List of Papers Browse by Subject Areas Author List
Abstract
Breast cancer imaging is multimodal but often incomplete: in routine care, patients may have CT, MRI, or digital mammography, and only a subset has all modalities. We study diagnosis from heterogeneous, partially observed imaging and propose ClinLoop, a human–AI workflow with two agents. A diagnosis agent predicts malignancy from the available modalities using modality-specific encoders and a missing-modality fusion module. A quality-assurance agent estimates uncertainty with an ensemble and defers low-confidence or discordant cases. Deferred cases are reviewed by a clinician, and the confirmed labels are fed back to update the model. This loop reduces manual review while keeping a safe fallback for uncertain and out-of-distribution inputs. We evaluate on a private cohort of 3108 patients with non-uniform modality availability (CT+MRI, CT-only, MRI-only, or mammography-only) and TNM staging. In five-fold patient-level cross-validation, ClinLoop improves AUROC and calibration compared with single-modality models and standard multimodal fusion baselines, with consistent gains across modality subgroups. ClinLoop supports practical deployment under missing data and limited expert time.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2960_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{XieHui_ClinLoop_MICCAI2026,
author = { Xie, Hui AND Wang, Ziyang AND Zhu, Sheng AND Wang, Zhiqiang AND Zhang, Zhuoneng AND Wang, Fangyijie AND Zhang, Tingfeng AND Ma, Jie AND Tan, Tao},
title = { { ClinLoop: Human-AI Collaborative Agents for Multimodal Breast Cancer Diagnosis Under Incomplete Clinical Data } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes ClinLoop, which is a human–AI framework for breast cancer diagnosis from incomplete multimodal imaging that integrates modality-specific learning with uncertainty-aware clinician feedback, improving performance and reliability across heterogeneous data settings.
- 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.
Overall, the idea presented in the paper is interesting and novel. The proposed ClinLoop framework addresses the problem of missing modality for breast cancer diagnosis through a collaborative AI–expert approach, and the reported performance is supportive of its effectiveness.
- 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) How is model performance affected by the experts’ level of expertise and years of experience? 2) Additionally, how many experts were involved in this study? Are multiple clinicians involved in annotating the same case, and if so, how are potential discrepancies between their assessments handled? 3) The paper does not provide information on whether the code will be released.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes ClinLoop, a novel and interesting approach that leverages human–AI collaboration to handle missing clinical information in breast cancer diagnosis. The experimental results support the effectiveness of the proposed method.
- 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.
Authors addressed my concerns.
Review #2
- Please describe the contribution of the paper
The paper proposes a simple two-agent human–AI system that can detect uncertain cases and send them to clinicians for review, and a flexible model that works even when some imaging types are missing, achieving better performance and reducing the workload for clinicians compared to existing 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 authors proposed ClinLoop, a human–AI collaborative workflow for multimodal breast cancer staging under real-world missing modalities. The idea is interesting and clinically relevant.
- 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.
Statistical analysis, including p-values, should be provided to support the comparisons between methods.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors proposed ClinLoop, a human–AI collaborative workflow for multimodal breast cancer staging under real-world missing modalities. The idea is interesting and clinically relevant.
- 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 #3
- Please describe the contribution of the paper
The paper proposes a ClinLoop, a human-AI collaborative agent framework for multimodal breast cancer diagnosis under scenarios of incomplete clinical data. Multimodal inputs include CT, MRI and mammography. They employ fusion methods to handle missing clinical data. The method consists of an annotator agent responsible for generating pseudo labels along with calibration and confidence weights, which are leveraged by the quality assurance agent for decision whether to defer to human review or not. The quality assurance agent measures uncertainty using an ensemble of K models and test time augmentation, selecting or deferring cases based on a non-conformity score. Deferred cases are reviewed a human fed back into the loop. The authors compare the performance of their method to off-the shelf LLMs and traditional classifiers like logistic regression and SVMs.
- 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.Realistic clinical setting: The paper leverages data that constitutes realistic clinical scenarios where clinical data is typically not available for all patients in the cohort. 2.Addresses model uncertainty: The paper proposes a method to address uncertainty of AI methods by explicitly calibrating and deferring to a human for review 3.Incremental/active learning: The paper provides a way to use human provided feedback to improve the subsequent performance of the system
- 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.Weak baselines: The paper assesses performance on single and multimodal models whose architecture or training scheme is not explicitly stated. The authors neither state if these are external baselines established in the literature. 2.Missing key methodological details: The paper does not provide much information of the various models used as components in the agentic framework. For instance, which encoders serve as the single and multimodal models. No details on how radiomic features are obtained for traditional classifiers. 3.No ablation of core contribution: The paper does not explicitly demonstrate which components of the framework contributes to the stated gains.
- 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
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 paper proposes an agentic framework with a human in the loop that explicitly accounts for model uncertainty. However, the experimental evaluation lacks rigorous due to insufficient baseline comparisons, absence of ablation studies of key components and limited methodological transparency. As a result it is difficult to attribute performance Gaines to the proposed contributions or assess their significance relative to existing approaches.
- 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 primary reason for the earlier weak reject recommendation was the lack of an ablation study evaluating which components were responsible for the reported performance gains. Without such analysis, it was difficult to clearly identify the paper’s specific contribution beyond incremental improvements achieved through combining existing methods. The authors have acknowledged this limitation in their rebuttal and indicated that they plan to include an ablation study in the revised version.
Author Feedback
We thank the reviewers (R1, R2, R3) and the Meta-Reviewer for recognizing the novelty and clinical relevance of the proposed ClinLoop framework. We are encouraged that R1 and R2 recommended acceptance, while R3’s weak-reject recommendation mainly concerns clearer experimental details and additional analysis. We are committed to releasing the code upon acceptance and reporting statistical analysis, including p-values, in the final paper.
R1: Clinician expertise and reviewer variability. The clinician-in-the-loop stage involved three breast imaging/oncology clinicians with 12, 8, and 5 years of experience, respectively. Disagreements were resolved by consensus; unresolved cases were assigned lower confidence tags rather than treated as high-confidence gold labels. As described in Sec. 3.4, all deferred cases and a budgeted subset of high-impact samples are reviewed, and confidence-tagged labels are fed back for iterative updates. Together with the ensemble disagreement check in Sec. 3.2, model-level uncertainty pre-filters ambiguous predictions for prioritized review, while confidence-weighted retraining down-weights uncertain judgments. This design mitigates both model uncertainty and inter-reader variability across different expertise levels.
R1/R2/R3: Code and reproducibility. We will release the training/evaluation code, configuration files, split-generation scripts, preprocessing instructions, and trained model recipes upon acceptance. The clinical images cannot be publicly released due to institutional governance, but we will provide sufficient implementation details to reproduce the method on comparable data.
R2: Statistical analysis. We agree that statistical testing should be reported. The manuscript uses 5-fold patient-level cross-validation with a held-out calibration split and reports discrimination, calibration, and selective-prediction metrics. We have now performed paired Wilcoxon signed-rank tests across the five folds comparing ClinLoop with the strongest non-ClinLoop baseline. The main improvements are statistically significant (p<0.01). We will add mean±std and p-values for Macro-F1/Joint-Acc, ECE, and selective risk.
R3: Baselines and methodological details. We agree that the baseline descriptions were too compressed. We will expand Secs. 4.1–4.3 to clarify that all deep baselines use the same patient-level folds, preprocessing, augmentation, optimizer, and training schedule. Single-modality baselines use 3D ResNet-50/MedicalNet for CT/MRI and EfficientNet-B4/ImageNet for DM. Missing-modality baselines include late fusion, MoE-style fusion, and modality dropout, all re-implemented under the same protocol. Radiomics baselines use an IBSI-compliant PyRadiomics pipeline with first-order, shape, and texture features (GLCM/GLRLM/GLSZM), ICC-based reproducibility filtering, LASSO selection, and 5-fold CV-tuned LR/SVM. LLM baselines use the same case inputs described in Sec. 4.3 and map outputs to TNM labels. For ClinLoop, CT/MRI use slice-based attention aggregation, DM uses a 2D CNN backbone, and multimodal fusion is conditioned on a modality mask to support arbitrary input subsets.
R3: Ablation and component contribution. We agree that a dedicated ablation should be made explicit. We will add component-wise ablations for mask-conditioned fusion, ensemble/TTA uncertainty, conformal QA, modality-consistency checking, calibrated pseudo-labeling, and closed-loop clinician feedback. Tables 2 and 3 already show the benefit of the full loop; the added ablation will make the source of the gains clearer.
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 novel human–AI collaborative agent framework for breast cancer diagnosis in the face of incomplete multimodal imaging data. Details that are missing include: statistical analysis (p-values) to support comparisons, ablation study to support which components are responsible for the reported performance gains, and the number of clinical experts involved or their expertise level.
- 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 have satisfactorily addressed the questions raised in the reviews.
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 addresses a clinically relevant problem with a novel human-AI collaborative framework. The authors have successfully addressed the concerns raised regarding statistical significance (e.g. p-values) and component contribution with an ablation study. All reviewers now recommend acceptance given the additions. The commitment to code release further strengthens reproducibility, making it a valuable contribution to the MICCAI community.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The overall reviewer rating is highly positive, highlighing the interesting and clinically relevant idea. Weaknesses such as statistical details were properly addressed in the rebuttal.
