List of Papers Browse by Subject Areas Author List
Abstract
High-quality Late Gadolinium Enhancement (LGE) MRI can be helpful for atrial fibrillation management, yet scan quality is frequently compromised by patient motion, irregular breathing, and suboptimal image acquisition timing. While Multiple Instance Learning (MIL) has emerged as a powerful tool for automated quality assessment under weak supervision, current state-of-the-art methods map localized visual evidence to a single, opaque global feature vector. This black box approach fails to provide actionable feedback on specific failure modes, obscuring whether a scan degrades due to motion blur, inadequate contrast, or a lack of anatomical context. In this paper, we propose Adversarial Concept-MIL (AC-MIL), a weakly-supervised framework that decomposes global image quality into clinically defined radiological concepts using only volume-level supervision. To capture latent quality variations without entangling predefined concepts, our framework incorporates an unsupervised residual branch guided by an adversarial erasure mechanism to strictly prevent information leakage. Furthermore, we introduce a spatial diversity constraint that penalizes overlap between distinct concept attention maps, ensuring localized and interpretable feature extraction. Extensive experiments on a clinical dataset of atrial LGE-MRI volumes demonstrate that AC-MIL successfully opens the MIL black box, providing highly localized spatial concept maps that specify causes of non-diagnostic scans. Crucially, our framework achieves this deep clinical transparency while maintaining highly competitive ordinal grading performance against existing baselines. The code is available at: https://github.com/arf111/AC-MIL
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5702_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/arf111/AC-MIL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{SulK._ACMIL_MICCAI2026,
author = { Sultan, K. M. Arefeen AND Hansen, Kaysen AND Orkild, Benjamin AND Morris, Alan AND Kholmovski, Eugene AND Bieging, Erik AND Kwan, Eugene AND Ranjan, Ravi AND DiBella, Ed AND Elhabian, Shireen Y.},
title = { { AC-MIL: Weakly-Supervised Atrial LGE-MRI Quality Assessment via Adversarial Concept Disentanglement } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes so-called “AC-MIL” which is a multiple instance learning framework for LGE-MRI quality assessment. The methods explicitly disentangles clinically meaningful concepts via bottlenecks, adversarial erasure (to prevent leakage), and spatial attention diversity (to prevent strong correlation between attention maps). The method aims to improve interpretability while maintaining competitive performance.
- 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.
Strengths:
- Well-motivated problem
- clear clinical relevance (interpretability for LGE-MRI)
- the concept seems to be novel
- thoughtful design: adversarial residual branch and spatial diversity constraint are conceptually sound and address known issues (leakage, entanglement)
- solid experimental protocol (10-fold CV, multiple metrics, ablations)
- qualitative analyses (UMAP, attention maps) support interpretability claims
- 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.
Weaknesses:
- empirical gains over prior work (e. g. , HAMIL-QA) are relatively modest and partly inconsistent (SAD)
- clarity could be improved: method description is dense and some design choices (e. g. , stop-gradient scheme) lack sufficient justification.
- also the choice of the SOTA methods [8] and [12] lacks justification
- evaluation is performed on a single data set only
- it is not clear whether the data set is public? there is no link provided, probably it is an in-house data set
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the method description is sound, the main point is the evaluation which is performed on a single in-house data set only. Also the quantiative improvements are partly not consistent. I suggest to extend the evaluation to include at least another public data set and further sota methods (with a justification why these methods are used). Why does the full model not achieve the best QWK—can this trade-off be better characterized?
- 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.
The rebuttal clarifies several methodological points, particularly the role of the SAD regularizer trade-off and the intended interpretability performance balance, which improves internal consistency. However, the evaluation remains limited to a single in-house dataset without external or multi-center validation, which substantially restricts conclusions about generalizability and benchmarking strength. As a result, despite a well-motivated and technically sound approach, the evidence is not yet sufficient for acceptance from my side.
Review #2
- Please describe the contribution of the paper
This paper proposes AC-MIL (Adversarial Concept Multiple Instance Learning) for quality assessment of atrial LGE-MRI (late gadolinium enhancement MRI). The method aims to address a key challenge in this task by decomposing the MIL feature space into several clinically interpretable concepts, such as sharpness, myocardial nulling, aortic enhancement, and residual factors. In addition, the framework uses spatial attention to generate heatmaps for localizing regions associated with image quality degradation.
- 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 problem is clinically meaningful. Image quality in atrial LGE-MRI directly affects downstream analysis and clinical usability, so reliable and interpretable quality assessment is of practical importance. 2.The method has a certain degree of novelty. The paper attempts to decompose the MIL feature space into clinically meaningful concept dimensions (e.g., sharpness, myocardial nulling, aortic enhancement, and residual factors), which goes beyond conventional black-box quality assessment models. 3.The method offers interpretability. The spatial attention heatmaps provide localization of potential quality-defect regions, which improves trustworthiness and makes the model output more aligned with clinical expectations.
- 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.Training and implementation details are insufficient. The manuscript does not provide a sufficiently complete description of the hyperparameter settings, which limits reproducibility and makes it difficult to assess the sensitivity of the reported performance to the training configuration. 2.The paper does not clearly specify the source institution(s), the total number of samples, the train/validation/test split ratios, or the sample distribution across quality levels. These details are essential for judging the reliability of the experimental conclusions and whether the dataset is balanced. 3.The manuscript does not adequately explain how the quality levels were defined by experts, nor whether inter-rater agreement was assessed. Since label reliability is central to a quality assessment task, the expert grading criteria and their consistency should be described much more clearly. 4.Ethics and compliance information is missing. As a medical imaging study, the paper should provide information regarding ethics approval, data access authorization, and related compliance procedures to improve transparency.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
the ethics approval and consent information are missing
- 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?
Overall, the paper addresses a clinically relevant problem and proposes an interpretable MIL-based framework for atrial LGE-MRI quality assessment. Its main strength lies in decomposing image quality into clinically meaningful concepts and combining this with spatial attention for local defect localization and potential feedback, which makes the work more attractive than conventional black-box approaches. However, the current version still has notable weaknesses in dataset description, label definition, experimental detail, which limit the methodological credibility and reproducibility of the work. With substantially stronger clarification in these aspects, the paper could become much more convincing.
- 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 authors have convincingly addressed my four key concerns (dataset details, label reliability, experimental clarity, and ethics compliance) in the rebuttal and promised to amend the manuscript accordingly. I therefore change my recommendation to Accept.
Review #3
- Please describe the contribution of the paper
The manuscript introduces AC-MIL (Adversarial Concept Multiple Instance Learning), a weakly supervised framework for assessing the quality of atrial LGE-MRI scans. The primary contribution is the integration of Concept Bottleneck Models (CBMs) into a MIL architecture to decompose a single global quality score into distinct, clinically meaningful concepts (Sharpness, Myocardium Nulling, Aorta/Valves). To handle latent quality variations not captured by these predefined concepts, the authors introduce an unsupervised residual branch regulated by an adversarial erasure mechanism to prevent information leakage.
- 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.Methodological Novelty: The combination of an unsupervised residual branch with predefined clinical concepts is an elegant solution to the completeness problem often found in Concept Bottleneck Models. The use of a Gradient Reversal Layer (GRL) to enforce feature orthogonality and prevent the residual branch from trivially relearning known semantics is well-executed.
2.Clinical Interpretability: By breaking down the opaque MIL pooling process into specific radiological failure modes, the framework provides highly actionable feedback for clinical technicians, directly addressing the “black box” criticism of standard deep learning approaches.
3.Strong Qualitative Validation: The manuscript provides compelling visual and statistical evidence supporting its architectural claims. The UMAP projections, Jensen-Shannon Divergence (JSD) boxplots, and visual attention maps (Figure 3) effectively demonstrate that the Spatial Attention Diversity (L_SAD) successfully forces spatial specialization.
4.Rigorous Ablation: The ablation study (Table 1) clearly isolates the impact of the CBM, SAD, and Adversarial components, providing transparency regarding the performance trade-offs inherent in enforcing 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.
1.Performance Trade-off: As shown in Table 1, the inclusion of the Spatial Attention Diversity (L_SAD) penalty actually decreases the overall Quadratic Weighted Kappa (QWK) from 0.68 to 0.66.While the authors logically justify this as a “necessary trade-off” to prevent shortcut learning and ensure true anatomical alignment, the manuscript would benefit from a deeper discussion on why spatial overlap inherently degrades the metric and how this trade-off might affect clinical deployment.
2.Lack of Expert Validation on Attention Maps: While the Jensen-Shannon Divergence proves the attention maps are distinct from one another, there is no quantitative validation that these maps accurately localize the correct artifacts. A small-scale reader study or expert scoring of the attention maps would strongly validate the clinical utility claim.
3.Dataset Diversity: The manuscript evaluates 775 LGE MRI scans but does not specify if this is a multi-center or single-center dataset. If it is single-center, claims of generalizability may be overstated, especially given the known variability in LGE-MRI acquisition protocols across different institutions.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
This is a very well-written and methodologically sound paper. To further strengthen your manuscript:
1- Consider clarifying the demographic and institutional makeup of your dataset (e.g., single vs. multi-center, scanner vendors).
2- It would be highly beneficial to include a brief discussion or supplementary analysis where a clinical expert grades the generated spatial attention maps to definitively prove they highlight the intended anatomical structures or artifacts.
3- In Section 3.2, you mention using a 100-epoch early stopping patience. Given the dataset size, please briefly comment on whether overfitting was observed prior to the early stopping trigger.
- 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 paper presents a highly original and methodologically rigorous approach to a pervasive problem in medical imaging: the opacity of weakly supervised models. The architectural choices—specifically the asymmetric stop-gradients and the adversarially-regularized unsupervised branch—are sophisticated and directly tackle the issue of concept leakage. Although there is a minor performance drop associated with enforcing spatial diversity, the resulting interpretability is highly valuable for clinical translation. The paper is well-organized, the claims are supported by strong qualitative and quantitative evidence, and the work provides a solid contribution to the MICCAI community.
- 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 authors provided a clear, direct, and convincing rebuttal that effectively addressed my initial concerns, as well as the misunderstandings raised by other reviewers.
First, they successfully clarified the nature of the QWK drop (0.68 to 0.66), confirming that it was an intentional, regularization-controlled trade-off for spatial disentanglement (SAD) rather than model instability. Providing the paired permutation p-values (p=0.008 for λSAD=0.1 vs. p=0.30 for λSAD=0.05) demonstrates a rigorous understanding of their model’s behavior and justifies their design choice to prioritize anatomical consistency over raw performance.
Second, the authors provided reassuring details regarding the expert validation of the concept labels, reporting solid inter-rater reliability scores (ICC/weighted κ of 0.74 for overall quality and 0.62–0.72 for concept categories) on a 50-case subset. This strengthens the clinical credibility of the weakly supervised labels.
Finally, while the dataset remains single-center and in-house, the authors reasonably point out the absence of public datasets with both ordinal QA and concept-level labels for atrial LGE-MRI. Their commitment to clarifying the scanner/source details and adding the omitted IRB/ethics statements in the final manuscript resolves the compliance concerns. The methodological contributions of AC-MIL remain strong and highly relevant.
Review #4
- Please describe the contribution of the paper
The paper focuses on image quality assessment for atrial fibrillation LGE-MRI, decomposing the overall quality score into several interpretable concepts. It adopts a Concept Bottleneck Model and incorporates an adversarial erasure mechanism to prevent the model from relying on a single dominant feature.
- 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.
This paper uses CBM to decompose the quality score into interpretable sub-concepts is clinically meaningful and the overall idea is clear. The adversarial erasure design is well thought out. By using a GRL, the model is forced not to rely on a single concept, which can theoretically improve robustness. The visualizations are intuitive and illustrate the contribution of different concepts to the final score. The use of 10-fold cross-validation is also reasonable given the small dataset size.
- 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.In Table 2, the full model (QWK = 0.66) performs worse than the version without the adversarial loss (QWK = 0.68). This means that according to the reported results, adding the paper’s core contribution actually decreases performance. The paper does not explain this phenomenon or provide statistical significance testing. This either suggests that the proposed design is ineffective or that the training process is unstable.
2.In Fig. 2a, the adversarial branch for the Sharpness concept achieves an AUC as high as 0.96, indicating that the other concepts still contain substantial blur-related information. This suggests that the adversarial erasure does not effectively remove the target concept information, which contradicts the paper’s claim that it encourages independence among concept features. This looks more like information leakage.
3.The paper describes the setting as weakly supervised, but it is unclear how the concept labels were annotated. Were the quality scores for each dimension manually assigned by radiologists image by image, or were they derived from the overall quality score? This is not clearly explained.
4.The paper only compares against several classification networks and does not compare with other interpretable methods. This weakens the evidence for its core claim of interpretability.
- 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.
(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 core contribution of the paper is the adversarial erasure mechanism, but the experimental results do not support this claim. In Table 2, the full model (QWK = 0.66) performs worse than the ablated version without the adversarial loss (QWK = 0.68), and no statistical significance testing is provided. In fig 2a, the adversarial branch for the Sharpness concept still achieves an AUC as high as 0.96, indicating that feature redundancy has not been effectively mitigated. These two pieces of evidence directly call into question the effectiveness of the proposed method, but no explanation is given. Also, the weak supervision labeling process is not clearly described, and the comparison baselines are limited to standard classification networks. The claimed contribution in interpretability is not sufficiently validated.
- 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.
My main concerns have been adequately addressed. The response provides sufficient clarification on the issues raised in my initial review and improves my confidence in the validity and presentation of the work.
Author Feedback
We thank the meta-reviewer and reviewers for their constructive feedback and address the major concerns below. M, R1, R3, R4: QWK drop, 0.68 -> 0.66.The 0.68 QWK row is not the model “without adversarial loss”; it is AC-MIL with CBM+adversarial erasure. The full model adds Spatial Attention Diversity (SAD), resulting in QWK=0.66.Thus, the reduction is due to SAD, not adversarial erasure. SAD penalizes overlap between concept-specific attention maps, reducing shortcut correlations and enforcing spatial disentanglement. As shown in Fig. 3, this yields more spatially separated and anatomically consistent attention maps. We performed paired analysis on existing out-of-fold predictions. For the reported setting (λSAD=0.1), the QWK reduction is statistically significant (paired permutation p=0.008<0.05), indicating a measurable cost. When SAD regularization is reduced (λSAD=0.05), the difference is not significant (p=0.30>0.05), showing that the trade-off is controlled by regularization strength. We chose λSAD=0.1 to prioritize interpretable, anatomically consistent concept separation while maintaining competitive QA performance relative to ABMIL and HAMIL-QA, rather than maximizing raw QWK. Thus, the full model reflects a principled interpretability–performance trade-off rather than instability. M, R4: Fig. 2a and “AUC = 0.96.” This appears to stem from a misinterpretation of Fig. 2a. We do not report a Sharpness adversarial AUC of 0.96.Fig. 2a shows adversaries predicting predefined concepts from the unsupervised branch; curves near chance indicate that predefined concept signals are not recoverable from the unsupervised branch above chance. Thus, Fig. 2a supports reduced leakage rather than demonstrating leakage. We will clarify the caption and text to avoid confusion. M, R2, R4: concept labels and weak supervision. The concept labels were manually assigned by expert readers at the volume level; they were not derived from the global quality score. The method is weakly supervised because AC-MIL uses only volume-level quality and concept grades, without slice-, patch-, or pixel-level supervision. Inter-rater agreement is available only on a subset. On a 50-case subset independently scored by 5 raters, ICC/weighted κ was 0.74 for fibrosis-assessment quality and 0.62–0.72 across the concept categories. We emphasize that these are limited subset estimates and not full-dataset reliability measurements for all 775 scans. R1: method clarity and stop-gradient. As reported in Sec. 2.4, the asymmetric stop-gradient prevents task-level gradients from forcing predefined concept branches to encode final QA information, reducing concept leakage. We will make this motivation more prominent. R2, R3: dataset, split, and ethics. The manuscript reports 775 LGE-MRI scans with stratified 10-fold CV: 10% test per fold and 10% of the remaining training portion for validation/early stopping. We will additionally clarify quality-grade distribution, scanner/source details under data-use restrictions, and the omitted IRB/ethics approval, consent/waiver, and data-use statements. This was a presentation omission, not a lack of IRB/data-use compliance. R1, R4: baselines and interpretability comparisons. ABMIL and HAMIL-QA were selected because they are directly aligned with weakly supervised MIL-based quality assessment, with HAMIL-QA being the closest prior atrial LGE-MRI QA method. Generic interpretable classifiers are not directly comparable because they do not address the hierarchical weakly supervised volume/slice/patch setting. We will clarify this choice. R1, R3: single-dataset evaluation. We agree that multi-center validation would strengthen generalizability. However, to our knowledge, no public dataset exists for atrial LGE-MRI quality assessment with both ordinal QA and concept-level labels. The dataset used in this work is in-house; code will be released upon acceptance, and dataset access will be explored through institutional MOU agreements.
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 agreed that the paper addresses a clinically relevant problem and proposes an interpretable framework for atrial LGE-MRI quality assessment with some interesting ideas. However, the main concerns centered on the strength of the empirical validation. In the rebuttal, the authors should (1) explain and justify the QWK drop from 0.68 to 0.66 when adding the SAD loss, with statistical significance testing; (2) clarify whether the sharpness adversarial branch AUC of 0.96 constitutes information leakage; (3) describe how concept-level labels were annotated and report inter-rater agreement if available.
- 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 rebuttal is sufficient in addressing the concerns of the reviewers, though R1 maintained rejection based on single-centre limitation. The limitation is noted but can be accepted for a MICCAI paper with techinical novelty, which is on bottleneck modeling and spatial attention diversity, and can be interesting for discussion at the conference.
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.
Although the paper proposes an interesting framework for atrial LGE-MRI quality assessment, the clinical motivation is not fully convincing: unlike real-time ultrasound QA, it remains unclear how the predicted quality concepts would be used to meaningfully adjust MRI acquisition or improve downstream clinical decision-making. In addition, the validation is limited to a single in-house dataset (considering there are so many public atrial LGE MRI resource in MICCAI challenge, such as LA challenge 2018, LAScarQS 2021, CARE 2023, etc), the gains over existing methods are modest, and the interpretability claims still lack strong external or clinician-validated evidence. Therefore, despite the positive aspects of the method, I am not convinced that the current evidence is sufficient for acceptance.
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 rebuttal addressed the main concerns about the SAD trade-off, the adversarial leakage interpretation, and the concept-label annotation process. In the final version, please add the promised details on dataset composition, splits, ethics approval, inter-rater agreement, and clearly state the limitation that the evaluation is based on a single in-house dataset.