List of Papers Browse by Subject Areas Author List
Abstract
Accurate cell-type annotation is a key prerequisite for downstream analysis in single-cell multi-omics studies, yet new assays such as scATAC-seq are often weakly labeled and exhibit substantial modality shift from well-annotated scRNA-seq references. This dual label scarcity combined with cross-modality discrepancy makes naive pseudo-labeling brittle, leading to error accumulation and systematic confusion among closely related cell states. We propose CADENCE, a unified semi-supervised transfer framework that builds on a divide-and-conquer strategy to separate source-like target cells from target-specific ambiguous cells. For source data, we expand supervision via optimal-transport-based pseudo-labeling under a curriculum and class-balanced selection. For target data, we introduce three reliability-enhancing components: (i) a gentle teacher–assistant–student scheme that isolates noisy supervision to an assistant branch while distilling robust representations to the student through exponential moving average updates; (ii) a lightweight bidirectional latent translator trained with cycle consistency, providing per-cell reliability signals to gate and reweight ambiguous target samples and to safely shrink their candidate label sets; and (iii) a confusing-pair correction mechanism that dynamically identifies highly confused class pairs and rectifies soft labels using prototype geometry, mitigating boundary collapse. Extensive experiments on publicly available single-cell multi-omics benchmarks demonstrate that CADENCE consistently improves annotation robustness under severe label scarcity and reduces misclassification among similar cell types, yielding more reliable cross-modality cell atlases without requiring paired measurements.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4247_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{LuZhe_CycleVerified_MICCAI2026,
author = { Lu, Zheng AND Yu, Peng AND Wang, Yan},
title = { { Cycle-Verified Gentle Teaching and Confusion Correction for Label-Scarce Cross-Modal Single-Cell Annotation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes the CADENCE framework to tackle the single-cell multi-omics cross-modality transfer problem under extremely scarce label scenarios. By integrating optimal transport and a reliability-aware strategy, it enhances the robustness of cell type annotation. The proposed method has certain novelty, and its experimental results outperform existing approaches. However, the paper exhibits obvious deficiencies overall, including non-standard writing, ambiguous experimental configurations, and insufficient experimental design.
- 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.CADENCE focuses on the practical and challenging scenario of extremely low label ratios, with a clear research motivation that aligns with real-world demands for single-cell data analysis. 2.CADENCE achieves better performance than multiple existing methods under this setting, which validates the effectiveness of the proposed strategies and provides certain reference value.
- 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 writing is poor, with missing explanations for mathematical notations and symbols, making the paper highly incomprehensible. 2.Only a 1% label ratio is tested; it is recommended to evaluate performance under several additional label ratios and conduct hyperparameter sensitivity analysis. 3.The evaluation metric is overly singular. More metrics should be used to assess experimental performance, or additional visualizations should be added to demonstrate the model’s effectiveness. 4.All four tasks, data preprocessing pipelines, and model parameter settings should be clearly and systematically defined in the experimental setup section of the main text.
- Please rate the clarity and organization of this paper
Poor
- 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.
(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?
CADENCE framework for single-cell multi-omics cross-modality transfer under extremely scarce labels, with a clear practical motivation and superior experimental performance over existing methods via optimal transport and reliability-aware strategies, showing certain novelty and reference value. However, it has obvious defects including non-standard and confusing writing, insufficient experimental settings with only a 1% label ratio tested, single evaluation metrics, and unclear definitions of tasks, preprocessing and parameters.
- Reviewer confidence
Very confident (4)
- [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.
N/A
Review #2
- Please describe the contribution of the paper
The paper introduces CADENCE, a robust semi-supervised framework designed for cross-modality cell-type annotation under extreme label scarcity (e.g., $\approx 1\%$ source labels). The framework addresses the issue of boundary collapse and error accumulation by shifting the focus from simply expanding pseudo-labels to strictly controlling their reliability. Key technical contributions include a gentle teacher-assistant-student architecture to isolate noisy supervision, a cycle-consistent latent translator (SWITCH) for dynamic sample gating, and a prototype-based confusing-pair correction (CPC) mechanism to rectify systematic misclassifications between similar cell states.
- 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 introduction and abstract are exceptionally well-written. The logical progression from the identified problem (boundary collapse) to the proposed components is clear and easy to follow.
2.The synergistic use of cycle consistency for reliability gating and prototype geometry for local boundary refinement provides a sophisticated solution to noise propagation in extreme low-label regimes.
3.CADENCE consistently outperforms state-of-the-art methods across diverse benchmarks.
- 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 paper focuses on mitigating boundary collapse, there is no qualitative visualization to demonstrate this effect. Relying solely on Accuracy is somewhat narrow. Including F1-score or other metrics would provide a more solid evaluation of annotation robustness.
2.The paper lacks sensitivity experiments for several critical hyperparameters, particularly the loss balancing weights $\lambda_{SL}$, $\lambda_{align}$, and $\lambda_{ent}$ in the final objective.
3.The notation in Eq. 3 $\arg \max \hat{y}(x)$ is theoretically redundant as $\hat{y}$ is already a label/coupling result; it should likely be $\arg \max p(x)$ to refer to the predicted probability. The parameters in Eq. 8 &10, $\sigma$, $\delta$, and $\gamma$ are used but not defined. The function $\rho(e)$ for the pseudo-label budget is mentioned without explanation or definition.
4.The CADENCE acronym (specifically the “CA” part) is not clearly explained in the text.
5.Figure 1 is overly crowded and difficult to read. The term GTA is used in the figure/caption but its full name (Gentle Teaching Assistant) should be explicitly stated there for clarity.The “Proto” label in the diagram lacks descriptive value regarding how the prototypes are actually utilized during the update process.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper tackles a realistic and challenging regime (1% label scarcity) in multi-omics with a very clean logical structure and strong quantitative gains. The “confusing-pair correction” is particularly insightful. However, the lack of sensitivity analysis and various notational oversights (undefined variables, confusing diagram, and missing full names like GTA in captions) need to be addressed.
- 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 resolved my concerns regarding the formulation and evaluation. The authors provided solid theoretical justifications for Eq. 3 and successfully added Macro-F1 metrics and clustering visualizations. I recommend acceptance.
Review #3
- Please describe the contribution of the paper
The paper proposes CADENCE, a semi-supervised framework designed to tackle the challenge of extreme label scarcity (only 1%) in cross-modal single-cell annotation. By introducing a “teacher-assistant-student” distillation mechanism and a prototype-based error correction method, it successfully prevents classification boundary collapse caused by noisy pseudo-labels.
- 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.This paper addresses the practically relevant and underexplored challenge of extreme label scarcity in RNA-to-ATAC multi-omic knowledge transfer. 2.The proposed method demonstrates meaningful empirical gains across tasks.
- 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 method-level novelty is limited. OT-based expansion is already central in DANCE, the teacher–assistant idea closely resembles Gentle Teaching Assistant, and cross-modality translation/cycle-style structure has clear precedents such as BABEL.
2.The paper argues that it prevents boundary collapse among similar cell types, but it mostly reports overall accuracy; confusion matrices, per-class metrics, or targeted analyses of rescued cell-type pairs would make that claim much stronger.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
CADENCE offers a practically motivated solution to extreme label scarcity in multi-omic transfer, demonstrating consistent empirical improvements over the DANCE baseline. However, the methodological novelty is limited and the claim regarding the prevention of boundary collapse lacks granular evidence.
- Reviewer confidence
Not confident (1)
- [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 clarifies my concerns regarding novelty. However, since it only provides CADENCE’s F1 score without comparisons to other methods, my experimental concerns remain. Therefore, I would maintain my score.
Author Feedback
We thank the reviewers and AC for their constructive feedback and for recognizing the importance of extreme label scarcity and CADENCE’s promising performance.We address the major concerns below. 1)@R1,R2,R3:Evaluation metric and visualization evidence.We used Accuracy as the primary metric because prior RNA to ATAC transfer studies,including DANCE (NeurIPS 2024) and scNCL (Bioinformatics 2023),mainly report Accuracy for their primary comparisons.To provide a more comprehensive evaluation,we additionally computed Macro-F1; on TASK4,CADENCE achieves a Macro-F1 of 58.27%,consistent with the reported Accuracy improvement.Our visualization further shows that,under extreme label scarcity,similar cell populations are easily confused,while CADENCE separates them more effectively.We will include the Macro-F1 and visualization results in the final paper. 2)@R1:Evaluation under different label ratios.Our work focuses on the most challenging regime of extreme label scarcity, so we emphasize the approximately 1% labeled-source setting.CADENCE is not limited to this setting:on Task 4,it achieves 57.26% Accuracy with 5% source labels and 62.68% Accuracy with 10% source labels, consistently outperforming DANCE and prior methods.We will include these results in the final paper to show its effectiveness beyond the extreme 1% case. 3)@R1, R2:Methodology,notation and missing definitions.We will go through the entire paper and add all missing explanations for mathematical notations, symbols, and hyperparameters. In particular,in Eq.3, $\arg\max \hat{y}(x)$ is intentional:Eq.2 normalizes the OT solution $Q^*$ into a soft pseudo-label $\hat{y} \in \Delta^C$, and Eq.3 maps this OT derived soft assignment to a hard class index for the source pseudo-label CE loss.Replacing it with $\arg\max p(x)$ would discard the global structural information introduced by OT and reduce source expansion to simple confidence-based pseudo-labeling.We will make this distinction explicit and define all previously unexplained quantities, including $\sigma$, $\delta$, $\gamma$, and the curriculum budget function $\rho(e)$. 4)@R1:Experimental setup and reproducibility.Our data preprocessing strictly follows DANCE,including the RNA/ATAC preprocessing pipeline such as TF-IDF transformation and scaling.In the final paper,we will provide a detailed description of the experimental setup to improve reproducibility. 5)@R2:Hyperparameter sensitivity.We conducted a sensitivity study on the three key loss weights by varying $\lambda_{SL}$ within $[0.05, 0.2]$, $\lambda_{align}$ within $[0, 0.2]$, and $\lambda_{ent}$ within $[0, 2.0]$,while keeping other settings fixed.These ranges cover weaker and stronger regularization around the default values and include zero-weight cases for the alignment and entropy terms.CADENCE maintains competitive performance across these settings, and the results will be added to the final paper. 6)@R2:Acronym and figure clarity.We will define CADENCE at its first occurrence as “Cycle-verified Assistant-driven DANCE with confusing pair corrEctioN” and revise Fig.1 by clarifying the workflow stages,spelling out “Gentle Teaching Assistant” (GTA),and replacing vague labels such as “Proto” with explicit annotations for prototype maintenance and confusing pair correction. 7)@R3:Novelty.We acknowledge that OT based source expansion follows DANCE. Different from GTA, our Assistant does not learn pseudo labels from all unlabeled samples.Instead,in the cross-modal RNA to ATAC setting,it specifically learns from ambiguous and low confidence target ATAC cells,preventing noisy target pseudo labels from directly affecting the student classifier.BABEL requires paired multi omic data to reconstruct the other modality,while CADENCE needs no paired data and focuses on target cell type annotation,using latent cycle consistency only as a reliability gate.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a semi-supervised framework for cross-modal single-cell annotation under extreme label scarcity. The reviewers’ scores are 2, 4, and 4, indicating mixed opinions. The reviewers agree that the problem is relevant and the proposed framework shows promising performance. However, concerns remain regarding clarity of presentation, completeness of experimental validation, and limited analysis supporting key claims. The AC therefore recommends inviting the paper for rebuttal.
The authors are encouraged to carefully address the reviewers’ comments in the rebuttal, particularly regarding clarification of the methodology, strengthening experimental analysis, and improving presentation quality.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes CADENCE, a semi-supervised framework for label-scarce cross-modal single-cell annotation, focusing on RNA-to-ATAC transfer with very limited labeled data. The method combines OT-based source expansion, teacher-assistant-student distillation, cycle-based reliability verification, and prototype-based confusing-pair correction to improve robustness under noisy pseudo-labels.
The initial reviews were mixed but slightly positive, with scores of 2 / 4 / 4.Reviewer 1 recommended rejection due to poor writing, unclear notation, limited evaluation beyond the 1% label setting, narrow metrics, and insufficient experimental details. Reviewers 2 and 3 were positive, appreciating the practical importance of extreme label scarcity, the empirical gains, and the confusing-pair correction idea, while noting concerns about clarity, sensitivity analysis, and moderate novelty relative to prior methods such as DANCE, GTA, and BABEL.
The rebuttal clarified several important points, including the formulation around Eq. 3, the role of the OT-derived pseudo-labels, missing notation, preprocessing assumptions, and the relationship to DANCE, GTA, and BABEL. These clarifications helped address several concerns about formulation, novelty framing, and reproducibility.
Some limitations remain. The manuscript still needs substantial improvement in clarity, notation, figure presentation, preprocessing/task descriptions, and reproducibility details. The novelty is moderate, and the submitted evaluation remains centered on the 1% label-scarce setting. The evidence for preventing boundary collapse among similar cell types should also be presented more clearly.
After rebuttal, the reviewer opinions are R / A / A. Reviewer 1 maintained reject, mainly due to clarity and reproducibility concerns. Reviewers 2 and 3 recommended accept after the rebuttal clarifications. I place greater weight on the two positive assessments, while recognizing that this remains a borderline paper.
I recommend Accept. The paper addresses an important label-scarce cross-modal single-cell annotation problem, and the proposed combination of reliability-aware teaching, cycle verification, and confusion correction is coherent and empirically useful. The initial review balance was already slightly positive, and the rebuttal clarified the main formulation and novelty-framing issues sufficiently to bring the paper above the acceptance threshold.
For the final version, the authors should substantially improve clarity, notation, reproducibility, and the presentation of existing per-class or confusion-pair evidence. They should also follow the MICCAI guidelines and avoid adding new experiments or new experimental settings that were not part of the original submission. Reporting additional metrics, such as Macro-F1, for the same experiments already included in the submitted manuscript is acceptable, but new label-ratio experiments or new sensitivity studies should not be added as new evidence in the final version unless explicitly permitted by the conference guidelines.
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.
This paper presents a well-motivated semi-supervised framework for cross-modal single-cell annotation under extreme label scarcity, demonstrating consistent empirical superiority. In the rebuttal, the authors effectively resolved the core concerns regarding mathematical notations, hyperparameter sensitivity, and evaluation metrics by introducing supplementary Macro-$F_1$ scores and multi-ratio validation. Consequently, despite residual reservations from one reviewer regarding presentation, the technical merits and thorough clarifications justify a recommendation 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.
Two reviewers’ concerns were solved after the rebuttal which also provided the 5% and 10% labeling experiments. After reading the paper, rebuttal and the reviews after the rebuttal, the AC recommends accepting this paper which should be revised based on the three reviewers’ feedback.