List of Papers Browse by Subject Areas Author List
Abstract
Multi-task learning is widely adopted for automated embryo grading. However the reliability of predicted confidence across heterogeneous grading components remains underexplored. We observed that blastocyst development stage, inner cell mass, and trophectoderm exhibit distinct uncertainty characteristics, and that overconfident predictions may become more evident when outputs are jointly evaluated. To improve robustness under heterogeneous uncertainty, we propose a confidence-aware soft-label learning framework that selectively refines supervision based on feature consistency and entropy stability. Unlike uniform regularization, the proposed approach preserves discriminative structure while reducing excessive confidence in unstable spaces. Experiments on 28,670 Day-5 embryo images from seven IVF clinics showed that our method achieves the highest overall mF1 (0.7218) while reducing overconfident error rates relative to baseline training. When combined with temperature scaling, calibration further improves without performance degradation. These findings highlight that uncertainty-aware supervision provides a practical strategy for enhancing reliability in multi-task medical AI systems.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2767_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{ParBog_ConfidenceAware_MICCAI2026,
author = { Park, Bogyu AND Jo, Yongwon AND Kang, Jiyeon AND Heo, Jin AND Kim, Hyung Min},
title = { { Confidence-Aware Supervision for Robust Multi-task Embryo Grading } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a multitask learning framework for embryo grading that incorporates a confidence-aware soft labelling scheme to improve model robustness under heterogeneous uncertainty. The framework leverages the internal feature representations of the neural network to determine the appropriate degree of label smoothing for each sample, assigning harder labels where the network is confident and softer labels where uncertainty is high. The decision of when to apply soft versus hard labels is governed by an entropy stability criterion, which monitors changes in the model’s predictive entropy across training to identify stable versus uncertain regions. The method is evaluated on a private dataset of day-five embryos, and the authors further demonstrate that temperature scaling further improves model calibration.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper is well-motivated and delivers on what it sets out to do. The multitask learning framework is studied in considerable detail, jointly learning across three tasks whilst remaining sensitive to the varying levels of confidence throughout training. The paper is clearly structured and well written.
The experimental evaluation is thorough and robust, examining performance across individual tasks as well as aggregated results, giving a comprehensive picture of the method’s behaviour. The discussion section is strong and demonstrates that the authors have a good understanding of their results and their implications.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The most significant limitation is the exclusive use of a private dataset, which severely limits reproducibility and generalisability. At a minimum, a publicly available dataset should have been included alongside the private one to allow for independent validation and fairer comparison with existing work.
The paper is unnecessarily long due to the inclusion of standard formulations; the full equations for temperature scaling, predictive entropy, and ECE are well established and need only be cited. Relatedly, the prominence given to temperature scaling is disproportionate to its contribution. That temperature scaling improves calibration across all methods is entirely expected behaviour; it need only be noted as a standard post-hoc calibration step.
Minor Points
- Standard ECE is used rather than KDE-ECE, which is the recommended calibration metric in the MICCAI community (Metrics reloaded: Recommendations for image analysis validation, 2022).
- ResNet-18 is a dated backbone; the authors should justify this choice.
- Several tables appear on incorrect pages.
- 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?
This is a well-motivated and clearly structured paper that delivers on its stated objectives, with thorough experiments and a strong discussion. However, the exclusive use of a private dataset is a significant concern that undermines reproducibility. The paper also overstates the contribution of temperature scaling and includes unnecessary formulations, resulting in a presentation that lacks focus.
The underlying method is promising, and these issues are addressable. With the addition of a public dataset evaluation and a more measured framing of contributions, this could be a strong submission.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
- This paper focuses on automated embryo grading in IVF by training on the Gardner classification system, which comprises three distinct metrics. The study addresses the challenges arising from multi-task learning when training on these metrics. It proposes a novel learning framework specifically designed to resolve the variability in uncertainty levels between these three metrics.
- While such reliability assessment is relatively uncommon in this field, the research introduces a comprehensive evaluation metric - joint task assessment that integrates all tasks - which more closely resembles clinical practice, complementing existing metrics like ECE (Expected Calibration Error) and OER (Observed Calibration Error).
- 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 construction of a large dataset comprising 28,670 images collected from seven clinics for training and evaluation
- The approach effectively addresses the biological distinction between stage-specific requirements (global structure) and ICM/TE-specific requirements (local structure), translating the empirical knowledge accumulated by embryologists into a technical solution through multi-task learning uncertainty separation
- 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.
- This study exclusively analyzes 2D static images captured at a single time point on the fifth day of culture, failing to leverage temporal developmental data or spatial embryonic information.
- While it employs ResNet18 as the backbone architecture, only the training methodology introduces novelty - the structural design itself shows no originality.
- No direct comparison is made with other multi-task learning (MTL) methods mentioned in the related works section.
- Although the separation between Stage and ICM/TE shows some improvement, there remains concern that the uncertainty separation between ICM and TE has not yet surpassed certain performance benchmarks.
- The paper’s explanations and presentation are reasonably clear, though it lacks an overarching figure that would provide better visual organization.
- 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
This research appears highly valuable. Notably, its evaluation approach focuses on reliability from a clinical perspective, and the constructed dataset holds exceptional value. I hope the authors will make their code and dataset publicly available - I believe this would have a profound impact on the industry.
- 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 problem formulation is particularly interesting, effectively reducing embryo grading issues to a reliability perspective. The dataset size and annotation quality are also exceptionally high. However, some weaknesses stand out, including poor overall paper organization and limited improvement in accuracy metrics.
- 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 clearly explain implementation details that were previously omitted due to space constraints in the paper (including 128-d projection, cluster purity calculation methods, K-means++ initialization, and the frequency of EMA updates), significantly addressing concerns about reproducibility. They also provide convincing and logical explanations regarding the rationale behind using ResNet18 as the backbone architecture (to isolate the effects of the architecture and purely evaluate the effectiveness of the proposed method), as well as the intended role of temperature scaling (TS).
Regarding the lack of direct experimental comparisons with other existing multi-task learning (MTL) methods, the authors argue that these methods have “orthogonal objectives,” and we can understand the theoretical basis for this argument. However, we cannot deny that having such benchmark comparison experiments would have strengthened the paper further. Additionally, while the dataset remains unpublished due to IRB and privacy restrictions—which is understandable given the medical domain—this does present certain limitations from a reproducibility perspective.
However, the significance of validating the approach using a large-scale multi-center dataset of 28,670 images, and the approach of enhancing credibility through “joint-task evaluation” that closely mirrors clinical practice, makes this a valuable contribution that “creates opportunities for discussion” in this field.
Due to the substantial improvement in technical clarity and resolution of questions through rebuttal, this paper is judged to fall within the weak acceptance range.
Review #3
- Please describe the contribution of the paper
The paper proposes a novel confidence-aware soft-labeling training framework to improve the reliability of multi-task learning in general, and IVF embryo grading in particular. The framework selectively refines labels for high-confidence samples based on feature consistency and cluster purity. Experimental evaluation on a multi-class 3-task embryo grading dataset shows that the proposed method achieved the highest mF1 score and reduced confidence miscalibration compared to the baseline.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper addresses a clinically relevant problem
- The method is an improvement over naive label smoothing, as it improves both confidence calibration and prediction accuracy
- Extensive calibration metrics are reported to analyze the confidence improvement
- 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.
- Increased training complexity by EMA and K-means purity clustering
- The method introduces several new hyperparameters that might have a sensitive effect on the training convergence
- The Methods section (§3) seems to be underspecified. I would not be able to faithfully reimplement the method with the provided information (see below)
- 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 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
- Purity computation: Are ground truth labels or model predictions used to calculate cluster purity?
- How is the 128-d projection of the ResNet features computed? Linear? MLP? PCA?
- How is the K-means clustering initialized?
- Some typographical inconsistencies in References section (inconsistent capitalization etc.)
- 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 presents a novel method that provides a clear advantage over simple label smoothing without sacrificing classification accuracy in MTL scenarios. However, the manuscript is missing some implementation details, which limits reproducibility.
- 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 clarified some of my concerns in their rebuttal. Without seeing the updated manuscript, it is difficult to assess the improved clarity and details in the methods description. Thus, my score of “weak accept” remains.
Author Feedback
We sincerely thank the reviewers for their constructive feedback and for recognizing the clinical relevance, reliability-oriented evaluation, and methodological contribution of our work.
A major concern raised by Reviewers #1 and #3 was reproducibility and missing implementation details. We agree that several aspects of the method can be clarified further. The 128-d feature projection is obtained using a learned linear layer from the 512-d ResNet backbone representation prior to clustering. Cluster purity is computed using ground-truth labels within each cluster, and K-means uses k-means++ initialization. Clustering is recomputed at each refinement stage using EMA features (decay=0.999). The entropy threshold was selected from τ ∈ {0.01, 0.02, 0.03}, and the number of refinement updates was searched within 3–7 based on validation performance. Due to space limitations, several implementation details were condensed in the current submission.
Regarding dataset generalizability (R1/R2), although the dataset cannot be publicly released due to IRB and patient privacy constraints, it consists of 28,670 embryos collected from seven IVF clinics across multiple years with patient-independent splits. We emphasize that this is a multi-center heterogeneous dataset rather than a single-site cohort. The goal of this work was to study confidence reliability under heterogeneous grading uncertainty in realistic clinical settings. Our focus was reliability-aware supervision under standard Day-5 static image settings, where confidence behavior becomes clinically meaningful when heterogeneous grading outputs are jointly considered rather than evaluated independently.
Reviewer #1 noted that temperature scaling (TS) was overemphasized. We agree that TS itself is a standard post-hoc calibration method and do not consider it a methodological contribution of this work. Rather, it was included to demonstrate that representation-level refinement and inference-time calibration act as complementary mechanisms. Accordingly, the primary objective of the proposed framework is not large classification accuracy gains, but reducing overconfident behavior while preserving discriminative performance under heterogeneous uncertainty.
Regarding the lack of comparison with other MTL approaches (R2), existing MTL balancing methods primarily focus on gradient or loss balancing across tasks, whereas our objective is different: improving confidence reliability under heterogeneous uncertainty. Our framework selectively refines supervision based on confidence stability rather than optimizing task weights. Therefore, the proposed method is orthogonal to conventional MTL balancing approaches.
For the choice of ResNet18 (R1/R2), we intentionally used a lightweight and well-established backbone to isolate the effect of the proposed confidence-aware supervision strategy from architectural improvements. Our goal was not to introduce a new architecture, but to demonstrate that the proposed framework consistently improves reliability behavior under a standard MTL setting.
Regarding method complexity and hyperparameter sensitivity (R3), we observed relatively stable behavior across the explored τ range and update counts. Since refinement is performed only a few times during training, the additional computational overhead remains modest.
Finally, we appreciate the suggestions regarding KDE-ECE, overall figure organization, formatting consistency, and clearer presentation of task-specific performance behaviors.
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.
Paper summary: This paper proposes a multitask learning framework for embryo grading that incorporates a confidence-aware soft labeling scheme to improve model robustness under heterogeneous uncertainty. The method is evaluated on a private dataset of day-five embryos, and the authors further demonstrate that temperature scaling improves model calibration.
Strengths: The reviewers identified several strengths, including:
- Well-motivated and clearly presented, with a principled confidence-aware multi-task learning framework.
- Clinically relevant and biologically grounded, effectively modeling global and local embryo characteristics. Large-scale multi-center dataset (28,670 images from seven clinics), enabling robust evaluation.
- Comprehensive experiments, including task-level, overall, and calibration analyses.
- Improved accuracy and confidence calibration, outperforming naive label smoothing approaches.
Weaknesses: The main concerns raised by reviewers include:
- Limited reproducibility and generalizability, as evaluation is conducted exclusively on a private dataset without validation on a public benchmark.
- Methodological novelty is moderate, with improvements stemming mainly from the training strategy rather than the backbone/model architecture, and lacking direct comparison with prior multi-task learning methods.
- Limited modeling scope, relying only on single-timepoint 2D images without leveraging temporal developmental dynamics or richer spatial information.
- Increased complexity and reproducibility concerns due to additional hyperparameters, EMA/K-means components, and insufficient methodological detail for faithful reimplementation.
- Presentation can be improved, including conciseness, clearer visual organization, and stronger justification of design choices (e.g., backbone selection, evaluation metrics).
Decision: The paper has clear strengths, but several substantive concerns remain unresolved, particularly regarding reproducibility and generalizability, comparison with prior multi-task learning methods, and the presentation of the paper. These points appear addressable through author clarification. I therefore recommend Invite for Rebuttal.
- 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.
Thank you for the detailed and constructive rebuttal. The responses addressed the main concerns raised during review and improved confidence in both the methodological clarity and experimental setup.
In particular, the authors clarified several previously missing implementation details related to feature projection, clustering, refinement strategy, hyperparameter selection, and training stability, substantially improving reproducibility. The rebuttal also clarified the motivation and scope of the work, improving confidence reliability under heterogeneous clinical uncertainty rather than maximizing classification performance or introducing a new architecture. The explanation of the multi-center dataset design and patient-independent evaluation further strengthened confidence in the clinical relevance of the study.
While some limitations remain, particularly regarding public data availability and broader comparisons with alternative MTL approaches, the current evidence supports the paper’s central claims and positions the contribution clearly within reliability-aware medical AI.
I recommend authors to incorporate the clarified implementation details and improve presentation and reproducibility in the camera-ready version.
My final recommendation is Accept.
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 confidence-aware supervision framework for multi-task embryo grading that aims to improve both predictive performance and confidence calibration under heterogeneous grading uncertainty. Reviewers highlighted the clinical relevance of the problem, large multi-center dataset, reliability-oriented evaluation, and comprehensive experimental analysis.
The primary concerns are related to reproducibility, implementation details, comparison with existing multi-task learning approaches, and dataset availability. The rebuttal provided substantial clarification regarding the methodological details, training procedure, clustering strategy, hyperparameter selection, and rationale behind several design choices. While limitations remain regarding public data availability and broader benchmarking, these issues are understandable given the clinical setting and do not outweigh the strengths of the work. Overall, the paper addresses an important problem, demonstrates consistent improvements in reliability and calibration, and provides a meaningful contribution to confidence-aware learning in medical AI.
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 paper aims to address an important problem with a well-motivated method and strong experimental validation. Strengthening the benchmark experiments and incorporating public dataset evaluations would significantly enhance the paper’s contribution.