Abstract

Magnetic resonance imaging (MRI) enhanced by the gado-linium-based contrast agents (GBCAs) is crucial in the assessment and management of cancer. However, the use of GBCAs introduces additional costs and raises potential safety concerns, including the risk of gadolinium accumulation in brain. Several generative learning methods based on GANs and diffusion models have been proposed to generate contrast-enhanced MRI from non-contrast-enhanced MRI. However, GANs face challenges such as gradient vanishing and mode collapse. Diffusion models also face several challenges, such as generation instability and long sampling times. In this paper, we propose a controllable flow matching (CFM) model for efficient synthesis of 3D contrast-enhanced brain MRI with fine-grained details of targets of interests. CFM adopts a straight-line generation path, enabling the generation of images in a single step. We design a multi-stage training strategy integrating controllable constraints to ensure that such a single-step sampling generating contrast-enhanced MRI meet specific controllable conditions. Our CFM model has been evaluated on both the BraTS2023 and an in-house datasets. Experimental results demonstrate that CFM led to state-of-the-art image generation and tumor delineation performance with promising generalizability. Our codes can be found at https://github.com/ladderlab-xjtu/CFM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1113_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ladderlab-xjtu/CFM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ChaHen_Controllable_MICCAI2025,
        author = { Chang, Heng and Shang, Yu and Wang, Haifeng and Liang, Yuxia and Wang, Haoyu and Wang, Fan and Niu, Chen and Lian, Chunfeng},
        title = { { Controllable Flow Matching for 3D Contrast-Enhanced Brain MRI Synthesis from Non-Contrast Scans } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15975},
        month = {September},
        page = {118 -- 127}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors proposed a controllable flow matching model for synthesizing contrast-enhanced brain MRI. The controllable generation is be implemented as a segmentation loss and is used in the proposed multi-stage training. The experiments show that the generation quality is better on BraTS2023 and in-house dataset and the segmentation performance is also higher.

  • 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 motivation behind the proposed method is intuitive and the result seems promising.
  • 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 reviewer would like to know what the segmentation model is used in Tab. 2. If it is the same model used in training CFM (i.e., SwinUNETR), the segmentation results are biased to the optimization of generating segmentor-prefered images. As a result, the authors should provide the result from other segmentation models (e.g., nnUNet).
    2. Since the downstream application is tumor segmentation, the authors should also present the generation results in the tumor area separately to demonstrate the superiority of the proposed method. Also, the proposed CFM only outperforms FM by a small margin on BraTS2023 (250 test samples). While the performance is better on the in-house dataset, there is only 73 test samples in it, rendering the performance gain not significant.
  • 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.

    (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?

    Even though the motivation seems intuitive, the performance of the proposed method needs complementary experiments to prove its effectiveness.

  • 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 an image translation network, to generate gadolinium-enhanced T1 from T1w MRIs. The authors turn to controllable flow-matching models, citing the instability of GANs and long inference times of diffusion models.

    The authors include an auxiliary segmentation task and a multi-stage training regime.

  • 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 method, according to the evaluation, outperforms the others. Generalization capabilities are assesed. The use of an auxiliary task is well motivated and segmentation results are convincing. The paper is structured well enough.

  • 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 paper suffers greatly from its obtuse notation. Notably, $v$ (a crucial part of the method) is often said to not change wrt $t$, yet is always noted as $v(X, t)$. This hampers the rest of the paper, where $t$ is either set to 0 or sampled from a uniform distribution. While this distinction seems important, since $v$ is invariant to $t$, the paper cannot let the reader develop the intuition why that is. Second 2 and 3 would greatly benefit from a rewrite. Section 3.2, especially, could be greatly simplified and its notation lightened. Algorithm 1 does not bring anything meaningful and, considering the limited format of MICCAI, the space could have been better utilized for more results, for example.

    One of the key arguments against diffusion models, their inference time, lacks motivation. Indeed, while timely diagnostic of brain tumors is important, I am doubtful of what benefits saving a few seconds (according to the results table) may add. Is there clinical motivation behind this claim that the inference time of diffusion models, in the order of seconds, is prohibitive ?

    The method also seemingly only bring marginal improvements over competing methods. Looking at qualitative results, the tumor reconstruction seems overly smooth compared to DDIM and ControlNet++.

    No ablation study is included. It is unclear what improvements the multi-stage training scheme or auxiliary task bring.

    The authors unfortunately do not provide a public implementation of their method.

  • 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 provide sufficient information for 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

    Section 2.2 mentions paired T1ce and T1w images. Were they co-registered ?

    What are the acquisition parameters of the in-house dataset ? What make and model of scanner ?

  • 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?

    While the proposed work is on the surface good enough, many aspects would require a major revision to the paper for it to be up to the standards of MICCAI. However, it is doable. Sections 2 and 3 need to be adapted to be easily followed for a reader not familiar with flow-matching models. The results need to be presented in a more convincing manner.

  • 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 authors have sufficiently adressed the reviewers’ comments. While the overall results are still fairly unconvincing (the main point driving for a rejection), the authors have justified them better in their rebuttal.



Review #3

  • Please describe the contribution of the paper

    The primary contribution of this work is the development of Controllable Flow Matching (CFM), a novel and highly effective method for synthesizing 3D contrast-enhanced brain MRI from non-contrast scans. The authors successfully adapt flow matching principles, introducing a straight-line path for efficient one-step generation and cleverly integrating segmentation-based controllability.

  • 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.
    • I like the idea of using flow matching to do cross-modality MRI image generation, including the way the authors involve segmentation mask into the training.
    • The paper is written well and easy to follow.
    • The experiments are solid.
  • 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 still gap (~0.2 in dice) between the proposed method and T1c_GT. Did authors did deeper analysis to see what is missing there?
  • 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 provide sufficient information for 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?

    It is good paper with nice idea. The results are solid as well.

  • 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.

    My concern was addressed in the rebuttal letter.




Author Feedback

We thank in-depth reviews and appreciate for affirming our contributions. The main concerns are addressed below.

  • [R1] - The gap (~0.2 in Dice) between the proposed method and T1c_GT.

As pointed out by the reviewer, we have analyzed the relevant factors. Some BraTS2023 images suffer from poor acquisition (e.g., blurriness), leading to distribution differences between generated images and T1c_GT. In addition, using a segmenter pre-trained on T1c_GT, some generated T1c images show suboptimal segmentation results.

  • [R2] - Choice of the segmenter

1) SwinUNETR is used in Table 2, as stated in Implementation Details. We also compare with ControlNet++, which incorporates segmentation optimization.

2) Since this work mainly focuses on integrating controllable constraints into flow matching, we consistently use the SwinUNETR model pre-trained on T1c_GT as the segmenter, with its parameters frozen during training.

3) We will demonstrate the inference performance of other segmentation methods in our future work.

  • [R2] - Evaluation of tumor generation quality

The reviewer’s forward-looking suggestion aligns closely with our considerations.

1) We have tried tumor-based evaluation via bounding boxes but found it hard to define a consistent region due to tumor variability, such as bounding boxes may include irrelevant pixels. Future work will include more suitable tumor-specific metrics.

2) Qualitative tumor results are shown in visualizations.

3) Tumor generation quality can be indirectly justified by the downstream segmentation performance.

  • [R2] - Overall generation evaluation 1) As the reviewer noted, CFM performs similarly to FM overall, but shows clear visual differences in tumor regions.

2) Tumor region segmentation improves significantly, suggesting our method better captures tumor structures.

3) Due to the difficulty in collecting glioma data, the in-house dataset contains only 73 samples. these are real clinical cases. Directly testing the trained model on it demonstrates CFM’s better generalization.

  • [R3] - Sign of $v$

Thank you for the careful observation regarding the notation.

The theoretical definition of $v$ remains unchanged as $X_1 - X_0$ , which is also the target of network optimization. The network outputs $v_θ(X_t, t)$, as stated in Equation (5): ” $v_θ(·,·)$ denotes a velocity prediction network”. Thus, the output varies depending on the input $X_t$ and time step $t$. We will revise the paper and clarify the notation to provide a more precise description of the flow matching framework.

  • [R3] - Comparison with diffusion models

1) Inference speed partially reflects the performance of deep learning approaches.

2) As shown in the visualization results, Diffusion models generate noisier results due to error accumulation over multiple sampling steps.

3) Similarly, the less smooth results of DDIM and ControlNet++ compared to FM and CFM are also caused by the presence of noise. Smoothness in FM/CFM may result from blurred data in BraTS2023.

  • [R3] - Ablation study explanation

Section 4.2 describes the baseline FM without controllable constraint, trained under the same framework (Algorithm 1), serving as an ablation. Due to MICCAI formatting limits, we couldn’t include experiments that only introduce the controllable constraint without Algorithm 1. We will present a more ablation study in our future work.

  • [R1, R2, R3] - Reproducibility

To ensure reproducibility, we will publicly release the PyTorch implementation code upon publication.

  • [R3] - Data Details

1) Section 2.2 mentions that the paired T1ce and T1w images are co-registered.

2) The in-house dataset was acquired using a 3.0-T whole-body MR scanner (Discovery MR750w; GE Healthcare; Milwaukee, USA) with a 24-channel head coil. The scanner parameters were: slice thickness = 1 mm, number of slices = 140, imaging matrix = 256 × 256, and voxel size = 1 × 1 mm. Such acquisition differences from the training data (BraTs) further demonstrate our method’s strong performance.




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’

    The paper presents a controlled flow matching method to generate contrast enhanced MRI images of brain tumors. Flow matching is a relatively new approach and using tumor segmentation task as an approach to constrain generation is novel. The presented results demonstrate feasibility of the approach. The reviewers identified a few weaknesses including improving the explanation of the method such as the segmentation method used for tumor segmentation, which we hope the authors will address in the revised submission. Overall, the paper is technically novel with respect to medical application, is well written, clearly motivated, and results show feasibility with strengths outweighing the weaknesses including more detailed ablation studies as the reviewers remarked.



back to top