pith. sign in

arxiv: 2605.11555 · v2 · pith:ZKEFVANTnew · submitted 2026-05-12 · 💻 cs.CV

ScribbleDose: Scribble-Guided Dose Prediction in Radiotherapy

Zhenxi Zhang , Yitao Zhuang , Yao Pu , Peixin Yu , Zirong Li , Yan Xia , Hui Li , Bin Li
show 2 more authors
Fuchen Zheng Ge Ren
This is my paper

Pith reviewed 2026-05-20 23:10 UTC · model grok-4.3

classification 💻 cs.CV
keywords scribble annotationdose predictionradiotherapyanatomical maskssparse supervisionmedical image analysisstructure-guided generation
0
0 comments X

The pith

Radiotherapy dose prediction maintains high accuracy using only sparse scribble annotations instead of full structure masks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a framework that predicts radiation doses for cancer treatment by starting from rough scribble marks on key anatomical structures rather than requiring complete boundary outlines. A completion step turns the scribbles into full masks while keeping important edges consistent, and these masks then guide a dose-generation network to focus high doses on targets and protect nearby organs. Experiments on an open dataset show that prediction quality stays strong even though the initial labeling effort drops sharply. This creates a more feasible path for using AI in radiotherapy planning when detailed annotations are hard to obtain.

Core claim

Sparse scribble labels on anatomical structures can be completed into dense, boundary-consistent masks that serve as effective conditioning input for dose prediction networks, enabling high-dose concentration inside target volumes while sparing organs-at-risk at performance levels comparable to fully annotated approaches but with substantially lower annotation cost.

What carries the argument

Scribble Completion Module that propagates sparse labels to dense masks via supervoxel regularization for geometric consistency, combined with Structure-Guided Dose Generation Module that conditions the prediction network on the completed masks.

If this is right

  • Annotation time for structure-guided dose prediction drops significantly while prediction quality holds.
  • The completed masks enforce the desired dose-structure coupling without needing exhaustive manual delineation.
  • Public release of code and reannotated scribbles allows direct replication and extension on the GDP-HMM dataset.
  • The framework offers a scalable route for dose prediction in settings where full annotations are impractical.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same scribble-to-dense-mask step could reduce labeling costs in other medical image tasks that rely on anatomical structure guidance.
  • If the completion module generalizes across scanners or patient populations, the method could be deployed with minimal additional fine-tuning.
  • Combining this approach with active learning might further lower the number of scribbles needed per case.

Load-bearing premise

The completed masks from scribbles have sufficiently accurate boundaries to provide reliable structural guidance without degrading the final dose predictions.

What would settle it

Direct comparison of dosimetric metrics or clinical plan quality between scribble-guided predictions and full-mask predictions on the same patient cases shows clear degradation.

Figures

Figures reproduced from arXiv: 2605.11555 by Bin Li, Fuchen Zheng, Ge Ren, Hui Li, Peixin Yu, Yan Xia, Yao Pu, Yitao Zhuang, Zhenxi Zhang, Zirong Li.

Figure 1
Figure 1. Figure 1: Comparison of existing dose prediction paradigms and the proposed scribble￾guided framework. (a) represents the conventional mask-based paradigm that relies on densely delineated structural masks. (b) and (c) aim to reduce annotation burden through CT-only modeling and two-stage foundation-model-based pipelines with box prompts, respectively. (d) presents the proposed unified scribble-guided framework. 1 I… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed scribble-guided dose prediction framework. The framework includes two main components: (a) the Scribble Completion Module (SCM) and (b) the Structure-Guided Dose Generation Module (SGDGM). 2 Method In this section, we present the proposed scribble-guided dose prediction frame￾work ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of predicted dose distributions. Top: DVH of PTVs and OARs. Solid lines represent the reference dose distributions, while dashed lines denote the predicted dose distributions. Bottom: The axial dose maps. achieves performance comparable to the mask-based C3D baseline, suggesting that sparse scribble guidance can recover effective structural cues for dose pre￾diction while substantial… view at source ↗
read the original abstract

Anatomical structure masks are widely adopted in radiotherapy dose prediction, as they provide explicit geometric constraints that facilitate structure-dose coupling. However, conventional manual delineation of these masks requires precise annotation of structure boundaries relevant to radiotherapy, which is time-consuming and labor-intensive. To address these limitations, we propose a scribble-guided dose prediction framework that relies solely on anatomical structures annotated with sparse scribbles. Specifically, we design a Scribble Completion Module (SCM) to generate dense anatomical masks by propagating sparse scribble labels to semantically similar voxels. During the propagation process, a supervoxel-based regularization is introduced to preserve geometric boundary consistency to ensure anatomical plausibility. Furthermore, we propose a Structure-Guided Dose Generation Module (SGDGM) to strengthen the correspondence between sparse structural cues and dose distribution. Herein, the completed dense masks derived from scribbles serve as structural guidance to condition the dose prediction network. This scribble-mask-dose consistency encourages high-dose concentration within target volumes while effectively sparing surrounding organs-at-risk. Extensive experiments on the open-source GDP-HMM dataset demonstrate that the proposed method maintains superior dose prediction performance while substantially reducing annotation cost, providing a practical paradigm for dose prediction under sparse structural annotation. The code and reannotated scribbles are made publicly available at https://github.com/iCherishxixixi/ScribbleDose.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces ScribbleDose, a framework for radiotherapy dose prediction that uses only sparse scribble annotations on anatomical structures instead of full manual delineations. It consists of a Scribble Completion Module (SCM) that propagates scribble labels to dense masks via supervoxel-based regularization to preserve boundary consistency, and a Structure-Guided Dose Generation Module (SGDGM) that conditions a dose prediction network on these completed masks to enforce structure-dose coupling. Experiments on the open GDP-HMM dataset claim that the method achieves superior dose prediction performance while substantially reducing annotation effort; code and reannotated scribbles are released publicly.

Significance. If the central performance claims hold under proper validation, the work could meaningfully lower the barrier to deploying AI dose prediction in clinical radiotherapy by replacing labor-intensive full-structure annotations with sparse scribbles. The public release of code and reannotated scribbles is a clear strength that supports reproducibility and further research on sparse-annotation regimes.

major comments (2)
  1. [§4 and §3.2] §4 (Experiments) and §3.2 (SCM): the central claim that SCM-generated dense masks provide reliable structural guidance equivalent to manual delineations rests on the untested assumption that supervoxel regularization produces anatomically plausible boundaries. No quantitative mask-quality metrics (Dice, Hausdorff distance, or boundary error) are reported for SCM outputs versus full annotations, nor is there an ablation measuring how completion inaccuracies propagate to downstream dose metrics such as DVH parameters or gamma pass rates on GDP-HMM.
  2. [Table 2 / Figure 4] Table 2 / Figure 4 (quantitative results): while superior performance over baselines is asserted, the manuscript does not include a direct comparison of dose prediction error when using SCM masks versus oracle full masks on the same scribble inputs. This leaves open whether the reported gains are driven by the guidance mechanism or by other architectural choices, weakening the attribution to reduced annotation cost.
minor comments (2)
  1. [Eq. (3)] Notation for the supervoxel regularization term in Eq. (3) is introduced without an explicit definition of the similarity kernel; a short appendix derivation would improve clarity.
  2. [§4.1] The GDP-HMM dataset description should explicitly state the number of patients, structures, and scribble density per case to allow readers to gauge the annotation reduction factor.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for the constructive and detailed feedback. The comments have prompted us to strengthen the validation of our intermediate results and clarify the contribution of the scribble-guided approach. We provide point-by-point responses below and have incorporated additional experiments and metrics in the revised manuscript.

read point-by-point responses

  1. Referee: [§4 and §3.2] §4 (Experiments) and §3.2 (SCM): the central claim that SCM-generated dense masks provide reliable structural guidance equivalent to manual delineations rests on the untested assumption that supervoxel regularization produces anatomically plausible boundaries. No quantitative mask-quality metrics (Dice, Hausdorff distance, or boundary error) are reported for SCM outputs versus full annotations, nor is there an ablation measuring how completion inaccuracies propagate to downstream dose metrics such as DVH parameters or gamma pass rates on GDP-HMM.

    Authors: We thank the referee for this observation. While the manuscript emphasizes end-to-end dose prediction performance, we agree that explicit quantification of SCM mask quality would better support the claim of anatomical plausibility. In the revised manuscript we report Dice similarity coefficients and Hausdorff distances between SCM outputs and the corresponding full ground-truth annotations on the GDP-HMM dataset. We have also added an ablation that measures the downstream impact of mask inaccuracies on DVH parameters and gamma pass rates, showing only modest degradation relative to perfect masks and confirming that supervoxel regularization preserves sufficient boundary fidelity for effective dose guidance. revision: yes

  2. Referee: [Table 2 / Figure 4] Table 2 / Figure 4 (quantitative results): while superior performance over baselines is asserted, the manuscript does not include a direct comparison of dose prediction error when using SCM masks versus oracle full masks on the same scribble inputs. This leaves open whether the reported gains are driven by the guidance mechanism or by other architectural choices, weakening the attribution to reduced annotation cost.

    Authors: We concur that an oracle comparison would more cleanly isolate the benefit of the SCM. In the revised version we therefore include a controlled experiment in which the SGDGM is conditioned on ground-truth full masks (derived from the identical scribble-annotated cases) versus the SCM-completed masks. The results demonstrate that performance with SCM masks approaches the oracle upper bound, indicating that the observed improvements stem primarily from the effective structural guidance obtained under sparse annotation rather than from unrelated architectural factors. revision: yes

Circularity Check

0 steps flagged

No significant circularity; supervised pipeline evaluated on held-out data

full rationale

The paper describes a standard supervised learning pipeline: scribble annotations are completed into dense masks via SCM (with supervoxel regularization for boundary consistency), then fed as conditioning input to SGDGM for dose prediction. All performance claims are empirical, measured against held-out GDP-HMM test cases rather than being algebraically forced by fitted parameters, self-referential definitions, or self-citation chains. No equations or derivations reduce the output to the input by construction; the method remains falsifiable via external dose metrics and mask quality ablations.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that sparse scribbles plus supervoxel regularization can recover masks whose geometric fidelity is adequate for dose prediction; no free parameters, axioms, or invented physical entities are described in the abstract.

axioms (1)
  • domain assumption Supervoxel-based regularization preserves geometric boundary consistency during label propagation
    Invoked in the description of the Scribble Completion Module to ensure anatomical plausibility.

pith-pipeline@v0.9.0 · 5794 in / 1157 out tokens · 26262 ms · 2026-05-20T23:10:30.091029+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    What is plan quality in radiotherapy? the importance of evaluating dose metrics, complexity, and robustness of treatment plans,

    V. Hernandez, C. R. Hansen, L. Widesott, A. Bäck, R. Canters, M. Fusella, J. Göt- stedt, D. Jurado-Bruggeman, N. Mukumoto, L. P. Kaplanet al., “What is plan quality in radiotherapy? the importance of evaluating dose metrics, complexity, and robustness of treatment plans,”Radiotherapy and Oncology, vol. 153, pp. 26– 33, 2020

    work page 2020

  2. [2]

    A practical approach to assess clinical planning tradeoffs in the design of individualized imrt treatment plans,

    R. Monshouwer, A. L. Hoffmann, M. Kunze-Busch, J. Bussink, J. H. Kaanders, and H. Huizenga, “A practical approach to assess clinical planning tradeoffs in the design of individualized imrt treatment plans,”Radiotherapy and Oncology, vol. 97, no. 3, pp. 561–566, 2010

    work page 2010

  3. [3]

    Automatic planning for functional lung avoidance radiotherapy based on function-guided beam angle selection and plan optimization,

    T.Xiong,G.Zeng,Z.Chen,Y.-H.Huang,B.Li,D.Zhou,X.Liu,Y.Sheng,G.Ren, Q. J. Wuet al., “Automatic planning for functional lung avoidance radiotherapy based on function-guided beam angle selection and plan optimization,”Physics in Medicine & Biology, vol. 69, no. 15, p. 155007, 2024

    work page 2024

  4. [4]

    A generalizable dose prediction model for automatic radiotherapy planning based on physics-informed priors and large-kernel convolutions,

    T. Xiong, G. Ren, Z. Chen, Y.-H. Huang, Z. Ma, Z. Li, Y. Sheng, Q. J. Wu, and J. Cai, “A generalizable dose prediction model for automatic radiotherapy planning based on physics-informed priors and large-kernel convolutions,”Medical Physics, vol. 53, no. 1, p. e70272, 2026

    work page 2026

  5. [5]

    A feasibility study for predicting optimal radiation therapy dose distributions of prostate cancer patientsfrompatientanatomyusingdeeplearning,

    D. Nguyen, T. Long, X. Jia, W. Lu, X. Gu, Z. Iqbal, and S. Jiang, “A feasibility study for predicting optimal radiation therapy dose distributions of prostate cancer patientsfrompatientanatomyusingdeeplearning,”Scientific Reports,vol.9,no.1, p. 1076, 2019

    work page 2019

  6. [6]

    Trdosepred: A deep learning dose prediction algorithm based on transformers for head and neck cancer radiotherapy,

    C. Hu, H. Wang, W. Zhang, Y. Xie, L. Jiao, and S. Cui, “Trdosepred: A deep learning dose prediction algorithm based on transformers for head and neck cancer radiotherapy,”Journal of Applied Clinical Medical Physics, vol. 24, no. 7, p. e13942, 2023

    work page 2023

  7. [7]

    A transformer- embeddedmulti-taskmodelfordosedistributionprediction,

    L. Wen, J. Xiao, S. Tan, X. Wu, J. Zhou, X. Peng, and Y. Wang, “A transformer- embeddedmulti-taskmodelfordosedistributionprediction,”International Journal of Neural Systems, vol. 33, no. 08, p. 2350043, 2023

    work page 2023

  8. [8]

    A novel mamba-based 3d dose prediction model with channel-aware scan for nasopharyngeal carcinoma radiotherapy,

    P. Zhou, Q. Peng, Y. Li, and C. Li, “A novel mamba-based 3d dose prediction model with channel-aware scan for nasopharyngeal carcinoma radiotherapy,”In- ternational Journal of Radiation Oncology, Biology, Physics, vol. 123, no. 1, p. e134, 2025. 10 Zhang et al

    work page 2025

  9. [9]

    Dosediff: Distance- aware diffusion model for dose prediction in radiotherapy,

    Y. Zhang, C. Li, L. Zhong, Z. Chen, W. Yang, and X. Wang, “Dosediff: Distance- aware diffusion model for dose prediction in radiotherapy,”IEEE Transactions on Medical Imaging, vol. 43, no. 10, pp. 3621–3633, 2024

    work page 2024

  10. [10]

    Currentadvancesinautomationinradiotherapy,

    J. Nijkamp, B. Knäusl, M. Aznar, D. Georg, D. Thorwarth, D. Thwaites, L. P. Muren,andU.A.vanderHeide,“Currentadvancesinautomationinradiotherapy,” Radiotherapy and Oncology, vol. 205, 2025

    work page 2025

  11. [11]

    Feasibility of ct- only 3d dose prediction for vmat prostate plans using deep learning,

    S. Willems, W. Crijns, E. Sterpin, K. Haustermans, and F. Maes, “Feasibility of ct- only 3d dose prediction for vmat prostate plans using deep learning,” inWorkshop on Artificial Intelligence in Radiation Therapy. Springer, 2019, pp. 10–17

    work page 2019

  12. [12]

    Transdose: Transformer-based radiotherapy dose prediction from ct images guided by super-pixel-level gcn classification,

    Z. Jiao, X. Peng, Y. Wang, J. Xiao, D. Nie, X. Wu, X. Wang, J. Zhou, and D. Shen, “Transdose: Transformer-based radiotherapy dose prediction from ct images guided by super-pixel-level gcn classification,”Medical Image Analysis, vol. 89, p. 102902, 2023

    work page 2023

  13. [13]

    Segment anything model (sam) for radiation oncology,

    L. Zhang, Z. Liu, L. Zhang, Z. Wu, X. Yu, J. Holmes, H. Feng, H. Dai, X. Li, Q. Liet al., “Segment anything model (sam) for radiation oncology,”arXiv preprint arXiv:2306.11730, 2023

    work page arXiv 2023

  14. [14]

    Generalizable and promptable artificial intelligence model to augment clinical delineation in radiation oncology,

    L. Zhang, Z. Liu, L. Zhang, Z. Wu, X. Yu, J. Holmes, H. Feng, H. Dai, X. Li, and Q. Li, “Generalizable and promptable artificial intelligence model to augment clinical delineation in radiation oncology,”Medical Physics, vol. 51, no. 3, pp. 2187– 2199, 2024

    work page 2024

  15. [15]

    Scribblesup: Scribble-supervised convolu- tional networks for semantic segmentation,

    D. Lin, J. Dai, J. Jia, K. He, and J. Sun, “Scribblesup: Scribble-supervised convolu- tional networks for semantic segmentation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 3159–3167

    work page 2016

  16. [16]

    Cyclemix: A holistic strategy for medical image seg- mentation from scribble supervision,

    K. Zhang and X. Zhuang, “Cyclemix: A holistic strategy for medical image seg- mentation from scribble supervision,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 11656–11665

    work page 2022

  17. [17]

    Rmt: Retentive networks meet vision transformers,

    Q. Fan, H. Huang, M. Chen, H. Liu, and R. He, “Rmt: Retentive networks meet vision transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5641–5651

    work page 2024

  18. [18]

    Automating high quality rt planning at scale,

    R. Gao, M. Diallo, H. Liu, A. Magliari, J. Sackett, W. Verbakel, S. Meyers, M. Zarepisheh, R. Mcbeth, S. Arberetet al., “Automating high quality rt planning at scale,”arXiv e-prints, pp. arXiv–2501, 2025

    work page 2025

  19. [19]

    Scribbles for all: Benchmarking scribble supervised segmentation across datasets,

    W. Boettcher, L. Hoyer, O. Unal, J. E. Lenssen, and B. Schiele, “Scribbles for all: Benchmarking scribble supervised segmentation across datasets,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 46002–46024

    work page 2024

  20. [20]

    Mednext: Transformer-driven scaling of convnets for medical image segmentation,

    S. Roy, G. Koehler, C. Ulrich, M. Baumgartner, J. Petersen, F. Isensee, P. F. Jaeger, and K. H. Maier-Hein, “Mednext: Transformer-driven scaling of convnets for medical image segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023, pp. 405–415

    work page 2023

  21. [21]

    Medsam2: Segment anything in 3d medical images and videos,

    J. Ma, Z. Yang, S. Kim, B. Chen, M. Baharoon, A. Fallahpour, R. Asakereh, H. Lyu, and B. Wang, “Medsam2: Segment anything in 3d medical images and videos,”arXiv preprint arXiv:2504.03600, 2025

    work page arXiv 2025

  22. [22]

    Sam-med3d:Avisionfoundationmodelforgeneral-purposesegmentationon volumetric medical images,

    H.Wang,S.Guo,J.Ye,Z.Deng,J.Cheng,T.Li,J.Chen,Y.Su,Z.Huang,Y.Shen et al.,“Sam-med3d:Avisionfoundationmodelforgeneral-purposesegmentationon volumetric medical images,”IEEE Transactions on Neural Networks and Learning Systems, 2025

    work page 2025

  23. [23]

    A cascade 3d u-net for dose prediction in radiotherapy,

    S. Liu, J. Zhang, T. Li, H. Yan, and J. Liu, “A cascade 3d u-net for dose prediction in radiotherapy,”Medical Physics, vol. 48, no. 9, pp. 5574–5582, 2021

    work page 2021