List of Papers Browse by Subject Areas Author List
Abstract
Detecting vertebral metastases on CT in real-world practice is challenging because high-quality lesion masks are scarce, while vertebra-level labels mined from radiology reports are plentiful but noisy. We propose U-ReLSE, an uncertainty-regularized log-sum-exp pooling method for multiple instance learning (MIL) that converts voxel-wise lesion predictions into vertebra-level malignancy probabilities. The key idea is to let the model “trust” voxels it is confident about and “down-weight” voxels it finds ambiguous: we learn voxel-wise evidential (Dirichlet) certainty under segmentation supervision and use it as deterministic weights inside a mask-aware pooling operator, softly suppressing unreliable regions while preserving interpretable lesion localization and uncertainty maps. We train a 3D vertebra ROI model with a shared encoder and segmentation/evidential heads using a unified hybrid objective (Dice loss, evidential losses, and vertebra-level BCE) that integrates strong voxel-level supervision and weak report-derived labels. On multi-institutional in-house data with two cross-national external test cohorts, U-ReLSE consistently improves vertebra-level detection over standard pooling baselines, including Spine-Mets-CT-SEG (AUC 0.861 vs. 0.838–0.839; DeLong p<0.05) and a US-private cohort (AUC 0.907 vs. 0.871–0.903), with favorable recall–false-positive trade-offs. Because U-ReLSE only replaces the pooling layer, it can be readily applied to other strong–weak supervised detection problems under label noise and domain shift.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0632_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{KoiTat_UReLSE_MICCAI2026,
author = { Koike, Tatsuki AND Ichinose, Akimichi AND Kudo, Akira AND Kitamura, Yoshiro AND Takei, Mizuki AND Kobayashi, Kazuma AND Nakatani, Fumihiko AND Shibaki, Ryota AND Ito, Kimiteru AND Matsumoto, Hiromi AND Hori, Masatoshi AND Tomiyama, Noriyuki AND Miyake, Mototaka},
title = { { U-ReLSE: Voxel-Wise Evidential Uncertainty-Regularized MIL for Vertebral Metastasis Detection with Hybrid Supervision } },
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
This study proposes a hybrid supervised learning method utilizing noisy vertebra-level labels via Uncertainty-Regularized Log-Sum-Exp (U-ReLSE) pooling. This approach injects the obtained certainty into MIL pooling to suppress unreliable voxels, enabling noise-robust training. The model was trained using multi-center, single-nation data and evaluated on an external, multi-national, multi-center cohort. In particular, it effectively addressed various challenges encountered in real-world clinical 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.
This approach injects the obtained certainty into MIL pooling to suppress unreliable voxels, enabling noise-robust training. The model was trained using multi-center, single-nation data and evaluated on an external, multi-national, multi-center cohort. In particular, it effectively addressed various challenges encountered in real-world clinical settings.
- 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.
However, although external validation was conducted, its value is limited due to the lack of quantitative results and ablation studies.
- 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?
This study proposes a hybrid supervised learning method utilizing noisy vertebra-level labels via Uncertainty-Regularized Log-Sum-Exp (U-ReLSE) pooling. This approach injects the obtained certainty into MIL pooling to suppress unreliable voxels, enabling noise-robust training. The model was trained using multi-center, single-nation data and evaluated on an external, multi-national, multi-center cohort. In particular, it effectively addressed various challenges encountered in real-world clinical settings.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes a framework for vertebra-level metastasis detection from CT that combines heterogeneous supervision sources, including both voxel-level and vertebra-level labels of varying reliability (e.g., manually verified versus report-derived). The key methodological component is an uncertainty-aware aggregation mechanism, Uncertainty-Regularized Log-Sum-Exp (U-ReLSE) pooling, which incorporates evidential uncertainty estimates into a mask-restricted pooling operation over vertebral regions. This allows the model to aggregate voxel-level predictions into vertebra-level decisions while accounting for uncertainty. The approach is evaluated on multiple datasets, including external cross-national cohorts, and demonstrates improved performance compared to evaluated baselines.
- 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 vertebra-level metastasis detection from CT, which is of clear clinical importance. The formulation at the vertebra level aligns well with clinical workflows and reporting. *The framework combines voxel-level and vertebra-level annotations with different levels of reliability (e.g., manually verified versus report-derived labels). This is a realistic setting, and the ability to leverage such heterogeneous supervision is a strong aspect of the work. *The proposed U-ReLSE pooling integrates evidential uncertainty estimates into a log-sum-exp aggregation over anatomically constrained regions. While conceptually related to existing weighted pooling and attention-based MIL approaches, incorporating uncertainty to modulate the aggregation is a reasonable extension in the presence of noisy labels. *The method is evaluated on multiple datasets, including cross-national cohorts, which is a significant strength. External validation beyond a single in-house dataset increases confidence in generalization. The paper compares against multiple baselines and demonstrates improvements, supporting the proposed design.
- 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.
*Some of the main claims are stronger than what is directly supported by the experiments. The abstract states that the method stabilizes learning under label noise and domain shift while retaining interpretable localization and uncertainty cues. While the external validation does provide some evidence of cross-dataset generalization, I did not find direct evidence for improved robustness to label noise, nor a quantitative validation of the interpretability of the localization/uncertainty outputs. Thus, these aspects seem more suggested than demonstrated. *The distinction from the evaluated baselines is not explained clearly enough, and the comparison to nGWP pooling, Attention MIL pooling, and ReLSE would benefit from more intuition. At present, it is hard to to understand what specific limitation of each baseline is being addressed and what is contributed by U-ReLSE. A clearer comparison would also help the reader better judge the degree of novelty. *The supervision terminology is confusing. The paper appears to use “strong” and “weak” supervision primarily to refer to label reliability, rather than the more common usage based on annotation granularity such as voxel-level versus vertebra-level labels. Since the method also mixes supervision at different spatial levels, this terminology makes parts of the method (e.g. the loss design) harder to follow than necessary. Clearer definitions early in the paper would improve readability and reduce ambiguity. *The notation around the losses is not fully clear. In particular, the transition from vertebra-level labels to notation of the form y_{v,i,c} (voxel-level labels?) is not introduced. This matters because it affects the interpretation of how the evidential head is supervised and how the claimed uncertainty mechanism is learned. *The method may be sensitive to the quality of the vertebra segmentation used in preprocessing. Since pooling is restricted to voxels inside the vertebra mask, errors in this mask could directly affect the final vertebra-level predictions. This is particularly relevant in a paper that emphasizes generalization under domain shift, because robustness may then partly depend on the cross-domain robustness of the preprocessing segmentation model. I would prefer to see an analysis of this dependency. *Some design choices would benefit from additional justification. For example, constraining the learnable γ to a fixed range may be reasonable, but the paper does not make it fully clear why this interval was chosen and how sensitive performance is to it. Since γ controls the pooling behavior (mean-like vs max-like aggregation), this choice is not entirely trivial. *A few formulations are unnecessarily vague. For example, phrases such as “reliable validation and test sets” are not fully explained. It would be helpful to clarify whether “reliable” refers to annotation quality, curation, or simply the use of fixed splits. This clarity is important in a paper that discusses noisy supervision. *The external validation is a strength, but some wording around it could be toned down. Referring to the evaluation as “benchmark-style external validation” seems overstated when one of the cohorts is private, since this limits direct comparison by future work. The external validation itself is valuable, but the framing could be more toned down.
- 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?
The paper addresses a clinically relevant problem and proposes a well-motivated framework that combines heterogeneous supervision sources. The use of uncertainty-aware pooling to incorporate voxel-level evidential estimates into vertebra-level predictions is a reasonable design, and the method is evaluated on multiple datasets, including external cohorts. At the same time, some of the central claims, particularly regarding robustness to label noise and interpretability, are only partially supported by the experiments, and certain aspects of the method and notation could be clarified further. Overall, the paper presents a solid contribution with convincing empirical results, but with some limitations in presentation and clarity. These factors place it above the acceptance threshold, but not among the strongest submissions. Releasing the source code would strengthen the reproducibility of the work.
- 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 rebuttal clarified several methodological and presentation-related issues, including the positioning relative prior approaches. I remain somewhat concerned that my comments W1 and W8 were not addressed, i.e. “Some of the main claims are stronger than what is directly supported by the experiments…” and “(…) Referring to the evaluation as ““benchmark-style external validation”” seems overstated when one of the cohorts is private, (…) the framing could be more toned down.” My concern was not merely about wording, but about ensuring that central claims are aligned with the empirical support provided. I still find some of the framing around robustness, interpretability, and external validation stronger than what is demonstrated. Nevertheless, I acknowledge the limited rebuttal space and appreciate the authors’ engagement with the broader set of comments. I therefore retain my recommendation, but would strongly encourage the authors, should the paper be accepted, to align the framing of their claims with the supporting evidence in the final manuscript.
Review #3
- Please describe the contribution of the paper
The paper studies a practical supervision regime for vertebral metastasis detection, where voxel-level lesion masks are limited but report-derived vertebra-level labels are more abundant and noisy. Its main technical contribution is U-ReLSE, which uses voxel-wise evidential certainty learned from strongly labeled cases to improve vertebra-level aggregation. The external validation on two cohorts is also a meaningful strength.
- 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 problem setting is clinically relevant and practically motivated. 2) The method is technically coherent in how it transfers information from strong voxel supervision to weak vertebra-level prediction. 3) The paper includes external validation rather than relying only on an internal split. 4) The model retains lesion-level interpretability instead of reducing the task to a purely black-box vertebra classifier. 5) The gains appear reasonably consistent across datasets.
- 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 novelty is practically useful, though somewhat incremental, and reads more as a careful extension of existing ideas than as a fundamentally new MIL framework. 2) The MIL comparison is reasonable but still somewhat limited; comparison against a broader set of more recent MIL variants would help clarify the robustness of the gain. 3) Vertebra localization/indexing is understandably not the main focus of the paper, but the downstream pipeline depends on it, so a brief discussion of how indexing errors might affect supervision and evaluation would strengthen the manuscript. 4) The weak-label construction could be described more clearly, especially for negatives, unmentioned vertebrae, and ambiguous reports. 5) Table 3 is informative, but it is not a fully matched supervision-only ablation because supervision type is partially entangled with differences in training data composition. 6) An additional controlled comparison that derives weaker vertebra-level labels from the same strongly annotated cases could help isolate the effect of supervision type more clearly.
- 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.
(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 paper addresses a clinically meaningful problem and presents a technically coherent approach for a realistic hybrid-supervision setting. I also view the external validation on two cohorts as a notable strength, and the empirical improvements are generally consistent. My reservations are mainly about experimental clarity rather than the overall direction: the comparison in Table 2 is largely fair but not a strictly pooling-isolated ablation, Table 3 is informative but not fully disentangled from differences in training data composition, and the manuscript would benefit from clearer discussion of the upstream vertebra localization/indexing dependency and the report-label generation procedure. Overall, I find the paper promising and practically relevant, but at present I remain slightly below the acceptance line.
- 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 reasonably addressed my main concerns regarding weak-label construction, ablation interpretation, and the dependency on vertebra preprocessing. I still view the novelty as moderate, and Table 3 should be interpreted as a comparison of realistic annotation regimes rather than a fully disentangled supervision-only ablation. However, the paper addresses a clinically relevant problem, presents a coherent method, and provides external validation with generally consistent gains.
Author Feedback
We thank all reviewers and the AC. We address the main concerns: novelty, ablation, clarity, and preprocessing dependency. 1.Novelty over prior work (R2-W2, R3-W1). Our novelty lies in using voxel-level certainty, learned primarily from a small set of voxel masks, as MIL aggregation weights, enabling more robust incorporation of noisy vertebra-level weak labels. Prior uncertainty-aware MIL methods learn instance uncertainty only from bag-level weak supervision, where uncertainty may reflect both incorrect vertebra labels and visual ambiguity between normal and abnormal structures. In contrast, U-ReLSE uses an evidential head trained with voxel mask supervision to estimate voxel-level lesion certainty that directly weights vertebra-level aggregation. This prevents BCE gradients driven by noisy vertebra labels from reinforcing uncertain voxels as lesions, thereby mitigating overfitting to label noise. The improvement from Hybrid ReLSE to U-ReLSE in Table 3 (AUC 0.831->0.861) supports this design. Thus, U-ReLSE does not merely add an uncertainty component to MIL. It uses voxel-mask-supervised lesion certainty to gate the local evidence that contributes to vertebra-level weak label aggregation and its back-propagated gradients. 2.Ablation design (R3-W5/W6, R1-W1). R3 notes that ablations are “entangled.” However, comparisons of nGWP, Attention MIL, and U-ReLSE in Table 2, and of Hybrid ReLSE vs. U-ReLSE in Table 3, are pooling-isolated ablations: training data, supervision, and encoder are held constant while only pooling varies. Notably, Hybrid ReLSE and U-ReLSE share the same LSE formulation, differing solely in uncertainty weighting, directly showing that certainty-weighted aggregation improves robustness under noisy supervision. Regarding supervision-type isolation (R3-W6): Table 3 was designed to compare realistic annotation scenarios. We did examine the suggested setting as part of our ablation design: on Spine-Mets, AUC was 0.742 (Verified-Vertebra-only) and 0.821 (Verified+Report-Vertebra), both well below Hybrid U-ReLSE (0.861); this was omitted for space. We will clarify comparison levels in the final manuscript. R1’s concern about missing quantitative results and ablations is addressed by Tables 2 and 3, which provide quantitative external validation and ablations. 3.Clarity. Terminology (R2-W3/W7): Our use of “strong/weak” for label reliability and “voxel/vertebra” for granularity caused confusion. We will adopt three categories: (a) Verified-Voxel (multimodality-confirmed masks), (b) Verified-Vertebra (confirmed vertebra labels), (c) Report-Vertebra (noisy vertebra labels). “Reliable validation and test sets” refers to annotation quality; we will clarify this. Notation (R2-W4): y_{v,i,c} denotes voxel-level labels, defined only for Verified-Voxel data. L_edl and L_kd apply exclusively to these. γ range (R2-W6): γ is constrained to [4,16], interpolating between average and max pooling. This range was selected via validation, within which performance was stable. The learned γ converged to 8.17.Weak-label construction (R3-W4): Only vertebrae explicitly named in reports were labeled as weak positives. Unmentioned vertebrae were excluded from BCE (not labeled negative) to avoid false-negative risk. Negatives were defined as vertebrae without positive masks in mask-annotated cases. Ambiguous or non-vertebra-specific reports were excluded. We will clarify in the final manuscript. 4.Preprocessing dependency (R2-W5, R3-W3). Vertebra segmentation [13] is a fixed model shared by all methods and does not affect relative comparisons. ROIs include margins around centroids, and pooling masks use dilation to mitigate boundary errors. Evaluation-set vertebra indices were curated by radiologists. Consistent gains on two distributionally distinct external cohorts support generalization of the pipeline. 5.Ethics (R3). This study is IRB-approved. All data were obtained with opt-out consent and anonymization per ethical guidelines.
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 addresses a relevant clinical problem. The use of uncertainty-aware pooling and external validation are strengths. However, the novelty over prior work is not fully clear, and some key claims are not well supported. There are also concerns about clarity and aspects of the evaluation. These points need to be clarified in the 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.
The paper includes evaluation on external cohorts and addresses a practically meaningful application. Reviewers raised questions regarding novelty, evaluation setup, and presentation clarity, and the rebuttal clarified several of these points. Based on the reviewer discussion and updated feedback, the paper is above the acceptance threshold.
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.
R2 and R3 were satisfied with the rebuttals and revised their recommendations (R2: Weak Accept to Accept, R3: Weak Reject to Accept). However, both reviewers suggested further clarification on certain parts of the paper that should be addressed in the final camera-ready version.
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 rebuttal addressed most key concerns, clarifying methodology and experimental design. The strengths outweigh the weaknesses, supporting acceptance.
