List of Papers Browse by Subject Areas Author List
Abstract
Quantitative $T_1$ mapping by MRI is an increasingly important tool for clinical assessment of cardiovascular diseases. The cardiac $T_1$ map is derived by fitting a known signal model to a series of baseline images, while the quality of this map can be deteriorated by involuntary respiratory and cardiac motion. To correct motion, a template image is often needed to register all baseline images, but the choice of template is nontrivial, leading to inconsistent performance sensitive to image contrast. In this work, we propose a novel deep-learning-based groupwise registration framework, which omits the need for a template, and registers all baseline images simultaneously. We design two groupwise losses for this registration framework: the first is a linear principal component analysis (PCA) loss that enforces alignment of baseline images irrespective of the intensity variation, and the second is an auxiliary relaxometry loss that enforces adherence of intensity profile to the signal model. We extensively evaluated our method, termed ``PCA-Relax’’, and other baseline methods on an in-house cardiac MRI dataset including both pre- and post-contrast $T_1$ sequences. All methods were evaluated under three distinct training-and-evaluation strategies, namely, standard, one-shot, and test-time-adaptation. The proposed PCA-Relax showed further improved performance of registration and mapping over well-established baselines. The proposed groupwise framework is generic and can be adapted to applications involving multiple images.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2857_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2857_supp.zip
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Zha_Deeplearningbased_MICCAI2024,
author = { Zhang, Yi and Zhao, Yidong and Huang, Lu and Xia, Liming and Tao, Qian},
title = { { Deep-learning-based groupwise registration for motion correction of cardiac T1 mapping } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15002},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work introduces a deep learning approach for motion correction of (pre and post contrast) T1 mapping cardiac MRI data. The approach uses both a PCA-based and physics model based loss to achieve template-free registration. A comparison is made between in-domain (pre-contrast) and out-of-domain (post-contrast) performance
- 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.
Interesting methodology, particularly the combination of a PCA-based groupwise registration loss with a physic-based model
- 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.
Small dataset (50 subjects total and only 10 subjects for testing) Limited evalutation (only standard deviation of T1 maps)
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 proposed methodology is interesting and very relevant to the field. However, it is not clear how much of this is novel or if it is oversold by the authors. Group-wise registation with a PCA-based loss function is not new so the exact contribution of this paper should be clearly stated
The title should mention deep learning. Sometimes it is not clear that you are proposing deep learning-based registration.
The evaluation is very limited, already the maps look quite well without motion correction and it is difficult to see the benefit of any motion correction method. The authors quantify this with the standard deviation of the maps but this is flawed. Registration involves interpolation so will always smooth the output maps relative to the raw maps – this does not necessarily show the effect of motion correction. The method is not compared to some more establised methods e.g. a non-deep learning approach. So it is hard to get a feeling for the relative performance and the evaluation is not clinically relevant.
The authors suggest it is a general purpose method for quantitative CMR but I suggest to reduce the focus of the writing to T1 mapping as this is the only data you show here.
- 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 Reject — could be rejected, dependent on rebuttal (3)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main limitation is the lack of evaluation and comparison to other methods
- 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 Reject — could be rejected, dependent on rebuttal (3)
- [Post rebuttal] Please justify your decision
My opinion is unchanged as the approach seems to be a relatively simple extension to deep learning registration using a loss that was well established for traditional optimisation-based registration, and there is very limited evaluation with only 10 patients in test set.
Review #2
- Please describe the contribution of the paper
The paper focuses on groupwise registration of cardiac MR images in order to perform relaxometry for quantitative imaging. The approach consists mostly in the definition of two losses, one for enforcing the low rank nature of intensity profiles and the second one to minimize the discrepancy in relaxometry parameter estimation
- 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 paper is fairy well presented, with a clear set of contributions.
- The two additional losses are easy to understand.
- The authors have performed a thorough set of experiments to evaluate the performance of the method including out of domain (post-Gd) ones. Statistically significant differences were reached for most configurations
- 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 PCA loss has been already proposed by Huizinga et al. in 2016 in paper whereas the relaxometry is a standard L2 norm loss of reconstructed parameters. Therefore the novelty of the 2 losses is still limited.
- The paper content can be improved in several ways (see below).
- The performance metrics based on standard deviation should be discussed (see below)
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 use SD as the performance metrics but they never say explicitly what it means. I suspect this is standard deviation of the relaxometry parameters. Please state this explicitly in the paper.
- The authors also should discuss why SD is a relevant metrics in addition to cite paper [9]. In particular, one expects that the metrics should be in terms of errors in the relaxometry parameters k, C and T1\star or in terms of errors of registrations. For instance, this type of metrics were used in [8] with simulated T1MOLLY datasets. By minimizing the standard deviation, it is questionnable whether the authors are minimizing the correct performance metrics.
- In p 5 it is mentioned that “The mapping module is pre-trained… in a fully self-supervised fashion”. It is not clear at all how this is done. This requires the knowledge of the relaxation parameters. How are they estimated ? How self supervision is performed ?
- Paper improvements :
- p 5 top “by non-differentiable least-square methods” least square methods are differentialble (see [8] for instance). Why is not differentiable ?
- 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?
- The paper methodological novelty is limited but the experiments are well done and they show a limited but significant gain with prior work.
- 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 Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
The authors have justified using the standard deviation criterion as performance metrics but they have not addressed the comparison with simulated ground truth data as done in other similar publications.
Review #3
- Please describe the contribution of the paper
A deep learning group-wise registration framework is proposed that does not rely on a template image. Instead, temporal coherence is regulated by low-rankness along the temporal direction. A PCA-based loss is thus proposed and paired with a fitting loss (optional). Experiments were conducted on in-house T1 MOLLI cardiac mapping data.
- 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 manuscript clearly describes the main contributions, conducted experiments, and parameters. An auxiliary task for aiding the motion correction can be beneficial if modelled in accordingly.
- 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.
Under the consideration of motion, the signal model – including the mapping – becomes a non-linear problem. A clear statement should be placed in the Abstract and Introduction, that a simplified decoupled and linearized model is studied here.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Do you have any additional comments regarding the paper’s reproducibility?
Source code will be made publicly available. Investigations are carried out on in-house datasets.
- 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 reference the works in the order of appearance.
- Please reference related works:
- https://pubmed.ncbi.nlm.nih.gov/35571217/
- https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7517224/
- Is the parameter “N” also optimized/learned or fixed?
- Please clearly distinguish the work to previous PCA-based group-wise registrations:
- Reference 8
- https://arxiv.org/abs/2001.03509
- 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
Accept — should be accepted, independent of rebuttal (5)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors provided a technical sound study and manuscript. The implicit template provides sufficient information for registration and has not been studied before in combination with a deep learning registration. The auxiliary task is also aiding the motion registration.
- 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
Strong Accept — must be accepted due to excellence (6)
- [Post rebuttal] Please justify your decision
The authors have addressed all my comments in the rebuttal.
Author Feedback
We thank all reviewers for their insightful and constructive comments. We appreciate the positive feedback that our work is “interesting and very relevant to the field” (R1), “technical sound” (R3), and that incorporating PCA and relaxometry losses in a DL framework is novel and “has not been studied before” (R4), supported by “a thorough set of experiments and significant gain” (R3). Other comments are addressed below:
Evaluation metric (R1,R3): We acknowledge the reviewers’ concerns regarding our evaluation metric’s definition and clinical relevance. We apologize for the confusion caused by assuming the reader’s familiarity with the “fitting SD” metric, which is commonly used in the CMR community to evaluate mapping quality [6,10]. Instead of evaluating tissue heterogeneity (SD across T1 in ROI), the fitting SD measures the quality of curve fitting at each pixel. It transforms the pixel-wise fitting residual error (of signal intensity) to that of fitted T1 value, using the inverse Hessian and analytical Jacobian of the signal model. Therefore, well-registered images yield low fitting errors thus low fitting SD, indicating good motion correction. This fitting SD is comparable to the pixel-wise R^2 metric in [7], but offers better clinical insight as it has the same physics unit (ms) as T1. Given the difficulty of obtaining in vivo ground-truth T1, fitting SD has been widely used as a surrogate metric of registration quality for CMR [22,24,Tilborghs(MEDIA2018)]. We will elaborate on this metric in the revision.
Contribution of PCA loss (R1,R3): Though initially introduced in [8], the PCA loss is integrated into a DL framework for the first time, to the best of our knowledge. This brings two major merits: first, the DL-based optimization of PCA loss is much faster (~20s) than conventional optimization on spline parameters (~600s); second, the amortized optimization in DL enables direct inference, as shown in our experiments.
Small dataset (R1): The dataset is a real-world clinical dataset of 50 subjects, where each underwent pre- and post-Gd scans at three ventricular levels, containing 300 slices (3300 images). Our one-shot and TTA tests showed the method’s efficacy with limited data.
Method comparison, applicability to other sequences (R1): We mainly focused on benchmarking with established deep learning-based methods [2,28]. We used Elastix [11] as an initial baseline and tested our method on other qMRI sequences from our physicist collaborators. We observed better quantitative performance compared with Elastix and broader applicability to other sequences, but for brevity we did not include the results in the paper. We will modify the claim in the revision.
Registration quality (R1): We agree that the improvement is visually subtle as we did not intentionally select cases with large motion. However, we could still observe improvement at the LV septum and RV wall in Fig. 4b, where high fitting SD values indicate uncorrected motion. The effect can be better appreciated in moving mode (supplementary avi files).
Title (R1): Following the suggestion we will change the title to “PCA-Relax: Template-free deep-learning-based groupwise registration for motion correction in quantitative cardiac MRI”.
Non-differentiability of LS (R3): The T1 loss [8] is itself differentiable, but fitting optimizers require up to thousands of iterations, lacking practical differentiability [5,27,Hallack,Tilborghs]. In contrast, our mapping module replaces the iterative optimizer and is directly differentiable. We will clarify and add references.
Pretraining of mapping module (R3): The self-supervision learning followed [29] (Eq. 5) without prior estimation of signal parameters.
Linear assumption (R4): The linearity of PCA on the signal model will be stated in the revision.
References (R4): Thank you for suggesting the references and we will include them.
Parameter N (R4): N is fixed during image acquisition and is independent of our pipeline.
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.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Reviewers note that the proposed method is a translation of well-established classical methods into a deep learning framework, and therefore lacks radical novelty. Furthermore, the approach only robustly beats baselines (with statistical significance) when test-time (re)training is used, which removes one of the main advantages of learning-based approaches, namely inference speed. Given that test-time training is required, a comparison to the classical (non deep-learning) variant of this approach is lacking (i.e., what if the deformations and quantitative parameters are directly regressed using autograd + LBFGS)?
- 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).
Reviewers note that the proposed method is a translation of well-established classical methods into a deep learning framework, and therefore lacks radical novelty. Furthermore, the approach only robustly beats baselines (with statistical significance) when test-time (re)training is used, which removes one of the main advantages of learning-based approaches, namely inference speed. Given that test-time training is required, a comparison to the classical (non deep-learning) variant of this approach is lacking (i.e., what if the deformations and quantitative parameters are directly regressed using autograd + LBFGS)?
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’
While the reviewers opiniion remain mixed for this paper, the overall score was raised. The problem is important and the solution is interesting. Though both novelty and validation are quite limited. Specifically a validation on pbulicaly available data which include also manual segmentation of the myocardium contours can help in evaluating the value of the proposed approach.
- 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).
While the reviewers opiniion remain mixed for this paper, the overall score was raised. The problem is important and the solution is interesting. Though both novelty and validation are quite limited. Specifically a validation on pbulicaly available data which include also manual segmentation of the myocardium contours can help in evaluating the value of the proposed approach.
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’
Overall the paper is certainly border-line, with several concerns left after the rebuttal process.
However, I don’t believe there is sufficient concern to over-ride the average reviewer opinion/score – overall there is sufficient interest and support (in some cases very strong) that the paper should be accepted.
- 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).
Overall the paper is certainly border-line, with several concerns left after the rebuttal process.
However, I don’t believe there is sufficient concern to over-ride the average reviewer opinion/score – overall there is sufficient interest and support (in some cases very strong) that the paper should be accepted.