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

Abstract

In resource-limited settings, portable ultra-low-field (uLF, i.e., 0.064T) magnetic resonance imaging (MRI) systems expand accessibility of radiological scanning, particularly for low-income areas as well as underserved populations like neonates and infants. However, compared to high-field (HF, e.g., ≥ 1.5T) systems, inferior image quality in uLF scanning poses challenges for research and clinical use. To address this, we introduce Super-Field Network (SFNet), a custom swinUNETRv2 with generative adversarial network components that uses uLF MRIs to generate super-field (SF) images comparable to HF MRIs. We acquired a cohort of infant data (n=30, aged 0-2 years) with paired uLF-HF MRI data from a resource-limited setting with an underrepresented population in research. To enhance the small dataset, we present a novel use of latent diffusion to create dual-channel (uLF-HF) paired MRIs. We compare SFNet with state-of-the-art synthesis methods by HF-SF image similarity perceptual scores and by automated HF and SF segmentations of white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). The best performance was achieved by SFNet trained on the latent diffusion enhanced dataset yielding state-of-the-art results in Fréchet inception distance at 9.08 ± 1.21, perceptual similarity at 0.11 ± 0.01, and PSNR at 22.64 ± 1.31. True HF and SF segmentations had a strong overlap with Dice similarity coefficients of 0.71 ± 0.1, 0.79 ± 0.2, and 0.73 ± 0.08 for WM, GM, and CSF, respectively, in the developing infant brain with incomplete myelination, and displayed 166%, 107%, and 106% improvement over respective uLF-based segmentation metrics. SF MRI supports health equity by enhancing the clinical use of uLF imaging systems and improving the diagnostic capabilities of low-cost portable MRI systems in resource-limited settings and for underserved populations. Our code is made openly available at https://github.com/AustinTapp/SFnet.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3995_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

https://github.com/AustinTapp/SFnet

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Tap_SuperField_MICCAI2024,
        author = { Tapp, Austin and Zhao, Can and Roth, Holger R. and Tanedo, Jeffrey and Anwar, Syed Muhammad and Bourke, Niall J. and Hajnal, Joseph V. and Nankabirwa, Victoria and Deoni, Sean and Lepore, Natasha and Linguraru, Marius George},
        title = { { Super-Field MRI Synthesis for Infant Brains Enhanced by Dual Channel Latent Diffusion } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15003},
        month = {October},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces a Super-Field magnetic resonance imaging enhancement approach aimed at improving the image quality of ultra-low field (uLF) MRI, making it comparable in perceived information to high-field (HF) MRI at a lower cost in clinical point-of-care settings.

  • 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 used a dual-channel latent diffusion process to generate paired (uLF-HF) synthetic images, addressing the typical issue of small datasets in infant 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.

    Although the application domain of the method is promising, the method takes good use of the latest technology. The paper has some concerns regarding methodology construction, data sets, and experiments, detailed details can be found below.

  • 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?

    The authors claimed to release the source code and/or dataset upon acceptance of the submission. Although the paper lacks a lot of details and hyperparameters, which are necessary for the paper’s reproducibility.

  • 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

    Considering the capacity of the limited content and the topic of the paper, my considerations will be limited to these aspects:

    1. The skeleton of the method consists of an LDM, a GAN learning strategy and SFNET. This undoubtedly increases the cost and complexity of learning. On the one hand, the author needs to mention the hyperparameters and setting during training and prediction more clearly. A table would be helpful. On the other hand, the improvements brought about by this complex structure will benefit from various aspects, and the article does not discuss this concern.

    2. The data of the paper is collected as a small-scale data set, which makes the experimental results and the design of the complex framework unconvincing. I suggest bringing the method to a public data set to get a more objective evaluation. I think such a scenario is not difficult to simulate. On the other hand, the proposed method is not exactly at the same level as the compared method, and a broader comparative experiment is expected.

    3. As mentioned before, the complexity of the method and the improvements it brings are not yet known. I’m surprised the authors didn’t complete a comprehensive ablation experiment to provide more insight. At the same time, such ablation experiments can also be put into the table 1 for reference.

  • 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?

    The task of the paper focuses on is very good and has practical significance. However, the description and experiments of the paper need to complete a major round of improvement, especially the concerns mentioned above.

  • 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

    I have no further concerns. Thank the authors for addressing my comments within the limited space, and looking forward to the camera-ready version of the revised paper.



Review #2

  • Please describe the contribution of the paper

    In this paper, the authors introduce Super-Field Network (SFNet), a custom swinUNETRv2 with generative adversarial network components that uses ultra-low field (uLF) MRI and generates super-field (SF) images comparable to HF MRI. The authors did an interesting work and the manuscript is well written and could be a valuable addition to the field.

  • 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.
    • Authors acquired a cohort of infant data (n = 30; aged 0–2 years) with paired uLF-HF data from a limited-resource area like Uganda with an underrepresented population in research.
    • SFNet is compared with state-of-the-art synthesis methods by HF-SF image similarity perceptual scores and by automated HF and SF segmentations of white matter (WM) gray matter (GM), and cerebrospinal fluid (CSF).
    • SF MRI supports health equity by enhancing the clinical use of uLF imaging systems and improving the diagnostic capabilities of low-cost portable MRI systems in resource-limited settings and for underserved populations.
  • 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 authors compared their approach SFNet, with only three methods based on quantitative perceptual and image segmentation metrics. The data size of n = 30 is small for generalization of the model.

    The proposed method also did not demonstrate any indication of significant clinical abnormalities that would be of concern to radiologists when examining pathological lesions. For more details, please see the following papers: 1 Lin, H., et al., (2023). Low-field magnetic resonance image enhancement via stochastic image quality transfer. Medical Image Analysis, 87, 102807. 2 Tanno, R., et al., (2021). Uncertainty modelling in deep learning for safer neuroimage enhancement: Demonstration in diffusion MRI. NeuroImage, 225, 117366.

    Also, a similar and an advanced work already has published where an AI method “SynthSR,” that enhances clinical brain MRI scans of various contrasts (T1, T2, etc.), orientations (axial, coronal, and sagittal), and resolutions into high-resolution T1 scans compatible with most human neuroimaging tools. The authors should critically discuss this work and describe how the proposed method has improved. Iglesias, J.E., et al., 2023. SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances, 9(5), p.eadd3607.

  • 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?

    None

  • 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
    1. What is the practical application of the SFNet generated super-field (SF) images in clinics?
    2. For generalization of the model, authors should compare their approach with more SOTA methods.

    3. “MAB include a spatial adaptation block and joint residual feature aggregation block”; need details.

    4. The authors should give a detailed description of the phenomenon in “Fig. 3. Segmentations obtained from various images, from left to right.”
    5. Moreover, it should be noticed that the clinical appliance has to be decided by medical professionals since the existing differences between the real image and the one generated by the proposed system could be substantial in the medical field.
  • 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?

    Authors acquired a cohort of infant data with paired uLF-HF data from a limited-resource area like Uganda with an underrepresented population in research. The authors introduce SFNet, a custom swinUNETRv2 with generative adversarial network components that uses ultra-low field (uLF) MRI and generates super-field (SF) images comparable to HF MRI. However, the proposed method did not demonstrate any significant clinical abnormalities that would be of concern to radiologists when examining pathological lesions.

  • 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

    Thank the authors for comprehensively addressing my comments within the limited space. I am quite satisfied with it and have no further comments.



Review #3

  • Please describe the contribution of the paper

    The paper introduces a novel MRI enhancement technique, the Super-Field Network (SFNet). This technique combines transformer attention mechanisms with CNN feature extraction in a GAN-style schema, aiming to enhance the quality of ultra-low-field MRI images to a level comparable to high-field MRIs. This may have significant implications for improving diagnostic capabilities in underserved and resource-limited areas.

  • 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 integration of dual-channel latent diffusion to enhance a small dataset of infant brain MRIs is particularly noteworthy.

    The paper reports substantial improvements in image quality metrics like FID, PSNR, and perceptual similarity.

    By enhancing uLF MRI, the technology could greatly benefit pediatric healthcare in low-resource settings.

  • 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 study’s limited sample size (30 infants) could impact the generalizability of the results. I acknowledge the latent diffusion-supplemented dataset (n = 90).

    While the paper compares SFNet with other state-of-the-art methods, a more detailed analysis of how this technique specifically outperforms each compared method could strengthen the paper.

  • 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?

    Provided Link: Redacted

  • 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

    Expanding on the potential impact of this technology in various low-resource settings globally could provide a stronger case for its utility and adoption.

    Additional validation using external datasets or in different clinical settings could help in understanding the robustness and reliability of SFNet.

    In the manuscript, you have compared the performance of SFNet with established methods such as LoHiResGAN and LF-SynthSR. To ensure a fair comparison, could you clarify if these comparison models were retrained on your specific dataset or if the evaluation was performed using pre-trained models? Retraining the comparison models on the same dataset as SFNet would help in assessing their relative performance accurately under identical conditions. Could you provide additional details on the training protocols used for these models in the context of your study? This information would greatly enhance the credibility of the comparative analysis presented in your paper.

    It appears there may be an inconsistency in the reference list related to the citations for LoHiResGAN [18] and LF-SynthSR [20]. It may listed as [19], and [21] in the References perhaps?

    1. Islam, K. T., Zhong, S., Zakavi, P., et al. (2023). Improving portable low-field MRI image quality through image-to-image transla-tion using paired low- and high-field images. Scientific Reports, 13(1), 1-13.
    2. Iglesias, J. E., Schleicher, R., Laguna, S., Billot, B., … Kimberly, W. T. (2023). Quantitative Brain Morphometry of Portable Low-FieldStrength MRI Using Super-Resolution Machine Learning. Radiology, 306(3), e220522. https://doi.org/10.1148/radiol.220522
  • 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 novel approach, significant improvements over existing methods, and potential for substantial impact in health equity for pediatric care in resource-limited settings are the primary factors influencing this recommendation.

  • 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

    Accept — should be accepted, independent of rebuttal (5)

  • [Post rebuttal] Please justify your decision

    The authors’ rebuttal has satisfactorily addressed my concerns, prompting me to change my decision to Accept.



Author Feedback

We thank all reviewers for their constructive feedback and for appreciating how our work supports health equity of underserved populations via developing technology of ultra-Low-Field (uLF, 64mT) MRI. Last week, 3 keynotes on uLF MRI were presented at ISMRM, emphasizing the value of this emerging technology. We are excited reviewers have commended our dual-channel latent diffusion and novel SFNet architecture. Below, we address concerns and indicate responses. Please note our paper was submitted to the ‘MICCAI for Health Equity’ category.

[R4] Reproducibility (Hyperparameters): Our repository [redacted] contains all training and testing code, pretraining and fine-tuning scripts, and hyperparameters for all networks, including the ablation studies. The LDM network hyperparameters are mentioned in section 2.1.

[R1, R3] Comparison to SOTA work; citation inconsistency: We apologize for our oversight in not explicitly citing SynthSR, 2023, by Inglesias, et al., which will be added. SynthSR was discussed in the introduction and is a compared method in Table 1 under ‘LF-SynthSR’. From our citation [21], Inglesias, et al. state “LF-SynthSR, builds on our previous method, SynthSR”. Both SynthSR and LHResGAN are pretrained and the only known methods affirmed by their authors as usable ‘out of the box’ for uLF-HF synthesis; we will correct this in section 3.3 and fix the introduction’s citations.

[R3, R4] Ablation studies; model training: Our ablation studies were outlined in paragraph 2, section 2.2, and guided our selection of the best networks, which were presented in section 3.3. Notably, Table 1 showed a pivotal architecture ablation study ‘SwinUNETR’, which is our SFNet (n=90) without GAN aspects (‘Ancillary Discriminator Components’ of Fig 2). Studies of sequential pretraining yielded SFNet (n=30, n=90) and ‘SwinUNETR’ results seen in Table 1. Omitted from Table 1 are networks trained from scratch: SFNet (n=30 and n=90), was slightly, but not significantly (p=0.17 and p=0.09) better than Table 1’s ‘SwinUNETR’ model. A ‘SwinUNETR’ trained from scratch had perceptual metrics similar to SynthSR (p=0.23) and LHResGAN (p=0.34) but achieved higher DSC and RVE (p=0.02 and p=0.04). As requested, section 2.2 will be updated to better explain our ablation studies and model selection process. We will include purposefully omitted (as stated in section 3.3) ablation study outcomes in Table 1 at submission.

[R1, R3, R4] Concerning dataset scope: In our limitations, we acknowledged the constrained size of our dataset, which is justified by two reasons. First, obtaining paired uLF-HF data, especially from underrepresented regions like Uganda, is challenging. To our knowledge, no public datasets of uLF-HF pairs exist. Second, uLF MRI is a new technology and emerging area in pediatric healthcare, making acquisition more difficult. In fact, given the rarity of data and interest in this promising field, our team released the first public pediatric uLF dataset from low-resource areas for a challenge at [redacted]. Despite the limited dataset, our novel dual-channel latent diffusion tripled the cohort and significantly enhanced model performance. These clarifications will be summarized in section 4.

[R1] Pathology and clinical application: We mentioned in future work we wish to include pathology. Note our data is acquired at 64mT (ultra LF) in contrast to 0.5-1T data (commonly LF). Currently, uLF MRI is not established as a diagnostic exam due to the field strength’s detrimental impact on image quality. Hence, our study is essential to improving uLF images to diagnostic level quality. Our initial work is performed on presumably healthy patients; uLF cases are from an ongoing research study of infant growth in underserved areas to identify developmental differences based on geography or socioeconomic status. Thus, images were not acquired with the expectation of identifying pathology. These clarifications will be made in sections 1 and 3.1.



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’

    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



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’

    The authors’ rebuttal has comprehensively addressed reviewers’ concerns within the limited space.

  • 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).

    The authors’ rebuttal has comprehensively addressed reviewers’ concerns within the limited space.



back to top