List of Papers Browse by Subject Areas Author List
Abstract
Diffusion MRI (dMRI) is sensitive to microstructural barriers, yet most existing methods either assume impermeable boundaries or estimate voxel-level parameters without recovering explicit interfaces. We present \textit{Spinverse}, a permeability-aware reconstruction method that inverts dMRI measurements through a fully differentiable Bloch–Torrey simulator. Spinverse represents tissue on a fixed tetrahedral grid and treats each interior face permeability as a learnable parameter; low-permeability faces act as diffusion barriers, so microstructural boundaries whose topology is not fixed a priori (up to the resolution of the ambient mesh) emerge without changing mesh connectivity or vertex positions. Given a target signal, we optimize face permeabilities by backpropagating a signal-matching loss through the PDE forward model, and recover an interface by thresholding the learned permeability field. To mitigate the ill-posedness of permeability inversion, we use mesh-based geometric priors; to avoid local minima, we use a staged multi-sequence optimization curriculum. Across a collection of synthetic voxel meshes, Spinverse reconstructs diverse geometries and demonstrates that sequence scheduling and regularization are critical to avoid outline-only solutions while improving both boundary accuracy and structural validity. Our source code is available at \href{https://github.com/BioMedAI-UCSC/spinverse}{https://github.com/BioMedAI-UCSC/spinverse}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6164_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/6164_supp.zip
Link to the Code Repository
https://github.com/BioMedAI-UCSC/spinverse
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KhoPra_Spinverse_MICCAI2026,
author = { Khole, Prathamesh Pradeep AND Brenes, Mario M. AND Petiwala, Zahra Kais AND Mirafzali, Ehsan AND Gupta, Utkarsh AND Li, Jing-Rebecca AND Ianus, Andrada AND Marinescu, Razvan},
title = { { Spinverse: Differentiable Physics for Permeability-Aware Microstructure Reconstruction from Diffusion MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a novel inverse-design framework that integrates a fully differentiable Bloch-Torrey PDE solver, allowing for the direct reconstruction of explicit microstructural boundaries from dMRI signals without the need for large-scale training data.
- 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 introduces a fully differentiable Bloch-Torrey simulator integrated into an inverse-design pipeline. This approach embeds the fundamental physics of diffusion MRI directly into the optimization loop.
- The method successfully unifies topological freedom with biophysical realism. By parameterizing boundaries as continuous face permeabilities, it not only allows complex, non-predefined shapes to emerge (topology-free), but also explicitly models semi-permeable water exchange—a critical physiological factor often ignored by traditional compartmental models. 3.The paper demonstrates a sophisticated staged optimization strategy. The authors provide a robust solution to the local minima problem inherent in complex inverse physics problems.
- 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 iterative optimization process is computationally expensive, taking approximately 33 minutes per voxel on a single 24GB NVIDIA L4 GPU GPU. Thus the method faces severe scalability issues when attempting to reconstruct higher-resolution grids or larger volumes, making it currently impractical for whole-brain or even small-ROI clinical applications. 2.Despite the proposed curriculum learning strategy, the inverse problem remains highly non-convex. The reconstruction quality is heavily dependent on the specific choice of b-values and diffusion times. 3.The evaluation is exclusively performed on noise-free or low-noise synthetic data. Clinical dMRI is frequently corrupted by Rician noise, eddy current distortions, and motion artifacts. It is unclear how the emergent permeability field would behave under realistic SNR conditions, where noise could be misinterpreted as phantom boundaries.
- 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.
- 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 integration of a fully differentiable Bloch-Torrey PDE solver into an optimization-based reconstruction pipeline is a significant technical achievement. The use of face-wise permeabilities on a fixed tetrahedral mesh is a nice solution to the long-standing challenge of topology-constrained deformation in mesh-based methods. The ability to recover complex structures like tori without a predefined template is a good contribution. However, the reconstruction time of 33 minutes per voxel presents a major hurdle for clinical adoption. Without a clear path toward acceleration (e.g., through hybrid learning-physics approaches), the method’s impact remains limited to small-scale research scenarios. The evaluation relies exclusively on synthetic data. The lack of robustness testing against clinical-grade noise and artifacts leaves the method’s real-world performance unverified.
- 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 #2
- Please describe the contribution of the paper
Spinverse proposes an inverse problem formulation for reconstructing microstructural boundaries from diffusion MRI (dMRI) signals. The core idea is elegant: rather than pre-segmenting a mesh into tissue compartments or optimizing vertex positions, the method parameterizes boundary location implicitly through a face-wise permeability field on a fixed tetrahedral grid. Low-permeability faces act as diffusion barriers, so boundaries emerge organically via optimization rather than by constraining mesh topology. The forward model is a differentiable re-implementation of the Bloch-Torrey PDE (via PyTorch/Autograd), and inversion proceeds via Adam with MRF-based spatial regularization and a staged multi-sequence curriculum. The paper is technically ambitious and addresses a genuine gap in the dMRI microstructure literature: existing differentiable solvers (ReMiDi) fix topology, while compartment models (NODDI, SANDI) do not reconstruct explicit boundaries.
- 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 topology-free boundary representation. 2.End-to-end differentiable pipeline. 3.Stronger boundary validity than learned baselines.
- 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.Evaluation scope is very limited. 2.Scalability and computational cost are prohibitive. 3.Multi-compartment performance is poor without clear diagnosis. 4.Threshold τ_b is not robustly validated. 5.Missing implementation details (The paper does not specify how many tetrahedra the ambient mesh contains, what the mesh resolution is relative to the ground-truth geometry, or how many eigenmodes are retained in the truncated eigendecomposition.) 6.compare against non-differentiable physics-based microstructure inversion approaches (e.g., ActiveAx, CHARMED) or recent learned inverse problem solvers (unrolled networks, deep equilibrium models) that could serve as alternative baselines for the inverse problem formulation
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Major Revision. The core formulation is novel and technically sound, but the evaluation is too narrow for a MICCAI publication. At minimum: (1) add a noise sensitivity study, (2) provide concrete mesh/scalability statistics, (3) deepen the multi-compartment failure analysis, and (4) include a threshold sensitivity experiment.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
The paper proposes a (to my knowledge) novel parameterization of tissue microstructure within a voxel as a collection of permeabilities between faces in a 3D mesh. An FEM discretization of the Bloch–Torrey PDE is reparameterized in terms of the field of permeabilities and some existing numerical techniques are employed to speed up the forward solves. Inversion targets recovery of the face permeability field and is a accomplished through regularized MSE, with regularizers related to the continuity + smoothness of the permeability field. The estimates are formed in a multistage procedure, first optimizing over long diffusion time acqusitions and then using this as an initialization for shorter times. The method is evaluated over simulated data.
- 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) This reviewer appreciates the novelty and ambition of the work. To my knowledge, this is a novel, rich parameteterization of the microstructure and this work lays the ground work of an inversion scheme that can be subsequently built upon.
2) The experiments, while simple and only use simulated data, are comprehensive, demonstrate the potential of the method and clearly highlight the effect of key methodological components, e.g. multi-stage optimization, priors, etc.
3) Authors explicitly identify configurations the method struggles in (multi-compartment), which is both refreshing and helps promote subsequent research.
- 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) This reviewer is skeptical on the statistical identifiability of model parameters (the kappas) in practical diffusion acqusitions with realistic noise levels. Perhaps some between voxel spatial regularzation could help here (extending priors to adjacent faces in neighboring voxels?). In any case, this is beyond the scope of the current work, but could be investigated in the journal version.
2) Related to 1, ways to select prior regularization strength hyperparameters will be needed for practical adoption, but again this is well beyond the scope.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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?
The work is well motivated, novel and ambitious. While it may be a ways away from practical use in dMRI inversion, it is a promising direction and provides foundational framework that can be built upon. The paper should be accepted.
- 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.
This work is a novel contribution to microstructure modeling and inversion in dMRI. It is likely ways away from practical applied use, but that is acknowledged by the authors and I do not see it as a drawback. I was strongly in favor of the paper being accepted, and nothing about the rebuttal has changed my position.
Author Feedback
We thank all reviewers for their careful reading and constructive feedback. We are pleased that the novelty of the topology-free permeability parameterization and differentiable Bloch-Torrey pipeline is recognized across all reviews, and address the major concerns below. [Noise Robustness (R1, R2)] Noise sensitivity analysis is an important validation direction we acknowledge. With space constraints in the submission, we chose to focus on establishing the theoretical framework and its viability against learned baselines. Regarding eddy currents and motion, these are preprocessing-level corruptions handled prior to microstructure estimation by standard pipelines (NODDI/SANDI). Comprehensive noise characterization is central to the planned journal extension. [Baseline Comparisons - ActiveAx and CHARMED (R2)] ActiveAx/CHARMED are parametric compartment models in the same category as NODDI/SANDI, which we already discuss in the paper as representative of this class. They produce no explicit boundary reconstruction; a Chamfer distance comparison is undefined for methods that output no surface. Our learned baselines (MLP, GNN) are the natural alternatives, they operate on the same tetrahedral mesh and output face permeabilities directly. Unrolled networks and deep equilibrium models are valid for inverse imaging problems broadly, but adapting them to mesh-based boundary reconstruction is non-trivial and constitutes additional work outside the scope of the paper. [Multi-Compartment Failure Diagnosis (R2)] The multi-compartment results provide an honest representation of where the method succeeds (single-axon) and where it struggles (two-axon, multi-compartment). We explicitly acknowledge in our paper that multi-compartment optimization remains an open problem. The paper identifies multiple candidate explanations: (i) rotational degeneracy collapses reconstructions toward symmetric envelopes, (ii) inferring multiple interfaces from the same acquisition is inherently more ill-posed, and (iii) single-compartment regularization strengths may collapse multi-compartment boundaries. We are actively pursuing solutions to this problem. [Missing Implementation Details (R2)] These details were omitted from Sec 4.The reconstruction and reference share the same ambient mesh (576 tetrahedra, 189 vertices, 1296 faces); they differ only in permeability values assigned to faces, so mesh resolution is identical by construction. The truncated eigendecomposition retains 500 eigenmodes per mesh. These will be added to the Setup paragraph. [Computational Cost (R1, R2)] We acknowledge that 33 minutes per voxel is a current limitation of the approach. Physics-based inversion requires a full PDE solve per optimization iteration, inherently more expensive than parametric methods. Unlike learned methods, Spinverse requires no training data and generalizes across acquisition protocols and mesh resolutions without retraining; a flexibility that comes at inference-time cost. Accelerating physics-based microstructure solvers is an active area of our research, including hybrid learning-based initialization and optimized eigensolvers to reduce spectral decomposition cost. [Threshold Analysis (R2)] Tau_b=10^-3 is the log-space midpoint of the optimized permeability range [10^-5, 10^-1], which was determined empirically from forward model experiments. The threshold was further informed by the distribution of optimized face permeabilities at convergence, where boundary faces consistently fall below this value. [Non-Convexity and Acquisition Sensitivity (R1)] Regarding acquisition sensitivity, non-convexity is mitigated but not eliminated by the staged curriculum(Fig3), which avoids gradient-conflict failure modes of joint optimization, and the regularizers(Eq3) constrain the search to physically plausible fields. Achievable accuracy is fundamentally bounded by signal information content, which is inherent to dMRI inverse problems and not specific to Spinverse.
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.
Reviews are mixed, authors should respond to issues raised specifically those concerning setting (noise free models being unrealistic to the point of being unusable), computational cost, and comparison baselines which may be missing (i.e., non-differentiable inverse problem solvers).
- 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.
Accepted, though admittedly seems borderline w.r.t. the stack. Clear technical advances, but far from clinical impact (as R1 and R3 note) due to problems in acquisition. Still, these ideas are valuable in my opinion.
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.
Original reviews were on the positive side already, and the rebuttal addresses the reviewer concerns to the extent possible without introducing new experimental results. Even though this physics modeling approach is certainly not the type of work that will yield immediate clinical benefit, I would view it as a welcome and valuable contribution to the conference program in terms of diversity of ideas.
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.
After reviewing the initial feedback, author rebuttal, and post-rebuttal comments, I recommend accepting this paper for presentation at MICCAI. All reviewers recognized the groundbreaking topology-free permeability parameterization and fully differentiable Bloch-Torrey PDE solver. This method overcomes the limitations of traditional compartment models by allowing complex microstructures to emerge from the dMRI signal while realistically modeling semi-permeable boundaries. The authors have addressed the reviewers’ major concerns, clarified methodological choices, and candidly discussed the method’s limitations. The reviewer endorsed the paper, noting its potential to invigorate the dMRI community. While initial concerns from Reviewers #1 and #2 have largely been resolved, the authors responsibly highlight current limitations, particularly regarding computational costs and multi-compartment performance. This foundational work aligns with MICCAI’s focus on medical image computing and presents a technically sophisticated solution to a fundamental problem in dMRI analysis.
