Paper Info Reviews Meta-review Author Feedback Post-Rebuttal Meta-reviews

Authors

Zhenghao Feng, Lu Wen, Peng Wang, Binyu Yan, Xi Wu, Jiliu Zhou, Yan Wang

Abstract

Currently, deep learning (DL) has achieved the automatic prediction of dose distribution in radiotherapy planning, enhancing its efficiency and quality. However, existing methods suffer from the over-smoothing problem for their commonly used L_1 or L_2 loss with posterior average calculations. To alleviate this limitation, we innovatively introduce a diffusion-based dose prediction (DiffDP) model for predicting the radiotherapy dose distribution of cancer patients. Specifically, the DiffDP model contains a forward process and a reverse process. In the forward process, DiffDP gradually transforms dose distribution maps into Gaussian noise by adding small noise and trains a noise predictor to predict the noise added in each timestep. In the reverse process, it removes the noise from the original Gaussian noise in multiple steps with the well-trained noise predictor and finally outputs the predicted dose distribution map. To ensure the accuracy of the prediction, we further design a structure encoder to extract anatomical information from patient anatomy images and enable the noise predictor to be aware of the dose constraints within several essential organs, i.e., the planning target volume and organs at risk. Extensive experiments on an in-house dataset with 130 rectum cancer patients demonstrate the superiority of our method.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43987-2_19

SharedIt: https://rdcu.be/dnwJB

Link to the code repository

https://github.com/scufzh/DiffDP

Link to the dataset(s)

N/A


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper uses a diffusion-based dose prediction model to generate dose volumes which consider both the planning target volume and organs at risk.

  • Please list the main strengths of the paper; you should write about 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 prediction of dose volumes for planning is an area of high interest in radiation therapy (eg. for knowledge based planning). The novel method in this paper incorporates both the OARs (where we want to limit the dose) and the PTV (where we want to prescribed does).

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    There are no details of the type of radiation technique used (VMAT?) There are quite a few English errors (eg. automatical in the first line of abstract) There is no discussion of the results compared with previous work in this area. There could be more discussion how the dose volume is useful in a clinical workflow (eg. link to the linac/MLCs)

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

    I would question the lack of ethics and patient informed consent.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html

    In addition to the comments mentioned above:

    • Are the dose constraints to OARs and prescribed dose to the PTV consistent for all patients?
    • How is the dose volume going to be used in a clinical workflow?
    • Page 2: What is “fair performance”
    • Page 2:”influential” isn’t the right word - critical is better
    • How accurate is the structure encoder?
    • Please include details of the CT’s - especially slice thickness.
    • There are no details of ethics or informed consent.
    • How were the volumes resampled to 256x256?
  • Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

    6

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This is a very good paper, proposing a novel solution to a real clinical problem.

  • Reviewer confidence

    Very confident

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper studies the problem of dose distribution map prediction in the cancer treatment using diffusion models.

  • Please list the main strengths of the paper; you should write about 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 problem itself is very interesting, and and the paper is well written.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    My biggest concerns about this paper is the clarity and the motivation.

    • The prediction of dose distribution is definitely a very interesting problem. However, the predicted dose distribution can be hard to achieve in practice, or even not possible due to the properties of radiation source and the clinic practice. Such practical constraints make the motivations insufficient for the proposed research problem

    • The definitions of some of the notations in the derivation of diffusion models principles and the metrics used to evaluate performance of dose distribution prediction are not clearly given (see section 9 for more info). This makes the submission less readable.

  • Please rate the clarity and organization of this paper

    Excellent

  • Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

    The authors promised to make the research implementation public

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
    • When the authos proposed to predict the dose distribution maps, how can we guarantee such dose distribution maps are practically acheivable? The prediction from the models may be excellent, however, it will be meaningless if it is not practically achievable.

    • How eq(4)/eq(3) is derived? It dose not look very obvious to me how to derive it from eq(1), eq(2), and the definition of $\gamma_t$. The authors are suggested to add the rigorous derivation in supplemental materials.

    • Are the $\alpha_t, \gamma_t$ learnable parameters or non-learnable hyperparameters?

    • How eq(7) is derived? It dose not look very obvious to me how to derive it from setup given in this paper. The authors are suggested to add the rigorous derivation in supplemental materials.

    • During the training via eq(9), will the pretrained encoder also be updated?

    • The authors should define the evaluation metrics they used such as $D_{98}, D_2$?

    • In figure 3, can the authors plot the zoom-in version with PTV Gy in [40,55]? For the vertical axis, does “1.0” means 1\% or 100\%?

  • Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

    4

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Interesting problem, interesting approach, but the motivations are not fully justified

  • Reviewer confidence

    Confident but not absolutely certain

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper proposes incorporating diffusion models to predict dose distribution in radiotherapy planning for rectal cancer, emphasizing the over-smoothing problem. With the implementation of the structure encoder, the proposed method achieved quantitative and qualitative improvement in dose prediction compared to the existing methods. It was applied to the dataset of 130 rectal cancer patients, indicating its usefulness as a real-world application for radiotherapy.

  • Please list the main strengths of the paper; you should write about 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 main contribution of the paper seems to be a novel application of diffusion models for dose prediction in radiotherapy planning.
    • Rather than simply applying the existing diffusion models, the paper also aims to incorporate the original module, the structure encoder, and demonstrates its effectiveness experimentally (Table 2).
    • The extensive comparison with existing methods in qualitative (Fig. 4) and quantitative (Table 1) manner rigorously proves that the proposed method achieves the better dose distribution prediction while mitigating the over-smoothing problem.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    • Considering the well-defined task and motivation of the paper, I did not find any major weaknesses in the paper. The paper contains a well-described methodology, extensive evaluation, and good results.
  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

    The in-house dataset of the rectal cancer will not be available to the readers. However, the training and evaluation codes will help readers reproduce the results on their own datasets or the public datasets.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html

    For the task of dose distribution prediction using deep learning (DL), the majority of current research seems to use the CT image and contours of PTVs and OARs. However, it can be argued that such DL-based methods may have some weaknesses. For example, DL-based methods may be biased by various factors such as imaging protocol, contour variation, and interstitial protocols for treatment planning, thus limiting their generalizability across institutions and imaging devices. Therefore, given the rapid advances in fast Monte Carlo dose calculation, DL-based methods may be more beneficial if they can further advance automation, e.g., to generate contours of PTVs and OARs from images. Therefore, such conceptual integration of both automated contouring and dose prediction from images will make DL-based methods more promising for actual clinical practice.

  • Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

    7

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Given the specific scope of the task (i.e., dose prediction for radiotherapy planning from CT images and contour data), the paper provides sufficient technical contribution (i.e., the proposal of modified diffusion models), comprehensive comparison with existing methods based on a large dataset, and significant performance improvement.

  • Reviewer confidence

    Confident but not absolutely certain

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A




Primary Meta-Review

  • Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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.

    Overall, reviews agree that the paper adresses an interesting problem, which has high clinical relevance in the radiotherapy field. The method proposed in the paper appears novel and the paper is clearly written. However before final acceptance, authors would need to address important points raised by the reviewers, particularly by R#2 on the practical constraints which would apply for the proposed research problem, and missing definitions which hinders the overall understanding of the paper.




Author Feedback

We sincerely thank all the reviewers (R1, R2, R3 & Meta-Reviewer) for their acknowledgment of our methodological contribution and their constructive comments for further clarification.

Q1: Whether the predicted dose distribution is practically achievable? (R1, R2 & R3) A1: Thanks for your comments. In practice, some parametric regression techniques can be utilized to directly generate the executable parameters from the predicted dose distribution which can be inputted into the treatment planning system (TPS) to form a clinically achievable radiotherapy plan. Considering the satisfactory performance of our proposed model, the predicted dose distribution can provide radiologists with an initial point close to the ideal plan. In this manner, the trial-and-error steps and planning time can be reduced. Therefore, our work of dose prediction is practically achievable and of vital importance to accelerate the radiotherapy workflow. Q2: There are no details of the type of radiation technique used (VMAT)? (R1) A2: Sorry for this negligence. All the patient data is acquired through the volumetric modulated arc therapy (VMAT) treatment. We will add the corresponding description in the final version. Q3: Details of ethics or informed consent. (R1) A3: Thanks for your comments. A total number of 130 patients with rectum cancer were included in this study. The collection of CT images followed the standard medical procedures and the ethical approval was granted by the local ethics committee which is omitted for blind review. We will add related statements in the final version. Q4: Details about the CT’s - especially slice thickness. (R1) A4: Thanks for your comments. The resolution of the CT volume is 0.9766mm0.9766mm3mm and there are about 130~248 2D slices inside each volume. The original size of the 2D slice is 512x512 which is then resized to 256x256. Q5: Details of resampling volumes to 256x256. (R1) A5: The original resolution of the CT images is 512x512 which brings the model training with a heavy computation burden. Therefore, we utilize a simple resize operation with bilinear interpolation to resample them to 256x256. Q6: Detailed derivations of diffusion models principles. (R2) A6: Thanks for your valuable comments. We have omitted some derivation processes in the diffusion model due to the page limit. To make the derivation and computation more understandable, we will provide a detailed process in the supplementary materials of the final version. Q7: Are the alpha_t, gamma_t learnable parameters or non-learnable hyperparameters? (R2) A7: Following [1], We set them as non-learnable for simplification. We will state it clearly in the final version. [1] Ho, J., Jain, A., & Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840-6851 (2020). Q8: During the training via eq (9), will the pre-trained encoder also be updated? (R2) A8: As stated in Section 2.2, the structure encoder is pre-trained to gain some ability of feature extraction and equip the training stage with a better initial point. Then, in the training process of diffDP, the parameters of the structure encoder are also updated according to eq (9). Q9: Detailed definitions of the metrics used to evaluate performance. (R2) A9: Sorry for this negligence. As an essential metric of dose prediction, Dm represents the minimal absorbed dose covering m% percentage volume of PTV. Consequently, we use several corresponding metrics, i.e., D_98, D_2, maximum dose (D_max), and mean dose (D_mean) to measure the performance of dose prediction. We will add a clear statement about Dm in the final version. Q10: In Figure 3, can the authors plot the zoom-in version with PTV Gy in [40,55]? For the vertical axis, does “1.0” means 1% or 100%? (R2) A10: Thanks for your comments. We will provide the zoom-in version of Figure 3 in our revised manuscript. Besides, the “1.0” in the vertical axis means 100% and we will fix this mistake.



back to top