List of Papers Browse by Subject Areas Author List
Abstract
Anatomical structure masks are widely adopted in radiotherapy dose prediction, as they provide explicit geometric constraints for structure-dose modeling.
However, conventional manual delineation of these masks requires precise annotation of radiotherapy-relevant anatomical boundaries, which is time-consuming and labor-intensive.
To reduce the dependence on dense structural masks, we propose ScribbleDose, a scribble-guided dose prediction framework that learns dose distributions from sparsely annotated anatomical structures.
Specifically, we design a Scribble Completion Module (SCM) to reconstruct dense anatomical priors by propagating sparse scribble labels to semantically correlated voxels.
During label propagation, a supervoxel-based regularization is introduced to preserve boundary consistency and improve anatomical plausibility.
Furthermore, we propose a Structure-Guided Dose Generation Module (SGDGM) to incorporate the reconstructed structural priors into dose prediction through mask-guided attention.
The completed masks derived from scribbles serve as intermediate structural guidance, forming an end-to-end scribble-mask-dose learning pipeline under sparse annotation.
Experiments on the GDP-HMM dataset demonstrate that ScribbleDose achieves competitive dose prediction performance using only scribble-level structural annotations.
The source code and reannotated scribble annotations are publicly available at https://github.com/iCherishxixixi/ScribbleDose.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1119_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/iCherishxixixi/ScribbleDose
Link to the Dataset(s)
GDP-HMM: https://github.com/RiqiangGao/GDP-HMM_AAPMChallenge/tree/main
Scribble annotation: https://zenodo.org/records/20110954
BibTex
@InProceedings{ZhaZhe_ScribbleDose_MICCAI2026,
author = { Zhang, Zhenxi AND Zhuang, Yitao AND Pu, Yao AND Yu, Peixin AND Li, Zirong AND Xia, Yan AND Li, Hui AND Li, Bin AND Zheng, Fuchen AND Ren, Ge},
title = { { ScribbleDose: Scribble-Guided Dose Prediction in Radiotherapy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of the paper is dose prediction framework that accepts only sparse scribbles of Planning Target Volume (PTV) and Organs at Risk (OARs), alongside CT images, as input. This approach reduces the labor-intensive process of annotating anatomical structures, which typically requires precise segmentations in traditional workflows. The framework consists of two primary stages: Scribble Completion Module, which reconstructs dense segmentations from sparse labels using a scribble-guided semantic clustering strategy for label propagation into the feature space. It further utilizes a super-voxel based boundary constraint to ensure anatomical consistency within the spatial domain. Second stage called Structure-Guided Dose Generation Module employes mask-guided attention mechanism to couple the reconstructed masks with the dose. A strength of this framework is that it allows for the joint optimization of structural reconstruction and dose generation. The proposed solution outperforms existing coarse-prompt methods and even masked-based C3D model across the evaluated metrics.
- 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.Novel Problem Formulation: The paper introduces novel approach by incorporating scribble-based inputs into the dose prediction workflow. This approach addresses the bottleneck of manual annotation. Furthermore, the joint optimization of structural reconstruction and dose generation allows the model to surpass the performance of traditional masked-based method. 2.Support for reproducibility: The authors facilitate the reproduction of their work by providing their code and the specific set of scribble annotations. Evaluating the framework on a large, open-access dataset further assists in the verification and benchmarking of the results. 3.Clear Presentation and Context: The paper is well-structured and clear. The literature review effectively situates the work by identifying current limitations in reducing manual segmentation effort and explaining how the proposed framework addresses these gaps. 4.Comprehensive Comparative Analysis: The evaluation includes a comparison against several prompt-based methods, including mask, box, point and prompt-less configurations. Additionally, the ablation provides useful insights into the specific impact of both the semantic and morphological guidance components.
- 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.Omission of Clinical Dose-Volume Metrics: While the inclusion of Dose-Volume Histograms (DVH) is informative, the paper lacks specific quantitative metrics for Organs at Risk (OARs), such as Dmax or Dmean. These metrics are important to verify plan safety and adherence to dose constraints. Additionally, the lack of Gamma Index analysis limits the evaluation of the spatial dose distribution feasibility. 2.Lack of Computational Efficiency Analysis: The paper does not provide data regarding the inference time of the proposed two-stage framework. Given that the approach involves both Scribble Completion Module and a Structure-Guided Dose Generation Module, it is important to understand how its efficiency compares to more traditional or prompt-less models to assess its practical utility in a clinical workflow.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
My recommendation is driven by the novel problem formulation, which offers tangible time-saving benefits for clinical annotation workflows. Additionally, the authors’ commitment to reproducibility was a significant factor in my positive assessment of this work.
- 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 #2
- Please describe the contribution of the paper
The main contribution of this paper is the proposal of an end-to-end, scribble-guided radiotherapy dose prediction framework. By seamlessly integrating a scribble completion module with a dose prediction network, this pipeline successfully achieves high-quality dose distribution predictions even under conditions of highly sparse structural annotation.
- 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 proposes a continuous end-to-end learning framework that allows the dose prediction loss to backpropagate directly to the upstream scribble completion module. This significantly enhances the deep coupling between structural priors and dose generation. 2.The authors introduce a dual regularization mechanism during the anatomical mask reconstruction process. By combining semantic distribution constraints in the feature space with supervoxel-based geometric constraints in the physical space, they establish highly effective orthogonal regularizations. 3.The authors provide and plan to open-source scribble level structural labels based on the large scale GDP-HMM dataset, offering crucial data support for future weakly supervised research in this domain.
- 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.While the authors claim the superiority of their end-to-end approach over two-stage pipelines, there is a lack of direct empirical evidence. The paper must include a two-stage scribble baseline (e.g., utilizing an off-the-shelf SOTA scribble segmentation model to generate dense masks, followed by a frozen dose prediction network). This is necessary to prove that the performance gain stems strictly from the end-to-end joint optimization rather than the difference in prompt modalities. 2.The semantic clustering mechanism described utilizes techniques common in metric learning and weakly supervised segmentation. The authors must clarify whether this module is entirely custom designed or adapted from existing literature. If borrowed, proper citations must be explicitly provided. 3.The paper lacks a robustness analysis on how the granularity and accuracy of supervoxel generation affect the final dose prediction. A discussion of failure cases in such challenging anatomies is recommended. 4.The initial scribbles used in the experiments are automatically generated and strictly confined within the boundaries of perfect masks. However, in real-world clinical practice, the manually drawn scribbles inevitably contain subjectivity. The paper does not evaluate or discuss the model’s tolerance to noisy or real-world scribbles, which weakens the argument for its immediate clinical applicability.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 makes a solid contribution to the field of radiotherapy dose prediction, and the proposed framework demonstrates impressive accuracy. However, the core innovation claimed by the authors is the superiority of an end-to-end pipeline over multi-stage methods, while it suffers from a significant logical gap in the experimental validation. My main consideration is the absence of two stage comparison utilizing the exact same scribble modality drastically diminishes the persuasiveness of their central claim. I hope the authors can thoroughly address and provide the missing baseline during the rebuttal phase.
- 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 main contribution of this paper is the proposal of ScribbleDose, a weakly-supervised radiotherapy dose prediction framework that relies solely on sparse scribble annotations, substantially reducing the clinical burden of dense mask delineation. The framework achieves this through two key innovations:A Scribble Completion Module (SCM) that reconstructs anatomically consistent dense masks from sparse scribbles using semantic feature propagation and supervoxel boundary constraints. A Structure-Guided Dose Generation Module (SGDGM) that explicitly couples the reconstructed structural priors with dose modeling via a mask-modulated attention mechanism. Together, these form an end-to-end “scribble-mask-dose” optimization pipeline that mutualizes structural reconstruction and dose generation.
- 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) Practical Clinical Motivation: The paper addresses a well-recognized bottleneck in radiotherapy planning. By utilizing sparse scribble annotations instead of traditional dense masks, the proposed method offers a pragmatic approach to significantly reduce manual contouring time, which holds good potential for clinical workflow integration.
(2) Unified Architectural Design: The integration of the Scribble Completion Module (SCM) and the Structure-Guided Dose Generation Module (SGDGM) provides a unified, end-to-end optimization pipeline. This design allows anatomical reconstruction and dose prediction to be trained jointly, which is a structural improvement over conventional decoupled two-stage foundation model pipelines.
(3) Competitive Empirical Results: The experimental validation on the GDP-HMM dataset demonstrates that the proposed framework achieves competitive dose prediction metrics (such as Dose Score and DVH Score). It effectively outperforms several coarse-prompt baselines (e.g., MedSAM2, SAMMed3D), verifying the feasibility of learning structure-dose relationships from sparse annotations.
- 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) Error Cascading and Lack of Uncertainty Modeling: The attention modulation map G in the SGDGM heavily depends on the pseudo-masks reconstructed by the SCM. Due to the lack of uncertainty estimation for these pseudo-labels, any over-segmentation or under-segmentation errors generated by the SCM in low-contrast soft tissues will be unconditionally amplified and forced into the downstream dose network via Equation (7). This poses a high risk of instability when dealing with complex clinical anatomical structures. (2) Misaligned Loss Formulation with Clinical Constraints: The ranking loss L_rank formulated in Equation (8) relies on the difference between the average predicted dose (D_OAR and D_PTV) within the PTV and OAR regions. However, in real-world radiotherapy planning, the dose limits for critical organs are typically constrained by the maximum dose (D_max) or specific volumetric doses (e.g., V_20). Merely constraining the average dose can easily mask localized, lethal “high-dose hotspots”, presenting a potential clinical risk. (3) Domain Gap in Simulated Scribbles: The scribble annotations in the dataset are algorithmically generated slice-wise from high-quality manual masks, ensuring they are strictly bounded inside the original contours. In contrast, real clinical scribbles are highly arbitrary (e.g., with positional shifts, or rough strokes on only a few key slices). The paper lacks robustness testing against non-contiguous slice scribbles or morphologically distorted scribbles, weakening its claims regarding practical clinical deployment. (4) Lack of a Fair Upper Bound in Baselines: In Table 1, ScribbleDose is compared against C3D, which uses dense mask inputs. However, C3D has a weaker backbone network compared to the Retentive Transformers (RMT) used in the proposed method, making it an unfair baseline. To rigorously quantify the actual performance degradation caused by using scribble annotations, the authors should include an experiment where Ground Truth (GT) Dense Masks are directly fed into the current RMT backbone to establish a proper performance upper bound. (5) Missing Fine-grained Ablations for SGDGM: Although Table 2 provides a thorough ablation for the components within the SCM, the SGDGM—the core module for integration—lacks independent validation. Readers cannot distinguish whether the performance gains originate from the inherent strength of the RMT architecture or specifically from the contributions of the attention modulation map G and the ranking loss L_rank. (6) Incremental Methodological Novelty and Missing Baselines: The algorithmic novelty of the Scribble Completion Module (SCM) is limited. Scribble-supervised segmentation is a mature field, with recent SOTA methods employing advanced pseudo-labeling techniques to address boundary adherence (e.g., enhanced foreground feature discrimination in EFFDNet [1], or knowledge distillation in ScribSD+ [2]). In contrast, the SCM relies merely on basic feature centroid clustering and traditional supervoxels. The authors fail to benchmark against these recent weakly-supervised methods. Consequently, the SCM appears more as a routine engineering adaptation for dose prediction rather than a fundamental methodological breakthrough. [1] Liu, Jinhua, et al. “EFFDNet: A Scribble-Supervised Medical Image Segmentation Method with Enhanced Foreground Feature Discrimination.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2025.[2] Qu, Y., Lu, T., Zhang, S., & Wang, G. (2024). ScribSD+: Scribble-supervised medical image segmentation based on simultaneous multi-scale knowledge distillation and class-wise contrastive regularization. Computerized Medical Imaging and Graphics, 116, 102416.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
My recommendation of a Weak Accept balances the clinical applicability and empirical results of ScribbleDose against its limited methodological novelty and specific experimental gaps. While the proposed method reduces manual contouring time by utilizing sparse scribbles, the Scribble Completion Module (SCM) lacks benchmarking against recent (2024/2025) weakly-supervised segmentation methods. Furthermore, the ranking loss (L_rank) uses average regional dose, which can mask localized high-dose hotspots (e.g., D_max in OARs), and the evaluation lacks an upper-bound baseline using ground-truth dense masks with the current RMT backbone. The paper sits marginally above the acceptance threshold due to its engineering utility, but the final score depends on the authors addressing the loss formulation and the missing baselines during the rebuttal.
- 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 sincerely thank all reviewers for their constructive comments and positive recognition of the clinical motivation, reproducibility, and empirical performance of our work. We carefully considered the concerns and have revised the manuscript accordingly within the MICCAI page limit.
First, regarding the concern about end-to-end optimization and two-stage baselines, we agree that a direct two-stage scribble-mask-dose comparison would provide additional evidence. Due to the strict page limit, we are unable to include a complete set of new experiments in the camera-ready version. Nevertheless, the current results already support the main claim of this work: sparse scribbles can provide effective structural guidance for dose prediction. In Table 1, ScribbleDose achieves the best performance among the evaluated sparse-prompt settings and remains comparable to the mask-based C3D baseline. We have therefore revised the wording to avoid over-claiming superiority over all possible two-stage pipelines. Specifically, we now state that separately optimizing structure extraction and dose prediction may limit joint adaptation, rather than definitively concluding that all two-stage designs are inferior.
Second, regarding the backbone fairness concern, we agree that an RMT model using full masks would likely provide a stronger upper bound. Our goal, however, is not to claim that scribble supervision can outperform dense-mask supervision under the same backbone, but to demonstrate that sparse scribbles can recover useful structural cues while reducing annotation burden. In addition, the ablation study shows that the proposed semantic and morphological guidance consistently improves the RMT-based scribble setting, suggesting that the gain is not solely due to the backbone. We have revised the result description from “surpasses the mask-based method” to “achieves comparable performance to the mask-based C3D baseline” to make the conclusion more precise.
Third, regarding the novelty of the Scribble Completion Module, we agree that semantic clustering and scribble-supervised segmentation are related to existing weakly supervised segmentation literature. We have added recent scribble-supervised medical segmentation studies and clarified that SCM is not intended as a standalone state-of-the-art scribble segmentation model. Instead, it is designed as a lightweight and differentiable structure completion component that can be jointly optimized with downstream dose prediction.
Fourth, regarding simulated scribbles, supervoxel sensitivity, and clinical applicability, we agree that real clinical scribbles may contain positional shifts, missing slices, and inconsistent stroke patterns. The current study uses simulated scribbles to provide a controlled and reproducible benchmark. We have added this point to the conclusion and explicitly describe the current work as an initial step toward scribble-guided dose prediction. We also added future work on real clinical scribbles, annotation-noise robustness, clinically interpretable dose objectives, and robustness of supervoxel-based priors under ambiguous anatomical boundaries.
Finally, we have improved presentation and reproducibility. The paper has been de-anonymized, the code and reannotated scribble annotations are now publicly available, recent related work has been added, and several over-strong claims, typos, figure labels, and table captions have been corrected. We believe these revisions better clarify the scope of the contribution and address the reviewers’ concerns within the available camera-ready space.
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 primary basis for the early accept recommendation is the novel problem formulation of using scribble-based guidance for radiotherapy dose prediction. By bypassing the bottleneck of labor-intensive pixel-wise annotations of OARs/PTVs, this approach offers significant clinical utility. This merit was consistently recognized by all three reviewers. To ensure a more comprehensive presentation, the authors are encouraged to use the camera-ready version to further clarify and discuss the reviewers’ concerns. Should the authors plan to extend this work for a journal submission, addressing the reviewers’ feedback regarding experimental design and evaluation in more depth will be essential.
