Abstract

4D Flow MRI is a promising imaging sequence that provides 3D anatomy and velocity along the cardiac cycle. However, hemodynamic biomarkers are susceptible to degradation due to the low resolution of the imaging modality, which can compromise vessel segmentation. In this study, we propose a novel deep-learning approach, named SURFR-Net, that combines both super-resolution and segmentation tasks, leading to a super-resolved segmentation. SURFR-Net is based on the RCAN super-resolution network, modified to handle a multi-task problem. A novel handcraft feature, named Weighted Mean Frequencies (WMF), has been introduced with the objective of assisting the network in differentiating between pulsatile and non-pulsatile fluid regions. Moreover, we demonstrate the use of WMF feature as input to enhance super-resolution and provide a more relevant segmentation on 4D Flow MRI images. The proposed solution has been shown to outperform the state-of-the-art solution, SRFlow, in terms of direction and quantification error on systolic and diastolic times with a maximum gain of 4.1% in relative error. Furthermore, this study demonstrates the benefit of combining the super-resolution with the segmentation in a multi-task framework on both outcomes. In conclusion, the proposed solution has the capacity to facilitate a super-resolved segmentation of the aorta, thereby potentially addressing the primary concern regarding 4D Flow MRI parietal biomarkers, such as wall shear stress.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4519_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)

N/A

BibTex

@InProceedings{PerSim_Superresolution_MICCAI2025,
        author = { Perrin, Simon and Levilly, Sébastien and Mouchère, Harold and Serfaty, Jean-Michel},
        title = { { Super-resolution and segmentation of 4D Flow MRI using Deep learning and Weighted Mean Frequencies } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors identify the issue of sensitivity of first-order differential biomarkers, in particular wall shear stress, to artery wall segmentation and imaging noise in 4D flow MRI images. To address this, they propose to fuse the tasks of segmentation and super-resolution for more finely-resolved 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.

    The presented evaluation suggests good performance, beating a previous baseline which is claimed to be state-of-the-art.

  • 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 section “2.1 Data” is hard to follow. It was unclear to me what the size of the training and test sets were. A schematic should be provided for the network layout. Furthermore, the presented method seems like incremental improvement, for which more baseline comparison would be good to clearly identify the superiority of the method.

  • 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 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
    • 2.1 Data:
      • What does it mean for CFD to be “downsampled directly in the k-space”?
      • There is a trailing “CFD” at the end of the second paragraph.
      • What is the factor 2 in the nearest neighbour interpolation?
      • What does it mean for the “five rotations [to] maintain the patch at a size of 32 × 32 × 16”?
    • 2.2 Weighted mean frequencies:
      • What does $*$ denote in (2)? If it is convolution, how can the energy functional be convolved with a frequency? Of which signal is $f^i$ the strictly positive frequencies?
      • What do u, v and w denote in (3) and how do they relate to u in (2)?
  • 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?

    The section “2.1 Data” is hard to follow. It was unclear to me what the size of the training and test sets were. A schematic should be provided for the network layout. Furthermore, the presented method seems like incremental improvement, for which more baseline comparison would be good to clearly identify the superiority of the method.

  • 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.

    Even though the authors address many of the concerns I raised and this is very welcome, at this point I cannot recommend acceptance. I encourage the authors to keep working on the manuscript (with the feedback given above) and submit it elsewhere.



Review #2

  • Please describe the contribution of the paper

    Combination of super-resolution and segmentation networks, and weighted mean frequencies for segmentation of 4D flow MRI datasets. Denoising non-fluid domain for extracting features guiding 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.

    Synergistic combination super-resolution and segmentation. Weighted mean frequency technique facilitating segmentation of 4D flow MRI.

  • 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.
    • Comparison to only one method SRFlow.
    • Many typos and grammatical errors.
  • 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

    There are many typos and grammatical errors, which decrease readability.

  • 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?

    Creative approach, yet the writing lacks clarity

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    Many typos and grammatical errors are fixed in this revised paper, improving its readability.



Review #3

  • Please describe the contribution of the paper

    The paper presents a novel deep neural network, whose architecture is named SURFR-Net, to super-resolve and segment 4D Flow CMR images of the aorta. The deep neural network leverages the use of Weighted Mean Frequencies (WMF) as input, which improves the capability to discriminate between voxels in the aortic lumen and those out of it.

  • 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 paper presents a novel deep neural network (SURFR-Net), which leverages on a novel method to improve the classification of voxels falling inside the aortic lumen (WMF). The presence of these two novel aspects is a major strength. Two further strengths are represented by i) the fact that the SURFR-Net performs two tasks, i.e., super-resolution and segmentation, and ii) the superiority vs. a state-of-the-art architecture. Finally, the analysis of the SURFR-Net performance with and without the contribution of WMF, as well as with different values of the weight that is relevant to the loss function, clearly shows the contribution of the WMF feature.

  • 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.

    I did not find major flaws nor major limitations. I only have some minor remarks:

    There are some typos here and there. For instance:

    Section 2.1, page 2 - In the sentence “MRI data are constituted of 9 1.5T images from patients whose presented a cardiovascular event”, “whose” should be “who”.

    Section 3.2, page 7 - In the sentence “SURFR-Net is 0.6 and and 4.13% better than …”, “and “ is repeated

    Section 3.2, page 7 - In the sentence “Thus, the network can be misled by low and steady velocity region …”, “region” should be “regions”

    In Table 1, it would be appropriate to indicate the unit of measures of the various metrics.

    Reference 1 is a paper under review by the same authors. It is unusual to cite a paper under review, especially if it the citation is a self-citation.

  • 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?

    To me, the paper seems the most innovative one among those I reviewed. It is well organized and well written. It does not have any significant weakness.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 final opinion is based on the level of novelty and the potential impact of the proposed method, which compensates for the limitations pointed out by the other authors. Of course, I hope the authors will fix typos in the text and in formulas, and will improve the clarity of the paper based on the feedback they received from the other reviewers.




Author Feedback

We are sincerely grateful to the reviewers for their insightful feedback, and particularly for their meticulous attention to detail in identifying typographical errors.

R#2: What does it mean for CFD to be “downsampled directly in the k-space”? Due to article length limitations, we could not fully detail the downsampling method for CFD. To elaborate, we calculate the complex image of the CFD using the following equation: $C=Me^\frac{iπ​}{VENC}V$, where M is the anatomic image marked by the velocity, C is the complex image, and V denotes the velocity image. Subsequently, this complex image is transformed into k-space via a Fourier transform. High frequencies are then cuted (specifically, half of the frequencies for a downsampling factor of two). Finally, an inverse transformation is applied to obtain the subsampled image. This approach is the same as that described by Ferdian et al. (2020).

R#2: What is the factor 2 in the nearest neighbour interpolation? In the sentence, “As preliminary work, low resolution segmentations are upsampled using nearest neighbour interpolation with a factor of 2.”, the “factor of 2” refers to the interpolation itself. Specifically, we are interpolating the segmentation reference by a factor of 2 using the nearest neighbour method. We agree that the current phrasing is unclear and will revise it for better comprehension.

R#2: What does mean the “five rotations [to] maintain the patch at a size of 32 × 32 × 16”? When stating “five rotations [to] maintain the patch at a size of 32 × 32 × 16,” we are referring to the fact that for a block of size n×n×m (where n corresponds to the x and y dimensions, and m to the z dimension), there are five possible rotations that preserve its original dimensions. Specifically, along the x-axis, three rotations (90°, 180°, 270°) maintain the dimensions. For both the y-axis and the z-axis, only one rotation (180° for each) achieves this.

R#2: What does $$ denote in (2)? If it is convolution, how can the energy functional be convolved with a frequency? Of which signal is $f^i$ the strictly positive frequencies? The $$ symbol in Equation (2) is a typographical error on our part. We are performing a multiplication, not a convolution, consistently with the used weighted average. This typographical error will be corrected.

R#2: What do u, v and w denote in (3) and how do they relate to u in (2)? u, v, and w refer to the velocity components along the x, y, and z axes, respectively, which is a standard naming convention in fluid mechanics. However, we recognize that the variable $u$ in Equation (2) was intended to denote an arbitrary velocity component, as specified beneath the equation. We fully understand the potential for confusion regarding these symbols and acknowledge the need for clearer differentiation and explicit definition of variables. To improve readability, we will modify the variable $u$ in Equation (2).

R#2: Network architecture schematic We had intended to include a schematic illustrating the network architecture. However, due to the paper’s length limitations, we are unable to do so.

R#3: Comparison to only one method SRFlow. Our comparative analysis is exclusively conducted with the SRFlow method. This choice is justified by its recent publication (2022), its demonstrated potential, and its architectural similarities to our own proposition, notably its use of channel attention and long skip connections. Although alternative recent 4D flow MRI super-resolution techniques are available, they typically belong to distinct deep learning paradigms, such as physics-informed or ensemble learning approaches. Furthermore, to our knowledge, no other existing methods simultaneously offer super-resolution and high-resolution segmentation for 4D flow MRI, which is the main contribution of our paper.




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’

    2 reviewers recommend accept and highlight that their concerns have indeed been addressed. While the third reviewer still recommends rejection, the addressed comments are highlighted in their post-rebuttal feedback with no clear new reasons raised for rejection.



back to top