Abstract

Fiber orientation distribution function (fODF) estimation from diffusion MRI is crucial for mapping brain connectivity but often requires extensive multi-shell acquisitions and complex computational methods. While supervised deep learning approaches have shown promise in accelerating this process, they typically require large training datasets and face challenges with domain shifts and interpretability. We present UFO-3, an unsupervised framework that combines a three-compartment biophysical model with deep learning for fODF estimation from single- shell data. The method leverages a U-Net architecture to simultaneously estimate fiber orientations and tissue microstructure parameters while maintaining physical constraints through an optimization-based reconstruction. Evaluated on synthetic data across 2500 test cases, UFO-3 achieves superior angular accuracy (MAE < 10◦ at infinite SNR) and correlation (ACC > 91 %) compared to existing methods, particularly in resolving challenging fiber crossings. On in vivo human brain data, UFO-3 produces fODF reconstructions comparable to multi-shell reference methods while providing interpretable tissue parameter estimates. The framework requires a one-time, subject-specific training of about 30 min on a single consumer GPU and enables fast inference (< 10 s per subject), improving throughput compared to other unsupervised approaches that require hours or days of training. Our results demonstrate that UFO-3 effectively balances reconstruction accuracy, biological interpretability, and computational performance without requiring extensive training data or multi-shell acquisitions.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4006_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/tensor2023/ufo-3

Link to the Dataset(s)

https://www.scidb.cn/en/detail?dataSetId=f512d085f3d3452a9b14689e9997ca94

BibTex

@InProceedings{GaoXue_UFO3_MICCAI2025,
        author = { Gao, Xueqing and Lin, Rizhong and Feng, Jianhui and Shi, Yonggang and Qiao, Yuchuan},
        title = { { UFO-3: Unsupervised Three-Compartment Learning for Fiber Orientation Distribution Function Estimation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {636 -- 646}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a framework to estimate Fiber orientation distribution functions from single-shell diffusion MRI. It evaluates the framework using synthetic data and human brain data showing similar fODFs to the ones obtained with multi-shell diffusion MRI.

  • Please list the major strengths of the paper: you should highlight 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 include an evaluation with synthetic data, where the ground truth fODF is known.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • One of my major concerns are the use of “rapid” in the title and the multiple claims regarding efficiency. As far as I understand the method requires subject-specific training of a network that estimates all tissue parameters. Authors say such network takes 30 min to be trained (depending on the machine). This is way less efficient and fast that conventional approaches such as multi-tissue CSD (SS3T).
    • I see no explanation to not include SS3T-CSD in the synthetic data evaluation.
    • Despite having in vivo data of 5 subjects, there are no quantitative results of the in vivo data evaluation, we only get to see one patch of one case in Figure 3.
  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
    • It is not mentioned how many instances of fiber configurations were simulated.
    • Is it a typo or did you really use 92 images (directions) acquired with b=1000s/mm^2 ?
  • 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.

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

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

    The gain in efficiency is not clear, there are no quantitative results to support paper claims regarding evaluation with invivo data.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    I am not convinced by the author’s response to my concerns:

    • Authors confirm that subject-specific training takes ~30 min, plus 10 seconds of “rapid” inference. So the fully automated process of estimating subject-specific FOD and compartment fractions takes ~30:10. Authors might claim that being fully automated is an advantage, but I do not consider it translates as being “efficient”, nor that replacing “rapid” with “efficient” will make such claims correct. Furthermore, it is not true that multi-tissue CSD can take up to 30 min, not even including the response estimation within the computation time.

    • I do not pretend the authors included a comparison against SS3T-CSD, but I wanted an explanation of why it was not included. The authors say that they did not include it to maintain consistent assumptions across compared methods. They say “SS3T-CSD is designed for three physical tissues (WM/GM/CSF) whereas our network learns three abstract compartments”. As the authors say in their manuscript, those compartments represent specific biophysical properties: intra-axonal fODF, extra-axonal, and trapped water, so these compartments are not abstract. Both methods have analogous assumptions (three compartments with specific physical interpretations), so I do not consider the author’s explanation satisfactory.

    • Regarding the results with in-vivo data, authors respond that they “follow common practice using qualitative assessments and indirect validation (Fig3) across five CHCP subjects”. Their qualitative results in figure 3 only show one patch of one case, not across five subjects. It is pointless to state that additional subjects were processed without presenting any corresponding results because the audience cannot evaluate the method’s performance across the claimed cohort when evidence is provided for only one case.



Review #2

  • Please describe the contribution of the paper

    This work tackles the problem of fODF reconstruction from single shell DWI acquisitions. The authors propose to build upon previous work from Tran and Shi (1), which posed fODF as an optimzation problem constrianed by intra and inter axonal and free water diffusivity. While (1) was limited to multi-shell dmri, the proposed work learns to estimate diffusivity parameters from single-shell dmri in an unsupervised, subject-specific manner.

    The authors test their framework in-silico and in-vivo dataset and report sharper ODFs compared to other methods.

    [1]: Tran, Giang, and Yonggang Shi. “Fiber orientation and compartment parameter estimation from multi-shell diffusion imaging.” IEEE transactions on medical imaging 34.11 (2015): 2320-2332.

  • Please list the major strengths of the paper: you should highlight 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 framework is certainly interesting and the reconstruction quality is appealing. The ability to estimate diffusivity from single-shell data is desirable and the unsupervised training regime makes the method generally applicable, which further improves clinical feasibility.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The subject-specific training regime, which as reported takes around 30 minutes, seems highly inneficient compared to CSD or SS3T-CSD (which also do not require GPUs) which according to the results in Figures 2 and 3, mostly results in very similar fODFs. The abstract mentions fast inference (< 10s per subject), but is negated by the half-hour training required beforehand for every subject. While the results are good, it is unclear if the time and computation requirements are worth it.

    It is also unclear if the results presented in Fig. 4a are desirable. While the frontal parts of the CST are “fuller” with UFO-3, they are overly full compared to the more “conservative” reconstructions from (SS3T-)CSD. As the ISMRM2015 dataset dmri was generated from the GT tracts, the extraneous volume cannot be considered positive.

    Finally, there are some issues with the manuscript itself. The text relies heavily on (1) and the manuscript can be confusing if the reader is not familiar with (1). The text right before equation 2 refers to P(s) as the projection of s, but the projection is not introduced (is it the projection onto the graph with V nodes ? if so, P should be introduced when the graph is introduced, earlier in section 2). Section 3.3 refers to the BPS but Fig. 4 refers to the CST, which are similar but different fiber bundles. Sections 3 and 4 mention a patch size of 3x3x3, but section 2.2 mentions point-wise 1x1x1 convolutions. Why patches then, what is the advantage of using patches ? The disadvantages ?

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • 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.

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

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

    While the idea of sharper fODF reconstruction from single-shell data is appealing, the presented results and the training limitations hamper the applicability of the method. Moreover, no public implementation is reported.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have provided a satisfactory rebuttal of the reviewer’s comments. While some concerns are still present wrt. the “timely” aspect of the method put forward by the authors, the rest of the comments were sufficiently addressed.



Review #3

  • Please describe the contribution of the paper

    This proposed framework UFO-3, an unsupervised method for estimating the fiber orientation distribution function from single-shell diffusion MRI. It combines a three-compartment biophysical model (intra-axonal, extra-axonal, and trapped water) with a U-Net–based deep learning approach, then enforces physics-informed constraints through optimization. Experiments on synthetic and in vivo datasets show that UFO-3 achieves accuracy on par with multi-shell reference methods.

  • Please list the major strengths of the paper: you should highlight 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 paper is well-structured and clearly presented.
    2. By incorporating a three-compartment biophysical model and enforcing physics-based constraints, the method leverages prior knowledge to enhance interpretability and robustness.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. The evaluation is being done on five subjects, within a limited gradient encoding scheme, the generalizability of such method should be evaluated on LARDI data or on higher b-values.
    2. Is dense sphere resample necessary? Should a uniformed distribution but fewer nodes reach on par performances and improve computational efficiency since the model is subject-specific and maybe gradient encoding specific.
  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    I hope the authors can open-source their code and provide a more detailed description of the training and testing processes.

  • 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.

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

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

    The compartment biophysical model and self supervise learning is interesting, and experiments are relatively solid.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A




Author Feedback

We sincerely thank the reviewers for their constructive feedback. Below we address the major concerns:

  1. “Rapid” claims and runtime (R1,R4) “Rapid” refers to inference (<10s per subject), which outputs fODFs+microstructure in one forward pass. While subject-specific training takes ~30min on a single NVIDIA RTX 3090, this is a fully automated one-time process per subject. For comparison, unsupervised learning baselines NODF[3] and RT-ESD[6] require training times of hours to days per subject. Also, classical CSD pipelines vary: SS3T-CSD typically takes 10-30min per subject, while standard CSD is faster but less accurate. Unlike these methods requiring pre-computed response functions, UFO-3 jointly estimates all parameters end-to-end, justifying its training time investment. We recognize the potential confusion and will replace “rapid” with “efficient” while clarifying these specific performance advantages in the revised manuscript.

  2. Comparisons and in vivo evaluation (R1,R3)
    • Synthetic: SS3T-CSD is designed for three physical tissues (WM/GM/CSF) whereas our network learns three abstract compartments. To maintain consistent assumptions when evaluating WM metrics (ACC/MAE), we compared against single-tissue CSD and three state-of-the-art unsupervised methods.
    • In vivo: Ground-truth fODFs are unavailable, so we follow common practice using qualitative assessments and indirect validation (Fig3) across five CHCP subjects. This cohort size (n=5) is comparable to related work (RT-ESD: 6; NODF: 1) and consistently yields anatomically plausible reconstructions of major tracts with competitive overlap scores.
  3. Grid density and generalization (R3) We use a 3072-point HEALPix grid following [6], balancing angular resolution and stability: lower densities yield unstable optimization, higher densities add little accuracy but increase memory/time costs. Extension to multi-shell/higher b-value data (e.g. LARDI,DiSCo) is noted as future work (§4.1).

  4. Manuscript clarity (R4)
    • P(s): We will revise §2.1 to clearly define P(s) as the projection of q-space signals onto a uniform HEALPix grid of V points and annotate Eq(2) accordingly.
    • Patch vs. pointwise convolutions: The network uses 1×1×1 convolutions, ensuring voxel-wise independence. Training and inference operate on 3×3×3 voxel patches for batching efficiency without introducing spatial context.
    • Fig4a caption: We will correct to BPS instead of CST.
    • Reliance on [30]: While we reference [30] for deeper theoretical foundations, we will expand our explanation of key principles to ensure standalone readability with the page limit.
  5. Tractography fullness (R4) The denser frontal pathways in UFO-3 recover high-curvature fibers that conservative GT tractography may omit. Per [22], even with perfect angular precision using GT orientations, the ISMRM2015 phantom underrepresents complex fiber configurations because tractography is fundamentally an ill-posed problem. Our quantitative overlap analysis (Fig4b) shows that UFO-3 maintains competitive tract-level consistency. These additional streamlines likely represent anatomically plausible pathways absent from the conservative GT. Future work will validate these pathways against ex vivo data.

  6. Others
    • Reproducibility (R4): Our codebase and synthetic dataset will be released upon acceptance. The final manuscript will include architectural details, hyperparameters, and preprocessing steps.
    • Simulation (R1): We generated 2500 test cases: 5 fiber configs×5 SNRs×100 orientations, including single-fiber and multi-fiber crossings (90°,60°,45°) consistent with [30].
    • Number of b1000 directions (R1): The 92 refers to directions in the raw acquisition. After distortion correction and preprocessing, 46 b1000 directions remain, which is what our model actually uses. We will clarify this in the revised manuscript.

We will incorporate all these changes in our revision and appreciate the reviewers’ guidance in strengthening our work.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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



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’

    This paper addresses an important topic in accurate FOD estimation using single-shell dMRI. Despite some minor ambiguities in the methods, the approach appears promising. However, I agree with the other reviewers that it cannot be described as a rapid method, as it requires subject-specific training that takes approximately 30 minutes.



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’

    N/A



back to top