List of Papers Browse by Subject Areas Author List
Abstract
Multi-shell diffusion MRI (dMRI) unlocks more expressive microstructural modeling than single-shell scans, yet its longer acquisition time hinders deployment in large-scale cohorts and time-constrained clinical settings. Synthesizing an unobserved shell from a single-shell input is fundamentally ill-posed and further complicated by protocol mismatch, where source and target gradient direction sets may not align. We propose DTI-SHNet, a single-to-multi-shell synthesis framework that operates in the real symmetric spherical harmonics (SH) coefficient domain and performs spatially aware volumetric regression. Given a source shell, we estimate diffusion tensor imaging (DTI) and use direction-agnostic parametric maps along with a brain mask as conditioning priors to guide a 3D U-Net regressor for source-to-target SH coefficient regression. To couple coefficient accuracy with signal fidelity, we introduce a signal consistency regularization that reconstructs signals on randomly sampled canonical directions from predicted coefficients and enforces agreement in the signal domain. Experiments on UK Biobank and Cam-CAN data for b=1000 to b=2000 dMRI synthesis show that DTI-SHNet achieves competitive visual quality compared to advanced methods, while better preserving downstream diffusion measures. Our code is available at https://github.com/xiaovhua/dti-shnet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4245_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xiaovhua/dti-shnet
Link to the Dataset(s)
UK Biobank: https://www.ukbiobank.ac.uk/
Cam-CAN: https://cam-can.mrc-cbu.cam.ac.uk/dataset/
BibTex
@InProceedings{LiBin_DTIGuided_MICCAI2026,
author = { Li, Binghua AND Andica, Christina AND Liang, Tong AND Chang, Ziqing AND Li, Chao AND Uchida, Wataru AND Takabayashi, Kaito AND Zhao, Qibin AND Tanaka, Toshihisa AND Sun, Zhe AND Aoki, Shigeki},
title = { { DTI-Guided Volumetric Spherical Harmonics Regression for Single-to-Multi-shell dMRI Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presented DTI-SHNet, a deep learning model for synthesizing multi-shell diffusion MRI from single-shell dMRI. Based on prior works, spherical harmonics (SH) representation was used to guide this synthesis pipeline. Experiments were conducted on two datasets, with different parameters and downstream tasks.
- 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.Novel method design for a well-motivated problem: synthesizing multi-shell dMRI using single-shell input would be valuable to support advanced dMRI analysis and modelling with fast scans. 2.Well-written texts and scientifically-decent figures. 3.Comprehensive comparison with prior methods, solid experiments, and convincing results showing improvement over other methods.
- 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.References seemed not to be linked to papers online (won’t jump to online full-text after clicking) 2.Worth exploring how the model performance degrades when input DTI are generated from a sparse single-shell acquisition (common in clinical practice) 3.In Fig. 2, the coefficient patch looks T1-weighted. This is worth discussing: is it only for illustration of patching operation (in that case the real patches need to be used) or are patches indeed have this T1-weighted looking (in that case the reason needs to be discussed)
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
A well-written paper with solid methodology and comprehensive evaluations
- Reviewer confidence
Very confident (4)
- [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 addressed my comments. My recommendation remains unchanged.
Review #2
- Please describe the contribution of the paper
The paper describes a method to extrapolate diffusion MRI signal from one b-value to a higher b-value, i.e. from one b-shell to another. The harmonic coefficients of source shell together with DTI-based quantities are fed to a U-Net to regress higher-shell coefficients.
- 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 paper is very readable, the method is clear, and the background is sufficiently explained.
- 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.Lack of clinical/methodology motivation It is not clear to the reviewer for what purpose would one extrapolate to multi-shell to then regress a NODDI model from these 2 signals (1 real signal and 1 extrapolated). Arguably this would be essentially equivalent to regress the NODDI maps directly from single shell
2.Weak contribution vs existing work This work essentially presents a U-Net to regress SH coefficients, which is almost the same as SH-CNN [4]. This is confirmed by the weak results on table 1
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
1.Stronger contribution I would suggest the authors to strengthen this work by taking advantage of more recent advances in deep neural nets, attention-based blocks, conditioned on DTI signals for instance, should give a boost in results
2.Stronger evaluation Another suggestion is to have less biased evaluation: Here you are comparing your GT multi-shell to a multi-shell consisting of 1 GT and 1 synthesised shell. You would need to remove that bias by comparing purely synthesised to pure GT. Similarly, I would have expected quantitative comparisons of downstream task multi-shell quantities (e.g. NODDI maps).
3.Presentation Don’t use Comic Sans font in figures for a scientific paper
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The major factors for this score are:
- Lack of clinical/methodology motivation
- Weak contribution
- Unconvincing results
- 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
Review #3
- Please describe the contribution of the paper
This paper presents an approach to synthesise multi-shell diffusion MRI from single-shell data. The approach uses diffusion tensor features (FA + MD) derived from the single-shell data as conditions on a generative model which operates in the spherical harmonics basis to synthesise spherical harmonic coefficients of a different shell. By operating in SH space, it allows synthesis of arbitrary directions in the new shell, and a loss term utilises this to promote signal consistency.
- 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.
- Well written, problem is clearly formulated and solution well presented
- The proposed method seems to be a sensible extension/combination of previous approaches to try to get the best of different methods
- The model shows good signal reconstruction and DTI metric performance, particularly in Cam-CAN dataset
- Ablation study is a nice addition, demonstrating the value of the proposed model elements.
- 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 the claimed benefits of this approach is better handling of gradient direction mismatch between shells. Explicitly, DirUNet is setup as a comparator which relies on gradient matching. However it appears that DirUNet performs similarly to DTI-SHNet on UKB from Table 1, even slightly better PSNR and NMSE, which suggests that at least in this example, the gradient dimension mismatch handling of DTI-SHNet is no better than DirUNet.
- Relatedly, do you have any comment on why performance is better in Cam-CAN than UKB even though it is the smaller of the datasets (so seen less in training)?
- I don’t really understand why there is a centre crop and then model is trained on smaller patches - is the centre crop chosen to make sure that it can cover the whole brain?
- I don’t think that the NODDI evaluation adds much - the figure is quite small so seeing the differences in the histogram that you mention is not that easy. Why not do a quantitative analysis similar to that for FA and MD done in Table 2?
- I find that there are details lacking in some places, for instance:
- “We also derive a brain mask M(x) from S0(x) via standard extraction” - it is not clear to me what ‘standard extraction’ means, at least provide a reference
- I would also add a citation for the “gradient distortion correction, susceptibility distortion correction, and eddy current and motion correction”
- I found the description of the training setup slightly confusing. You say “We adopt a subject-wise split with 90% of subjects for training and 10% for evaluation”. I take this to mean that the 10% are the final test-set that you’re reporting the results on, but then what is happening with the 90%? Is there any train/validation split to help monitor/avoid overfitting and what is your training stopping condition? Even just stating how many epochs are used would help.
- I appreciate that there is limited space for discussion but a clear limitation of this approach is that it can only work on the specific b-value pair it’s trained for which limits generalisability.
- 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 submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a new method for simulating multi-shell diffusion MRI from single shell. The method is well presented and justified and shows an improvement over competitors in some regards. I have some concerns, detailed above, around some of the results but overall the method shows a modest improvement on the state-of-the-art.
- 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.
I am satisfied by the authors’ comments. They propose to fix the few issues I raised and I am happy with the responses to other more open questions I asked.
Author Feedback
We sincerely thank reviewers for their constructive comments. [R1] Reference links Fixed. [R1] Worth exploring … sparse single-shell acquisition We explored sparse source directions in Cam-CAN. For PSNR/SSIM/RMSE/ΔFA/ΔMD, 15 directions give 28.7/0.966/0.028/0.039/0.49, while 6 directions give 27.7/0.959/0.035/0.047/0.58, showing stable performance under moderate sparsity and gradual degradation as expected. [R1] Figure presentation Thank you for pointing out the typo in Fig. 2.We only use dMRI and DTI priors, and will replace the misleading T1 patch with an SH coefficient patch. [R2] Motivation: Why synthesize multi-shell instead of directly predicting NODDI? Multi-shell synthesis is more flexible than predicting one specific map. The synthesized shell can support multiple analyses beyond NODDI, e.g. DTI, DKI and free-water imaging, whereas a NODDI-specific model is less reusable. We will clarify this motivation. [R2] Weak contribution vs. [4] (DirUNet) We respectfully disagree. [4] is not an SH-CNN; it performs low-to-high b-value regression in raw volume space and uses nearest-direction matching for unpaired directions. In contrast, we regress in the SH domain, enabling synthesis for arbitrary target directions. We further introduce direction-agnostic DTI conditioning and signal-consistency regularization to address ambiguity and signal fidelity in SH regression. [R2] More advanced architectures, such as attention-based blocks We partially agree that advanced modules may further improve performance. To examine this, we tested bottleneck spatial self-attention on a Cam-CAN subset, which slightly improved the results from 28.9/0.968/0.027/0.036/0.44 to 29.1/0.969/0.025/0.033/0.44, but increased parameters by 4.7% and required more 3D memory. Thus, we believe the current design provides a simple yet reasonable trade-off. We will explore advanced architectures in future work. [R2] Compare purely synthesised to pure GT, and quantitative evaluation for NODDI maps We agree and will add quantitative NODDI evaluation in revision. Regarding evaluation bias, our downstream evaluation already follows [4]: the acquired source shell and synthesized high-b shell form the synthesized multi-shell dMRI for downstream evaluation. This matches the intended case, where only the missing high-b shell is synthesized. Since NODDI requires multi-shell dMRI, evaluating only the synthesized high-b shell is not applicable. [R2] Don’t use Comic Sans font Fixed. [R3] Competitive DirUNet performance on UKB This is mainly due to the 50 source directions in UKB, as DirUNet’s nearest-direction matching benefits from dense references. In 6-direction UKB experiments, visual drops are similar for DTI-SHNet and DirUNet: 0.5/1.4/0.014 vs. 0.5/1.6/0.011 in PSNR/SSIM/RMSE. However, downstream drops are smaller for DTI-SHNet: 0.014/0.08 vs. DirUNet’s 0.035/0.40 in ΔFA/ΔMD, suggesting better robustness under sparse directions. [R3] Cam-CAN vs. UKB gap The gap is due to data distribution, not size alone. With matched-size training subsets of 200 subjects, the gap remains in our model, with differences of 3.8/4.7/0.059/0.005/0.10.This suggests gradient directions and cohort difference are major factors. [R3] Center crop and patch training Center cropping standardizes scans to a unified shape while preserving the whole brain. Patch training improves memory efficiency and augmentation for 3D regression, as in nnU-Net and 3D MedDiffusion. Full-volume inference uses overlapping sliding windows. [R3] Quantitative analysis on NODDI We will add quantitative NODDI errors in revision. [R3] Reproducibility details We will add the brain-mask definition, preprocessing references, train/test split, epochs, and training details. Fixed epochs are chosen from preliminary convergence experiments without using the test set. Final checkpoints are evaluated. [R3] Fixed b-value pair limitation We agree that the current model is trained for a fixed mapping, and will discuss carefully.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. 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 reviews show clear disagreement: one strong accept contrasts with a firm reject citing weak motivation and unconvincing results, while another leans weak accept with noted concerns. Given this variance, a rebuttal is necessary to clarify the paper’s clinical/methodological motivation and to strengthen the evidence supporting the reported improvements. The authors should particularly address concerns about the significance and robustness of results, and better justify the novelty and practical impact of the proposed method.
- 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.
All reviewers recommended acceptance. The authors have adequately addressed the reviewers’ questions and concerns. The paper is therefore recommended for acceptance.
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.
Significant concerns could be addressed and were well received by two of the reviewers. The presented deep learning model for synthesizing multi-shell diffusion MRI from single-shell dMRI has the strengths of a novel methodological contribution for a key problem (allowing advanced dMRI analysis with short scans), comprehensive comparison and a well written, clear manuscript - which now that other concerns were addressed - might result in positive discussions and stimulating poster conversations.
Meta-review #3
- 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.
The authors have produced a technically sound piece of work that addresses an important problem in dMRI analysis. However, the marginal improvements over existing methods and the lack of compelling clinical and methodological justification mean that the paper does not meet the rigorous standards required for acceptance at MICCAI. While the authors have effectively addressed all major technical concerns from the review process and clarified aspects of the method’s novelty and practicality, the paper still falls short of MICCAI’s high acceptance standards. MICCAI requires submissions to demonstrate significant contributions that advance the state-of-the-art in medical image computing and computer-assisted intervention. In this case, the marginal performance improvements and unresolved motivation issues prevent the paper from meeting this bar. It is also important to note that, per MICCAI guidelines, the rebuttal should not introduce new experimental results. Based on the above evaluation, I recommend rejection of this paper. I encourage the authors to strengthen the motivation for their approach, conduct more extensive comparisons with state-of-the-art methods, and explore additional clinical applications to demonstrate the practical value of their work for future submissions.