List of Papers Browse by Subject Areas Author List
Abstract
Class-incremental learning (CIL) is essential for medical imaging systems that incorporate new diseases over time, as it avoids sacrificing performance on previously learned conditions. We first revisit a broad range of CIL methods on medical benchmarks and find, unexpectedly, that a previously proposed classifier-based analytic paradigm can outperform many widely used state-of-the-art alternatives. Motivated by this observation, we examine this paradigm more closely and observe that existing classifier-based methods fail to adequately separate features of similar examples and leave them too close in the representation space, resulting in insufficient margins and reducing separability as new classes arrive. To address these limitations, we specifically propose a training-free analytic margin-expansion framework that builds the classifier using an eigen-directed design, enlarging discriminative margins along principal data directions while preserving complementary detail. Across many commonly used medical benchmarks, our method consistently improves over a wide range of CIL baselines, and extensive ablations and theoretical analysis further support the proposed design. Code will be released at https://github.com/hmep313/EDRP_MICCAI.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3555_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hmep313/EDRP_MICCAI
Link to the Dataset(s)
Skin8 dataset: https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-ham10000
MedMNISTv2 dataset: https://medmnist.com
PBC dataset: https://www.kaggle.com/datasets/unclesamulus/blood-cells-image-dataset
ChestX dataset: https://www.kaggle.com/datasets/nih-chest-xrays/data
BibTex
@InProceedings{WuYon_EigenDirected_MICCAI2026,
author = { Wu, Yongyi AND Wu, Yichen AND Li, Quanzheng AND Ma, Jianhua AND Wang, Hong},
title = { { Eigen-Directed Random Projection for Medical Image Class Incremental Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes EDRP (Eigen-Directed Random Projection), a replay-free class-incremental learning method for medical image classification that replaces purely random analytic feature expansion with a partially data-aligned projection derived from the principal eigenspace of buffered features. The key idea is if medical classes are fine-grained and easily entangled, then standard random projections may not expand margins sufficiently, so the classifier should emphasize principal directions while retaining a random complement.
- 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 proposed method is conceptually clean and lightweight, it is essentially a structured modification of the analytic projection/classifier pipeline rather than a heavy continual-training framework. 2.The authors provide reasonable ablations and some analysis of the role of eigen-directed projection, including the effect of the number of principal components and a storage comparison. 3.The storage-efficiency claim is potentially useful, since the paper reports substantially lower storage than RanPAC while preserving the appeal of a training-free analytic classifier.
- 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 core change is replacing part of RanPAC’s random projection with a principal-direction-aligned projection, which is intuitive but not especially deep, and the theoretical margin argument is fairly basic. 2.Fig. 1 discusses different pretraining regimes (IN1K/BioMed/UniMed), but the main experiments are conducted with a single frozen ViT-B/16-IN1K backbone, limiting the strength of claims about medical foundation models more broadly. 3.The evaluation protocol raises questions about clinical realism, especially the constructed “MedMNIST-Sub” benchmark formed by shuffling classes across four datasets/modalities, which may not reflect how diseases are introduced incrementally in practice.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I would lean reject. The paper is technically sound. However, I do not find the contribution strong enough for acceptance because the core idea is a fairly modest extension of the RanPAC-style analytic classifier paradigm, the theoretical insight is limited, and the empirical gains over the strongest baselines are generally small. More importantly, the evaluation does not fully establish that the method is a meaningful advance for medical continual learning specifically, the benchmarks are somewhat synthetic, the backbone setting is narrow, and the paper does not convincingly demonstrate superiority over the strongest emerging medical CIL directions beyond a selected set of baselines. Overall, this feels more like a careful incremental improvement than a sufficiently compelling new step for MICCAI.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal does clarify most of my concerns, so I’m happy to change my rating to a more positive one.
Review #2
- Please describe the contribution of the paper
The paper addresses class-incremental learning (CIL) for medical imaging. The authors first conduct a systematic benchmark of existing CIL methods on medical datasets and find that classifier-based analytic methods (e.g., RanPAC) unexpectedly outperform more recent CLIP-based and prompt-based alternatives. They then identify a key limitation of RanPAC’s random projection: it is agnostic to the anisotropic structure of medical feature embeddings, leaving similar classes poorly separated. To address this, they propose EDRP (Eigen-Directed Random Projection), a training-free method that constructs a data-aware projection by aligning with the principal eigenvectors of the feature covariance. Experiments on four medical benchmarks (Skin8, MedMNIST-Sub, PBC, ChestX) show consistent improvements over baselines with reduced memory overhead.
- 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 initial finding that simpler classifier-based analytic methods outperform more sophisticated CLIP-based and prompt-tuning CIL methods on medical benchmarks is a useful contribution as it challenges the assumption that newer, more complex methods are always better in domain-specific settings.l.
- The motivation behind the propsoed approach (separate in a better way the features) makes sense and motivates EDRP
- The method is training-free and buffer free (even if there is a buffer of features) and reduces memory consumption substantially compared to RanPAC (e.g., from ~411 MB to ~40 MB on Skin8)
- Results validate the hypotesis and the proposed method. Additionally the paper is well written and easy to follow- weak baselines on ChestX: while four datasets are used, the absolute performance on ChestX is very low (29.76% Last Accuracy), and the joint training upper bound is only 55.51%, raising questions about whether the task setup or feature extractor is appropriate for this dataset.
- the method requires maintaining a feature buffer and performing eigendecomposition on its correlation matrix. The paper does not provide runtime comparisons or discuss how the buffer size B affects computational cost and scalability
- Common CIL metrics such as Forgetting, forward transfer and backward transfer are not reported
- 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.
- weak baselines on ChestX: while four datasets are used, the absolute performance on ChestX is very low (29.76% Last Accuracy), and the joint training upper bound is only 55.51%, raising questions about whether the task setup or feature extractor is appropriate for this dataset.
- the method requires maintaining a feature buffer and performing eigendecomposition on its correlation matrix. The paper does not provide runtime comparisons or discuss how the buffer size B affects computational cost and scalability
- Common CIL metrics such as Forgetting, forward transfer and backward transfer are not reported
- 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?
From a scientific point of view, the paper and the research are well conducted, and I find that the main message can be useful for the research community working on CIL in the medical domain. The results confirm their claims.
- 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.
The rebuttal adequately addresses my three main concerns. The improvements over the strongest analytic baselines remain modest, but the contribution is methodologically coherent and the message is useful for the medical CIL community.
Review #3
- Please describe the contribution of the paper
This paper proposes Eigen-Directed Random Projection, a training-free class-incremental learning method designed to improve fine-grained separability in medical image classification. The key idea is to augment random projection with eigen-directed projections that are better aligned with the principal directions of pretrained medical image features. This is motivated by the observation that medical image features, particularly from ViT-based encoders, can be anisotropic, with discriminative information concentrated along specific feature directions.
Compared with RanPAC, which uses Gaussian random projection, the proposed method combines random projections with eigen-directed components that target feature directions associated with class confusion. The classifier is then learned analytically using ridge regression. The paper evaluates the method across several medical imaging benchmarks and compares it with a broad set of class-incremental learning methods, including analytic classifier-based methods, CLIP-based methods, prompt-based methods, and LoRA-based methods. The results suggest that EDRP improves average and final accuracy, especially in challenging medical class-incremental settings where visually similar classes are difficult to separate.
- 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 addresses an important and underexplored problem in medical image analysis: class-incremental learning in settings where new disease or cell categories may be introduced over time without full retraining. This is clinically relevant because medical datasets are often expanded gradually, and models need to preserve prior knowledge while learning new categories.
A major strength is the clear methodological motivation. The authors identify a mismatch between standard random projection and the geometry of medical image features. Random projection treats all feature directions equally, whereas medical image features may have an anisotropic spectrum in which only some directions are highly discriminative. The proposed eigen-directed projection is a reasonable and technically interesting way to exploit this structure.
The experimental evaluation is broad. The paper compares 14 class-incremental learning methods across 4 medical datasets, which makes the empirical contribution useful even beyond the proposed method itself. The finding that analytic classifier-based methods outperform several recent CLIP-based, prompt-based, and LoRA-based methods in this setting is valuable and worth reporting.
The ablation study is also a strength. The results suggest that eigen-directed projection and random projection provide complementary benefits. In particular, the ablation showing that EDP alone improves performance over the baseline, and that adding RP further improves performance, supports the claim that the two components are not redundant.
The ChestX results are particularly notable. EDRP appears to substantially outperform several competitors on this difficult benchmark, with relatively low variance. This suggests potential robustness in a challenging medical imaging setting where many methods perform poorly.
The method is computationally attractive because it is training-free and uses analytic classifier learning. This makes it potentially practical for medical imaging scenarios where repeated retraining may be expensive or difficult.
- 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.
A key weakness is that some implementation details are underspecified. The feature buffer is central to the proposed method, but the paper does not clearly explain how it is constructed, how many samples it stores, whether it remains fixed or grows as new tasks arrive, and how memory scales with the number of tasks. Since the paper claims memory efficiency, this point needs to be clarified. If the buffer grows with the number of tasks, the memory advantage over RanPAC or related methods may be less convincing in longer class-incremental sequences.
The ChestX comparison table appears incomplete. Table 2 includes RanPAC, InfLoRA, PROOF, and joint training, but it does not include some of the closest competitors from Table 1, such as SLDA and LoRanPAC. Since the ChestX results are one of the strongest empirical claims in the paper, the comparison should include the most relevant competing methods, or the authors should explain why they were omitted.
The dual-view mapping algorithm needs more detail. The authors do not clearly specify how the projection dimension is split between the random projection branch and the eigen-directed branch. This is important for reproducibility and for understanding how sensitive the method is to this design choice.
The paper would benefit from a clearer analysis of forgetting over time. A plot showing accuracy after each new task is introduced would help readers understand whether the method preserves old knowledge better than competitors, rather than only improving final or average accuracy.
The experiments appear to rely on a ViT-B/16 backbone. The anisotropic spectrum argument is plausible for ViT features, but it may not generalize to CNN-based encoders or other pretrained feature extractors. Testing the method with an additional backbone, or explicitly acknowledging this as a limitation, would strengthen the paper.
There are also minor presentation issues. Figure 5 could be improved by making the class mean markers more visible. There are small writing issues, including the typo “dirven” on page 5.Some explanatory sentences, such as the statement about discriminative cues lying outside the dominant subspace, should be rewritten for clarity.
- 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?
I recommend weak accept. The paper presents a genuine empirical and methodological contribution, and the proposed eigen-directed projection is a reasonable and interesting extension of random projection for medical class-incremental learning. The experimental evaluation is broad, and the ablation studies support the main design choices. The ChestX results are particularly promising and suggest that the method may be useful in difficult medical imaging settings.
However, I view the contribution as incremental rather than landmark. The method builds on existing analytic classifier-based class-incremental learning approaches, especially RanPAC, and the novelty lies primarily in the eigen-directed projection component. The paper also has some reproducibility limitations due to missing implementation details about the feature buffer and projection dimension split. The incomplete ChestX comparison and lack of a task-by-task forgetting analysis weaken the empirical case. Overall, the paper is marginally above the acceptance threshold, but the rebuttal should address these concerns.
Furthermore, the improvements with the proposed method are still marginal compared to the other methods.
- 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 #4
- Please describe the contribution of the paper
The manuscript reviews recent continual learning algorithms and finds classifier-based methods to perform best. While those methods rely on random feature projection, the paper proposes a data-driven projection based on the eigenvectors of the embedding space.
- 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 manuscript is well written and pleasant to read.
- The paper compares with a large number of baseline algorithms.
- The proposed method outperforms all of its baselines.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The classification of CL algorithms in Fig. 1 is unclear to me. Classifier-based algorithms can also be pre-trained ViTs, can’t they?
- As far as I understand the proposed method relies on a historical buffer. What does the buffer contain? What implications does it have on privacy preservation?
- In the experimental setup you mention that categories are shuffled, what is that supposed to mean? Also, you mention the application of the “first-session adaptation strategy”. What is that?
- The different styles (Last Acc and Avg Acc) in Fig. 1 are not clearly visible.
- Fig. 3 was never referenced.
- The point of Fig. 2 remains unclear. Maybe you could use the remaining space to mention that the SOTA fails to disentangle the classes.
- In the caption of Fig. 5, you mean to visualize class embeddings but wrote class means.
- The running title is too long.
- Typos: random projections (page 2), Our method is driven by (page 5)
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is well motivated, clearly explained, and outperforms all its baselines. I have mentioned a few minor weaknesses for the authors to strengthen their manuscript, but they do not constitute a reason to reject the paper for me.
- 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 accepting this paper due to the same reasons in my initial review. The only concerning criticism mentioned by other reviewers is limited novelty, which the authors have clarified in the rebuttal.
Author Feedback
We sincerely thank all reviewers. We’re glad to receive three posititve scores of 5/4/4.We carefully address concerns below and code will be released in revision. Q1: Modest extension of RanPAC (R1). A1: We respectfully clarify that EDRP is not a drop-in replacement. Our work follows a complete pipeline: (1)benchmarking that reveals classifier-based methods unexpectedly outperform alternatives in medical CIL; (2)margin-theoretic analysis showing separability is bottlenecked by the weakest projection directions; (3)diagnosing the mismatch between isotropic projection and anisotropic medical spectra; (4)a dual-view design where each component contributes non-redundantly, with a memory reduction. Q2: More backbones (R1, R3). A2: We first want to kindly clarify that our framework is backbone-agnostic, and the benefit of eigen-directed projection is orthogonal to backbone choice, since Fig. 2b is a general property of ViT embeddings regardless of pretraining data. We have conducted the evaluation on medical-domain backbones on Skin8, EDRP consistently outperforms RanPAC and LoRanPAC (e.g., +2.56% and +2.30% Last Acc on BioMed, +1.51% and +1.75% Last Acc on UniMed). Q3: Non-shuffled setting on MedMNIST-Sub (R1). A3: The shuffling protocol we initially designed was intended to simulate a complicated multidisciplinary consultation scenario. In the non-shuffled configuration, our EDRP still maintains its advantage over both RanPAC and LoRanPAC (e.g., +1.08% and +14.98% Acc), further indicating the generality of our method. Q4: Low performance of most methods on ChestX (R2). A4: Prior work Star-Prompt reports joint training with only 48.92% Acc on ChestX, consistent with our Table 2.The reason for such low performance lies in intrinsic dataset difficulty. Even so, our EDRP still achieves the highest accuracy. Q5: Buffer scalability (R2, R3). A5: Unlike prior medical CIL methods that store raw images for replay (e.g., ACL), our feature buffer stores only compact ViT embeddings (768-dim per sample), which is approximately 196 times more compact than a single image (224×224×3). This efficiency allows us to buffer 80% of per-class samples’ embeddings for more accurate eigenspace estimation. Although RanPAC’ buffer size does not change with the number of tasks, it needs to store a full M²-dimensional covariance matrix (e.g., M=10^4), which ovbiously exceeds our total storage as shown in Table 5.Q6: Missing forgetting metrics (R2, R3). A6: EDRP achieves a BWT of −12.50% vs. RanPAC’s −13.84% on Skin8, indicating less forgetting. We will include forgetting metrics in revision. Thanks. Q7: Incomplete comparison on ChestX (R3). A7: In our original experiments, EDRP consistently outperformed SLDA by +11.62%, RanPAC by +17.47%, and LoRanPAC by +3.70%. However, SLDA and LoRanPAC were omitted from Table 2 as they were unstable across seeds and showed no obvious advantage. We’ll clarify in revision. Thanks. Q8: Projection dimension split (R3). A8: M is evenly split: M_f = M_a = M/2; we’ll clarify this point in revision. Q9: Algorithm classification in Fig.1 (R4). A9: Thanks for carefully reading our work. The Classifier-based category refers to methods keeping the backbone frozen with only classifier updates. While these methods do use pre-trained ViTs, they differ from the Pre-trained ViT-based category, which adapts the backbone via prompts/LoRA. We will clarify this distinction. Q10: Buffer contents and privacy (R4). A10: The buffer stores compact ViT embeddings that cannot be inverted to reconstruct original images, offering privacy advantages. Q11: Shuffling and first-session adaptation (FSA) (R4). A11: Category shuffling means we randomly reorder the 36 classes from MedMNIST’s four sub-datasets across tasks. The FSA strategy, originally proposed in RanPAC, applies a lightweight adapter on task 1 before freezing the backbone to improve downstream feature quality. Q12: Presentation issues (R3, R4). A12: We’ll fix all these minor issues. Thanks.
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.
Reviewers think that the paper has clear methodological motivation, broad experimental validation, and promising results. However, there is some concern regarding single frozen ViT backbone and evaluation clinical realism that authors are encouraged to address.
- 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 unanimously agreed that the proposed method offers a solid methodological advancement for continual learning in medical imaging by addressing the anisotropic structure of feature embeddings. During the rebuttal phase, the authors successfully resolved the primary concerns regarding evaluation realism, memory buffer scalability, and missing baseline comparisons.
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.
Most of the major concerns have been addressed by the rebuttal. The authors are encouraged to carefully follow the reviews and revise the draft accordingly, if it is eventually accepted by this round of MICCAI.
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 authors have addressed most of the concerns raised by reviewers. I recommend Accept.