List of Papers Browse by Subject Areas Author List
Abstract
Accurate identification of cleavage stages is crucial for evaluating embryo development quality. However, due to occlusion and morphological confusion caused by continuous changes in the embryonic core region, it is challenging to autonomously learn discriminative features for different cleavage stages directly from image sequences alone. This challenge is particularly pronounced in classifying adjacent stages. This paper proposes a Multi-Image Frame Pair-Driven Cell-Division-Aware Category-Enhanced Contrastive Learning framework for Embryonic Cleavage Stage Classification (3CL-ECSC). The cell-division-aware module (CDA) takes concatenated image frame pairs as input, and outputs a binary indication of whether cell division occurred between them. Meanwhile, a category-enhanced contrastive learning classification (2CLC) module is constructed based on supervised contrastive learning, taking image sequences as input. The cell division predictions from the CDA serve as an explicit prior to perform consistency correction and confidence recalibration on the classification predictions. This process enhances discriminability at stage transitions and improves robustness to sparse stages. During training, we designed a two-stage cell-division-aware learning mechanism. This mechanism learns global cell division cues from multi-order adjacent stage frame pairs while utilizing 1st-order adjacent stage frame pairs to extract local cell division information. This refines stage boundaries and yields more stable, consistent predictions of the stage sequence. The proposed method showed improvements in Accuracy (+3.60%) and Transition Accuracy (+2.99%) for stage classification of human embryos, and gains of +2.46% in Accuracy and +17.86% in Transition Accuracy for mouse embryos, over state-of-the-art methods. Our implementation is publicly available at https://github.com/Yukun-Lee/3CL-ECSC.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3661_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Yukun-Lee/3CL-ECSC
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiYuk_CellDivisionAware_MICCAI2026,
author = { Li, Yukun AND Zhao, Yang AND Wang, Xu AND Fu, Longlong AND Wu, Jiahui AND Yang, Xuan AND Pei, Jihong},
title = { { Cell-Division-Aware Category-Enhanced Contrastive Learning Model for Embryonic Cleavage Stage Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper tackles the well-examined question of embryo stage classification through the novel lens contrastive learning. The method appears to out-perform other state-of-the-art methods.
- 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 contrastive learning approach is quite novel for embryo stage classification work.
- The evaluations against other methods is well-presented and gains convincing.
- Overall, the paper is easy to follow.
- 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.
- (Introduction) The authors mention the importance of abnormal cleavage patterns. However, the monotonicity constraint means that one such abnormal cleavage pattern - reverse cleavage - does not seem to be detectable in this framework. The authors should discuss (perhaps later in the paper) the limitations of imposing monotonicity.
- (Introduction) Another similar line of work would be approaches that explicitly count the cells in the embryo using segmentation (e. g. https://academic. oup. com/bioinformatics/article/41/4/btae617/7826612 or https://link. springer. com/chapter/10.1007/978-3-031-72083-3_8)
- (Methods) It would aid reproducibility to specify the set of augmentations used in the 1OCDA stage.
- (Experiments) While the dataset is publicly available, I do not believe that the authors have mentioned releasing their code anywhere. If this truly is the case, it would be good to see an explanation as to why (e. g. in a data availability statement)
- (Experiments) The authors mention each embryo sequence appearing in only one of the train, validation or test set. Was any additional stratification carried out (e. g. on patient, transfer status. . . )
- (Experiments) I think the paper would benefit from discussion specifically on robustness to abnormal cleavage patterns such as - in addition to the aforementioned reverse cleavage - direct uneven cleavage which can be a red flag in embryo development.
- (Experiments) The paper would benefit from details on the training hardware, memory footprint and runtimes.
- (General) There are a number of typos throughout the paper. For example ‘patterns-provide’ in the introduction should be ‘patterns provide’; ‘CDA-Enhanced Classification Heade’ should read ‘CDA-Enhanced Classification Head’.
- 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 contribution is reasonably novel and seems to outperform the state of the art. This really is more of a “strong weak-accept”. It’s a solid paper, but I think it needs to address the points raised in the weakness to move into hard “accept”.
- 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 addressed most of my points in their response and the paper will be good to go, pending actioning their responses. I think the abnormal cleavage patterns aren’t a fatal flaw, but the limitation must be acknolwedged.
Review #2
- Please describe the contribution of the paper
CDA looks through the embryo image sequence and detects when cell division happens. That division information is then used to help the classifier decide which embryo stage each frame belongs to.
Frame-by-frame predictions may be noisy and may jump in unrealistic ways. So the paper uses cell division timing as prior knowledge to correct those predictions. If the sequence of predicted stages does not match normal biological progression, the method adjusts the labels or their confidence. Use multi-order adjacent-frame information, so not looking at just the immediately next frame for short-term and broader temporal patterns.
- 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.
A novel application.
The method appears to offer moderate, task-specific innovation through the integration of cell-division-aware priors with sequence classification and correction.
The model strengthens its understanding of cell division in Global learning from multiple-frame pairs and Local learning from close frame pairs.
- 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.
Since most building blocks are established, the novelty seems more system-level and incremental than fundamentally methodological. Good ablation justify its usefulness, not sure novelty. Using cell-division-aware predictions to correct neighboring frame logits appears partly rule-based.
- 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 novelty seems more system-level and incremental than fundamentally methodological.
- Reviewer confidence
Somewhat confident (2)
- [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 do not see how the rebuttal addressed the weakness mentioned in the earlier review.
Review #3
- Please describe the contribution of the paper
This work proposed a framework for enhancing the performance of cleavage stage classification under environment noise and appearance similarity. To extract fine-grained features at single-cell scale, a cell division-aware module was utilised to counting the number of cells that were undergoing division process, which closely reflects the stage of embryo. Besides, category consistency was applied across sequences with multiple squeuences. Reinforcement learning was designed to boost the accuracy of 3CL-ECSC in cell division prediction. Experimental studies deliver the superiority of this framework in stage classification.
- 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.
Instead of classifying the embryo image into different stages directly, the authors proposed to predict the number of dividing cells, subsequently supporting the accuracy and robustness of the framework in noisy or indistinctive scenarios. The authors additionally improve the performance division prediction with advanced learning algorithms, e.g., distributive learning and reinforcement learning.
- 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 few comments are listed as follows for authors’ reference: 1.The organisation of this framework needs to be improved, especially regarding the usage of mathematical symbols. For example, what does “j-2” mean when “I” and “j” represents the “categories”. 2.What’s the “order” of adjacent stage frame, e.g., multi-order and 1st-order frame pairs; 3.For the identification of core regions, it is challenging to to extract the connected regions of the embryo with simple morphological operations, especially when the boundary is weak. 4.For the “CDA-enhanced Classification Heade”, it is verbose to express a simple classification task as two logits for “splitting” and “not splitting”. 5.While this work focuses on enhancing the performance of stage class-action with cell division-aware modules, how do the authors confirm the model does deliver the final decision based on the distribution of dividing cells.
- 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.
(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 methodology details for cell-aware modules.
- 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.
Thank the authors for addressing my concerns point-by-point. I think this work is interesting and ready for the presentation at MICCAI.
Author Feedback
We thank reviewers for their constructive feedback. We are encouraged that they find our model to be quite novel (R1), and can provide innovation for specific tasks (R2). They all agree on the effectiveness and promising results (R1, R2, R3). We address reviewer comments below.
Innovation(@R2, R3): Frames at stage boundaries show continuous, subtle local morphological changes, making single-frame stage identification highly challenging in both training and testing. In contrast, detecting cell cleavage between two frames is a simpler binary task (same vs. cross stage; ACC=95.96%, F1=95.98%), which better learns stable discriminative information with limited medical images and imbalanced categories for assisting stage recognition. Our core innovations include: 1) The 3CL-ECSC framework, which uses the proposed CDA to assist complex embryonic cleavage stage classification. 2) A novel stage category correction and confidence relabeling mechanism guided by CDA. The CDA module does not predict exact embryonic stages and only performs binary detection of cell division between frame pairs (@R3). Rather than serving merely as an auxiliary branch or an additional binary loss, CDA’s learned division probability calibrates per-frame stage predictions (@R2). Same-stage pairs are positively weighted to enhance temporal consistency, while cross-stage pairs receive negative weighting to suppress interference and highlight stage transitions. This embeds biological cell dynamics into both training and inference, turning basic binary division detection into a learnable stage transition prior for final classification. 3) Two-stage learning mechanism of CDA: MOCDA and 1OCDA. MOCDA constructs frame pairs across multiple stage intervals to learn global division cues and cross-stage transition patterns, while 1OCDA focuses on same-stage or adjacent-stage frame pairs to enhance boundary-sensitive discrimination. Regarding “order” (@R3), it represents the number of cleavage stage intervals between two frames in a frame pair. (The augmentations used in the 1OCDA: rotation, brightness and contrast adjustments, horizontal flipping, and color perturbation. (@R1))
Usage of CDA(@R2, R3): We thank R2 and R3 for their comments. CDA is not intended for cell segmentation or counting, but judges cell division events between embryonic frame pairs. The logits correction via CDA is not strictly rule-based post-processing, nor do we manually assign labels with fixed heuristics. Instead, its correction direction and strength are adaptively determined by CDA’s output, enabling biologically guided probabilistic recalibration. Incorporating biological priors effectively improves model interpretability and clinical credibility. Rather than overriding model inference with artificial rules, we embed biological priors into CDA-learned continuous weights to softly calibrate logits. Our method belongs to biologically guided probabilistic learning, instead of handcrafted rule-based decision making.
Monotonicity and abnormal patterns(@R1): The abnormal patterns considered in our setting mainly refer to prolonged intermediate-stage retention, unusually long odd-cell stages, or abnormal stage duration. These patterns still follow an overall progressive direction, so they do not conflict with the monotonicity constraint. We acknowledge the clinical importance of reverse and direct uneven cleavage (non-monotonic abnormal events), which will be addressed in future work.
Other details(@R1): We use two NVIDIA RTX 4090 GPUs(24GB). There is no additional stratification. We will add more related work in the introduction, and correct all spelling mistakes.
Other details(@R3): We will add detailed explanations for all subscripts in the formulas. CRE is a preprocessing step, and the extracted core region refers to the complete area containing all cells.
Code(@R2,R3): We will release the source code publicly upon paper acceptance.
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 reviews are mixed. All reviewers agree on the validity of the proposed method and the promising results. Below are key points to address:
- clarification on technical novelty
- address the monotonicity assumption and abnormal patterns
- evidence for the usage of cell-division information
- 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 addresses a meaningful embryology task and proposes a biologically motivated framework that uses cell-division-aware information and category-enhanced contrastive learning to improve cleavage-stage classification. The camera-ready should also include the clarifications promised in rebuttal.
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.
The paper proposes a cell-division-aware, category-enhanced contrastive learning framework for embryonic cleavage stage classification. The key idea is to use temporal cell-division information between embryo image frames to improve stage classification, especially in challenging cases where frame-level appearance is noisy, subtle, or ambiguous near stage boundaries. The method introduces a cell division-aware module, multi-order and first-order frame-pair learning, and category-enhanced correction/relabeling mechanisms to improve temporal consistency and stage discrimination.
The initial reviews were mixed but overall positive, with scores of 4 / 4 / 3.Reviewers appreciated the practical importance of embryo cleavage stage classification, the use of temporal cell-division cues, the task-specific contrastive learning formulation, and the reported improvements over prior methods. Reviewer 1 viewed the contribution as reasonably novel and close to a strong weak accept. Reviewer 2 also found the application and cell-division-aware prior interesting, though somewhat incremental. Reviewer 3 was more critical, mainly due to unclear methodological presentation and insufficient detail about the cell-aware components.
The main concerns raised before rebuttal were as follows. Reviewer 1 noted that the monotonicity constraint may not handle abnormal cleavage patterns such as reverse cleavage or direct uneven cleavage, and requested clearer discussion of related cell-counting/segmentation work, augmentation details, data/code availability, data split stratification, hardware/runtime, and robustness to abnormal cleavage patterns. Reviewer 2 questioned whether the novelty was mainly system-level and incremental rather than fundamentally methodological, and whether the correction of neighboring-frame logits was partly rule-based. Reviewer 3 raised concerns about unclear notation, the definition of frame-pair “order,” the reliability of core-region extraction when boundaries are weak, the verbosity and interpretation of the “splitting / not splitting” logits, and whether the model’s final decision was truly supported by cell-division information.
The rebuttal clarified several important points. The authors explained that the CDA module is not intended to perform cell counting, localization, or segmentation, and does not predict exact embryonic stages. Instead, it performs binary detection of cell-division events between frame pairs and uses the resulting probabilities as biologically guided transition priors to recalibrate stage predictions. They also clarified the roles of multi-order CDA and first-order CDA, the use of same-stage and cross-stage frame pairs, and the augmentations used in the first-order CDA stage. The rebuttal further clarified that the monotonicity assumption mainly concerns overall progressive cleavage-stage development, while acknowledging that reverse and direct uneven cleavage are important non-monotonic abnormal events that remain outside the current scope. The authors also provided hardware details, stated that there was no additional stratification, committed to releasing code upon acceptance, and promised to clarify notation, CRE, and related work in the final version.
Some limitations remain. The handling of truly non-monotonic abnormal cleavage patterns remains unresolved and should be explicitly acknowledged. The contribution is still somewhat task-specific and incremental, and several clarity/reproducibility issues must be fixed in the camera-ready version. In particular, the final paper should clearly explain the CDA module, the meaning of frame-pair order, the role of CRE, the distinction between binary division detection and stage classification, and the limitations of the monotonicity prior. The code release is also a useful commitment, but the final decision should be based primarily on the submitted manuscript and rebuttal clarifications rather than future promises.
After rebuttal, the reviewer opinions are A / R / A. Reviewer 1 moved from weak accept to accept, stating that most concerns were addressed and that abnormal cleavage patterns are not a fatal flaw if the limitation is acknowledged. Reviewer 2 moved from weak accept to reject, stating that the rebuttal did not adequately address the earlier concerns about the incremental/system-level nature of the method. Reviewer 3 moved from weak reject to accept, mainly because the authors addressed the concerns point-by-point and clarified that CDA was not a cell-counting or spatial localization module.
I recommend Accept. The paper is borderline but above the acceptance threshold. The problem is relevant, the use of cell-division-aware temporal priors is well aligned with the biological structure of embryonic cleavage stage classification, and the rebuttal resolved or mitigated several key misunderstandings about the CDA module. Although the novelty is moderate and several limitations remain, I view these as issues to be clarified in the final version rather than fatal flaws. The camera-ready version should clearly acknowledge the limitation regarding non-monotonic abnormal cleavage patterns, improve the method presentation and notation, clarify the role of CDA/CRE, and include the promised reproducibility details.
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. One reviewer who initially gave ‘weak accept‘ decision lowered the opinion to ‘reject’. 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.