Abstract

In this paper, we propose a Memory-Guided Random-Direction Feature Disentanglement (MG-RDFD) framework for multi-phase CE-MRI translation. Our method learns an interpretable and parameter-efficient encoder-decoder model without resorting to multiple domain-specific encoders or additional content discriminators. Given posteriors from paired inputs encoded by a shared VAE encoder, MG-RDFD selects a shared content anchor and phase-preferred style anchors along a random 1D latent direction using density-based criteria, which explicitly couples content and style within a unified latent geometry. We further introduce a segmentation-guided style memory bank to retrieve region-wise style prototypes and enhance organ/lesion fidelity. Experiments on multi-phase CE-MRI demonstrate competitive translation quality with favorable inference parameter efficiency, and the memory bank reduces regional texture confusion. The implementation is available at \url{https://github.com/QianMuXiao/MG-RDFD}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0347_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/QianMuXiao/MG-RDFD

Link to the Dataset(s)

https://github.com/LMMMEng/LLD-MMRI-Dataset

BibTex

@InProceedings{XiaQia_MemoryGuided_MICCAI2026,
        author = { Xiao, Qianmu AND Kui, Xiaoyan AND Ji, Zexin AND Gao, Yang AND Pan, Wengang AND Lu, Wenqi AND Duan, Jinming AND Zou, Beiji},
        title = { { Memory-Guided Random-Direction Feature Disentanglement Model for Multi-phase MRI Translation } },
        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 paper proposes a content-style disentanglement framework for multi-phase CE-MRI translation, featuring a shared VAE-based latent space, a random-direction anchor selection strategy for separating shared content and phase-specific style, and a segmentation-guided style memory bank to improve region-level fidelity.

  • 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 presents an interesting and technically coherent framework, with a reasonably novel random-direction disentanglement design and a segmentation-guided memory mechanism that is well motivated for improving region-level fidelity in medical image translation.

  • 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 model does not explicitly incorporate the physical properties of multi-phase imaging, such as temporal dynamics across phases. Since phase progression in contrast-enhanced MRI follows an underlying physiological and temporal pattern, ignoring this characteristic may limit the model’s ability to capture more realistic cross-phase relationships.
    2. It is unclear whether segmentation is required at test time. If so, the method has a clear dependence on segmentation priors and a pre-trained segmentation model, which reduces its general applicability and weakens the claim of being a lightweight translation framework.
    3. The method is designed in a 2D manner, whereas medical imaging data are inherently 3D. This may limit the model’s ability to preserve volumetric consistency and exploit inter-slice contextual information, which are often important in medical image translation.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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. The visualization in Fig. 3 appears to be based on simulated data. While this is helpful for illustrating the intuition of the method, a similar analysis or visualization on real data would make the evidence much more convincing.
    2. The proposed disentanglement strategy appears to rely on the Gaussian posterior statistics produced by the VAE encoder, such as the latent mean and variance. While this is an interesting use of the probabilistic latent space, these quantities may not directly correspond to disentangled semantic factors, such as shared content and phase-specific style. The manuscript would be stronger if it could better justify why posterior density overlap and discrepancy are reliable indicators for content and style separation.
  • 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 appreciate the interesting idea, the well-organized framework, and the promising results, but my overall score reflects some remaining concerns regarding the discussion and clarity of the methodological novelty relative to prior work, such as HiFi-Syn; I would suggest briefly clarifying this point.

  • 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 thank the AC for the comments and the authors for their response. The authors have addressed my main concerns, and I therefore maintain my original assessment.



Review #2

  • Please describe the contribution of the paper

    This paper proposes MG-RDFD, a memory-guided random-direction feature disentanglement framework for multi-phase contrast-enhanced MRI translation. The main contributions are twofold:

    First, the authors introduce Random Direction Feature Disentanglement (RDFD), which selects content and style anchors along a random 1D line connecting two samples drawn from paired latent Gaussian posteriors encoded by a shared VAE. The content point is defined as the maximum joint density point, while style points are identified via density discrepancy extrema. This provides an explicit geometric coupling between content and style within a unified latent space, eliminating the need for separate content/style encoders or additional content discriminators.

    Second, a segmentation-guided style memory bank is introduced, which maintains region-level style prototypes (for 7 organs, 7 lesion types, abdomen, and border) indexed by content keys. During translation, the memory retrieves dense style maps per region to enhance local texture consistency.

    The authors validate their approach on three cross-phase translation tasks (Pre→CA, Pre→CV, CA→CV) using the LLD-MMRI 2023 dataset, reporting competitive full-slice PSNR/SSIM and improved ROI-level fidelity for liver and lesion regions compared to several end-to-end and disentanglement-based baselines, while maintaining a lightweight inference footprint (13.15M parameters).

  • 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.

    Elegant geometric formulation for disentanglement: The core idea of selecting content and style points along a random 1D line using density-based criteria is mathematically clean and intuitive. The content point as the joint density maximizer and style points as density discrepancy extrema provide a principled geometric interpretation of content-style separation. The closed-form solution for the content point (Eq. 2) and the differentiable soft-argmax for style points (Eq. 4) make the approach both analytically grounded and end-to-end trainable.

    Architectural simplicity and parameter efficiency: By using a single shared VAE encoder instead of multiple domain-specific encoders and content discriminators, the method achieves a compact inference footprint (13.15M parameters), approximately 8× smaller than TransUNet or ResViT while achieving competitive or superior performance. This is practically meaningful for medical imaging deployment scenarios.

    Well-motivated memory bank design with region-level evaluation: The segmentation-guided style memory bank addresses a real limitation of global style representations, as different anatomical regions respond differently to contrast agents. The ablation study supports its effectiveness with consistent improvement from RDFD-only to MG-RDFD. Furthermore, the authors go beyond standard full-slice metrics by reporting ROI-level PSNR and MAE for liver and lesion regions, which is clinically relevant.

  • 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.

    Lack of repeated experiments despite stochastic nature of the method: RDFD fundamentally relies on random sampling from the two posteriors to construct the 1D line along which content and style points are selected. Despite this inherent stochasticity, all results in Table 1 are reported as single values without standard deviations from multiple runs. Given that the reported improvements over baselines are often marginal (e.g., 0.2–0.5 dB in PSNR), it is difficult to assess whether these gains reflect genuine methodological advantages or simply favorable randomness in a single training run.

    Data split strategy is not specified: The paper states that approximately 32k paired slices are split into training/validation sets at a 4:1 ratio, but does not clarify whether this split is performed at the subject level or the slice level. If slices from the same patient appear in both sets, data leakage could inflate the reported metrics. This is a critical detail for assessing the validity of the experimental results.

    Insufficient ablation studies: The ablation only demonstrates the incremental effect of adding RDFD and the memory bank on top of FDA-VAE. However, the method introduces numerous critical hyperparameters — κ (search range), τ (soft-argmax temperature), G (grid resolution), the number of memory slots per region Mk, and momentum μ — none of which are analyzed for sensitivity. Similarly, the memory bank design choices (K=16 regions, the role of the global cross-entropy loss Lglob) and the loss weighting scheme in Eq. 10 (8 terms) are not ablated. Without such analysis, it is unclear how robust the method is to these design decisions.

    Inconsistent SSIM results are not adequately discussed: Adding RDFD to FDA-VAE causes a notable SSIM drop on the Pre→CA task (80.40 → 78.59), and even the full MG-RDFD model does not recover the baseline SSIM on this task (79.99 vs. 80.40). This suggests that the proposed disentanglement may introduce structural distortions in certain translation scenarios. The paper does not analyze or explain this trade-off, which undermines the claim of consistent improvement.

    Validity of the single-input inference assumption is not analyzed: At inference time, the method constructs a pseudo-target posterior by flipping the mean of the source posterior while keeping the variance unchanged. This is a strong assumption that relies on the symmetry encouraged by FDA during training. However, the paper provides no analysis of how well this pseudo posterior approximates the true target posterior, nor does it discuss under what conditions this assumption might break down.

    Evaluation on a single dataset: All experiments are conducted on the LLD-MMRI 2023 dataset only. While the limited availability of multi-phase paired CE-MRI data is acknowledged, evaluation on a single dataset makes it difficult to assess the generalizability of the proposed method to other anatomical regions, imaging protocols, or institutional settings.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    The paper presents an interesting and mathematically elegant idea — using density-based criteria along a random 1D line to separate content and style within a unified latent geometry. The architectural simplicity and parameter efficiency (13.15M vs. 100M+ for some baselines) are also appealing. However, the experimental validation does not sufficiently support the methodological contribution, which is the primary factor behind my score.

    Most critically, the method inherently depends on random sampling, yet all results are from a single run without standard deviations. The reported margins over baselines (often 0.2–0.5 dB) could fall within run-to-run variance, making it impossible to confirm the claimed improvements. Additionally, the data split strategy (subject-level vs. slice-level) is not specified, raising concerns about potential data leakage.

    The ablation study demonstrates that each module contributes, but does not analyze the sensitivity to the numerous hyperparameters introduced (κ, τ, G, slot count, loss weights). The inconsistent SSIM results — where RDFD-only and even MG-RDFD underperform the baseline on some tasks — are not discussed, leaving open questions about when and why the method may introduce structural distortions.

    In summary, the core idea has merit, but the current experimental evidence is insufficient to validate the claims with confidence. A well-addressed rebuttal clarifying the data split, explaining the SSIM trade-offs, and justifying key design choices could shift my assessment toward acceptance.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    Adequately addressed. The authors clarified that the train/validation split was performed at the patient level (4:1) with lesion-distribution stratification, eliminating my concern about data leakage. The justification for using LLD-MMRI as a single but lesion-rich dataset (498 patients, 7 lesion types, paired multi-phase acquisitions) is also reasonable. Partially addressed. On the lack of variance estimates, the authors claim “internal repeated trainings varied mainly at the second decimal place.” The geometric argument — that the content anchor as a joint density maximizer in the overlap region is more sampling-robust than raw VAE samples — has theoretical merit, but remains an informal claim. Since reported gains are 0.2–0.5 dB (first decimal place), statistical significance cannot be confirmed. The SSIM explanation — that residual Pre-CA misalignment under CV-space registration penalizes Pre→CA — is plausible but post-hoc, and does not explain why MG-RDFD specifically underperforms the FDA-VAE baseline (79.99 vs. 80.40), since baselines use the same data. The argument addresses absolute task difficulty, not relative degradation from the proposed components. Insufficiently addressed. Hyperparameter sensitivity (κ, τ, G, memory slots, 8 loss weights) was answered with descriptions of what each hyperparameter controls, not with robustness analysis. Most importantly, my concern about the inference-time pseudo-posterior — constructed by flipping the source posterior’s mean while keeping its variance — was not addressed at all, despite being a core deployment mechanism relying on a strong symmetry assumption. Overall. The random-direction density-based disentanglement formulation is mathematically elegant, and the parameter efficiency (13.15M, ~8× smaller than TransUNet/ResViT) is practically appealing. However, the statistical significance of marginal gains, the validity of the pseudo-posterior at inference, the SSIM trade-off on Pre→CA, and sensitivity to numerous introduced hyperparameters remain unjustified. For a primarily methodological contribution, the experimental evidence is insufficient.



Review #3

  • Please describe the contribution of the paper

    The paper introduces MG-RDFD, a memory-guided random-direction feature disentanglement framework for multi-phase contrast-enhanced MRI translation. A shared VAE encoder produces paired Gaussian posteriors for two phases; along a randomly sampled 1D latent line between posterior samples, the method selects a content anchor by maximizing joint density and domain-specific style anchors by extremizing density discrepancy. A segmentation-guided memory bank stores region-wise style prototypes and retrieves dense, organ/lesion-aware style maps for decoding. On the LLD-MMRI dataset across three phase-translation tasks, MG-RDFD achieves competitive full-slice PSNR/SSIM with improved liver/lesion ROI fidelity and favorable parameter efficiency relative to end-to-end and disentanglement baselines.

  • 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 random-direction latent geometry to jointly determine content and styles via density overlap/discrepancy along a sampled 1D subspace is an elegant and interpretable mechanism that explicitly couples shared and domain-specific factors, avoiding multiple encoders or explicit content discriminators. -A closed-form solution for the content anchor (via maximizing joint log-likelihood) and a bounded, differentiable soft-argmax/min search for style anchors are technically sound and differentiable. -The segmentation-guided memory bank with momentum-updated, region-wise content keys and domain-specific style values (plus regional and global CE regularizers) is a principled way to stabilize regional appearances and mitigate texture confusion. -Evaluations include three clinically relevant phase-translation directions, full-slice and ROI metrics (liver/lesions), and parameter counts for inference. -Comparisons span both high-capacity end-to-end translators and disentanglement models; ablations (FDA-VAE backbone vs RDFD-only vs memory-augmented MG-RDFD) isolate the contribution of the proposed components.

  • 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 clinical applicability and practical value of the work remain insufficiently demonstrated. Multi-phase liver MRI translation is clinically meaningful only if the synthesized phases preserve diagnostically relevant enhancement patterns and can support real downstream interpretation. The current evidence is still mainly based on PSNR, SSIM, and ROI-level MAE/PSNR, which measure pixel-level fidelity but do not establish whether the generated images are clinically reliable or useful. -The paper does not provide downstream task validation. Given that the dataset contains liver-lesion annotations, it would be important to evaluate whether synthesized phases improve or preserve performance in lesion segmentation, lesion-type classification, or other clinically relevant tasks. Without such evidence, it is difficult to determine whether the reported ROI improvements translate into meaningful diagnostic utility. A segmentation Dice evaluation, lesion-wise detection/recall, enhancement-pattern consistency analysis, or classification performance under missing-phase settings would substantially strengthen the paper. -The methodological novelty, while interesting, should be better positioned against recent generative modeling literature. The paper builds upon an FDA-VAE-style framework, and the gains over this strong backbone appear relatively modest. In the current medical image synthesis landscape, diffusion models and latent diffusion models have become strong baselines for structure-preserving image translation and conditional generation. The paper only briefly discusses such methods and does not provide a direct comparison or a convincing rationale for excluding them. This makes it difficult to assess whether the proposed latent-geometry and memory-bank design remains competitive against more recent diffusion-based approaches. -Several practical aspects are underexplored. The inference process may involve stochastic sampling, yet the paper does not clearly state whether deterministic latent means are used at test time or whether output variability is evaluated. The per-pixel grid search used in RDFD and the memory-retrieval mechanism may introduce non-trivial computational overhead, but runtime, latency, and GPU memory usage are not reported. The method also appears to rely on segmentation masks during inference, but the source of these masks and robustness to segmentation errors are not quantified. Since the experiments are conducted on a single dataset with 2D slice-based training, cross-site, cross-scanner, and 3D volume-level generalization remain open questions. Overall, the paper is technically promising and addresses a relevant missing-phase MRI synthesis problem, but its current form does not fully establish clinical utility or practical deployability. The contribution would be much stronger if the authors added downstream clinical task evaluations, clarified inference-time dependencies, reported computational costs, and compared against or more carefully justified the exclusion of recent diffusion/LDM-based MRI translation baselines.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    Overall, MG-RDFD addresses an important and clinically motivated multi-phase MRI translation problem, and the proposed latent-geometry coupling with a segmentation-guided memory bank is technically thoughtful, with encouraging ROI/lesion-level improvements and attractive parameter efficiency. However, my main concerns are that the paper has not yet convincingly demonstrated real clinical applicability: the evaluation remains largely based on pixel-level fidelity metrics, without downstream validation such as lesion segmentation, lesion classification, or missing-phase diagnostic performance. In addition, the methodological novelty and competitiveness are not fully established, since gains over the FDA-VAE backbone appear modest and recent diffusion/LDM-based image translation models are not directly compared or sufficiently discussed. Practical issues such as inference determinism, runtime overhead, dependence on segmentation masks at test time, robustness to segmentation errors, and generalization beyond a single 2D slice-based dataset also remain underexplored. Therefore, while the work is promising and technically interesting, stronger clinical-task evidence, clearer methodological positioning, and more complete practical validation are needed to substantiate its impact.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    not suitable for the conference paper



Author Feedback

We thank the reviewers for their recognition and suggestions. Our responses are as follows. Phase-time information and RDFD rationale (R1): Explicit temporal modeling could introduce stronger clinical priors. However, LLD-MMRI does not provide precise acquisition time steps, and we focus on a general content-style disentanglement mechanism rather than temporal modeling. For the RDFD rationale, the content anchor is not merely a visually overlapping point: density-overlap selects the point that best represents both phase posteriors, while density discrepancy selects phase-preferred style anchors within the same latent geometry. Mask dependence and patient-level data split (R1&R2): Our model needs organ and lesion labels from pretrained models during both training and inference. It’s a low-cost structural/clinical prior. Mask downsampling to feature-map resolution attenuates pixel-level errors. We split patient volumes 4:1 with lesion-distribution stratification before slicing, avoiding patient-level leakage. We will state this in the Experimental Setup. 3D training and computation (R1&R4): 3D training can improve volumetric consistency. Our code can be easily extended to 3D, but we used 2D training considering computational cost and a fair 2D-baseline comparison. RDFD has no learnable parameters and runs on low-resolution feature maps, so its overhead is much smaller than learnable modules. Since FDA-VAE reports a positive active-parameter/latency relation, active inference parameters are a practical efficiency proxy. Stochastic sampling and variability (R2&R4): RDFD is stochastic, but the content anchor is less variable than raw VAE samples because it maximizes joint density in the overlap of two close posteriors(Fig.3); sampling mainly defines the search direction. Internal repeated trainings/validations varied mainly at the second decimal place, close to/below Table 1 precision. We agree formal multi-seed std would be stronger and will clarify this limitation without overstating small gains. Ablation scope and hyperparameter choices (R2): Our ablation tests overall RDFD and memory-bank effectiveness. Preliminary grid search selected RDFD hyperparameters: G trades style search resolution for compute/memory; κ sets the style search range, too small misses extrema and too large dilutes grid precision; τ controls soft-argmax sharpness, too small is hard/unstable and too large over-smooths extrema. K=16 comes from automatically available regions: major organs, seven lesion types, abdomen, and border. Loss/memory settings follow FDA-VAE/MGUIT. SSIM discussion (R2): We will revise the analysis to clarify possible causes of the Pre->CA SSIM drop. All phases are registered to CV, and all predicted segmentation labels are also defined in the CV reference space; thus, Pre/CA->CV benefits from direct alignment, whereas residual Pre-CA misalignment can be penalized by SSIM and may affect RDFD/memory via structural anchors and labels. FDA gain and LDM rationale (R2&R4): Compared with FDA-VAE, RDFD provides statistical disentanglement and alleviates late-training over-overlap (Fig. 3(c)); 0.2-0.5 dB PSNR gains are meaningful in this domain. Diffusion/LDMs are strong but rely on iterative sampling with high computational costs and long inference time, while our goal is a lightweight, interpretable, single-pass disentanglement framework with low active parameters. Single-dataset and clinical tasks (R4): LLD-MMRI is a lesion-rich paired multi-phase MRI dataset with 498 unique patients covering 7 different liver lesion types and multiple MRI phases/sequences. We use three clinically relevant CE phases (Pre, CA, CV), providing paired lesion-level supervision. This is richer for our task than many two-modality translation datasets with only tens of subjects. The dataset lacks expert pixel-level gold labels, making downstream evaluation difficult; we will consider lesion classification/detection and enhancement-pattern consistency in future work.




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 reviewers pointed out several concerns. Key issues include the lack of incorporation of multi-phase imaging’s inherent temporal dynamics, the unclear dependency on segmentation at inference time, and the limitations of the 2D model design for inherently 3D medical data. Furthermore, ablation studies are insufficient, and critical details such as data split strategies and computational costs are underexplored. The reviewers also noted that the single-dataset evaluation limits the generalizability of the method, while clinical utility remains inadequately demonstrated.

  • 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 proposed method is technically thoughtful and shows parameter efficiency with promising results. The key concerns raised by reviewers focus on the clarity and novelty of the methodological contributions relative to prior work, the modest experimental gains without statistical significance, and insufficient practical validation for clinical applicability.



Meta-review #2

  • 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.

    While R1 suggests acceptance, praising the random-direction disentanglement design, the majority of the reviewers raised critical concerns regarding empirical rigor and clinical validation. R2 and R4 highlighted that the statistical significance of the marginal gains cannot be verified, as the authors provided no run-to-run variance estimates despite the framework’s inherently stochastic sampling design. Furthermore, the core inference-time deployment mechanism remains unvalidated, with no analysis supporting the strong symmetry assumption used to construct the pseudo-posterior. Finally, the clinical utility of the synthesized multi-phase MRIs remains unproven, as the evaluation relies entirely on pixel-level metrics rather than downstream clinical task validation (e.g., lesion segmentation or classification performance), while the model’s sensitivity to its numerous hyperparameters and test-time dependence on segmentation masks remain unquantified.



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.

    The paper presents an interesting idea of random-direction disentanglement. Although two reviewers gave rejection after the rebuttal, the main concern is about the lack of comprehensive validation and implementation details. So I think this is a broadline paper with an interesting methodology contribution. I am leaning toward accepting it.



back to top