List of Papers Browse by Subject Areas Author List
Abstract
Voxel-wise 3D tumor segmentation is clinically essential yet prohibitively expensive to annotate at scale. In clinical workflows, radiologists can instead provide a single coarse 3D bounding box as a fast region-of-interest cue. However, such boxes are typically noisy, introducing background contamination near the inner boundary and tumor leakage beyond the box. Naive approaches that enforce hard box-as-mask constraints or use rigid pseudo-labels in self-training amplify confirmation bias and degrade boundary accuracy. To overcome this, we propose a safeguard frontier paradigm that learns 3D tumor segmentation from a noisy box by constructing a soft frontier supervision target. Specifically, we partitions 3D space into a far-field safe background and an inner/outer uncertainty band around the box. Within such band, we mine reliable foreground-core and safe-background seeds via entropy-based reliability, then generates a detached soft frontier target through reverse adversarial diffusion integrating the affinity-driven diffusion evolution with annealed prediction gate to jointly build the safeguard frontier between outward-diffusing foreground evidence and inward-diffusing background evidence. The model training combines a distance-decayed safe-background constraint with a adversarial diffusion consistency constraint. Importantly, the approach is plug-and-play with any 3D backbones. Experiments on MSD-Lung and MSD-Pancreas demonstrate that our method outperforms state-of-the-art (SOTA) box-supervised approaches by at least 8.0% in Dice score and exhibits strong robustness to controlled looseness and leakage perturbations, highlighting the practical value of scalable single-box annotation. Codes are at https://anonymous.4open.science/r/Noibox-0xAEE499A7.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1665_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{JiKai_NoiBox_MICCAI2026,
author = { Ji, Kailun AND Zhai, Shaoqing AND Lian, Cheng AND Su, Yixin AND Huang, Baoru AND Xiao, Feng AND Zhou, Quan},
title = { { NoiBox: Safeguard Frontier via Reverse Adversarial Diffusion for Noisy-Box-Supervised 3D Tumor Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a 3D tumor segmentation framework named NoiBox, addressing the widespread noisy bounding box supervision problem in medical imaging. To overcome the confirmation bias and boundary deterioration issues arising from directly treating bounding boxes as hard masks or using pseudo-labels, the authors introduce a safeguard frontier paradigm. This method first partitions the 3D space through a Box-induced Uncertainty Field Partition (BUFP) module, extracting high-confidence foreground and background seed points. Subsequently, a Reverse Adversarial Diffusion (RAD) module is employed to solve the Dirichlet minimization problem driven by affinity, generating a soft boundary target. This soft target effectively preserves the uncertainty of the boundary during training, avoiding irreversible error accumulation.
- 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 method proposed in this paper avoids the cost of dense voxel-level annotation, making it possible to train models using large-scale single-box annotation data. 2.Utilizing the RAD module to handle uncertainty bands represents a novel perspective. The adversarial mechanism, where foreground evidence diffuses outwardly and background evidence diffuses inwardly, ingeniously replaces traditional pseudo-label self-training, effectively mitigating the lack of global consistency in patch-based training.
- 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 experimental section only selected two datasets, making it difficult to verify the generalization of the method proposed in the paper. It is suggested to conduct experiments using additional datasets such as LIDC-IDRI and KiTS19, as in Reference 16.Meanwhile, the analysis of the experimental part is relatively simple and lacks reasonable explanations. 2.Figure 1 introduces the pipeline of the proposed method, but the process is difficult to understand. It is recommended to provide a detailed introduction in the method section. 3.The meanings of some formula variables and symbols in the article are not explained, such as formulas (3), (5), (7), and (8). Additionally, there is a lack of logical explanation and analysis for why formula (8) is expressed in the form of CspCse multiplication. 4.In Fig. 4, we can find that background evidence propagates inward, and can not find foreground evidence propagates outward. It is recommended to select instances where only part of the foreground is covered within the bounding box.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
This paper proposes a practical and high-performance solution for the 3D bounding box supervision problem with noise, which holds great potential for clinical translation. The idea of using bidirectional adversarial diffusion to generate soft labels is ingenious. However, the paper only selects two datasets for validation in the experimental section, making it difficult to verify the generalization performance of the model. Additionally, the paper does not provide a clear description of the methodological process. Based on its excellent experimental results, it is given Weak Accept. However, I am open to raising my score if the authors can thoroughly address these issues during the rebuttal phase.
- 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 authors are proposing a novel way of performing box-supervised 3D tumor segmentation. They propose using a safeguard frontier to construct soft targets of both foreground and background voxels that then propagate towards on another to enable soft boundary aware learning.
- 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.
- great concise summary of prior work in box-supervised segmentation
- great figure 1 - explains the two portions clearly
- Compared against a good number of other methods in the box segmentation space
- figure 3 is a good visualization of the process
- 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.
- I’m not sure that the evaluation on public datasets is necessarily a major contribution. The contribution is the methodology, and it’s great that you are using public data for evaluation but it is also expected that to support a contribution, one would evaluate it.
- Thorough comparisons were indeed made to other 3D box-supervised methods, but there are a lot of other 3D segmentation methods out there that are reporting / trying to achieve the same goal of tumor segmentation. Did you compare it at all to one of the point prompt based methods like Vista3D ? They are not box-supervised but have similar goals of getting fast low annotation segmentations.
- There is limited discussion in terms of impact / how it compares across the domain of tumor segmentation as a whole and no mention or discussion of limitations
Smaller comments:
- In the introduction, the first time you use “safeguard frontier” it would be great to explain what you mean by that. It is a term used repeatedly throughout the work, so it would be good to be clear on what exactly you mean by that.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
In a space that is developing so quickly I think it’s important to discuss / address the proposed contributions in context of where 3D segmentation will be going. Much of the work seems to be moving towards point or noun style prompts. 3D Tumor segmentation remains a difficult task and this work aims to improve low annotation segmentation as well as makes interesting contributions to computer vision work. Therefore it could still be a good contribution to the literature. However, without any comparison to state of the art methods that attempt to produce one or few shot segmentations it is difficult to see how relevant the work will be.
- 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 #3
- Please describe the contribution of the paper
The authors present a 3D tumor segmentation framework that takes coarse, noisy bounding boxes as input and learns a soft, probabilistic boundary, avoiding expensive voxel-wise annotation. The framework consists of two modules: (a) BUFP partitions the image into confident background, confident foreground, and an uncertainty band near the box edges. Each voxel is assigned a confidence score based on geometry and model-consistency, yielding trustworthy seed voxels (“anchors”) for foreground and background. (b) Starting from the seed voxels, RAD assigns each voxel a value between 0.0 (“background”) and 1.0 (“foreground”) via affinity-guided Dirichlet diffusion. Where both evidence streams meet, voxel value ~ 0.5, is considered the frontier. By avoiding hard class assignments, this yields a soft supervision target that preserves boundary ambiguity during training. Segmentation results on two Medical Segmentation Decathlon tasks (MSD-Lung and MSD-Pancreas) consistently outperform state-of-the-art box-supervised 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.
- Strong clinical motivation. Bounding box annotation is genuinely fast and scalable, and the paper’s focus on noisy boxes reflects real clinical practice.
- Conceptually elegant and well-motivated. The core idea of letting foreground and background evidence compete and meet at a natural boundary, rather than imposing hard constraints from a noisy box, is intuitive and well-grounded. It directly addresses the confirmation bias problem of pseudo-label self-training methods. Their soft frontier target is a principled way to avoid premature label commitment near uncertain regions.
- Works across different backbones. Demonstrating consistent performance gains across two different 3D backbones (SwinUNETR and UNETR) is a meaningful strength. It suggests the method is not tailored to a specific architecture and can be adopted broadly.
- Good ablations. The ablation in Table 2 clearly isolates the contribution of each component, and the qualitative visualization in Figure 3 effectively illustrates how the frontier evolves and tightens over training epochs.
- 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.
- Questionable generalizability of the affinity-based boundary mechanism. The affinity graph implicitly assumes tumor boundaries coincide with regions of high feature dissimilarity. This holds for the compact, circumscribed tumors in MSD-Lung and MSD-Pancreas, but is a meaningful limitation for infiltrative tumors (e.g., glioblastoma, peritoneal carcinomatosis) that spread gradually across tissue boundaries. The paper neither discusses this limitation nor evaluates on datasets with more infiltrative morphologies.
- Clarity of presentation. The ideas are communicated almost entirely through dense formalism. Key intuitions, particularly what the Dirichlet minimization computes and why it produces 0.5 at the frontier, are absent from the main text. A single intuitive paragraph per key equation would substantially improve readability without sacrificing rigor. Figure 1 is dense and hard to parse on first reading.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
The motivation of the paper is easing radiologist annotation burden. This is valid, but arguably too narrow. Any existing tumor segmentation model producing a rough prediction could serve as the noisy input, which means the method does not require a human-drawn box at all. This reframes their method as a refinement layer on top of existing segmentation pipelines, potentially enabling fully automatic iterative self-improvement.
- 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 core idea is novel and well-motivated, the results are consistent, and the clinical relevance is clear. The two main concerns, limited generalizability discussion and presentation quality, are addressable and do not undermine the contribution.
- 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
Author Feedback
1.Generalization and experimental analysis. (R#1) We have conducted additional experiments on LIDC-IDRI. Across two methods, our method improves the average Dice by 2.39 over the second-best baseline and reduces HD95 by 0.17, showing generalization beyond original datasets. Future updates will include dataset-specific observations, failure cases, and discussions on tumor size/boundary ambiguity. Results will be released on our anonymous GitHub.
2.Pipeline clarity. (R#1) We have added a clearer overview in the method section before introducing the formal equations to improve readability.
3.Formula variables and symbols. (R#1) We have re-examined and defined all variables and symbols upon first use. As for Eq. (8), (C_{sp}) and (C_{se}) are complementary constraints. Their multiplication acts as a conservative gate, where propagation is encouraged only when both spatial and semantic/evidence consistency are high. This reduces erroneous diffusion across ambiguous boundaries.
4.Foreground evidence propagation. (R#1) The evolution of the RAD in Figure 3 is used to verify the effectiveness of strategy selection. Foreground evidence indeed expands outward from reliable seeds toward the frontier. Due to MICCAI space limits, we will put the case with partial foreground coverage by the bounding box on our anonymous GitHub.
5.Contribution statement. (R#2) We agree that public-dataset evaluation should be viewed as validation. Our method demonstrates superior quantitative and qualitative performance against mainstream methods across different datasets. We have revised the contribution statement to focus on our methodological contributions including noisy 3D box-induced uncertainty field partition and safeguard frontier generation via reverse adversarial diffusion.
6.Comparison with Vista3D. (R#2) We have tested Vista3D with different backbone (SegResNetDS2 and Swin-UNETR). They achieve Dice scores 4.83 and 5.66 lower than our method. As our work focuses on box supervision, we plan to extend comparisons to more prompt-based methods in the future work.
7.Impact and limitations. (R#2) Compared to fully-supervised methods, our approach targets weakly-supervised tumor segmentation where only bounding boxes are available. It is not intended to outperform strong fully-supervised baselines, but offers a practical alternative when voxel masks are costly. Limitations (boundary ambiguity, infiltrative tumors, clinical verification) were omitted due to space limits, and will be evaluated in future work.
8.Definition of safeguard frontier. (R#2) We have defined “safeguard frontier” at its first occurrence in the introduction. Specifically, it refers to an estimated conservative boundary between reliable background evidence and uncertain tumor regions, designed to prevent background evidence from incorrectly propagating into tumor areas.
9.Infiltrative tumor limitation. (R#3) We agree that infiltrative tumors may violate the compact-boundary assumption. However, this limitation primarily affects local boundary smoothness, not overall segmentation accuracy. Reliable seeds affect the tumor segmentation. Our case analysis shows that Dice/HD95 remain robust to minor local deviations. We will explicitly discuss it with representative examples and update results on our anonymous GitHub.
10.Presentation and Dirichlet intuition. (R#3) We have added intuitive explanations after key equations, clarifying that Dirichlet minimization computes a smooth evidence field on the affinity graph with foreground/background seeds fixed to high/low values. The 0.5 level set marks the balance point between the two, motivating it as the safeguard frontier. Moreover, we have added a clearer overview in the method section to improve readability.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 reviewers provide consistently positive evaluations (all scores: 4, 4, 5), highlighting the strong methodological contribution, clear clinical motivation, and solid empirical performance of the proposed approach . In particular, the safeguard frontier paradigm and reverse adversarial diffusion mechanism are regarded as novel and well-motivated solutions to noisy box supervision, with consistent improvements demonstrated across datasets and backbones. While some concerns are raised regarding generalization, clarity of presentation, and broader contextual comparisons, these issues are relatively minor and can be addressed in revision. Based on the overall positive feedback and my own assessment, I recommend acceptance.