Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Adaptive radiotherapy (ART) for head-and-neck (H&N) cancer requires reliable longitudinal delineation of primary and nodal gross tumor volumes (GTVp/GTVn) to update treatment plans under evolving anatomy. Clinically, mid-treatment (midRT) contouring anchors to stable anatomical landmarks while interpreting deviations in tumor-bearing regions as treatment response; however, many learning-based methods either decouple pre-treatment (preRT) from midRT or impose overly global temporal consistency, thereby under-modeling the heterogeneous, target-dependent evolution of GTVp/GTVn. We propose RaD-Seg, a clinically grounded framework that mirrors the plan–verify–adapt workflow via three modules: (i) Region-Aware Dual-Stream Contrastive Learning uses region prototypes to disentangle stable anatomical anchors from response-sensitive features, maintaining spatial reliability and sensitivity to regression; (ii) Temporal-Difference Response Encoding encodes cumulative radiation effects as an explicit preRT–midRT residual code to modulate midRT representations, rather than conflating response with incidental appearance variation; and (iii) Plan-Guided Asymmetric Regularization anchors midRT semantics to the preRT planning definition in anatomically stable regions while permitting clinically plausible target evolution. Patient-level cross-validation on HNTS-MRG 2024 shows significant gains in midRT GTVp/GTVn delineation over strong baselines.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2420_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{KhoHee_RaDSeg_MICCAI2026,
        author = { Khor, Hee Guan AND Chen, Fang AND Wang, Yingni AND Ma, Longfei AND Liao, Hongen},
        title = { { RaD-Seg: Region-Aware Disentangled Learning for Plan-Guided Longitudinal Tumor Segmentation in Head-and-Neck Adaptive Radiotherapy } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16893},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    In this paper, the authors introduce a three-module approach called RaD-Seg for GTVp/GTVn longitudinal segmentation in head-and-neck adaptive radiotherapy treatment planning. These modules are inspired from the clinical paradigm of plan-verify-adapt and are validated on the HNTS-MRG 2024 dataset (N=150). In this study, RaD-Seg has been found to outperform five baseline approaches significantly. Although the clinical rationale behind the work is quite obvious and the ablation study is fairly thorough, the current version of this manuscript lacks in technical accuracy, experimental validation, and methodology.

  • 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.
    • One of the most valuable strengths of the paper is the association between specific technical modules and individual workflow phases of ART. Asymmetric focus on stable and regressing anatomy is highly justified.
    • Evaluating all modules and module pair combinations listed in Table 1 and conducting statistical significance tests is highly responsible research-wise. That each module works in addressing a different type of failure mode is an evident conclusion.
    • Although the Dice score improvements compared to the best existing solution (Dual Flow UNet) are relatively small (+0.20% for GTVp and +1.31% for GTVn), the Hausdorff distance reductions are impressive: GTVp from 6.45 to 4.25 mm, and GTVn from 9.81 to 7.55 mm. The boundary accuracy matters clinically much more than volume overlap in contouring tasks.
  • 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 paper fails to mention τ (contrastive temperature in Equation 2), m (margin in Equation 4), r (dilation radius in Equation 10), and any of the loss weighting factors in the compound objective function (Equation 13). These parameters are crucial and significantly influence the learning process and how the different losses interact with each other. Their absence makes the replication and fair comparison of the method impossible. As a multi-objective training is the primary contribution of the paper, this is a significant gap. There is no sensitivity analysis performed on these factors.
      • The anchor mask Ω_anc is derived from the ground-truth preRT label Y_pre using equation (10). It has been stated in the paper that “Inference requires only (I_pre, I_mid). “ This suggests that Y_pre cannot be obtained during inference. In the case where the anchor mask Ω_anc is defined using the predicted preRT label map, it should be stated clearly. This is because an error in estimating the anchor mask will affect the plan constraint. In addition, in the case where PGAR only performs during training but not at inference time, then this is very significant information that should be stated clearly, since it shows that plan-guided constraints do not incorporate boundary awareness at inference time.

    • Using a single encoder E for both preRT and midRT assumes the two timepoints live in a shared feature space, despite known appearance differences due to radiation-induced edema, tumor regression, and weight loss. The lack of an ablation study or comparison with a dual encoder approach leaves us wondering about their impact.

    • VoxelMorph is a registration algorithm and not a segmentation technique. It is listed in Table 2 without further description on how the midRT segmentations are created based on VoxelMorph (for example, deform the preRT label maps using the deformation fields computed). The comparison itself is not flawed. However, without this information, it is impossible to determine if the baseline is unfairly treated or disadvantaged.
  • 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 manuscript offers a highly clinically motivated approach with HD95 improvements compared to the best baseline being significant and relevant. In addition, the three proposed modules complement each other and exhibit strong ablative results. However, there are two reasons why I cannot confidently accept this work. First, all important hyperparameters such as contrastive temperature τ, margin m, dilation radius r, and composite loss coefficients are not specified, meaning that the work is not reproducible without them, and it remains unknown whether its performance is stable for different parameter values. Second, the role of PGAR during inference is unclear since Ω_anc is calculated based on Y_pre during training, while the authors claim that no labels are required during inference.

  • 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 present a clinically grounded framework that mirrors the plan–verify–adapt workflow in head-and-neck tumor segmentation targeting adaptive radiotherapy. They encode pre-RT MR and mid-RT MR images using a shared encoder into anatomy and response features, respectively, and enhance segmentation performance using three modules and corresponding losses:

    • L_ana: To enforce temporal consistency of background prototypes
    • L_resp: To encourage low temporal drift for background anchors and temporal signatures for target structures (i.e., GTVp and GTVn)
    • L_plan: To anchor midRT prediction masks to preRT prediction masks (i.e., plan semantics in stable anatomy) without forcing preRT features to explain midRT changes. The authors evaluate their method called RaD-Seg on the HNTS-MRG 2024 cohort (MICCAI Challenge 2024) for midRT segmentation of GTVp and GTVn, achieving statistically significant improvements over SOTA 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 is generally well written, and it provides convincing performance analyses.
    • The use of a public dataset (HNTS-MRG 2024) enables reproducibility (upon source code publication) and comparison with future approaches.
    • The authors thoroughly analyzed their proposed segmentation framework using a comprehensive ablation study.
    • Statistical tests show that all modules statistically contribute to superior performance over baseline and SOTA methods.
    • The presented approach constitutes an important approach to longitudinal tumor segmentation in head-and-neck adaptive radiotherapy
  • 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.
    • It is not entirely clear what your baseline model is in Table 1, it was never introduced
    • Implementation details are missing e.g. what type of image encoder do you use? It is not clear from the text and there is no code repository which gives more information about that
    • Caption of Figure 2: Top? Bottom? does not make sense!
    • Figure 2: “TDME”? you refer to your method in the text as “TDRE”
    • Figure 2: Why do you show a “Pre-RT” image here although the task is mid-RT segmentation? Please clarify.
    • Figure 2: Do all images actually show the same axial slice? I don’t think so. Please clarify.
    • Figure 2 t-SNE: it looks a bit overloaded, my understanding is that the task is to predict the midRT mask, why does it matter to also plot the preRT embeddings? If you want to plot them, I’d recommend plotting them sperate for better visualization and clarity.
    • Figure 3: again preRT? Acutally preRT or midRT? And again, does it really show the same axial slice across all figures? Please check.
    • For both Figure 2 and Figure 3: Why do you show both ground truth masks (GTVp and GTVn) but only GTVn for your model prediction? Please clarify.
    • How did you choose margin parameter m and radius r in the methods section? How does the choice of these parameters impact the learning process? Please clarify.
    • lack of discussion and analysis why proposed strategy is better than baseline and state-of-the-art methods. what are the hypothesis for each building block
    • why is the baseline method used already better than the standard nnUNet? Is the model bigger in terms of model parameters?
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (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 presents an innovative/novel approach with substantial empirical validation, clearly demonstrating superior performance in head-and-neck tumor segmentation targeting adaptive radiotherapy compared to existing approaches and has promising implications for clinical application. Inconsistencies mentioned in the weaknesses section NEED to be addressed 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 #3

  • Please describe the contribution of the paper

    The RaD-Seg work, contributes to the field by introducing three components that will help the segmentation of Head and Neck (H&N) Gross Tumor Volumes (GTVs). A region-aware component that disentangles anatomical features that will remain the same across two different image acquisitions, a temporal component that takes into account the tumor differences between the first radiotherapy acquisition and a middle one where the treatment already started. And, finally a component that anchors mid-treatment semantics to the pre-treatment planning definition through a teacher-student mechanism, while permitting clinically target evolution such as regression or deformation in the tumor-bearing areas.

  • 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 excels in its organization and clear statement of the novel components that it uses to improve the segmentation of the GTV. The logic and reasoning behing the architecture design are solid and the explanation of most of the components is understandable. Although, if more space was available, it would be easier to expand and understand. The paper is reproducible due to its use of a public dataset and easy to compare since the dataset belongs to a Grand Challenge. The wide variety of experiments in the ablation study and the comparison study allows us to understand the contribution of each change to the segmentation performance. And the statistical testing is adequate and useful to see the relevance of the architectural changes.

  • 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 things I would like to point out are related to the visibility of qualitative images that could be a little bigger to understand the changes between the different models and components. There is as well a lack of clinically useful discussion to see if the changes and achievements of this work are clinically relevant. At least the mentioning of the impact on dose distribution and so on could be useful.

  • 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.

    (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?

    I think this paper should be accept has it is very scientifically sound, presents novel contributions for the specificity of the automated segmentation field and opens the path for future work where a higher number of scans and volumes are available and need this kind of feature encoding.

  • 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

We thank the reviewers and meta-reviewer for their constructive assessment and for recognizing the clinical motivation, ablation design, and boundary improvements of RaD-Seg. We address the main concerns below.

-Reproducibility and hyperparameters. We agree that these details should be explicit. In our implementation, the contrastive temperature in Eq. (2) was τ=0.1, the response-margin in Eq. (4) was m=1.0, and the dilation radius in Eq. (10) was r=5 voxels after resampling to 1-mm isotropic spacing. The full objective in Eq. (13) used equal weights of 1.0 for all four terms, i.e., L=Lseg+Lana+Lresp+Lplan, with Lseg=CE+Dice. These values were fixed for all folds, ablations, and comparisons, with no test-set tuning. The choice of r corresponds to a 5-mm exclusion margin around preRT targets, avoiding excessive plan consistency near potentially regressing boundaries; m=1.0 encourages non-trivial response separation without over-enforcing temporal divergence. We will add these details, the backbone/channel configuration, training protocol, and the value for τ, m, r, and loss weights.

-Role of PGAR and use of labels. PGAR is a training-time regularizer. Ωanc is derived from the preRT planning label only during training to define reliable non-target regions and avoid using midRT labels for the plan constraint. At inference, neither Ypre, Ωanc, Ymid, nor the frozen plan teacher is required; the trained student directly predicts the midRT mask from (Ipre, Imid). Thus, the plan prior is distilled into the representation and decoder during optimization, while deployment follows the stated image-only input setting. We will explicitly state this distinction.

-Baseline and backbone clarification. The “Baseline” in Table 1 is not an external method; it is the same encoder-decoder backbone as RaD-Seg, trained with concatenated preRT–midRT input and only Lseg, without RDC, TDRE, or PGAR. Therefore, the ablation isolates the proposed modules rather than model size. Its stronger performance than nnU-Net mainly comes from paired longitudinal input and the same task-specific training setup, not from additional RaD-Seg losses. We will introduce this baseline clearly and report the exact encoder type.

-Shared encoder. We agree that a dual-encoder design is a meaningful alternative. We chose a shared encoder to maintain longitudinal feature comparability between preRT and midRT and avoid extra capacity that could confound the ablation. Since the scans are from the same MRI sequence and ART workflow, a shared feature basis is appropriate, while regression, edema, and deformation are explicitly modeled by the response stream and TDRE rather than forced to be invariant. We will clarify this rationale.

-VoxelMorph comparison. VoxelMorph was used as a registration-based contour-propagation baseline: the preRT image was registered to the midRT image, the deformation field was applied to the preRT GTVp/GTVn labels, and the warped labels were evaluated as midRT segmentations. We will add this description to clarify that VoxelMorph is not treated as a direct segmentation network.

-Figures and presentation. We will correct “TDME” to “TDRE” and clarify that preRT images are only longitudinal references, while predictions target midRT. Green denotes predicted GTVn to highlight false-positive nodal over-delineation near prior/preRT GTVp; RaD-Seg suppresses this implausible leakage. We will revise captions/legends, align slices, enlarge panels, and declutter t-SNE plots.

-Clinical interpretation. We will strengthen the discussion by emphasizing that the main clinical benefit is improved boundary fidelity, reflected by HD_95 reductions for both GTVp and GTVn. Although dose-distribution analysis is beyond this study, more accurate midRT boundaries are directly relevant to target coverage, margin adaptation, and sparing of nearby organs-at-risk.



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.

    Reviews were generally positive and acknowledge the work’s original contribution, however the authors should address reproducibility issues with hyper parameters in the final version.



back to top