditional generative adversarial networks, aiming to reduce reliance on contrast agents and address the issues of high computational cost and unclear details in breast DCE-MRI synthesis. The model leverages multi-parametric MRI (T1-weighted, multi b-value DWI, and ADC) to synthesize contrast-free enhanced images. By introducing a multi-scale enhancement consistency (MSEC) loss, it improves the fidelity of lesion structures and dynamic features, while also supporting multi-phase generation to enhance clinical efficiency. Experiments on the BMMR2 and internal datasets demonstrate that this method outperforms existing approaches in terms of PSNR and RMSE.

  • 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)This paper proposes a contrast-agent-free breast DCE-MRI synthesis method using multi-parametric MRI, which can generate multiple post-contrast phases. (2) A multi-scale enhancement consistency (MSEC) loss is designed to optimize the detailed structure of virtual contrast enhancement. (3) The method’s performance and generalizability are validated on both public and private datasets.

  • 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 core contribution claimed by the authors is a conditional generative model that “leverages T1w MRI together with multi-b-value DWI and ADC.” However, the preprocessing section completely omits any mention of spatial alignment. In breast MRI (prone position), DWI acquired via EPI sequences suffers from severe, non-linear susceptibility-induced geometric distortions, particularly near the chest wall. Furthermore, patient motion between the T1w and DWI acquisitions is inevitable. Stacking spatially misaligned T1w, DWI, and ADC slices as input channels violates the basic premise of multi-modal learning. The network is forced to learn from anatomically contradictory inputs, rendering the synthesized DCE-MRI outputs physically uninterpretable for clinical tasks like lesion characterization. (2)The authors attempt to strengthen their paper by including an in-house dataset for external validation. However, the experimental design of this validation introduces uncontrolled physical domain shifts that invalidate any conclusions drawn from it: b-value Mismatch: Training on b-values of 0/100/600/800 and testing on 0/80/500/1000 is physically flawed. The signal attenuation and SNR characteristics at these varying b-values are substantially different. Temporal Resolution Discrepancy: Comparing DCE kinetics from an 80–100s protocol to a 30s protocol without accounting for the vastly different pharmacokinetic phases (e.g., arterial vs. late parenchymal enhancement) is methodologically incorrect. Because the external validation fails to isolate the scanner domain shift from these confounding b-value and temporal shifts, any performance drop observed on the in-house cohort cannot be reliably attributed to a lack of model generalization. (3)The paper only states in Section 2 (Method) that it is based on the Pix2PixHD framework, e.g., “The architecture is based on the Pix2PixHD framework [16], consisting of …”, but does not provide a detailed description of the specific structures of the generator and discriminator, and key network architecture details are missing. (4) In Section 2 (Method), when describing the total loss function, the paper introduces four hyperparameters. Although it explicitly states that the weight coefficient for the MSEC loss is set to 50, it does not provide a detailed explanation of the values of the other weights or the rationale behind their selection. (5) A clear spelling error appears in Section 3.1 “Evaluation Protocol”: “representative contrast enhancement generatiion metho,d.”

  • 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 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.

    (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?

    Generative models in medical imaging must be grounded in physical reality. Because the inputs are not spatially aligned, and the evaluation is confounded by mismatched acquisition physics, the reported quantitative metrics (e.g., SSIM, PSNR) do not reflect the model’s true capability to synthesize clinically viable DCE-MRI. Addressing these issues would require a complete overhaul of the data preprocessing pipeline and a redesign of the validation experiments, which is beyond the scope of a rebuttal.

  • Reviewer confidence

    Very confident (4)

  • [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.

    The rebuttal does not adequately address key concerns regarding multimodal spatial misalignment and protocol-induced domain shifts in external validation; therefore, I recommend rejection.



Review #3

  • Please describe the contribution of the paper

    The paper proposed a conditional GAN-based network to synthesize breast DCE-MRI from multi-parametric MRI input including T1w, multi-b-value DWI and ADC maps. In the network, the authors introduced a multi-scale enhancement consistency (MSEC) loss to enforce the prediction coherency at both global and local image levels by computing the phase enhancement using the full image and downsampled images. The proposed method reported improved qualitative and quantitative results compared to baseline models without MSEC loss and other state-of-the-art methods including diffusion and an attention-based 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.

    1.The paper addresses an important clinical problem: eliminating gadolinium contrast in breast DCE-MRI while preserving diagnostic information. The use of multi-parametric MRI (T1w, DWI and ADC) is well motivated as T1w provides pre-contrast structural images and DWI/ADC provides tumor microstructure information. 2.A notable technical strength is the introduction of the MSEC loss, which explicitly calculates enhancement maps (post-pre subtraction) instead of absolute intensities. It forces learning of contrast dynamics rather than appearance and the multi-scale calculation helps balance global consistency and local detail. The formulation is simple yet well-motivated, and directly tied to the physics of DCE imaging (enhancement relative to baseline). It’s shown to be effective by the superior performance shown in Table 1.3.Another strong point is that the model generates multiple post-contrast phases simultaneously rather than a single phase. This is an important step toward modeling DCE as a temporal process and the paper also evaluated the dynamic enhancement curves in lesion regions in the results (Fig. 3 and Fig. 4). The qualitative and quantitative results suggest improved alignment with the ground truth enhancement trends, which is more clinically relevant than image similarity alone.

  • 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.While the paper claims to model multi-phase DCE, the temporal aspect is still weak. The model generates multiple phases jointly, but there is no explicit temporal modeling across phases. The evaluation of enhancement curves is limited to intensity trends in ROIs, without demonstrating that the model captures true physiological kinetics (e.g., wash-in/washout behavior). This raises the concern that the method is still primarily learning correlated appearance across phases, rather than true DCE dynamics. In addition, 3 out of 5 phases were selected from the BMMR2 dataset for training, and it is unclear at what times these phases were acquired, which makes it difficult to interpret the temporal consistency of the results. 2.The claimed importance of multi-parametric inputs is also not carefully analyzed. While ablations show overall improvements, there is no detailed investigation of which b-values contributed the most or whether ADC alone would be sufficient. A more systematic analysis would help clarify the role of diffusion information in the proposed framework.

  • 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

    1.The network is designed to predict three post-contrast phases and the authors selected three out of five available phases from the BMMR2 dataset. It would be helpful to clarify the rationale behind this choice. Additionally, how does the model scale with a larger number of phases? Does performance degrade when predicting more phases, and is there an observed break point? 2.The paper demonstrates the benefit of incorporating multi-b-value DWI, but a more detailed ablation would strengthen this claim. Specifically, how many b-values are necessary to achieve optimal performance? Is performance primarily driven by the number of b-values, or by the inclusion of higher b-values (i.e., stronger diffusion weighting)? 3.In Fig. 3 (right), please clarify how the normalized intensity is computed. How is the region of interest defined? 4.The qualitative comparisons across figures are somewhat inconsistent. For example, Fig. 3 shows only baseline and SSEC, while Figs. 4 and 5 show only AAD and diffusion-based methods. It would be beneficial to present all comparison methods consistently across figures/tables, or clearly justify the selection shown in each case. 5.In table 1, the baseline model using only T1w appears to outperform the same model using all inputs (T1w, ADC, DWI). This is somewhat counterintuitive, as multi-parametric inputs are expected to provide complementary information. Could the authors provide insight into this behavior? 6.There is a typo in Section 3.1 - “generatiion metho,d” should be “generation method,”.

  • 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 addresses a clinically important problem and presents a well-executed conditional GAN framework using multi-parametric MRI inputs. The proposed multi-scale enhancement loss is well motivated and showed improved performance compared to baseline models. The paper also showed the multiple post-contrast phases and evaluated the enhancement curves. Although there are limitations, particularly the lack of explicit temporal modeling, the work presents a useful and practical contribution. Overall, I believe the strengths outweigh the weaknesses and recommend acceptance.

  • 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 authors addressed the concerns in my review.



Author Feedback

Response to Reviewer 1 1.Lack of discussion against SOTA: We only included AAD and diffusion methods in the comparison because those were the only ones from the ones referenced in the introduction that provided code. As many implementation details are not provided in the other papers, it is difficult to reproduce their results, making any comparison unfair. We will share our code upon acceptance to ease comparisons of methods. 2.Unfair comparison with AAD: We used only T1w and ADC for the AAD baseline because its authors used those inputs in their implementation. We agree that adding DWI to AAD is worth testing in the future. 3.Missing ablation results for the in-house dataset: We only chose our best-performing method for external validation since we were interested in generalizability. Other methods are expected to perform worse. 4.Comparison with normalized enhancement: As shown in Fig. 3, our method matches the ground truth data. We agree that using normalized values could ease the interpretation of results. 5.Statistical Evaluation: We found a highly significant difference in PSNR, RMSE, and SSIM across all the evaluated methods using the Friedman test, followed by post-hoc analysis using the Nemenyi test (p < 0.05). 6.Typos: Will be corrected in the camera-ready version.

Response to Reviewer 2 1.Spatial Alignment and DWI distortions: Unfortunately, BMMR2 does not provide blip-up blip-down data for correcting distortions. Instead, we performed registration to mitigate this issue, but we acknowledge that this might not be enough in some cases, especially when the tumor is close to the chest wall. 2.Generalization using datasets with different protocols: Unfortunately, breast protocols are not standardized, so different datasets usually have different DWI b-values and DCE acquisition times. There are two options to tackle this issue: a) using the closest b-values and times for external testing, as done in the paper, or b) using models to predict the signal at other b-values and acquisition times. For b-values, one could fit the IVIM model and predict the signal at other b-values. Unfortunately, BMMR2 only provides 4 b-values, which can lead to noisy estimates of IVIM parameters. For predicting DCE signal at different acquisition times, one could use the extended Tofts model, for example, but we could introduce a potential bias in the analysis. For that reason, while we recognize option b) is plausible, we believe the selected approach is appropriate for the data at hand. We acknowledge that the decrease in performance on the external dataset may be partially due to the selected approach. 3.Network Architecture Details: We use the exact same architecture from (Ref. 4), but we adapted the loss function as described in the paper. The code Ref. 4 is available at https://github.com/RichardObi/SimulatingDCE. 4.Hyperparameters: For \lambda_{ms-sub}, we tested different values: 1, 10, 20, 50, and 100, with 50 yielding the best results. For the remaining weights \lambda_{adv}, \lambda_{fm}, \lambda_{per}, we used the empirical values suggested in Ref. 4. Response to Reviewer 3 1.Temporal Modeling & Phase Timings: For BMMR2, phases 1, 3, and 5 were acquired at ca. 80, 240, and 400 s after the contrast injection. Thus, they are expected to be representative of DCE on early wash-in, plateau, and late washout stages. To promote the method’s focus on enhancement, we used enhancement maps for training rather than the DCE raw data. We did not add any time modeling to avoid model-related biases in the prediction. 2.Ablation of b-values: This is a good suggestion for future work. We expect low and high b-values to be relevant for DCE prediction: low b-values can capture perfusion effects, and high b-values can capture tumor consistency through diffusivity. BMMR2 covers b-values on both ranges.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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.
    • Concerns about spatial misalignment across different MRI sequences should be addressed.
    • MRI protocol mismatching between experimental validation should be accounted for.
    • Includes more up-to-date methods for baseline comparisons.
  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The main concerns remained, including spatial misalignment and various MRI protocols.



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.

    The paper addresses the clinically important goal of reducing or eliminating gadolinium contrast in breast DCE-MRI while preserving diagnostic value. The problem is clinically relevant, although the results as pointed out by the reviewers must be presented in context due to lack of comparison to state of the art methods and limitations of the data, lack of spatial alignment and how the approach might handle domain shift and limitations should be acknowledged in the discussion in the final paper.



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.

    The paper addresses an important clinical problem and proposes a well-motivated framework for contrast-free breast DCE-MRI synthesis using multi-parametric MRI inputs. Reviewers generally agreed that the proposed MSEC loss is a meaningful contribution and that the experimental results are promising. The main concerns were related to spatial alignment, protocol mismatch in external validation, and comparison against recent methods. The rebuttal clarified several implementation details and reasonably addressed the motivation behind the experimental design and preprocessing choices.



back to top