List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Junaid R. Rajput, Tim A. Möhle, Moritz S. Fabian, Angelika Mennecke, Jochen A. Sembill, Joji B. Kuramatsu, Manuel Schmidt, Arnd Dörfler, Andreas Maier, Moritz Zaiss
Abstract
Chemical exchange saturation transfer (CEST) is an MRI
method that provides insights on the metabolic level. Several metabolite
effects appear in the CEST spectrum. These effects are isolated by
Lorentzian curve fitting. The separation of CEST effects suffers from the
inhomogeneity of the saturation field B1. This leads to inhomogeneities
in the associated metabolic maps. Current B1 correction methods require
at least two sets of CEST-spectra. This at least doubles the acquisition
time. In this study, we investigated the use of an unsupervised
physics-informed conditional autoencoder (PICAE) to efficiently correct
B1 inhomogeneity and isolate metabolic maps while using a single CEST
scan. The proposed approach integrates conventional Lorentzian model
into the conditional autoencoder and performs voxel-wise B1 correction
and Lorentzian line fitting. The method provides clear interpretation of
each step and is inherently generative. Thus, CEST-spectra and fitted
metabolic maps can be created at arbitrary B1 levels. This is important
because the B1 dispersion contains information about the exchange
rates and concentration of metabolite protons, paving the way for their
quantification. The isolated maps for tumor data showed a robust B1 correction
and more than 25% increase in structural similarity index (SSIM)
with gadolinium reference image compared to the standard interpolationbased
method and subsequent Lorentzian curve fitting. This efficient correction
method directly results in at least 50% reduction in scan time.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43993-3_44
SharedIt: https://rdcu.be/dnwNP
Link to the code repository
https://git5.cs.fau.de/rajput/picae
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposed a physics-informed conditional autoencoder (PICAE) approach to estimate a multi-pool model of CEST spectra with multiple B1 values. This proposed PICAE approach can provide B1 inhomogeneity correction and Lorenzian curve fitting. It also requires much fewer B1 values for separate the CEST spectra to different pools, which can provide at least 50% reduction in the scan time. The performance of the approach was evaluated using contrast-enhanced images.
- 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) novel formulation Physics-informed conditional autoencoder and bounded loss provide a reliable estimation of the 5-pool models. 2) strong evaluation Gd images were used in the evaluation to provide convincing results. 3) clinical feasibility The reduced scan time make this approach feasible in clinical research.
- 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.
A (very) minor weakness is that the approach still requires prior knowledge about the lower and upper bounds of the parameters and the number of pools.
- Please rate the clarity and organization of this paper
Excellent
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
No concerns about reproducibility.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
Overall it is a very good paper. A minor comment is that adding some explanation about the 5-pool model, especially NOE, MT, will help readers that are not familiar with CEST physics.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper has a novel method, strong evaluation and clinical feasibility.
- Reviewer confidence
Very confident
- [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
Review #6
- Please describe the contribution of the paper
The paper presents a deep learning-based method to correct B1 inhomogeneity in CEST MRI at 7T. The method utilizes two subnetworks, combining Conditional Autoencoders (CAE) and Physics-Informed Autoencoder (PIAE) to generate corrected CEST images, and outperforms conventional Lorentzian fitting methods in terms of the structural similarity to the Gd post-contrast T1-weighted images.
- 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.
- A novel approach to address B1 inhomogeneity correction in CEST MRI, which is an important problem in the field.
- The use of Physics-Informed Autoencoder, a potentially innovative strategy for reliable correction.
- Demonstration of the method’s clinical feasibility through quantitative evaluation on real patient data.
- 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.
- Lack of clarity and detail in the explanations, making some aspects hard to understand for a broader audience.
- Complicated deep learning model with potential redundancy in the two sub-networks, warranting further exploration through an ablation study.
- Insufficient evaluation and validation of the method on larger and more diverse datasets.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The author provide some code in the supplementary materials, but it is not easy to reproduce the results since the data are not to be released.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- Provide a more in-depth explanation of B1 inhomogeneity correction in CEST MRI and why simpler B1 correction methods used for structural imaging are not suitable.
- Enhance the illustration of the deep learning model in Fig. 2 for better understanding.
- Conduct an ablation study to understand the roles and contributions of CAE and PIAE sub-networks.
- Extend the evaluation and validation of the method on larger diverse datasets to demonstrate the robustness of the proposed method.
- Reconsider the use of SSIM as a similarity metric, as it may not be appropriate for comparing CEST maps and contrast-enhanced structural images. Instead, use cross-modality metrics such as mutual information.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a novel approach to correct B1 inhomogeneity in CEST MRI using a deep learning model, which is a contribution to the field. The use of the physics informed network is innovative, and the clinical feasibility of the method is demonstrated. However, the paper suffers from a lack of clarity in the research background and insufficient validation/investigation of the current model design.
- Reviewer confidence
Confident but not absolutely certain
- [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
Review #5
- Please describe the contribution of the paper
This paper proposes an auto-encoder-based approach that converts CEST spectra from 7T MRI to spectra at an arbitrary B1 field strength.
- 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 proposed model may solve several important issues in CEST imaging.
- 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.
- Language describing which subjects were scanned at 7T is unclear, and overall description of specific acquisitions used for training and validation, is unclear.
- Performance evaluation is largely visual. There’s not enough information provided for the reader to judge what data was used for evaluation of various CEST spectra.
- There’s no explanation of symbols used in equations.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The authors provide verbal description of their neural network model, which may require significant effort to reproduce. Formal equation derivations are poorly annotated, which also may potentially reduce reproducibility.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
Abstract. Define “SSIM”.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work presented may be important to improve SSIM imaging. However, several key information pieces are missing from the text as outlined above.
- Reviewer confidence
Somewhat confident
- [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
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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.
The work presented in this paper is a framework propose a physics-informed conditional autoencoder (PICAE) method for evaluating 7T-CEST MRI that accounts for B1 inhomogeneity in the input to estimate a multi-pool model of CEST spectra with multiple B1 values. The paper has innovative technical merits (PICAE), investigates an important application (address B1 inhomogeneity), has a reasonable evaluation cohort, and the results showed the advantage of the proposed method. However, the paper has some weak aspects. The first is the small cohort used for evaluation and validation of the method, a larger and more diverse datasets should be used (R3) (a discussion at least should be added) with more quantitative metrics (R2). I agree with R3 that an ablation study should be conducted in the experiments to signify the roles/ contributions of CAE and PIAE sub-networks. Please define/explain all symbols (R2) and acronyms (e.g., SSIM) to make the paper more clear and easy to follow (R2, R3) as well as the missing details (R1,R2)
Author Feedback
We thank all reviewers (R1, R2, R3, Meta-R) for acknowledging our contribution and their constructive comments for further clarification. We will add the results of the sub-networks (CAE and PIAE) and explain the symbols/acronyms in our final paper. Q1: Prior knowledge and 5-pool model. (R1) A1: we know from the literature that the 7T-CEST spectrum contains a contribution from 5 different pools. These pools are modeled by Lorentzian distributions and the CEST-spectrum is a linear combination of these distributions. Including this knowledge in the model ensures consistency with previous literature. Due to the unsupervised nature of the training, bounds must be introduced, otherwise the positions of the molecules at -3.5 and +3.5 ppm could be interchanged with -3.0 and 2.0 ppm, such that the output layer of the PIAE encoder would not represent the same molecules for all voxels. In the article, we briefly try to define that the different pools in the CEST spectrum provide information about different molecules in the human brain and that they are correlated with specific diseases. Further description of the specific molecules is omitted in this article because (i) the length of the article is limited and (ii) it does not interfere with the concept of the study. Q2: small cohort, ablation study, and why simpler B1 correction methods don’t work. (R3,R2&Meta-R) A2: Although we had only seven subjects for the study, training and evaluation of the model were performed voxel-wise, resulting in approximately 1 million samples for training and validation and half a million samples for evaluation. Moreover, training and validation consisted only of healthy subjects, and we have shown in the article that the tumor ring shown in the exogenous contrast image matches much better with the CEST maps reconstructed by our approach than with the conventional method. We also verified this result using another tumor patient (figure in the supplementary material), in which PICAE shows better separation of tumor and healthy tissue compared with baseline. PIAE is not redundant as its role is to separate pools from the B1-corrected CEST spectrum provided by CAE. However, the reviewers are correct that the role of sub-networks should be investigated further, and we will publish this result in our final article. The reason why the simpler B1 correction does not work is that it uses linear interpolation between two CEST spectra and equation 7 in the article shows that the relationship between the amplitude of these CEST maps and B1 saturation is quadratic if we ignore the Zref term in equation 7, and with the inclusion of Zref it gets even more complicated. Q3: “The authors provide verbal description of their neural network model” (Reproducibility), Similarity metric, and Clarity. (R3,R2&Meta-R) A3: We have shared the code for both networks exactly defining the architectures, the acquisition pipeline we used to collect data is described in [9], and the description of the training and evaluation process is detailed in our article to reproduce the results. Unfortunately, since the data are confidential, we cannot share them. As suggested by the reviewers, we tested various similarity metrics, and PICAE performed better than the conventional method in PSNR, gradient correlation coefficient, and SSIM. In contrast, there was no significant difference between tumor region mutual information with PICAE maps and conventional maps, although visualization and other metrics suggest significant improvement. Thorough analysis indicates that the distribution of tumor region in the reference image and conventional CEST maps had skewed distributions, whereas the distribution in PICAE maps was somewhat more symmetrical, hence the comparable mutual information despite superior quality. In the final article, we will define the symbols in equations (2,3,7) in more detail, include the correlation coefficient of the gradients, and also mention the missing acronym (SSIM).