List of Papers Browse by Subject Areas Author List
Abstract
Concept Bottleneck Models (CBMs) have emerged as a promising direction for interpretable medical image analysis by grounding predictions on human-understandable concepts. However, training effective CBMs typically requires dense concept annotations by clinical experts, which is prohibitively expensive and scarce. While semi-supervised CBMs attempt to leverage unlabeled data, they often suffer from label noise inherent in pseudo-labeling and fail to preserve the pathological correlations between concepts (e.g., symptom co-occurrences). In this paper, we propose a novel Structurally-Aware and Noise-Tolerant
CBM (SANT-CBM) framework for semi-supervised medical image
classification. To address the unreliability of pseudo-labels, we introduce a GMM-based Dynamic Weighting mechanism that models the loss distribution of unlabeled samples to separate clean supervision from noise, effectively “denoising” the training process. Furthermore, we propose a Structural Consistency Loss that aligns the correlation matrix of predicted concepts on unlabeled data with the reference correlation derived from clinical priors. This ensures that the model learns biologically plausible concept combinations rather than independent features. Extensive experiments on two public dermatology datasets (Fitzpatrick17k and 7-point) demonstrate that our method significantly outperforms state-of-the-art semi-supervised CBM baselines. Specifically, our approach achieves diagnostic accuracies of 77.83% (vs. 71.52%) on Fitzpatrick17k and 73.27% (vs. 67.52%) on 7-point (based on 10% labeled), establishing a new benchmark for label-efficient and interpretable medical diagnosis.
Keywords: Concept Bottleneck Models · Semi-supervised Learning ·Medical Image Classification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4448_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuHon_SANTCBM_MICCAI2026,
author = { Liu, Hongwei AND Liu, Jia AND Lai, Songning},
title = { { SANT-CBM: Structurally-Aware and Noise-Tolerant Semi-supervised Concept Bottleneck Models } },
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 addresses a practically important problem in concept bottleneck models, namely how to train interpretable CBMs under sparse concept annotations in a semi-supervised medical setting. The paper identifies two limitations of existing semi-supervised CBMs: noisy pseudo-labels on unlabeled data and the lack of structural constraints over concept co-occurrence. To address these issues, the authors propose SANT-CBM, which combines a GMM-based dynamic weighting mechanism to down-weight unreliable pseudo concept supervision and a structural consistency loss to align predicted concept correlations with a reference correlation matrix derived from labeled clinical concepts. Experiments on two dermatology datasets show improved task accuracy over prior semi-supervised CBM baselines, especially in the low-label regime.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper tackles an important bottleneck in CBMs, namely the heavy reliance on dense concept annotations, which is especially costly in medical imaging. Extending CBMs to the semi-supervised setting is highly relevant. 2.The discussion of pseudo-label noise and the neglect of concept-to-concept pathological structure is well motivated, particularly for medical diagnosis where concepts are often correlated rather than independent. 3.The two main components—GMM-based denoising and structural consistency regularization—are intuitive, and well integrated into the CBM pipeline. 4.The reported gains over SSCBM in task accuracy are substantial, especially with 10% labeled data, suggesting that the proposed regularization is useful for label-efficient diagnosis.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.A notable issue is that task accuracy improves clearly, but concept accuracy does not consistently improve across datasets and settings, compared to SSCBM (Table 1) . This makes it unclear whether the method truly improves the bottleneck concepts themselves, or mainly benefits the final classifier downstream of the concepts.
2.The paper does not sufficiently compare with or discuss recent label-free CBM methods, such as:
- Label-Free Concept Bottleneck Models
- Explain via Any Concept: Concept Bottleneck Models with Open-Vocabulary Concepts
- 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 lean weak accept. The paper addresses an important and underexplored problem in semi-supervised CBMs for medical imaging. The proposed method is well motivated, technically clear, and demonstrates meaningful gains in task accuracy under limited concept supervision. However, the evidence for improved concept quality is not yet fully convincing.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes a semi-supervised concept bottleneck model that leverages unlabeled data by generating pseudo-concept labels via kNN retrieval from a labeled feature bank, followed by per-concept Gaussian mixture model (GMM)–based denoising using agreement between pseudo-labels and model predictions. To further improve reliability, the method introduces a structure-aware regularization that aligns predicted concept correlations with a reference correlation matrix computed from labeled data. The approach aims to learn concept representations that are both noise-tolerant and structurally consistent, thereby improving downstream task performance under limited concept supervision.
- 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.
- Methodology. Effectively combines psuedo-label denoising (GMM) with structure aware learning, addressing key challenges in semi-supervised CBMs. And Uses per-concept reliabiilty weighting leveraging noisy pseudo labels effectively instead of discarding them.
- Demonstrates improved task performance in low-label regimes showing practical value when concept annotations are scarce .
- 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.
- Concept accuracy remains largely uncahnged, suggesting tha t improvements stem from enforcing structured interactions ratehr than improving individual concept predictions, which raises questions about interpretability benefits typically what CBMs are motivating.
- 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
-A brief discussion clarifying why task performance improves despite minimal gains in concept accuracy would bring more clarity to the results, particularly in relation to the limitations of standard concept evaluation metrics not capturing the structural consistency in combinations, but can still help improve task accuracy?
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important limitation of concept bottleneck models, scaling to limited concept supervision through a combination of denoising and structural constraints. The approach is promising and well-motivated, however further analysis can help understand the relationship between concept quality and downstream performance, as well as the robustness of the learned structure.
- 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 response clarifies that the proposed structured objective is intended to improve relationships among concepts that helps improve downstream task classification rather than independendt concept prediction accuracy. This helps explain the observed discrepancy between steady concept performance and improved task performance. While some uncertainty remains regarding the learned concept correlations estimated from relatively sparse labeled datasets (10-20%) consistently reflect clinically meaningful structure, I believe the paper presents a meaningful contribution to semi-supervised concept bottleneck learning.
Review #3
- Please describe the contribution of the paper
The paper proposes SANT-CBM, a semi-supervised concept bottleneck model for medical image classification that aims to improve both robustness and plausibility of concept learning under scarce concept annotations. The method combines two components on unlabeled data: (1) a GMM-based dynamic weighting strategy that down-weights potentially noisy pseudo-label supervision at the concept level, and (2) a structural consistency loss that aligns the batch-wise correlation matrix of predicted concepts with a reference correlation matrix computed from labeled concept annotations. Experiments on Fitzpatrick17k and the 7-point Checklist dataset show improved diagnostic accuracy over prior semi-supervised CBM baselines, especially in the low-label regime.
- 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 studies an important problem how to make concept bottleneck models more label-efficient while preserving interpretability in medical image classification. The motivation is well aligned with clinical deployment scenarios. The core strength of this research lies in its highly effective, label-efficient framework that synergizes probabilistic denoising with clinical priors to establish a new state-of-the-art in medical diagnosis. By introducing a GMM-based Dynamic Weighting mechanism, the model robustly filters noise from unlabeled data, while a novel Structural Consistency Loss explicitly forces the network to learn biologically plausible concept combinations rather than isolated features. This dual innovation enables SANT-CBM to significantly outperform existing methods, achieving remarkable diagnostic accuracy improvements of 6.31% and 5.75% on the Fitzpatrick17k and 7-point datasets, respectively, using only 10% labeled data.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.While the motivation to introduce clinical constraints is commendable, the proposed Formula (7) is fundamentally a standard feature correlation matrix alignment, a highly common approach for consistency loss. Shifting the computation target to a batch-wise predicted concept Pearson correlation matrix ($\hat{R}_{batch}$) lacks substantial theoretical novelty. 2.The use of the term “Clinical Priors” is heavily exaggerated. In the current methodology, the reference matrix is derived purely from the Pearson correlation of concept annotations within the labeled training set, rather than integrating external clinical knowledge graphs or expert-defined ontological structures. Consequently, this mechanism is essentially dataset-level concept correlation regularization, not an independent clinical prior. 3.Section 4.1 specifies a batch size of merely 16.Computing a Pearson correlation matrix on such a remarkably small batch size is statistically prone to extreme variance and instability. The authors do not address how calculating sample-to-sample correlations within such small batches avoids introducing highly noisy gradients during training. 4.The paper motivates the method partly through “noise tolerance,” but it is not always clear whether the targeted noise is pseudo-label noise (inherent in the semi-supervised learning process) or challenging real-world dermatology image quality. This distinction matters significantly because the proposed GMM weighting method operates on pseudo-supervision loss distributions rather than directly modeling or addressing physical image corruption. 5.Figure 1 is not provided in a vector graphics format. It exhibits significant blurring and pixelation when zoomed in, which fails to meet the standard publication quality of top-tier journals/conferences. 6.The visual representation of the “Clinical Priors” in Figure 1 only displays three features (Asymmetry, Color, Blue-Gray Veil). Given that the experiments utilize 22 and 19 concepts for the Fitzpatrick17k and 7-point datasets respectively, this excessive simplification prevents readers from intuitively grasping the actual computational complexity of the high-dimensional matrix alignment. 7.In the “Reference Matrix” portion of Figure 1, there is a complete absence of a legend or colorbar. Furthermore, it is never explicitly explained what the irregularly darkened areas within the orange boxes actually represent, leaving the visualization visually ambiguous.
- 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
he paper addresses a relevant and interesting problem, and I appreciate the attempt to combine robustness and structural plausibility in semi-supervised CBMs. For the rebuttal, I would especially appreciate clarification on three points: (1) why the reference correlation matrix is described as a “clinical prior” when it appears to be estimated from the labeled training set, (2) how the structural consistency term on unlabeled data should be interpreted relative to more standard consistency-style regularization, and (3) several implementation details that are currently hard to infer from the manuscript. Improving the explanation of Figure 1 would also make the paper easier to follow.
- 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 studies an important problem, namely semi-supervised concept bottleneck models for medical diagnosis under scarce concept annotations. The empirical results are encouraging: in the 10% labeled setting, SANT-CBM achieves clear task-level improvements over SSCBM on both Fitzpatrick17k and 7-point, which suggests that the proposed framework has practical value for label-efficient interpretable learning. However, I currently rate the paper as a Weak Reject because I am not yet fully convinced by the strength of the methodological contribution and by several aspects of the presentation and experimental justification. First, while the motivation for introducing “clinical priors” is appreciated, the proposed structural consistency loss in Eq. (7) appears, in essence, to be a standard correlation-alignment regularizer applied to the batch-wise Pearson correlation matrix of predicted concepts. Relative to prior semi-supervised CBM work, this is a meaningful extension, but the degree of theoretical novelty seems moderate. Second, the term “clinical prior” appears somewhat overstated in the current manuscript. The reference matrix is computed from concept correlations within the labeled training set, rather than derived from an external clinical knowledge source or expert-defined prior structure. As a result, the method seems closer to dataset-level concept correlation regularization than to the incorporation of an independent clinical prior. Third, I also have some concern about the stability of estimating a batch-wise Pearson correlation matrix with the reported batch size of 16, especially early in training when concept predictions on unlabeled data may themselves be noisy. This issue is not sufficiently discussed. In addition, Figure 1 is low-resolution and difficult to parse, and its simplified visualization does not clearly convey the actual high-dimensional matrix alignment used in the experiments. The lack of clear legend/color explanation further affects readability. Overall, I think the paper has promising empirical performance, but the current version still has important weaknesses in novelty calibration, conceptual framing, and clarity. For the rebuttal, I would particularly appreciate clarification on the exact type of “noise” the method is intended to address, the intended interpretation of the “clinical prior,” and the authors’ perspective on the stability of the batch-wise structural loss.
- 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.
I thank the authors for their rebuttal. However, after considering the response, I would still recommend rejection. The rebuttal clarifies that the proposed GMM module targets pseudo-label noise in unlabeled concept supervision rather than directly modeling physical image corruption or real-world dermatology image artifacts. This clarification is useful, but it also weakens the original “noise-tolerant” framing. In the submitted manuscript, the GMM branch is described as “denoising” the training process, but technically it is a small-loss / loss-distribution based reweighting mechanism applied to pseudo-concept supervision. I therefore still find the denoising claim somewhat overstated. Similarly, the rebuttal acknowledges that the so-called “clinical prior” is not derived from external clinical knowledge, expert-defined ontology, or a medical knowledge graph, but from concept correlations in the labeled training set. This supports my original concern that the manuscript overstates the clinical-prior aspect. The method is more accurately described as labeled-set concept correlation regularization rather than the incorporation of an independent clinical prior. I also remain unconvinced by the methodological novelty of the Structural Consistency Loss. As formulated in the paper, it computes a batch-wise Pearson correlation matrix of predicted concepts on unlabeled data and aligns it with a reference correlation matrix using a Frobenius-norm penalty. While this may be a reasonable regularizer, the operation itself is relatively simple and closely resembles standard correlation-alignment or consistency regularization. The rebuttal reframes the contribution as concept-level structure regularization for unlabeled predictions in semi-supervised CBMs, but I still view this as an incremental extension rather than a sufficiently strong methodological contribution. The response regarding batch stability is also only partially convincing. The authors mention epsilon normalization, a moderate structural-loss weight, and three independent runs, but they do not provide direct evidence that estimating a concept correlation matrix from a batch size of 16 is statistically stable, especially early in training when pseudo-concept predictions may be noisy. Finally, the concerns about the figures remain. The submitted Figure 1 is low-resolution, visually ambiguous, lacks adequate legends/color explanation, and does not clearly convey the actual high-dimensional concept-correlation alignment used in the experiments. The authors promise to improve the figure in the final version, but the current decision should be based on the submitted manuscript and rebuttal rather than promised future revisions. Overall, although the empirical gains are interesting, the rebuttal does not sufficiently resolve my concerns regarding overstated terminology, limited methodological novelty, insufficient stability justification, and presentation quality. I therefore maintain a reject recommendation.
Author Feedback
We thank the reviewers and meta-reviewer for their recognition of our work. We address the common concerns below. (1) Why task accuracy improves more than concept accuracy. We agree that SANT-CBM should not be interpreted as uniformly improving every marginal concept AUC. Our claim is more specific: under sparse concept labels, bottleneck quality has both an individual-concept aspect and a joint-configuration aspect. Standard concept accuracy evaluates each concept independently, whereas the downstream classifier also depends on whether the concept vector forms plausible co-occurrence patterns. Table 2 separates these effects. On Fitzpatrick17k with 10% labels, GMM denoising improves concept AUC from 86.31 to 87.81 and task accuracy from 71.51 to 74.38.Structural loss alone changes marginal concept AUC only slightly (86.47) but improves task accuracy to 75.13, indicating that it helps mainly through joint concept structure. Combining both gives the best reported pair, 88.84/77.83.We also note that under the low-label regime, further improving marginal concept AUC is inherently challenging, as even sota semi-supervised CBMs (SSCBM) achieve only moderate gains beyond the supervised baseline. In this context, the improvement from 86.31 to 87.81 (GMM alone) is non-trivial. More importantly, as shown in Table 2, the structural loss alone yields a substantial task gain (71.51 → 75.13) with almost no change in marginal concept AUC, confirming that task improvement primarily comes from better concept co-occurrence structure rather than from isolated concept accuracy. We will revise the discussion to state this distinction explicitly rather than implying that all gains come from marginal concept accuracy. (2）Noise type and batch stability. The noise targeted by the GMM is pseudo-label noise in unlabeled concept supervision. Image artifacts may induce unreliable pseudo-label losses, but the method does not claim to directly model physical image corruption. The GMM models per-concept pseudo-supervision loss distributions and down-weights unreliable pseudo-concept targets. For the batch-wise structural loss, the batch correlation is used as a stochastic training regularizer rather than as a one-shot estimate of clinical statistics. The Pearson denominator includes epsilon normalization, the structural term is weighted moderately (lambda_s=0.3), and the reported results are means of 3 independent runs. The structural-loss-only ablation improves task accuracy from 71.51 to 75.13, suggesting that the signal is useful rather than destabilizing in this setting. (3) “Clinical prior” and novelty. We agree with R3 that the current wording can overstate the source of the prior. The reference matrix is computed from labeled clinical concept annotations in the training set; it is not an external knowledge graph or expert ontology. We will rename it as an “empirical clinical-concept correlation prior” or “labeled-set concept reference matrix” throughout the final version. Our novelty claim is not Pearson correlation alignment itself is new. The contribution is concept-level structure regularization on unlabeled predictions in a semi-supervised CBM, complementing SSCBM’s local visual alignment and the GMM denoising branch. This is the mechanism behind the low-label task gains recognized by the reviewers. For R1’s label-free/open-vocabulary CBM concern, these methods address a different setting: discovering or naming concepts without a fixed clinically annotated concept set. SANT-CBM instead assumes a small labeled set of clinically defined concepts and a larger unlabeled pool. For R3’s Figure 1 concerns, we agree and will improve the final version with vector-quality artwork and a legend explaining the matrix values and highlighted regions.
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 feedback is mixed. However, the novelty contribution is recoganized by all. The authors are encouraged to address the common concerns raised by the reviewers, like “why task performance improves despite minimal gains in concept accuracy”, and clarify the points like “the exact type of ‘noise’”, in the final version.
- 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 addressed several main concerns raised by the reviewers, and are strongly encouraged to avoid overclaiming as per Reviewer #3’s comments.
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.
Started with diverse comments and concerns and diverging ratings, the rebuttal is with limited effectiveness in addressing the major concerns, including but not limited to issues over novelty concern, over-claims of clinical prior and others, and presentation concerns. Overall it requires significant revision efforts that goes beyond minor nature to get into MICCAI in this round.
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.
This is a borderline paper with mixed reviews. Reviewers agreed that the work addresses an important and under-explored problem—extending Concept Bottleneck Models to semi-supervised and label-scarce medical imaging settings—and that the integration of GMM-based pseudo-label denoising with structural consistency regularization is well motivated and yields substantial gains in task accuracy under low-label regimes. However, they also raised serious concerns about the paper’s soundness and significance: first, concept accuracy did not consistently improve alongside task accuracy, raising questions about whether the method enhances interpretability or merely optimizes downstream classification; second, reviewers highlighted overstated claims regarding “denoising” and “clinical priors,” noting that these are more accurately described as standard loss reweighting and dataset-derived correlation regularization rather than genuine noise modeling or external clinical knowledge integration; third, the methodological novelty of the structural consistency loss was viewed as limited, with close resemblance to existing correlation-alignment techniques, and concerns were raised about statistical instability due to small batch sizes and inadequate empirical justification. The rebuttal addressed some, but not all, concerns—clarifying design choices while falling short of resolving issues around terminology, novelty, and robustness. After careful deliberation, I lean towards accepting this manuscript.