List of Papers Browse by Subject Areas Author List
Abstract
Accurate tumor segmentation in CT imaging is important for clinical diagnosis and treatment planning. However, achieving precise segmentation is still heavily challenging due to highly ambiguous boundaries. Also, conventional flat classification frameworks are not anatomically consistent, as they treat organs and tumors at the same semantic level, ignoring the fundamental anatomical prior that tumors physically reside within their corresponding organs. In this paper, we propose a novel Hierarchical Evidential Deep Learning framework (HiEDL) to accurately and efficiently segment tumors by explicitly enforcing this anatomical containment prior. Specifically, to restructure the segmentation task, we introduce a hierarchical organ-to-tumor decoupling strategy that divides the process into superordinate (organ) and subordinate (tumor) levels. Furthermore, we adopt evidential deep learning to isolate predictive tumor uncertainty and optimize the network using a novel hierarchical boundary loss combined with distance maps, directly guiding the model to resolve ambiguous boundary predictions. Experimental results demonstrate the superiority of HiEDL compared with other state-of-the-art 3D segmentation networks, by achieving the highest region-level (Dice) and boundary-level (NSD) scores on the MSD Pancreas and MSD Liver datasets. We further validate the effectiveness of the hierarchical formulation and the significant benefits of the proposed boundary loss through ablation studies.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6475_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/laibi-yonsei/HiEDL
Link to the Dataset(s)
MSD dataset : http://medicaldecathlon.com/
BibTex
@InProceedings{ParYou_HiEDL_MICCAI2026,
author = { Park, Younghyun AND Jeong, Jin Gyo AND Yang, Sejung},
title = { { HiEDL: Hierarchical Evidential Deep Learning for Uncertainty-Aware Tumor Segmentation from 3D CT via Boundary Regularization } },
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
The paper makes a methodological contribution to uncertainty aware medical image segmentation by proposing HiEDL, a hierarchical evidential deep learning framework for tumor segmentation from 3D CT. The central idea is that conventional flat segmentation formulations treat background, organ, and tumor as labels at the same semantic level, which ignores an important anatomical prior: tumors are physically contained within their corresponding organs. The paper’s main contribution is to reformulate the segmentation task so that this containment prior is built directly into training. Instead of solving a single flat three class problem, the method decomposes the task into a superordinate level, which separates background from organ tumor foreground, and a subordinate level, which distinguishes organ from tumor only inside the foreground region. This hierarchical organ to tumor decomposition is intended to make the optimization problem more anatomically consistent and to focus the model more directly on the difficult tumor boundary regions that are often responsible for segmentation failure.
A second important contribution is the integration of evidential deep learning into this hierarchical formulation. Rather than relying only on standard softmax probabilities, the model produces evidential outputs that define a Dirichlet distribution over class probabilities, allowing the method to estimate predictive uncertainty in addition to segmentation scores. The paper argues that this is particularly important in safety critical tumor segmentation settings, where organ tumor interfaces are ambiguous and conventional deterministic networks can become overconfident in precisely the regions where mistakes matter most. By combining hierarchical decomposition with evidential modeling, the method is designed not only to improve segmentation quality but also to localize uncertainty in a way that is more aligned with the underlying anatomy and with the actual failure modes of tumor segmentation.
A third major contribution is the introduction of a hierarchical boundary regularization strategy based on signed distance functions. The paper argues that region based losses alone are insufficiently sensitive to local boundary misalignment, especially at organ tumor interfaces. To address this, it introduces separate superordinate and subordinate boundary losses, each matched to the corresponding hierarchical subproblem. This gives the framework a more targeted way to constrain predictions at the surface level, rather than relying entirely on volumetric overlap terms. The method is therefore not simply a hierarchical relabeling of the segmentation task, but a structured combination of hierarchical supervision, evidential uncertainty modeling, and boundary aware optimization.
Finally, the paper contributes an empirical study on the MSD Pancreas and MSD Liver datasets showing that the proposed framework improves both region level accuracy and boundary quality relative to several strong 3D segmentation baselines, including UNETR, UNETR++, SegResNet, and Swin UNETR. The paper also includes ablation experiments indicating that the proposed boundary regularization is an important part of the performance gain, and qualitative uncertainty maps suggesting that the hierarchical formulation helps confine uncertainty to more relevant organ tumor boundary regions. Overall, the paper’s main contribution is best described as a hierarchically structured, uncertainty aware, boundary regularized segmentation framework that explicitly encodes anatomical containment for tumor segmentation from 3D CT.
- 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.
Clear and clinically meaningful problem formulation. The paper targets tumor segmentation in CT, where boundary ambiguity between the tumor and its surrounding organ is a real and clinically relevant challenge. The motivation is well articulated: inaccurate tumor delineation can affect diagnosis, treatment planning, and downstream intervention, and uncertainty is especially important in these safety critical regions. The paper also correctly identifies that flat segmentation formulations do not reflect the anatomical reality that tumors reside within organs. This gives the work a strong application-driven motivation rather than being only a technical variation. Interesting hierarchical reformulation of the segmentation task. A major strength is the explicit modeling of the anatomical containment prior through a hierarchical organ to tumor decomposition. Instead of treating background, organ, and tumor as equally independent semantic classes, the method separates the problem into a superordinate background versus foreground task and a subordinate organ versus tumor task within the foreground. This is an intuitive and clinically meaningful formulation, and it is more anatomically consistent than standard flat segmentation. The hierarchy is not merely descriptive; it is embedded directly into the loss design, which makes it a substantive methodological contribution. Principled integration of evidential uncertainty modeling. The use of evidential deep learning is a strength because it gives the model a direct mechanism for representing predictive uncertainty in ambiguous regions, rather than relying purely on deterministic softmax outputs. In the context of tumor segmentation, where organ tumor interfaces are often low contrast and morphologically variable, this is a well motivated design choice. The paper is also careful to connect this uncertainty modeling to the stated clinical problem of overconfidence in difficult regions, which makes the use of evidential learning feel purposeful rather than simply appended. Well motivated boundary regularization strategy. The paper makes a convincing case that region overlap losses alone may be insufficient to resolve subtle boundary errors. The proposed hierarchical boundary loss based on signed distance functions is therefore a meaningful strength of the paper. It directly targets boundary alignment at both the superordinate and subordinate levels and is integrated coherently with the hierarchical formulation. This is particularly relevant for tumor segmentation, where clinically important errors often occur at the surface rather than in bulk volumetric regions. Strong empirical results on two public datasets. The evaluation on both MSD Pancreas and MSD Liver strengthens the paper considerably. The method outperforms several competitive 3D segmentation baselines, including UNETR, UNETR++, SegResNet, and Swin UNETR, on both Dice and NSD. It is especially positive that the improvements are shown not only in region overlap but also in a boundary-sensitive metric, since that aligns directly with the paper’s central motivation. Using two public datasets also improves the credibility and relevance of the empirical results. Useful ablation study supporting the main design choices. The paper includes component analysis showing that the proposed boundary loss contributes meaningfully to the performance gains, and that region loss alone does not provide the same benefit. This is important because it demonstrates that the improvement is not just due to using a different architecture label, but is connected to the actual hierarchical boundary-aware design. The associated uncertainty map visualizations also help illustrate how the proposed method changes the uncertainty distribution in a way that is consistent with the intended anatomical behavior. Clear presentation and coherent paper organization. The method section is well structured and the figures are helpful, especially the diagrams explaining the hierarchical formulation and the uncertainty maps in the ablation study. The technical narrative is easy to follow from motivation, to method, to evaluation. This clarity makes the contribution easier to assess and strengthens the paper overall. Good balance between methodological contribution and application relevance. The paper is not only an application of a standard segmentation model to a medical dataset, nor is it a purely abstract methodological paper detached from clinical significance. Instead, it combines a clinically meaningful problem, an anatomically informed formulation, uncertainty-aware modeling, and practical evaluation on relevant public datasets. That balance is a notable strength and makes the work more compelling overall.
- 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 methodological novelty is meaningful but still somewhat incremental relative to the prior literature. The paper combines three ingredients that are each individually motivated by prior work: evidential deep learning for uncertainty estimation, hierarchical or anatomically structured segmentation ideas, and boundary loss based on signed distance functions. The paper’s value lies in how these are combined for tumor segmentation, but the submission does not fully clarify what part of the framework is genuinely new beyond this integration. In particular, the evidential formulation builds on prior evidential segmentation work such as Li et al. on region based evidential learning, the boundary loss follows the signed distance formulation of Kervadec et al. , and the idea of hierarchical segmentation guided by anatomical structure is also related to recent hierarchical segmentation frameworks such as SALT. The current presentation would be stronger if it more explicitly positioned the precise technical novelty of HiEDL relative to these existing directions. The uncertainty story is not evaluated as rigorously as the title and framing suggest. A central motivation of the paper is uncertainty aware segmentation, yet the empirical evaluation focuses mainly on Dice and NSD. The uncertainty discussion is supported primarily through qualitative uncertainty maps, but there is no quantitative evaluation of uncertainty quality, calibration, or failure detection behavior. Since the paper explicitly argues that evidential modeling is important for addressing overconfidence in safety critical settings, it would be helpful to assess whether the uncertainty estimates are actually better calibrated or more clinically useful than those of the baselines. Without this, the uncertainty component feels less thoroughly validated than the segmentation accuracy improvements. The baseline and ablation study could be stronger and more complete. The paper compares against several strong segmentation architectures, which is good, but the component analysis is relatively limited. In particular, the ablation table compares a baseline, a baseline with region loss, and the full boundary loss version, but it does not isolate all major ingredients as clearly as it could. For example, it would be useful to separate the effect of the hierarchical formulation itself from the evidential component and from the boundary regularization component. As currently presented, it is somewhat difficult to determine how much of the gain comes from the hierarchy, how much from evidential learning, and how much from the boundary term. A more complete ablation would strengthen the causal interpretation of the results. The clinical significance is plausible, but the paper remains primarily a benchmark style segmentation study rather than a strong clinical translation study. The application is clinically relevant, but the experiments are limited to public retrospective datasets and standard segmentation metrics. There is no evaluation of downstream clinical impact, no robustness analysis across acquisition settings, and no user facing or workflow oriented assessment showing how the uncertainty estimates would actually be used in practice. This does not invalidate the work, but it does mean that the clinical claims should remain modest. The paper is strongest as a method paper with medical motivation rather than as evidence of translational readiness. Some implementation and experimental details that would strengthen reproducibility and interpretation are missing or underemphasized. While the paper provides the main preprocessing, training, and loss settings, some details remain limited. For example, the exact architectural backbone used for HiEDL is not foregrounded as clearly as it could be, and the fairness of the comparison to all baselines would be easier to assess with more detail on whether their training configurations were fully harmonized or reproduced from reference implementations. Additional reporting of variance across runs or confidence intervals would also improve the rigor of the empirical comparison. As presented, the results are promising, but more experimental detail would make the conclusions more robust. The gains, while consistent, are moderate in magnitude, so statistical support would help. The method achieves the best Dice and NSD values on both datasets, which is a clear positive. However, the absolute improvements over the strongest baseline are not very large, especially on the pancreas task. In this setting, reporting standard deviations across runs, significance testing, or more extensive robustness experiments would help determine whether the observed improvements are consistently meaningful rather than being potentially sensitive to training variability. The paper would benefit from clearer discussion of limitations. The current conclusion emphasizes the benefits of the hierarchical formulation and uncertainty aware boundary loss, but there is limited discussion of where the method may still fail, such as extremely small tumors, poorly contrasted lesions, or cases where organ boundaries themselves are uncertain. A more explicit limitations discussion would make the paper more balanced and would help readers understand the practical scope of the contribution.
- 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
Thank you for the clear and well motivated submission. I think the paper addresses an important tumor segmentation problem and proposes a coherent combination of anatomical hierarchy, evidential uncertainty modeling, and boundary aware optimization. The use of both Dice and NSD is appropriate for the stated motivation, and the experiments on two public datasets strengthen the paper. My main suggestion for improvement would be to sharpen the positioning of the contribution relative to prior work in evidential segmentation, hierarchical segmentation, and boundary loss methods, since the novelty lies more in the integration and task specific design than in any single completely new component. I also think the paper would be stronger with a more complete ablation that more explicitly disentangles the effect of the hierarchical formulation, the evidential component, and the boundary regularization. Finally, because uncertainty awareness is a central theme of the paper, a more quantitative assessment of uncertainty quality or calibration would strengthen the overall argument. Overall, I found this to be a thoughtful and promising submission with a good balance between methodological development and application relevance.
- 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?
I rated this paper as Weak Accept because it presents a coherent and clinically meaningful method for tumor segmentation that combines an anatomically motivated hierarchical formulation, evidential uncertainty modeling, and boundary aware optimization. The paper is well written, technically clear, and evaluated on two public datasets, where it consistently outperforms several strong 3D segmentation baselines on both Dice and NSD. I also view the inclusion of ablation results as a positive aspect, since it provides at least some support for the claim that the proposed boundary regularization contributes materially to performance improvements. Overall, the paper offers a good balance between methodological contribution and medical relevance.
The main reason this is not a stronger accept is that the novelty is solid but not fully decisive. The submission combines several existing ideas, including evidential deep learning, hierarchical anatomical structuring, and signed distance based boundary loss, and the exact novelty of the full method relative to prior work could be articulated more sharply. In addition, although uncertainty awareness is central to the framing, the uncertainty component is validated mostly qualitatively through uncertainty maps rather than through a more quantitative analysis of calibration or failure detection. I also think the ablation study could be more complete in disentangling the roles of the hierarchy, evidential formulation, and boundary loss.
Even with these limitations, I find the paper to be above the acceptance threshold because the method is well motivated, anatomically sensible, empirically competitive, and likely to be of interest to readers working on uncertainty aware medical image segmentation. The contribution is not merely an incremental benchmark application, and the improvements are aligned with the stated clinical problem of resolving ambiguous tumor boundaries. For these reasons, I would support acceptance, while recognizing that the paper is closer to the margin than to the top tier of the accept range.
- 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 HiEDL, a hierarchical evidential deep learning framework for uncertainty-aware tumor segmentation. HiEDL explicitly integrates the anatomical hierarchy into the evidential learning process while adaptively regularizing ambiguous tumor boundaries.
- 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 an important and clinically relevant issue: uncertainty-aware tumor segmentation, which is crucial for reliable computer-aided diagnosis. The motivation of modeling uncertainty together with anatomical hierarchical structure is sound and meaningful for medical image analysis.
- 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.Clarity of baseline and necessity of loss components are unclear. Page 7, Table 2: The definition of “Baseline” is not clearly specified in the text. In addition, Table 2 shows that adding hierarchical region loss alone (Baseline w/ Region loss) leads to obvious performance degradation. The authors should clarify the necessity of retaining this loss term in the final objective function. Since the full HiEDL achieves the best performance while region loss alone is harmful, it is strongly suggested to evaluate a simplified variant that only uses hierarchical boundary loss without region loss. If such a simplified version yields comparable or even better performance, the current four-term loss function in Eq. 12 would raise serious concerns about over-parameterization and redundant design. 2.Inappropriate choice of comparative methods. The paper claims a novel framework for tumor segmentation, yet its core segmentation backbone is still a 3D UNet. However, the comparison in Table 1 mainly focuses on Transformer-based models, while the most fundamental and critical baseline vanilla 3D UNet is missing. This makes the experimental comparison logically inconsistent and difficult to interpret. Furthermore, the referenced competing methods are from 2018 and 2022, lacking comparisons with recent state-of-the-art approaches from the past three years. Up-to-date comparisons are essential to validate the practical advantages of HiEDL under the current technical state of the art. 3.Insufficient experimental support for the claimed contribution. From the results shown in Table 1 on the MSD liver and pancreas tumor segmentation tasks, the performance of HiEDL is considerably lower than that of widely recognized methods such as nnUNet reported in the literature. Although the paper proposes a hierarchical evidential learning framework with uncertainty awareness, the experimental results fail to justify the effectiveness and superiority of the method. The discrepancy between methodological motivation and empirical performance weakens the overall contribution significantly. 4.Eq. 12 involves multiple weighting hyperparameters, yet their specific values are not provided. Moreover, the authors do not analyze how these hyperparameters (especially λ) affect training dynamics and final performance. It is strongly recommended to visualize the evolution of λ during training and conduct sensitivity analysis to clarify whether each loss term brings positive or negative impacts to model learning. 5.By integrating evidential learning and uncertainty modeling on top of 3D UNet, the proposed method inevitably introduces additional computational complexity and inference overhead. The authors do not provide any analysis regarding model complexity, runtime, or memory consumption. A quantitative comparison of computational costs against baseline methods is necessary to justify whether the performance change (even if improved) is worth the extra computational burden.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Actually the idea to improve segmentation performance by evidence DL sounds good. However, the performance shown is not that impressive and the selected competitive methods are classic but out of date. Due to the above issues, even though the idea is promising, the method requires careful tuning and further validation before publication.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The main contribution of this paper is HiEDL, a Hierarchical Evidential Deep Learning framework for uncertainty-aware tumor segmentation in 3D CT. It decouples the three-class (background/organ/tumor) task into a hierarchical organ-to-tumor formulation at the loss level to explicitly enforce the anatomical containment prior (tumors reside inside organs). It further integrates evidential deep learning (Dirichlet-based uncertainty modeling) with a novel hierarchical boundary loss based on Signed Distance Functions (SDF), applied separately at superordinate (organ) and subordinate (tumor) levels, to regularize ambiguous boundaries while quantifying predictive uncertainty.
- 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.
1Explicitly incorporates a strong anatomical prior (hierarchical containment) that is often ignored in flat classification, leading to more anatomically consistent predictions. 2Combines evidential uncertainty modeling with a dedicated hierarchical boundary regularization (SDF-based), effectively reducing over-confidence at ambiguous organ-tumor interfaces. 3Achieves consistent state-of-the-art results on two challenging MSD benchmarks (Pancreas and Liver), outperforming strong baselines (UNETR, UNETR++, SegResNet, Swin UNETR) in both region-level (Dice) and boundary-level (NSD) metrics. 4Thorough ablation studies and uncertainty map visualizations clearly demonstrate the contribution of the hierarchical formulation and boundary loss. 5Practical focus on safety-critical tumor segmentation with uncertainty awareness.
- 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.
1The hierarchical decoupling is implemented only at the loss level rather than in the network architecture, limiting its ability to guide feature learning. 2Gains over the strongest baseline (Swin UNETR) are solid but relatively modest (Pancreas: Dice +0.0273, NSD +0.0095; Liver: Dice +0.0324, NSD +0.0459). 3No open-source code, pretrained models, or uncertainty evaluation for clinical decision-making (e.g., failure detection) is mentioned. 4Novelty is incremental: it builds on existing EDL and boundary-loss literature without introducing a fundamentally new network component or training paradigm.
- 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?
This is a well-motivated and technically sound methodological contribution that meaningfully addresses over-confidence and boundary ambiguity in 3D tumor segmentation by leveraging anatomical hierarchy and evidential learning. Results are consistent and superior on two public benchmarks, with clear ablations and visualizations. While the core ideas build incrementally on EDL and boundary-loss literature, the overall integration is effective and clinically relevant, making it suitable for acceptance at MICCAI.
- 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
Author Feedback
We thank the meta-reviewer and all reviewers for their constructive feedback and recognition of HiEDL.
[Novelty: R1.1/R3.4] HiEDL is not a re-use of EDL [14], SDF loss [12], and hierarchical labeling. (i) Existing EDL segmentation operates under a flat 3-class Dirichlet, forcing organ/tumor evidence to compete with anatomically irrelevant background; we re-parameterize evidence at two semantic levels via α_sup=(α0, α1+α2) and α_in=(α1, α2) restricted to the foreground (Eqs. 5-6) — to our knowledge the first hierarchical evidence aggregation in dense prediction. (ii) The SDF boundary loss [12] is defined for flat foreground labels; we re-derive it at both levels (Eq. 11) using p1+p2 against ϕ(y_sup), and the foreground-renormalized tumor score r_tum against ϕ(y2) with mask m_fg, aligning surface supervision with the hierarchical structure. (iii) Fig. 4 confirms this is not relabeling: standard EDL spreads tumor uncertainty across extra-organ regions, whereas HiEDL confines it to the organ-tumor interface.
[Baseline & validation: R2.2/R2.3] Our backbone is an advanced U-Net variant that inherently outperforms vanilla U-Net, so the vanilla U-Net comparison was omitted. We selected representative 3D segmentation networks spanning CNN- and Transformer-based designs and trained them under the same experiment settings.
[Ablation: R1.3/R2.1/R2.2] We acknowledge that the ablation could better disentangle hierarchy, evidential modeling, and boundary regularization. We clarify that the “3D U-Net” label in Fig. 2 denotes a generic 3D backbone, not the ablation baseline. In Table 2, the baseline is the Swin UNETR-based setting without the proposed hierarchical objectives, as reflected by identical Dice/NSD to Swin UNETR in Table 1.The region-only variant is the hierarchical region-based EDL objective, providing volumetric anatomical supervision under the containment structure. Its lower performance does not imply redundancy; rather, region-level EDL alone is insufficient for ambiguous tumor surfaces. The boundary term adds complementary surface-sensitive regularization. This synergy is reflected by gains over baseline: +2.7%/+3.2% Dice and +1%/+4.6% NSD on Pancreas/Liver.
[Loss-level design: R3.1] The hierarchy is intentionally loss-level, keeping HiEDL backbone- and pipeline-agnostic; gradients from both objectives propagate through the shared backbone and shape feature learning.
[Uncertainty: R1.2/R3.3] The manuscript evaluates Dice and NSD, but not calibration or failure-detection metrics. Our intent is to use evidential uncertainty to support boundary-aware segmentation by highlighting ambiguous organ-tumor interfaces, not to claim a clinical failure-detection system. Fig.4 qualitatively shows this behavior. Calibration/failure analyses will be included in future studies. We plan to release code after review.
[Hyperparameters: R2.4] Loss weights are in Sec. 3.1: w_sup=w_in=1, λ_sup=λ_in=0.2, t=20; the ramp prevents strong early boundary supervision while region predictions are unstable.
[Cost: R2.5] Although explicit benchmarking of runtime and memory overhead is not the primary focus of this work. HiEDL requires no Monte Carlo sampling, ensembling, or repeated inference; uncertainty is computed from Dirichlet evidence in one forward pass. Added cost mainly comes from voxel-wise operations and training-time losses. Future work will include cost analysis.
[Statistics & scope: R1.5/R1.6/R3.2] While the numerical gains may appear incremental, HiEDL demonstrates highly consistent improvements in Dice and NSD across two distinct datasets. Under strictly controlled conditions—sharing identical splits, preprocessing, and training—this consistent ranking firmly validates the robustness of our boundary alignment method.
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 hierarchical evidential deep learning framework for uncertainty-aware CT tumor segmentation, showing promising performance. However, the reviewers raised critical concerns regarding the novelty of the proposed method as well as the convincingness of experimental validations.
- 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.
Overall, the authors’ rebuttal could addressed most concerns. Incremental improvements over relatively limited competing methods hamper the strength of the proposed method.
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.
Reviewers have a partial agreement on the merit. One reviewer raised the major concern, I see a similar problem in this paper, and the limited novelty on the segmentation task. I recommend acceptance, but I think the paper is on the borderline of acceptance.
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 presents HiEDL, a hierarchical evidential deep learning framework for uncertainty-aware tumor segmentation in CT images enhanced by anatomical prior knowledge constraints. Two public datasets and four classic competing methods are involved in the experimental evaluation, and consistent empirical gains in boundary-sensitive metrics (NSD) are achieved. I agree with the reviewers’ concerns regarding the incremental technical novelty and the relatively dated competing methods (2018-2022). On the other hand, I also agree with the reviewers’ recognition regarding the underlying motivation and clinical relevance of this study. Uncertainty-aware segmentation remains a critical need in clinical practice. While dominant approaches mainly rely on computation extensive MonteCarlo dropout for uncertainty estimation, this work used a single-forward-pass way through evidential deep learning. Overall, I think the merits of this paper slightly outweigh its limitations and bring this paper to a relatively upper position in my cohort. Therefore, I recommend accept.