List of Papers Browse by Subject Areas Author List
Abstract
Diffusion MRI microstructure fitting is nonconvex and often performed voxelwise, which limits fiber peak recovery in narrow crossings. This work introduces PRISM, a differentiable analysis-by-synthesis framework that fits an explicit multi-compartment forward model end-to-end over spatial patches. The model combines cerebrospinal fluid (CSF), gray matter, up to K white-matter fiber compartments (stick-and-zeppelin), and a restricted compartment, with explicit fiber directions and soft model selection via repulsion and sparsity priors. PRISM supports a fast MSE objective and a Rician negative log-likelihood (NLL) that jointly learns σ without oracle information. A lightweight nuisance calibration module (smooth bias field and per-measurement scale/offset) is included for robustness and regularized to identity in clean-data tests. On synthetic crossing-fiber data (SNR=30; five methods, 16 crossing angles), PRISM achieves 3.5° best-match angular error with 95% recall, which is 1.9× lower than the best baseline (MSMT-CSD, 6.8°, 83% recall); in NLL mode with learned σ, error drops to 2.3° with 99% recall, resolving crossings down to 20°. On the DiSCo1 phantom (NLL mode), PRISM improves connectivity correlation over CSD baselines at all four tracking angles (best r=.934 at 25° vs. .920 for MSMT-CSD). Whole-brain HCP fitting (~741k voxels, MSE mode) completes in ~12 min on a single GPU with near-identical results across random seeds.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5506_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/m-agour/Prism
Link to the Dataset(s)
DiSCo1 phantom (Rafael-Patino et al., 2021), fetched via DIPY: dipy.data.fetch_disco HCP subject 100307: https://www.humanconnectome.org/ Stanford HARDI, fetched via DIPY: dipy.data.fetch_stanford_hardi
BibTex
@InProceedings{AboMoh_PRISM_MICCAI2026,
author = { Abouagour, Mohamed AND Shah, Atharva Jaydeep AND Garyfallidis, Eleftherios},
title = { { PRISM: Differentiable Analysis-by-Synthesis for Fixel Recovery in Diffusion MRI } },
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
The work proposes a methodology for joint inversion of multi-compartment stick and zepplin models and nuisance bias field, for both Gaussian and Rician noise models. The microstructure parameters, multiplicative bias field and noise model parameters are estimated jointly using gradient based optimization, along with spatially prior to promote smoothness in fiber directions and compartment volume fractions.
- 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 found the methodology novel, ambitious and well motivated. Particularly the integration of components typically considered “pre-processing” directly into the microstructure inverse problem is a good direction.
2) The quantitative results are promising in comparison to standard techniques, if not state of the art techniques at this point (see item 3 on weaknesses).
- 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) A key component of the method is spatial regularization. Particularly, the authors mention several priors used to regularize the peaks, including a “direction continuity”. What is this penalty function? How it handle/associate multiple peaks? More explanation would be helpful.
2) How are the prior regularizer penalties (the lambdas) set? Presumably, these will depend on acquisition parameters and SNR. Can the authors provide guidance on how to tune/set them in practice for different datasets. How sensitive is the method to their values?
3) There are a variety of recent (deep) learning-based approaches to FOD estimation which have reported to largely surpass the performance over CSD family of methods (see Section 4.5 of [1] for a non exhaustive list). The numerical experiments would be improved by comparing with one (or several) representative examples from this class of methods.
4) The method fixes the diffusivities. This can lead to bias in peak identification when microstructure differs significantly from nominal values [2]. This should at least be acknowledged, and ideally evaluated via simulation.
5) Some visual comparison of the recovered FOD fields between the methods would be a very helpful supplement to the quantitative results for evaluation.
[1] Davood Karimi, Simon K. Warfield; Diffusion MRI with machine learning. Imaging Neuroscience 2024. [2] Guo F, et. al. Fiber orientation distribution from diffusion MRI: Effects of inaccurate response function calibration. J Neuroimaging. 2021.
- 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.
(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?
Paper is borderline. Method is well motivated and novel and results are promising. Concerns remain about the number of tuning parameters and the sensitivity of the methods performance to them, as well as lack of qualitative results on the FOD reconstructions, particularly in crossing fiber regions, to supplement the quantitative results (see weaknesses for more details);
- Reviewer confidence
Very confident (4)
- [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.
Promising paper on the borderline. Main cited weakness of original submission were lack proper evaluation of sensitivity to model hyper-parameters/clear guidance for choosing them across different datasets, lack of qualitative evaluation on real data, and questionable relevance of comparisons/omissions. These issues are not adequately addressed by rebuttal. Paper would benefit from larger revision.
Review #2
- Please describe the contribution of the paper
The paper proposes PRISM, a fitting framework for diffusion MRI microstructure that combines an explicit multi-compartment analytic signal model (CSF, GM, K white-matter stick-and-zeppelin fibers, and a restricted compartment) with a nuisance calibration module (bias field, per-measurement affine correction, learned noise level). All parameters are optimized jointly end-to-end using Rprop. The method supports two data-fidelity modes: MSE and a Rician NLL that also learns the noise level self-supervised. It is evaluated on synthetic crossing-fiber data, the DiSCo1 phantom, and an HCP subject against CSD, MSMT-CSD, and ODF-FP.
- 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 end-to-end joint optimization of tissue parameters together with nuisance calibration (bias field, per-measurement affine, noise level) is a clean and to my knowledge novel reformulation of a pipeline that is typically split into separate preprocessing and fitting stages.
- The Rician NLL formulation with self-supervised sigma recovery is a useful addition and produces a meaningful improvement over MSE.
- The crossing-fiber benchmark is designed to be in the author’s disfavor: PRISM uses a deliberately mismatched D_perp (0.4 versus the true 0.3), while CSD and MSMT-CSD receive oracle response functions. Outperforming oracle-initialized baselines under this setup is a strong result.
- The related work section is quite thorough and detailed to my knowledge and correctly situates PRISM against voxelwise classical fitters, learning-based voxelwise mappings, mesh-based differentiable simulators, and dictionary approaches.
- Evaluation is quite through since it spans synthetic, phantom, and in-vivo data with multiple metrics.
- 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 forward model and the source of differentiability are not clearly specified. Equations 3-5 define a closed-form analytic signal, so “differentiability” here appears to mean standard autodiff over analytic compartments rather than a custom differentiable solver. The term “differentiable” often implies a custom solver in related work, so stating explicitly that the forward model is closed-form with gradients from autodiff (and naming the framework such as PyTorch, Julia or JAX) would resolve the ambiguity and distinguish the contribution from differentiable simulator-based methods.
- The DiSCo1 ablation shows the restricted compartment drops r by 9.4 pp when removed, versus 1.9 pp for topology and 0.7 pp for spatial priors. Without the restricted compartment PRISM scores r = 0.840, below MSMT-CSD at 0.920.Since the baselines do not include this compartment, part of the headline gain is model capacity, not the analysis-by-synthesis framework. Either give MSMT-CSD an equivalent compartment or report PRISM without the restricted compartment alongside baselines.
- No sensitivity analysis for regularization. The method uses at least six tissue regularizers (spatial smoothness, direction repulsion, sparsity, orphan-WM, directional continuity, fiber ordering) plus calibration L2 and Total Variation, all with fixed global weights. It is not clear which are load-bearing, and the calibration regularizer weights and the 8^3 bias-field grid size are not justified or discussed.
- No qualitative results. Only quantitative metrics are reported. Visualizations of fiber orientations, ODFs, or tractography on the phantom or HCP against baselines would support the numbers better and show failure modes.
- The “differentiable analysis-by-synthesis” framing somewhat overclaims. In practice this is gradient-based fitting of an analytic multi-compartment model with joint nuisance estimation, which is close to what the NODDI or SANDI toolbox already do. Comparing against those methods, not only simulator-based ones, would make the actual difference (joint nuisance calibration, Rician NLL with learned sigma, patch-based spatial regularization) clearer.
- 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
Visualizations: No qualitative figures are provided. Fiber orientation overlays on synthetic crossings, ODFs or tractograms on DiSCo1, and tissue fraction maps on HCP would support the quantitative claims and expose failure modes. A recovered bias field on HCP would also show the calibration module is doing real work rather than self-regularizing to identity. Regularizer values, their usage and effects: Six tissue regularizers plus calibration L2 and TV are used with fixed global weights and no sweep. A short sensitivity analysis on the main weights (lambda_sp, lambda_rep, lambda_sparse, calibration L2, bias-field grid size) would clarify which terms are load-bearing.
- 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 framework with joint tissue and nuisance calibration, Rician NLL with self-supervised sigma, and solid results across synthetic, phantom, and in-vivo data. The handicapped-but-winning synthetic benchmark supports the accuracy claim, and the Rician NLL with learned sigma is broadly applicable beyond this setting and thus a good contribution. That said, the method is closer to prior analytical fitters like NODDI than the paper makes it sound, the DiSCo1 gain largely comes from the restricted compartment the baselines lack, the source of differentiability is not stated, and there are no qualitative results or regularizer sensitivity analysis. A clearer positioning against NODDI, a compartment-matched ablation, and qualitative visualizations would address the main concerns.
- 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 recommend accept with revisions, contingent on how many of the points below can be addressed in the camera-ready. Point 1: I agree with the clarification, but the mode of differentiability should be stated explicitly in the paper. Without naming the framework (PyTorch autograd in this case), a reader cannot tell whether gradients come from autodiff over the analytic compartments, a dedicated backward pass, or a differentiable simulator component. Stating this also addresses the “differentiable analysis-by-synthesis” framing, which otherwise reads as ambiguous. Point 3: The rebuttal acknowledges the issue and revises the framing, which partially addresses my concern. The fact remains that PRISM without the restricted compartment scores 0.840 versus MSMT-CSD at 0.920 on DiSCo1, so part of the headline gain comes from added model capacity rather than from the analysis-by-synthesis framework itself. I would like to see the camera-ready be explicit about this in Sec. 4 and the discussion, or include a short study isolating the restricted-compartment effect in an extended version. A short discussion contrasting restricted versus unrestricted diffusion behavior would also help. Point 6: This answers most of my questions about regularizer weights and their fixed values across datasets. For the extended version, I would like to see a proper sweep over regularizer effects, which terms each setting is sensitive to, and which are load-bearing. Point 7: Qualitative figures within the camera-ready additional page would substantially support the quantitative claims. I recognize space is limited, so a fuller set in the extended version would be appropriate. The idea is novel, and the joint nuisance calibration together with the Rician NLL formulation make this a good contribution. My recommendation is contingent on the differentiability clarification, an explicit acknowledgment of the restricted-compartment capacity effect in the camera-ready, and at least some qualitative figures being included.
Review #3
- Please describe the contribution of the paper
The authors introduce PRISM, a method for fiber recovery in voxels of diffusion MRI data. It performs a global, rather than voxel-wise, fitting of the multi-compartment forward model. PRISM also estimates nuisance parameters (due to spatial bias field, signal intensity variation due to gradient direction, noise standard deviation). By being fully differentiable, the microstructure and nuisance parameters are estimated in one end-to-end optimisation.
The forward model includes up to K white matter fiber compartments, and includes soft model selection for the their number. This is done by incorporating a direction repulsion penalty term as well as a sparsity penalty term on minor fiber fractions.
Authors validate their method using simulated, digital phantom and in-vivo data. In simulated data, PRISM outperforms the baselines in terms of fiber detection and best-match angular error. In phantom data, they improve connectivity correlation.
- 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 novelty of this paper mainly resides in the optimisation of the nuisance calibration parameters simultaneously to the microstructure parameters, correcting for nuisances that may survive preprocessing.
The forward model includes many compartments (CSF, gray matter, up to K white matter fiber compartments and a restricted compartment), which enables to capture many different microstructural environments.
Experiments to evaluate the benefits of the calibration module are performed, showing it reduces angular error in the presence of intensity nuisance factors, but also that it does not affect microstructure parameter estimation in unperturbed data.
- 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.
PRISM performs global rather than voxel-wise optimisation, but its performance is compared solely against voxel-wise methods. A comparison with other global optimisation algorithms should be made. (e.g. Daducci A, Dal Palù A, Lemkaddem A, Thiran JP. COMMIT: Convex optimization modeling for microstructure informed tractography. IEEE Trans Med Imaging. 2015 Jan;34(1):246-57.doi: 10.1109/TMI.2014.2352414.)
Furthermore, spatial information via regularisation has previously been employed, but no comparisons have been made. (e.g. A. Ramirez-Manzanares, M. Rivera, B. C. Vemuri, P. Carney and T. Mareci, “Diffusion Basis Functions Decomposition for Estimating White Matter Intravoxel Fiber Geometry,” in IEEE Transactions on Medical Imaging, vol. 26, no. 8, pp. 1091-1102, Aug. 2007, doi: 10.1109/TMI.2007.900461.)
Qualitative results are not presented, impeding visual verification of the reported improvements to fiber recovery.
No comparisons to other methods are made for the real data experiment.
- 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
There seems to be a typo in the first term of the sum in equation 9.
- 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?
The experiments performed in this paper do not compare performance against other global optimisation methods, or methods that also employ spatial information. Furthermore, qualitative results are not present.
- Reviewer confidence
Not confident (1)
- [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 concerns and have stated that qualitative figures will be included.
Author Feedback
We thank all reviewers and area chair for their constructive feedback.
1.Differentiability and positioning (R2, meta). The forward model (Eqs. 2-5) is a closed-form analytic multi-compartment expression; “differentiable” refers to gradient propagation via PyTorch autograd, not a custom solver, distinguishing PRISM from FEM-based simulators. In contrast to voxelwise fitters (CSD, NODDI, SANDI), PRISM jointly estimates bias field, per-measurement affine correction and noise level with microstructure; optimises 3-D slabs with neighbour-aware regularisation; and supports Rician NLL with self-supervised sigma.
2.Direction-continuity penalty (R1, meta). The summed term is K-permutation-invariant and avoids explicit cross-voxel peak matching: for each fiber k we sample its own fraction along its own direction and penalise along-fiber changes, L_dc = (1/K) sum_k E_x f_k(x) (|f_k(x) - f_k(x + e d_k(x))| + |f_k(x) - f_k(x - e d_k(x))|), with e=0.5 voxel and trilinear interpolation. The f_k weighting localises the penalty to voxels carrying that fiber; +/- handles antipodal symmetry. This formula will replace the vague description. Thank you for pointing this out.
3.Restricted-compartment (R2, meta). R2 is right that part of the DiSCo1 gain reflects model capacity. Sec. 4 ablation reports PRISM without the restricted compartment at r=0.840 (full 0.934), below MSMT-CSD (0.920). PRISM’s contribution is enabling this compartment with nuisance calibration in one optimisation; we will temper Sec. 4/Discussion.
4.Fixed diffusivities (R1). Kernel diffusivities are fixed by design, following NODDI/SANDI/MSMT-CSD practice: joint estimation of diffusivities, fractions and directions is ill-conditioned at clinical SNR/shell counts (Jelescu 2016). PRISM still learns per voxel S_0, all K+3 fractions, K fiber directions and f_intra (Sec. 2). The synthetic benchmark probes this: PRISM uses D_perp=0.4 (Eq. 5) while baselines receive oracle/ground-truth-derived response functions (Sec. 3); despite this mismatch PRISM achieves 1.9x lower angular error at SNR=30 (Table 1). We will acknowledge in the revised version that a larger mismatch can bias peaks.
5.Missing comparisons (R1, R3, meta). (a) Many learning-based FOD methods (Karimi & Warfield 2024) need labelled training/protocol matching; PRISM is training-free. (b) COMMIT (Daducci 2015) weights streamlines from a precomputed tractogram, complementary to local FOD/fixel recovery. (c) DBF (Ramirez-Manzanares 2007) is relevant prior work on intra-voxel basis-function FOD; differs in scope from PRISM’s multi-compartment setting. (d) NODDI/SANDI are voxelwise analytic fitters, not K-crossing FOD baselines. Positioning will be revised.
6.Regularisation sensitivity (R1, R2, meta). Because signals are b0-normalised, all tissue weights are fixed when enabled (lambda_sp = lambda_rep = 0.01, lambda_sparse = 0.02; orphan/order = 0.01, dc = 0.005; Sec. 2), with no per-dataset tuning. Per Sec. 3, synthetic uses repulsion only, DiSCo1 adds topology, HCP the full suite; weights do not change. Sec. 4 ablation isolates each block: spatial 0.7 pp, topology 1.9 pp, restricted 9.4 pp. Calibration weights and the 8^3 bias-field grid were also fixed; the grid limits bias to low spatial frequencies. A systematic sweep is future work.
7.Qualitative figures and HCP scope (all reviewers, R3). HCP lacks ground-truth fiber orientations, so it is not claimed as a ground-truth fiber benchmark; head-to-head comparisons use synthetic and DiSCo1 data because they provide ground truth. If accepted, the camera-ready will add fiber-peak overlays from existing fits, within the additional page.
8.Eq. 9 typo (R3). The -log y term was dropped in transcription; it is a data-only constant independent of all optimised parameters, so gradients, fits and reported numbers are unaffected. We will correct it.
We commit to releasing source code upon acceptance. Thank you again for these thoughtful suggestions.
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.
All three reviewers recognize the clear motivation, the novel joint nuisance-calibration framework, and the promising quantitative results of the proposed PRISM method, yet persistent critical concerns place the paper near the acceptance threshold. The authors must prioritize addressing these core points in their rebuttal. 1) Clarify the concrete formulation of spatial regularizers (especially direction continuity penalty) and provide sensitivity analysis plus practical tuning guidance for regularization weights; 2) Reframe the claimed “differentiable analysis-by-synthesis” contribution to avoid overstatement and clarify technical distinctions from prior arts; 3) Extend experimental comparisons with learning-based FOD estimation methods, global optimization baselines, and classic analytical models like NODDI; 4) Justify fixed diffusivity settings, explain unbalanced multi-compartment designs against baseline methods, and specify the implementation of differentiable optimization.
- 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.
After careful review of the initial assessments, the author’s rebuttal, and post-rebuttal responses, I recommend rejecting this paper for several reasons. The primary concern is the lack of sensitivity analysis for the numerous regularization parameters, raising doubts about the method’s applicability to new datasets with varying characteristics. The authors’ claim that weights are fixed globally is problematic, as it does not provide guidance for practitioners. Additionally, the paper fails to compare PRISM with relevant state-of-the-art learning-based FOD estimation techniques, which are vital for contextualizing its performance. The absence of qualitative results also hinders the assessment of the recovered fiber orientations. Although reviewers expressed willingness to accept based on proposed changes, acceptance should rely solely on the current manuscript, which lacks essential components for a complete scientific contribution. The reviewer, confident in their assessment, maintained a Reject recommendation, highlighting substantive concerns about hyperparameter sensitivity and incomplete evaluations. The framing of “differentiable analysis-by-synthesis” remains overstated; the method is more of a gradient-based fitting approach similar to existing models like NODDI and SANDI. While the joint nuisance calibration is a valuable contribution, it does not represent a significant paradigm shift in the field.
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.
The reviewers are generally positive after rebuttal, with two recommending acceptance and one maintaining a borderline rejection. The remaining concerns mainly relate to clearer reporting of differentiability, hyperparameter sensitivity, qualitative examples, and acknowledging the restricted-compartment capacity effect. Since these issues appear addressable in the camera-ready rather than fatal to the contribution, I recommend Accept.
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.
Even though reviewer 1 is not wrong in suggesting that this paper would benefit from a larger revision, assuming that all the changes that have been promised in the rebuttal will be made to the final version, and code will be released, the paper is clearly among the stronger ones in my rebuttal batch already. Some of the suggested additional experiments might not fit within the MICCAI page limit anyway, and might thus be left for future work.
