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

Abstract

Deep unrolling networks have established a strong foundation for accelerated MRI reconstruction, yet they allocate uniform network capacity across all spatial regions, disregarding the inherently heterogeneous nature of reconstruction difficulty. In practice, this heterogeneity is governed by two distinct sources: physics-induced factors such as spatially varying aliasing and encoding instability determined by the acquisition operator, and anatomy-driven factors arising from structural complexity. We propose Physics-informed Mixture-of-Experts (PiMoE), a framework that replaces the monolithic learned prior with a set of specialized experts whose routing is guided by acquisition physics. Specifically, we derive three physical descriptors, aliasing susceptibility, phase alignment, and magnitude imbalance, directly from the undersampling mask and coil sensitivities without requiring ground-truth images. These descriptors are unified into a difficulty score that adaptively modulates the coupling between physics-based and feature-based routing. As local ill-posedness intensifies, routing authority shifts toward the physics-informed pathway, while feature-based routing dominates in well conditioned regions. Experiments on the fastMRI brain and CMRxRecon cardiac datasets demonstrate that PiMoE achieves consistent improvements over state-of-the-art methods across $4{\times}$ and $8{\times}$ acceleration, with zero-shot generalization to unseen anatomical domains.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4108_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KimKyu_PiMoE_MICCAI2026,
        author = { Kim, Kyuri AND Na, Yoonho AND Ye, Sung-Joon},
        title = { { PiMoE: Physics-Informed Mixture-of-Experts for Accelerated MRI Reconstruction } },
        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

    This manuscript introduces PiMoE (Physics-informed Mixture-of-Experts) for accelerating MRI reconstruction. The core problem addressed in this work is that MRI reconstruction difficulty is spatially uneven, while traditional unrolling networks use a uniform learned prior. This work derives three categories of physical descriptors—AS, PA, and MI—from the sampling mask and coil sensitivities to characterize local ill-posedness. By adaptively assigning weights to the feature router and physics router based on a “difficulty score,” it drives expert specialization. Experiments were conducted on fastMRI brain and CMRxRecon cardiac datasets. 4x and 8x acceleration factors were utilized, and results regarding zero-shot generalization to cross-domain data were included. Nevertheless, the current presentation of the paper exhibits several shortcomings that must be addressed before its claims can be considered fully convincing. Particularly in the following aspects: (i) Insufficient innovation intensity and physical soundness; (ii) Incomplete experimental evidence and ablation studies; (iii) Baseline fairness and efficiency analysis; (iv) Statistical rigor. In summary, the research questions and methodological design of this paper are adequate, and the experiments possess a certain degree of persuasiveness.

  • 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.

    Major strengths: 1.The three physical descriptors (AS, PA, MI) are derived directly from acquisition parameters without requiring ground-truth images. Decomposing coil conditioning into PA and MI rather than a single g-factor preserves discriminative structure for routing. 2.The proposed adaptive fusion between physics-based and feature-based routing is conceptually clean, and the monotonicity constraint enables an emergent two-phase restoration process (physics-driven first, then anatomy-driven). 3.Experts spontaneously specialize without explicit supervision, which is rare in MRI reconstruction literature and offers a degree of interpretability.

  • 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.

    Major concerns：

    1. The primary innovation of this paper lies in replacing specific layers of the monolithic prior with physics-informed routing, rather than proposing a fundamentally new paradigm for the MRI reconstruction prior itself. It is necessary to clarify what unique advantages this approach offers compared to general MoE methods for image restoration.
    2. What exactly are the advantages of the AS, PA, and MI metrics proposed in this paper, compared to more classic measures such as the g-factor or the encoding condition number? Are they more interpretable, more differentiable, or better suited for token-wise routing—or are they simply more effective in terms of empirical results?
    3. The main experiments cover only retrospective random Cartesian undersampling, and the zero-shot generalization evaluation is limited to just two sub-domains, which is insufficient to demonstrate generalizability. Further validation is required across different devices, centers, and noise levels, as well as with more complex sampling patterns.
    4. The ablation studies are incomplete. Table 2 is helpful; however, for fastMRI, results are provided only for cases (a), (c), and (e), making it difficult to determine which component contributes most significantly to performance on this dataset. Furthermore, due to the absence of ablation experiments for individual descriptors, it is currently impossible to rigorously answer the following questions: Are all three categories of descriptors necessary? Are they mutually complementary?
    5. The paper highlights that sparse specialization yields structural efficiency; however, while it reports that the number of active parameters is comparable to that of Dense-B, it omits key metrics such as inference time, memory footprint, and training costs. I am concerned about whether this approach holds practical utility within the context of medical imaging.
    6. The statistical rigor requires supplementation. The paper’s results primarily present mean-based metrics, and the performance improvement over HUMUSNet on the fastMRI dataset is not substantial. It is necessary to provide additional metrics—such as significance tests or variance measures—to further validate the robustness of the proposed method.
    7. Does the method depend on the quality of coil sensitivity estimation? If the sensitivity map estimation is biased, how robust are the three classes of physical descriptors?
    8. Expert specialization is currently primarily supported by visualized results. Is it possible to provide additional quantitative metrics—such as routing entropy, expert utilization balance, the correlation between specific experts and different descriptor regions, or utilization patterns under varying coil counts?
  • 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?

    PiMoE provides an interpretable integration of acquisition physics with MoE routing, achieving appealing emergent expert specialization and zero-shot generalization. However, the innovation is more of an engineering-level improvement; the advantage of AS/PA/MI over the classical g-factor is not rigorously justified; and the experimental validation is limited in scope with insufficient ablation analysis. The method also lacks practical deployment metrics.

  • 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 #2

  • Please describe the contribution of the paper

    PiMoE proposes a spatially adaptive regularizer for unrolled multi-coil MRI reconstruction. The monolithic prior in each cascade is replaced by a sparse MoE module (E=8, top-2) whose router takes both image features and physics-derived descriptors of local ill-posedness. Three descriptors are computed directly from M and {S_c} without ground truth: Aliasing Susceptibility (AS), built by convolving PSF sidelobes with the RSS of sensitivities; Phase Alignment (PA), the magnitude of the mean unit phasor across coils; and Magnitude Imbalance (MI), a participation-ratio scalar over coil power. These are aggregated into a per-token difficulty score d ∈ [0,1] via a monotonicity-constrained linear projection, which interpolates between a feature router π_f and a physics router π_p. Built on a HUMUSNet backbone, evaluated on fastMRI brain and CMRxRecon at 4× and 8× plus a zero-shot test on unseen contrasts/views.

  • 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 core framing is genuinely novel and well motivated. To my knowledge this is the first paper to derive MoE routing signals directly from the MR forward operator rather than from latent features alone. The observation that physics-induced difficulty is deterministically known at acquisition time, while existing adaptive methods only look at latent features, is correct and underexploited. The three descriptors are physically sensible and ground-truth-free. AS is a clean PSF-sidelobe energy map, PA captures the phase degeneracy that drives g-factor amplification, and MI is a normalized participation ratio. The authors correctly note that a single collapsed g-factor would lose the structural information the router needs, so they keep the three disentangled for routing while collapsing them only for the scalar authority d. This is a thoughtful design. The softplus on w^(t) enforces monotonicity of d in each descriptor, a small but meaningful inductive bias that keeps the authority shift interpretable. Per-cascade learnable parameters then allow different cascades to learn different difficulty thresholds. The expert-specialization analysis in Fig. 3 is the most compelling part. d starts high on fastMRI (0.77, variable coil count) and low on CMRxRecon (0.18, standardized 10 virtual coils), and both converge toward small values in late cascades. This emergent two-phase behavior (early physics-driven artifact suppression, late anatomy-driven refinement) is exactly what the framework predicts and was not hard-coded. The dataset-dependent ablation pattern is internally consistent with the hypothesis. Feature-only MoE helps on CMRxRecon (standardized coils, anatomy dominates) but actually hurts on fastMRI (variable 2 to 24 coils, physics dominates), and the adaptive d rescues fastMRI. This is more informative than a bare PSNR table. Active parameter count (115M) is comparable to the dense baseline (109M), and the parameter-matched Dense-L (config b) is the right control. Zero-shot results (+1.44 dB on fastMRI T1-Post, +1.20 dB on CMRxRecon Aorta-Tra over HUMUSNet) are encouraging.

  • 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.

    fastMRI gains are small and not significance-tested. At 8×, PSNR goes 36.87 to 37.07 (+0.20 dB), SSIM 0.9172 to 0.9228, while typical per-slice std on fastMRI brain is well above 1 dB. No std or paired test anywhere in Table 1 or 2.The CMRxRecon 8× gain (+1.44 dB) is more convincing, but the “consistent improvements” framing oversells the fastMRI side. Baseline set is thin. U-Net, E2E-VarNet (2020), and HUMUSNet (2022) are reasonable but the paper omits more recent unrolled / attention-based methods. PromptMR+ [20] is the most glaring omission since it is cited and also targets adaptive capacity allocation (via prompts rather than experts), making it the most direct competitor. Without this comparison the MoE-specific contribution is hard to isolate. No ablation of the three descriptors. This is the single most important missing experiment. The paper builds substantial physical justification for AS, PA, and MI but never shows whether all three are needed. AS-only, PA-only, MI-only, and pairwise combinations on CMRxRecon 8× would directly answer whether MI adds anything over PA (both are coil-geometry quantities and could be redundant), or whether AS alone is doing most of the work. Easy to run before rebuttal. Mainlobe-nulling threshold τ in Eq. 3 is undefined. AS is sensitive to it: too small and the mainlobe contaminates AS, too large and useful sidelobes are removed. The value of τ and a short sensitivity check should be in the paper. MI definition has a scaling issue. Eq. 5 gives MI = 0 for balanced coils (correct) but MI → 1 − 1/C, not 1, in the single-coil-dominance limit. This contradicts the text. More importantly, MI’s effective scale depends on C, which is exactly the regime fastMRI emphasizes (variable 2 to 24 coils). Either renormalize (e.g., (C − 1/Σp²)/(C − 1)) or clarify. Spatial vs cascade-wise routing claim is conflated. The intro promises spatial adaptation (harder regions get more physics routing), but Fig. 3(a) only shows cascade-wise evolution of mean d. A per-token check (e.g., scatter of d vs expert entropy within a single cascade, or KL between π_p and π_f as a function of d) would directly support the spatial claim. The current evidence supports the temporal story but not the spatial one. AS extension to non-Cartesian is hand-waved. AS is computed along the phase-encoding axis, which is well-defined for 2D random Cartesian. For radial or spiral the PSF sidelobes are 2D and anisotropic, so the extension promised in the conclusion is non-trivial. A short sketch would help. “Zero-shot generalization” is weakly defined. The held-out sets are still the same datasets and vendors, just unseen contrasts/views. This is within-dataset distribution shift, not true cross-dataset/cross-vendor generalization. Either soften the wording or add a genuine OOD test. Compute overhead of descriptor computation is not reported. AS needs an FFT of the mask plus a convolution with RSS per slice. For 3D or real-time settings this could matter. A timing comparison vs one cascade forward pass would help. Fig. 3(b) is too dense to read. 8 cascades × 3 routing modes × per-expert maps is too much for the printed size.

  • 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.

    (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 contribution is genuinely novel: MoE routing for MRI reconstruction grounded in the forward operator rather than in image features, instantiated through three principled, ground-truth-free descriptors. The monotonic difficulty scorer and dual-routing interpolation are clean design choices. The emergent cascade-wise transition from physics-dominated to anatomy-dominated routing in Fig. 3(a), and the dataset-dependent ablation pattern (feature-only MoE helping CMRxRecon but hurting fastMRI, with adaptive d rescuing fastMRI), together provide more convincing mechanism validation than a typical PSNR table. What holds me back from a stronger score: fastMRI gains are small (+0.20 dB at 8×) with no significance testing, and the paper oversells “consistent improvements”; there is no ablation isolating which of AS, PA, MI is doing the work, which is critical given that descriptor design is the central claim; the MI definition has a scaling issue that bites hardest on the variable-coil fastMRI regime the paper emphasizes; the baseline set is missing recent prompt/attention-based adaptive methods, particularly PromptMR+, which is cited but not benchmarked; and the spatial-routing claim from the intro is not directly measured. All of these are addressable in rebuttal without new training: a descriptor ablation on CMRxRecon 8×, a clean MI definition, a τ value with sensitivity note, paired significance tests on existing runs, and at least an architectural-delta discussion versus PromptMR+. If the descriptor ablation shows all three contribute and MI is fixed cleanly, I would move to 5.If AS alone explains most of the gain, the framing would need to change but the core contribution would remain interesting. On balance, the methodological idea is novel, the mechanism validation is convincing, and the weaknesses are mostly clarification-level.

  • 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 introduces PiMoE, a physics-informed mixture-of-experts framework for accelerated multi-coil MRI reconstruction. Instead of using a single shared prior across all spatial regions, the method routes tokens to specialized experts according to local reconstruction difficulty. The key idea is to derive three physics-based descriptors directly from the acquisition setting, aliasing susceptibility, phase alignment, and magnitude imbalance, and use them to modulate the balance between a feature-based router and a physics-based router. The paper applies this design within an unrolled reconstruction backbone and evaluates it on both brain and cardiac MRI datasets, with an additional zero-shot test on unseen anatomical domains.

  • 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 targets a real limitation of current unrolled MRI reconstruction models. Many existing methods use a single monolithic learned prior and assign the same model capacity to all spatial regions. This paper starts from a sensible observation: reconstruction difficulty varies across space because sampling artifacts and encoding stability vary across space. That problem setup is well motivated for MRI reconstruction. 2.The physics-guided routing idea is interesting and technically well aligned with the task. The method does not rely on latent features alone for adaptation. It derives routing cues from the acquisition process itself, using descriptors computed from the undersampling mask and coil sensitivities. That design gives the routing mechanism a clearer physical basis than a purely feature-driven gating scheme. The distinction between global interference and local conditioning also makes sense from an MRI perspective. 3.The dual-router design is more thoughtful than a standard MoE plug-in. The paper does not just add experts to the network. It combines a feature router and a physics router, then uses a learned difficulty score to shift authority between them. That gives the method a concrete mechanism for deciding when physics should dominate and when image features should dominate. This part is more interesting than a generic sparse expert formulation. 4.The ablation table supports the value of the full routing design. Table 2 shows that feature-only routing helps on CMRxRecon but does not consistently help on fastMRI, while the full adaptive dual-routing model improves both datasets. That pattern supports the authors’ argument that explicit physics guidance becomes more important when acquisition conditions vary more strongly.

  • 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.

    Please list the major weaknesses of the paper. 1.The empirical gains over the strongest baseline are modest in some settings. On fastMRI, the improvement over HUMUSNet is quite small at 4× and 8×. For example, the PSNR gain is only 0.13 dB at 4× and 0.20 dB at 8×. The gains on CMRxRecon are stronger, but the paper should discuss more clearly when the added routing complexity is worth it. Right now the practical impact looks uneven across settings. 2.The comparisons do not isolate whether the gain comes from MoE capacity or from the proposed physics-informed routing. Table 2 helps, but it still leaves an important ambiguity. The paper compares against a dense baseline, a parameter-matched dense model, feature-only routing, fixed dual routing, and the final model. That progression is useful, but it does not fully isolate the value of each proposed ingredient. In particular, I would like to see a stronger comparison against a more competitive non-physics adaptive routing baseline within the same backbone, beyond the feature-only case. 3.The zero-shot generalization section is too brief. The paper claims robust zero-shot generalization to unseen anatomical domains, which is an important claim, but the evidence is limited to a short paragraph with two held-out subsets and no table of full metrics, variance, or qualitative comparisons. That section needs more detail to support the strength of the claim. 4.The paper reports no statistical testing or variance across runs. The reported improvements are often small, especially on fastMRI. Without repeated runs, confidence intervals, or statistical testing, it is hard to judge how stable those gains are. This matters because the method adds substantial architectural complexity.

  • 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.

  • 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?

    I based my score on the balance between a clear and interesting idea, solid but not decisive empirical gains, and a few missing pieces in the evidence chain. The strongest part of the paper is the core idea. The authors identify a real limitation in standard unrolled MRI reconstruction, namely that local reconstruction difficulty is spatially heterogeneous and depends on both acquisition physics and anatomical structure. Building a routing mechanism that uses descriptors derived from the mask and coil sensitivities is a meaningful step beyond purely feature-driven adaptation. The dual-router design and the difficulty-aware fusion mechanism are also well matched to the problem. The experiments show that the method works and improves over strong baselines on both fastMRI and CMRxRecon, especially at higher acceleration and on the cardiac dataset. The ablation study also supports the claim that the full adaptive routing design matters more than a plain MoE plug-in. The routing visualizations further help the paper because they make the expert behavior easier to interpret.

  • 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 appreciate all valuable comments from reviewers and meta-reviewers. We address the main concerns and provide additional analyses below.

Per-descriptor ablation (R1, R2). We agree that ablating AS, PA, and MI individually would further strengthen the analysis, and will include such per-descriptor ablations with statistical validation in future work.

Advantage over g-factor (R1). The g-factor compresses the per-pixel noise amplification of inverting the encoding matrix into a single scalar. However, ill-posedness from phase redundancy and from coil-power dominance both collapse into similarly large values, so the router only receives the information that a pixel is “difficult” without being able to identify what kind of difficulty is present. This directly conflicts with the expert-specialization benefit that MoE is intended to exploit. The g-factor also reflects only the coil geometry, missing the global interference induced by the undersampling mask, which our AS descriptor is designed to complement. Disentangling these three axes provides the router with cause-specific input.

Scaling of MI (R2). We thank R2 for the careful observation. R2 is correct that MI in Eq. (5) saturates at (C − 1)/C rather than at 1, with 1 reached only asymptotically as C → ∞, and we will revise the surrounding text to make this explicit. In practice, however, MI is not used as a direct router input. The difficulty score (Eq. 6) and the physics router (Eq. 7) apply learnable cascade-wise parameters w(t), b(t), and W_g, into which the coil-count-dependent upper bound is naturally absorbed during training. To verify this empirically, we partitioned the fastMRI test set by coil count (C ≤ 8, 9 ≤ C ≤ 16, 17 ≤ C) and observed PSNR gains over the baseline of 0.24, 0.20, and 0.16 dB, respectively. If the C-dependent scale of MI were distorting routing, the gain should shrink in the low-C bin where MI’s range is most compressed, yet the largest gain is observed precisely there. We agree the normalization proposed by R2 is principled and note it as a future-work direction, while retaining Eq. (5) in this revision for consistency with the trained model.

Scope of zero-shot claim (R1, R2, R3). We thank the reviewers for noting that, given the importance of the zero-shot generalization claim, our current presentation is too brief to fully substantiate it. Our experiments simulated domain-shift by holding out modalities/views from training (FLAIR/T1/T1Pre/T2 → T1-Post on fastMRI; SAX/LAX/LVOT views → Aorta-Tra on CMRxRecon), but these remain within the same datasets and do not constitute cross-dataset or cross-vendor generalization. Following R2’s suggestion, we will reposition these experiments as a within-dataset domain-shift evaluation, temper the related terminology, and more clearly delineate the scope of the claim.

Non-physics adaptive routing baseline (R3). Configuration (c) in Table 2 is designed precisely for this comparison, under the same HUMUSNet backbone, the same number of experts, and the same top-k setting, the router in (c) is reduced to a feature-based softmax gating, identical in form to the canonical vanilla MoE baseline established in the MoE literature, and thus serves as a standard non-physics adaptive routing baseline. Thus the gap (c) → (e) isolates the effect of introducing physics-informed routing, holding MoE capacity and adaptive routing itself constant, while the gap (d) → (e) further isolates the effect of the learnable d(t). We will clarify this point in the description of Table 2.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 reviewers are overall positive, with scores of 4, 4, and 5, and I agree with the provisional accept recommendation. The paper presents a well-motivated physics-informed MoE framework for accelerated multi-coil MRI reconstruction. Its main strength is the use of acquisition-derived descriptors to guide expert routing, rather than relying only on latent image features. The dual-router design and difficulty-aware fusion are technically sensible and provide an interpretable mechanism for spatially adaptive reconstruction.

    The main limitations are that some fastMRI gains are modest and not supported by statistical testing, the individual contributions of the AS/PA/MI descriptors are not fully ablated, and the comparison to recent adaptive reconstruction methods could be stronger. The authors should also clarify descriptor definitions, computational overhead, and the scope of the zero-shot generalization claim. Overall, the idea is novel and relevant, and the evidence is sufficient for provisional acceptance.



back to top