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

Abstract

Diffusion tensor based cardiovascular magnetic resonance (DT-CMR) offers a non-invasive method to visualize the myocardial microstructure. With the assumption that the heart is stationary, frames are acquired with multiple repetitions for different diffusion encoding directions. However, motion from poor breath-holding and imprecise cardiac triggering complicates DT-CMR analysis, further challenged by its inherently low SNR, varied contrasts, and diffusion-induced textures. Our solution is a novel framework employing groupwise registration with an implicit template to isolate respiratory and cardiac motions, while a tensor-embedded branch preserves diffusion contrast textures. We’ve devised a loss refinement tailored for non-linear least squares fitting and low SNR conditions. Additionally, we introduce new physics-based and clinical metrics for performance evaluation. Access code and supplementary materials at: https://github.com/ayanglab/DTCMR-Reg

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2286_paper.pdf

SharedIt Link: https://rdcu.be/dV1P6

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72069-7_60

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2286_supp.pdf

Link to the Code Repository

https://github.com/ayanglab/DTCMR-Reg

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Wan_Groupwise_MICCAI2024,
        author = { Wang, Fanwen and Luo, Yihao and Wen, Ke and Huang, Jiahao and Ferreira, Pedro F. and Luo, Yaqing and Wu, Yinzhe and Munoz, Camila and Pennell, Dudley J. and Scott, Andrew D. and Nielles-Vallespin, Sonia and Yang, Guang},
        title = { { Groupwise Deformable Registration of Diffusion Tensor Cardiovascular Magnetic Resonance: Disentangling Diffusion Contrast, Respiratory and Cardiac Motions } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15002},
        month = {October},
        page = {640 -- 650}
}

Reviews

Review #1

  • Please describe the contribution of the paper
    1. The authors employ a group-wise registration strategy and propose a tensor-embedded module that can generate pseudo diffusion images to conserve textural information.
    2. The authors employ a combined loss for non-linear least squares fitting to increase high quality.
    3. The authors conducted extensive experiments on two datasets.
  • Please list the main strengths of the paper; you should write about 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 authors employ a group-wise registration method with an implicit template, which is constructed in Turker form. The template can guide the b-spline registration network.
    2. The authors propose a tensor-embedded module to generate pseudo diffusion images with the same anatomy but different contrasts to guide the group-wise registration.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. Lack of novelty. Almost all mentioned methods exist. Firstly, the group-wise registration is proposed in [1]. Secondly, the tensor-embedded contrast conservation is similar to the Diffusion Encoding Suppression as well as the B-spline-based registration in [2]. Thirdly, the mentioned evaluation criteria, such as the number of negative eigenvalues and helix angle gradient line profile already, are utilized in [2].
    2. Lack of comparisons with state-of-the-art methods, such as [1] and [2].

    [1] Magat, Julie, et al. “A groupwise registration and tractography framework for cardiac myofiber architecture description by diffusion MRI: an application to the ventricular junctions.” Plos one 17.7 (2022): e0271279. [2] Wang, Fanwen, et al. “Efficient post-processing of diffusion tensor cardiac magnetic imaging using texture-conserving deformable registration.” Medical Imaging 2024: Clinical and Biomedical Imaging. Vol. 12930. SPIE, 2024.

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    Please see the details on question 6.

  • 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

    Reject — should be rejected, independent of rebuttal (2)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Almost all mentioned methods exist.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • [Post rebuttal] Please justify your decision

    Thanks the authors for their rebuttal. After reading the rebuttal, I would like to raise my score to 3. The main reason is still the missing comparisons to its closely related work. I hope the authors can improve the manuscript on this end. Thank you.



Review #2

  • Please describe the contribution of the paper

    Registration methods based on intensity similarity or theoretical likelihood for MRI prove inadequate for cDTI datasets (diffusivity information is encoded as signal loss across images). Thus, registration has to be performed for different contrasts depicting the same image content. The authors propose using an implicit template as the anchor for two concurrently trained models: 1) group-wise registration network with diffeomorphic transformations, and 2) tensor-embedded generator model creating pseudo-diffusion images to guide registration. Their hypothesis suggests that a model generating diffusion contrasts performs better concerning contrast variations. The authors explore the utility of segmentation labels as auxiliary loss for improved registration performance. The method outperformed traditional and deep learning registration approaches wrt two novel evaluation metrics (rel. number of negative entries in diffusion tensors, quality of the linear fit of the Helix angle profile).

  • Please list the main strengths of the paper; you should write about 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.
    • Language and Structure: The manuscript is well-written and structured.
    • Motivation: The motivation for the work is articulated clearly. Adaptations compared to similar works are clearly delineated. Novel contributions are identified in the tensor generation branch, the application of adapted loss functions, the use of a physics-informed quality metric, and the application of self-supervision using segmentation masks.
    • Data: The registration method is evaluated using real cDTI data acquired from different field strengths, mirroring scenarios commonly encountered in clinical practice.
    • Results: The proposed method is compared against both traditional and deep learning-based registration methods, providing a comprehensive assessment of its performance against existing approaches. Furthermore, an ablation study is conducted, enabling the assessment of the impact of individual changes to the overall approach. Differences in performance with and without denoising suggest potential achievement of the goal of enhanced fitting under low SNR conditions.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    • While the novel metrics show higher performance for the proposed approach, the lack of traditional registration quality metrics limits interpretability and comparability.
    • While the selection of negative eigenvalues is a physics-informed prior that is straight forward with respect to its meaning, this is less clear for the clinical marker, the helix angle gradient. First, helix angle gradient linearity may be preserved in pathological conditions and is by no means indicative of a healthy heart. Secondly, this registration will have to perform on patient data and not on healthy subjects. The introduction of a metric for “healthy” hearts therefore appears counter intuitive. Clinical cDTI studies further suggest that other tensor parameters like MD, FA, tensor shape metrics or sheetlet mobility may be more relevant, while being rather unspecific. The sole application on data from healthy volunteers may thus be considered as a limitation of this work. When considering the helix angle gradient as not necessarily representative in this evaluation, the negative eigenvalue results of the proposed method compared to standard rigid registration depicts only a minimal difference (1.244% to 1.251%), which questions the usefulness of the proposed approach.
    • SNR directly effects the accuracy of tensor reconstruction and therefore both metrics used. Using data from two different scanners with different field strengths may likely have led to SNR differences in the data. While the authors claim that the proposed approach is meant to enhance frame fitting performance under low SNR conditions, this aspect is only superficially assessed in the ablation study
    • Data is not made public, thus the reproducibility is limited.
  • Please rate the clarity and organization of this paper

    Excellent

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    • Helix angle metric: Please consider adding limitations of the metric as stated in the weaknesses section.
    • Negative Eigenvalues: Why is the NE metric shown in median plus standard deviation? Please use e.g. quantiles (p25, p75) instead of the standard deviation when reporting median values. NE% (arrow up) should be NE% (arrow down). Please use concise notation (table 1: NE%, table 2: Percentage of NE).
    • “conventional abd deep-learning based techniques” should be: conventional and deep-learning based techniques
    • Figure 3: Caption: Please clearly state what is seen in which row. Colorbars: Colorbars are vertically misaligned in row 2, 3 and 4. Colorbars in row 1 can be summarized to one colorbar on the right.
  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Despite the limitations such as the questionable usefulness of the HAG metric and the absence of reproducibility due to the lack of a public dataset, the manuscript’s strengths lie in its very good organization, the usage of real cDTI data, and comprehensive ablation study as well as results (comparison to DL and non-DL methods). These aspects contribute to the overall assessment of the manuscript.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    In part, my comments were sufficiently addressed. The authors’ commitment to adding a discussion on the limitations of the HAG metric is acceptable. However, the response regarding the usefulness of the proposed method lacks clarity. The acquisition scheme of STEAM uses two separate heartbeats to encode diffusion, relying on consistent myocardial positions across heartbeats. This results in diffusion times that enable short echo times, minimizing the impact of motion. Any motion during the scan would be encoded through the diffusion gradients, rendering the images unusable. Consequently, STEAM data is acquired in breath holds and at cardiac cycle sweet spots where strain influence is negligible. Even if deformable image registration is still necessary, it remains unclear why this is the case given the data acquisition process (potential offsets in breath hold positions?). Furthermore, this does not address the question why rigid registration achieves almost identical results to the proposed method under these conditions (w.r.t. the percentage of negative eigenvalues).

    Additionally, while the explanation regarding SNR is generally acceptable, the difference between the original (SNR = 9.28) and denoised (SNR = 9.38) is minimal. Further clarification is needed on how such a small difference impacts the registration process.

    All in all, I’ll not change my initial rating.



Review #3

  • Please describe the contribution of the paper

    The authors’ application scenario is a novel imaging modality. The authors propose groupwise registration and motion disentangled approach to solve the DT-CMR image problem. The authors introduce new evaluation metrics, which are relevant to the task.

  • Please list the main strengths of the paper; you should write about 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 authors identified problems in DT-MRI images and proposed solutions accordingly. The authors introduce novel physics-based and clinical metrics for performance evaluation, and present them graphically, as in Figures 2/3/4.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    The results of the experiments are somewhat confusing. The authors propose ‘Optional semi-supervised version’, is this result shown in S4.b in Table 1? Then, why the result is not as good as the ‘Proposed’ one? The authors propose two new metrics, HAG and NE%, but detailed calculations are not given. Why not use the popular DSC and smoothing metrics? Why is there no GROUND TRUTH in the Figures 3 and 4? There are some explicit errors. For example, in Tables 1 and 2, the ‘Percentage of NE’ indicator arrow is in the wrong direction? There are some grammatical errors, such as the appearance of two consecutive ‘at’ in the fifth line of section 2.3.

  • Please rate the clarity and organization of this paper

    Very 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    The authors used a more complicated training framework, but the results do not show a significant improvement compared to Transmorph. What if compared to more advanced methods?

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    New applied image modalities, new evaluation metrics, but some details are unclear.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Author Feedback

We appreciate positive comments from reviewers: “novel groupwise registration network with an implicit template for motion separation, a tensor-embedded module for contrast separation, and a refined loss for noisy frames (R3, R4, R5)” and “novel physics-based and clinical metrics were proposed (R3, R5)”. To address other comments:

Not using DSC as evaluation metrics/Semi-supervised version fails(R3, R5) We didn’t use DSC or similar metrics because DTCMR is not a proper modality for anatomical structure delineation and unreliable ground truth segmentation could be produced due to low SNR and unclear boundaries. Moreover, manual delineations of 60+ frames per case were impractical. Although we used automatic segmentation (10.1002/mrm.28294) with denoising, inaccurate delineations were produced due to subtle deformations. Thus, segmentation-based metrics were not used, and we had also shown that incorporating DSC loss could not improve performance (see S4.b in ablation study).

Limitation/Lack of details of metrics(R3, R5) The proposed work was a proof-of-concept study on healthy cohorts. Due to page limits, detailed metric calculations were omitted, but the HAG limitation will be added in the final version.

Usefulness of proposed method(R3) Despite small differences between Rigid and proposed method, deformable motion occurs during the cardiac cycle (10.1002/mrm.28294), and heart margins move (10.1148/radiology.209.2.9807578), making deformable registration essential.

Impact on SNR(R3) We didn’t measure SNR directly and further studies can quantify SNR’s impact on registration accuracy. In literatures (P2), structural MRI (e.g., T1 mapping in ref [7]) generally has higher SNR. Hence, we claimed that working with low SNR DTCMR is more challenging. Due to page limits, we conducted ablation studies on the original (SNR (mean(std) = 9.28(4.88)) and denoised images (SNR = 9.38(4.96)) with SNR defined as myocardium mean over background corner mean. Denoising improved registration performance, indicating higher SNR led to better accuracy. This underscored the importance of enhancing SNR in DTCMR image registration.

Data availability(R3) We don’t have ethical approval to share data openly for this retrospective study.

Lack of novelty(R4) R3 and R5 acknowledged our innovations. R4 mentioned [1] for groupwise, but it applied to ex-vivo hearts only, without signal loss or deformable motion, focusing on atlas-based cardiac analysis, which differed from the motion correction task we faced. [1] collected high-resolution structural images (0.15mm) to support diffusion (0.6mm), while in our more challenging case with in-vivo low SNR diffusion images (2.8mm), we didn’t use structural counterparts. R4 referenced [2] a contrast conservation step; however, disentangling anatomy and contrast remained questionable. Additionally, [2] chose the brightest frame as fixed, which was not optimal (ISMRM 2024, Abstract #2141). Our proposed method embedded the tensor information as a constrain for grouped frames with refined loss, unlike [2] used five different losses with extensive hyperparameter tuning just for disentangling.

Lack of comparison(R4) R3 praised our comprehensive performance assessment against existing approaches. As mentioned, [1] need paired structural images and aimed at high-resolution cardiac analysis, which differed from our task. According per MICCAI policies, additional experimental results are not allowed in the rebuttal. We refer to the outputs in the SPIE Arxiv version of [2]: for R2 (mean(std)), RMSE, and NE% (median [p25, p75]), [2] reported 0.895(0.054), 6.865(3.481), and 1.7[0.6, 3.1], respectively. All metrics were worse than our proposed method.

Groundtruth (GT) in Fig 3/4(R5) No GT labels were available for single-frame DTCMR, and segmentation labels were not evaluated. Similarly, no GT labels were available for our registration task.

Minor issues We apologize and will correct the typos in the final version.



Meta-Review

Meta-review #1

  • 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’

    I agree with the positivie statements of the reviewers that this paper provides an interesting, novel methodology to address an important issue. While not fully satisfied with the rebuttal, one score was raised further and they comment on being partially satisfied. All in all this paper was positively received by the reviewers and particularly their agreement on novelty and interest for the MICCAI community lead me to recommend acceptance.

    Groupwise Deformable Registration of Diffusion Tensor Cardiovascular Magnetic Resonance: Disentangling Diffusion Contrast, Respiratory and Cardiac Motions

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    I agree with the positivie statements of the reviewers that this paper provides an interesting, novel methodology to address an important issue. While not fully satisfied with the rebuttal, one score was raised further and they comment on being partially satisfied. All in all this paper was positively received by the reviewers and particularly their agreement on novelty and interest for the MICCAI community lead me to recommend acceptance.

    Groupwise Deformable Registration of Diffusion Tensor Cardiovascular Magnetic Resonance: Disentangling Diffusion Contrast, Respiratory and Cardiac Motions



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

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    N/A



back to top