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

Abstract

The Orientation Distribution Function (ODF) characterizes key brain microstructural properties and plays an important role in understanding brain structural connectivity. Recent works introduced Implicit Neural Representation (INR) based approaches to form a spatially aware continuous estimate of the ODF field and demonstrated promising results in key tasks of interest when compared to conventional discrete approaches. However, traditional INR methods face difficulties when scaling to large-scale images, such as modern ultra-high-resolution MRI scans, posing challenges in learning fine structures as well as inefficiencies in training and inference speed. In this work, we propose HashEnc, a grid-hash-encoding-based estimation of the ODF field and demonstrate its effectiveness in retaining structural and textural features. We show that HashEnc achieves a 10% enhancement in image quality while requiring 3x less computational resources than current methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2292_paper.pdf

SharedIt Link: https://rdcu.be/dV5Da

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72104-5_30

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2292_supp.pdf

Link to the Code Repository

https://github.com/MunzerDw/NODF-HashEnc

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Dwe_Estimating_MICCAI2024,
        author = { Dwedari, Mohammed Munzer and Consagra, William and Müller, Philip and Turgut, Özgün and Rueckert, Daniel and Rathi, Yogesh},
        title = { { Estimating Neural Orientation Distribution Fields on High Resolution Diffusion MRI Scans } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {307 -- 317}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This manuscript proposes HashEnc, a grid-hash-encoding-based estimation of the ODF field. The proposed method is an accelerated version of NODF in [5]. However, the training time is still computationally inefficient. Moreover, I think the proposed method and NODF have serious problems on motivations and implementations.

  • 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 method is an accelerated version of NODF in [5].

  • 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.

    This manuscript proposes HashEnc, a grid-hash-encoding-based estimation of the ODF field. The proposed method is an accelerated version of NODF in [5]. However, the training time is still computationally inefficient. Moreover, I think the proposed method and NODF have serious problems on motivations and implementations.

    Detailed comments:

    • it is not clear to me why the authors use INR to obtain continuous ODF field with such high computationally cost. We could use a very simple yet efficient solution to obtain continuous ODF field without INR.
    • The simple way is: 1) use traditional methods to estimate discrete ODF field in grids. 2) use interpolation on the discrete ODF field to obtain the continuous ODF field. Then done. Here we could simple linear interpolation, or use Gaussian Process for interpolation. It is very simple and more computationally efficient than the proposed method.
    • Note that the authors have already described the Gaussian process prior in the manuscript (Eq. 3 and 4). Then, why cannot we just use Gaussian process for interpolation to obtain a continuous ODF field?

    • Note that if we use the above simple way (estimation + interpolation), the loss on each grid voxel is minimized, which has much lower loss than the proposed method.

    • Therefore, I cannot see the motivation of the proposed method. It seems to me that the authors try to solve a simple problem using a difficult way.

    • In page 3, the authors define ODF based on FRT. However, there are many ways to define ODF (diffusion ODF, or fiber ODF).
  • 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 does not provide sufficient information for reproducibility.

  • 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

    However, the training time is still computationally inefficient. Moreover, I think the proposed method and NODF have serious problems on motivations and implementations.

  • 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

    Strong Reject — must be rejected due to major flaws (1)

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

    However, the training time is still computationally inefficient. Moreover, I think the proposed method and NODF have serious problems on motivations and implementations.

  • 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 Reject — must be rejected due to major flaws (1)

  • [Post rebuttal] Please justify your decision

    I still suggest rejection after rebuttal. The authors did not clarify the motivation of INR. I agree with the authors that INR has somehow denoise effect, compared with pure estimation without considering denoise. However, if your goal is denoise, then you could try estimation with denoise (MAP or spatial regularization, as shown in papers). Therefore, denoise is not the goal or motivation of INR. Considering INR is very time consuming, I cannot agree that it is a correct and efficient way to obtain continuous ODF field, compared with estimation (with denoise) + linear or non-linear interpolation, or other so called continuous super-resolution methods.



Review #2

  • Please describe the contribution of the paper

    Building on previous work SIREN, this paper proposes to utilize HashEnc, a grid-hash-encoding technique, suggested by NVIDIA, to estimate neural orientation distribution fields on high-resolution diffusion MRI scans. This approach accelerates the training process and enhance image quality.

  • 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 previous work used SIREN to estimate neural orientation distribution fields, but it faced long training times, making it impractical. The authors use a grid-hash-encoding technique suggested by NVIDIA to enhance the training speed.

  • 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.

    Some problems should be solved before it is considered for publication. Firstly, there is an inadequate explanation of how the grid-hash encoding method enhances training speed. Secondly, this work lacks comparisons with other methods. Thirdly, the results provided by SIREN is somewhat counterintuitive as the estimation become worsen with more numbers of diffusion gradients.

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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

    Several details need improvement. Firstly, explain the rationale for adopting the grid-hash-encoding method to enhance training speed. Secondly, present quantitative results from traditional and other deep learning-based methods. Thirdly, provide a thorough explanation for the obtained results from SIREN.

  • 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 motivation of this work is reasonable and meaningful. But the feasibility of method is not analyzed very clearly and experiments carried out in this work failed to strongly validate the effectiveness of the proposed method.

  • 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 motivation is valid and significant. And the results show the enhancement in training speed and estimation accuracy.



Review #3

  • Please describe the contribution of the paper

    The authors propose the use of HashEnc to obtain a computationally more efficient implicit neural representation (INR) of ODF fields. Based on experiments on a publicly available dataset, they claim that, compared to SIREN, which a recently proposed alternative INR, this approach reduces the time consumption for high-resolution data by a factor of three, while simultaneously increasing the image quality.

  • 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 deals with a timely topic.
    • The paper is well-written and easy to follow.
    • The contribution and benefit in terms of computation times is clear.
    • The qualitative comparison is convincing in terms of the perceptual quality of the derived GA maps.
  • 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.
    • Novelty is limited. The work essentially takes existing methodology from Mueller et al. (2022) and applies it to ODF fields.
    • The focus on FSIM as the only quantitative measure for evaluation is problematic. Even though a measure of perceptual similarity is relevant to the human interpretation of derived maps, another important type of downstream analysis would be quantiative, e.g., statistical testing of group differences. To assess the suitability of the proposed methodology in such a context, a more direct quantification of the bias and variance that various INRs introduce into metrics such as GFA would be informative.
    • Even though tractography is a primary use case for ODFs, no tractography results are presented.
    • Since high-resolution data as it is considered here is rare in practice, I would have liked to see a discussion (and supporting experiments) of whether the benefits of HashEnc are limited to this rather narrow use case, or whether it should also be considered for more standard setups.
  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

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

    Even though the key points seem clear, and the approach largely relies on a previously described method, the reported speed and accuracy of INRs could depend on details of the implementation. Therefore, it would be highly appreciated if the authors could make a reference implementation available along with their published paper.

  • 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

    If possible, I would like to see a more comprehensive quantitative evaluation, as well as results showing how the different ODF representations affect downstream tractography. From visually inspecting results in Fig. 3, I would expect that the increased noise level in the HashEnc ODFs might lead to instable tractography. If possible, I would also like to see a discussion of the applicability to dMRI data with a more standard spatial resolution.

    The current discussion of “INRs in Medical Imaging” in the related work section is missing works that are arguably among the most closely related, since they also model dMRI data:

    Hendriks et al., “Neural Spherical Harmonics for Structurally Coherent Continuous Representation of Diffusion MRI Signal”, CDMRI 2023 https://link.springer.com/chapter/10.1007/978-3-031-47292-3_1

    Ewert et al., “Geometric deep learning for diffusion MRI signal reconstruction with continuous samplings (DISCUS)”, Imaging Neuroscience 2024 https://direct.mit.edu/imag/article/doi/10.1162/imag_a_00121/120017

  • 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?

    Despite the above-mentioned reservations with respect to novelty, and remaining questions regarding practical utility (for statistical analysis, tractography, use outside the rather specialized domain of high-res data), I believe that this manuscript satisfies the usual expectations for a MICCAI conference paper.

  • 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

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

  • [Post rebuttal] Please justify your decision

    Even though I agree with the two other reviewers that a more extensive evaluation is desirable, I still think it could be left to a potential journal version of this work, and that the work is interesting enough for presentation at MICCAI already in its current form, especially considering the limited space available in MICCAI papers. Therefore, I stick to my original score.



Author Feedback

We thank the reviewers for taking the time to review our submission and for giving a thorough feedback. We are glad that the reviewers find our paper “well-written and easy to follow” (reviewer 3) and that the “contribution […] is clear” (reviewer 1, 3, and 4). In the following, we address the concerns of the reviewers in separate sections.

Reviewer 1:

  1. INR vs other methods: Linear interpolation on the per-voxel ODF estimates is reasonable for moderate to high SNRs, but fails at high resolutions due to the low SNRs, resulting in poor estimates with extremely high variance (see SHLR-Raw results in [5]). Our deep neural network learns continuous spatial basis functions, promoting spatial regularity and smoothing in the ODF field, thereby improving estimates for noisy high-resolution data.

  2. Using the Gaussian process prior in Eq. (3): This GP is in fact used for prediction. The mean and covariance of the Gaussian in Eq. (3) are both a function of the deep network parameters, which learns the spatial correlation structure of the ODF field. Therefore, inference using the resulting GP model will result in flexible spatially regularized inference + interpolation that is robust to noise + sparse angular samples.

  3. Models beyond the FRT: Our methodology is adaptable beyond the Funk-Radon Transform (FRT). It works with any fixed, linear, and rotationally invariant forward model, affecting only the G matrix while maintaining the rest of the approach.

Reviewer 3:

  1. FSIM as the only quantitative measure: We chose the FSIM measure due to its sensitivity to image “visual quality”, fine-grained details and texture, which are of the utmost importance for high-res imaging. The reviewer rightly points out the importance of analyzing bias/variance in downstream diffusion quantities. We will include this analysis in the supplementary material.

  2. HashEnc on standard spatial resolution: HashEnc is suitable for lower resolution dMRI scans such as the ABCD dataset [1], but our focus is on high-resolution data where traditional methods fail to perform. For the performance of SIREN on the ABCD dataset, see the supplementary section of NODF [5].

  3. Related references & tractography: We appreciate the feedback and will include the references in the related work, in addition to a tractography figure.

Reviewer 4:

  1. How HashEnc enhances training speed: For inference at a single voxel, HashEnc requires only 13k parameters, speeding up both forward and backward passes compared to SIREN’s 10 million parameters. Increasing HashEnc’s embedding vectors from 2 to 4 doubles its capacity but adds only 1792 parameters to the forward/backward pass (size of input feature vector increases from 28 to 56). In contrast, doubling SIREN’s parameters means that all 20 million parameters are required for any voxel inferences, greatly increasing computation time. We will add this additional explanation to the method section.

  2. Comparison with other methods: We focus solely on SIREN to benchmark against similar implicit neural network methods. Other approaches aren’t resolution agnostic and use different inputs, such as raw signal data. At submission, to the best of our knowledge, the only published INR method for dMRI data with publicly available code was NODF [5]. We will add the 2 additional suggested references from reviewer 3 to the related work.

  3. Counter intuitive results of SIREN: We thank the reviewer for the feedback and recognize that the results for SIREN M=70 are unintuitive compared to M=40 and M=20. As we were rerunning the experiment to validate the numbers, we realize a mistake in the selection of the regularization strength. We will fix the FSIM scores for SIREN M=70 and update the table accordingly.

References: [1]: Casey et al. “The adolescent brain cognitive development (ABCD) study: imaging acquisition across 21 sites.” [5]: Consagra et al. “Neural Orientation Distribution Fiel



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’

    This paper received a mixed recommendation from all reviewers. The major concern raised from R1 is the motivation of using implicit neural representations to estimate a continuous Orientation Distribution Function (ODF) field from diffusion MRIs. It was not clear whether the proposed method would perform better than simpler and computationally more efficient interpolation methods. There are relatively moderate concerns from other reviewers regarding the limited novelty of the proposed method, and lack of comprehensive quantitative evaluation as well as a thorough comparison with other works that estimate ODF fields. While this paper has some good merits, the current manuscript could be strengthened and considered for future submissions to MICCAI.

  • 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).

    This paper received a mixed recommendation from all reviewers. The major concern raised from R1 is the motivation of using implicit neural representations to estimate a continuous Orientation Distribution Function (ODF) field from diffusion MRIs. It was not clear whether the proposed method would perform better than simpler and computationally more efficient interpolation methods. There are relatively moderate concerns from other reviewers regarding the limited novelty of the proposed method, and lack of comprehensive quantitative evaluation as well as a thorough comparison with other works that estimate ODF fields. While this paper has some good merits, the current manuscript could be strengthened and considered for future submissions to MICCAI.



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 there is some disagreement between the reviewers, the authors have addressed some of the reviewers’ concerns resulting in one reviewer changing their score from slightly negative to slightly positive. Overall, this is an interesting paper that can lead to interesting discussions in the community.

  • 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 there is some disagreement between the reviewers, the authors have addressed some of the reviewers’ concerns resulting in one reviewer changing their score from slightly negative to slightly positive. Overall, this is an interesting paper that can lead to interesting discussions in the community.



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’

    This is a good quality paper. The use of InstantNGP (which the authors recoin as ‘HashEnc’) is not necessarily new, but the combination with this application is. The paper can be of interest for the MICCAI community. The authors have addressed most comments by the reviewers and will likely find the remaining comments and suggestions for additional experiments useful for future work.

  • 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).

    This is a good quality paper. The use of InstantNGP (which the authors recoin as ‘HashEnc’) is not necessarily new, but the combination with this application is. The paper can be of interest for the MICCAI community. The authors have addressed most comments by the reviewers and will likely find the remaining comments and suggestions for additional experiments useful for future work.



back to top