List of Papers Browse by Subject Areas Author List
Abstract
In radiation therapy (RT), the reliance on pre-treatment computed tomography (CT) images encounters challenges due to anatomical changes, necessitating adaptive planning. Daily cone-beam CT (CBCT) imaging, pivotal for therapy adjustment, falls short in tissue density accuracy. To address this, our innovative approach integrates diffusion models for CT image generation, offering precise control over data synthesis. Leveraging a self-training method with knowledge distillation, we maximize CBCT data during therapy, complemented by sparse paired fan-beam CTs. This strategy, incorporated into state-of-the-art diffusion-based models, surpasses conventional methods like Pix2pix and CycleGAN. A meticulously curated dataset of 2800 paired CBCT and CT scans, supplemented by 4200 CBCT scans, undergoes preprocessing and teacher model training, including the Brownian Bridge Diffusion Model (BBDM). Pseudo-label CT images are generated, resulting in a dataset combining 5600 CT images with corre-sponding CBCT images. Thorough evaluation using MSE, SSIM, PSNR and LPIPS demonstrates superior performance against Pix2pix and CycleGAN. Our approach shows promise in generating high-quality CT images from CBCT scans in RT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1234_paper.pdf
SharedIt Link: https://rdcu.be/dVZed
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72378-0_12
Supplementary Material: N/A
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Hwa_Improving_MICCAI2024,
author = { Hwang, Joonil and Park, Sangjoon and Park, NaHyeon and Cho, Seungryong and Kim, Jin Sung},
title = { { Improving cone-beam CT Image Quality with Knowledge Distillation-Enhanced Diffusion Model in Imbalanced Data Settings } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15001},
month = {October},
page = {123 -- 132}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work adapt the Brownian Bridge Diffusion Models to synthesis diagnostic-CT-like image based on daily check up CT acquired in radiation therapy. It has compared to other state-of-the-art generative techniques such as DDPM and GAN
- 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 paper is structured well and easy to follow. Moreover, the clinical motivation is described sufficient and it adapt a novel diffusion technique, which demonstrated advantages in tackling computer vision tasks, into the medical domain. Last but not least, it also used clinical data for the evaluation of the algorithm.
- 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.
The major weakness of this approach is its clinical feasibility. The core idea of diffusion model is to take an unknown probability distribution then gradually convert it to a known probability distribution. Adapting this principle to synthesis medical image, especially for the purpose of intervention (such as RT-planning) is somewhat risky, as the anatomical correctness is crucial in such a clinical application. As the correctness of the generated images can not be per se guaranteed, it is doubtful that such an approach will ever be successfully translated into clinical use. Furthermore, the proposed method learns to synthesis the new images based on the intensity values of the images, which is not always the true HU value. In fact, different medtech vendor haven their own method to generate pseudo HU value to improve the visual quality of the CT images. This aspect has not been take into the consideration of this paper in general. Depending on type and vendor of the scanner that are used for data acquisition, the result of the method might be different.
- 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 provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
It seems to me, that the approach was not discussed with clinical experts in depth, as a generated image without any guarantee of its anatomical correctness and generalizability across the different CT scanners will rarely be accepted by clinical users for RT planning purpose. I suggest the author to work more closer with the clinicians to have a in-depth understanding of the clinical requirement
- 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
Weak Reject — could be rejected, dependent on rebuttal (3)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method applied to solve this problem is not fundamentally new, although the adaptation to medical domain has certain novelty. However, adaptive RT planning was chosen as the target problem space, where a diffusion model based an approach is generally sort of ill-posed (refer to 6)
- Reviewer confidence
Very confident (4)
- [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
The paper presents a method for generating synthetic CT from CBCT 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.
Clinical motivation is provided. The authors present useful qualitative results in addition to quantitative results. It’s great to see the authors discuss a failure case with some discussion.
- 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 is strangely no mention of the MICCAI Synthrad challenge ( https://arxiv.org/abs/2403.08447 ) or validation against the challenge data.
- What would be the actual clinical benefit if a perfect CT quality volume could be generated from CBCT? Would the patient be replanned each treatment fraction?
- Just a comment about the introduction and data imbalance: In many/most EBRT instances there is only a single planning CT used. Margins may be added to account for internal motion between fractions. In addition many centres do not use daily CBCT (partly to reduce additional ionizing radiation delivery) and many centres still use portal imaging instead (EPID).
- This is probably my ignorance, but the use of pseudo labels (eg. Fig 2) is confusing to me. Can you explain this in a bit more detail?
- Section 2.3 - How were the images normalised? This is important for CT, where the HU values have physical meaning.
- Table 1. reports metrics based on normalised images. Clinically the metrics must to be based on HU (and probably only MSE and SSIM are useful). For RT, a dosimetric comparison would be even more useful. You could check the metrics used for the recent MICCAI SynthRad challenge for some suggested metrics.
- Figs 3 and 4 and 5 - the authors must ensure the window levels are same for the CT and the synthetic CT. This can easily be done in a open source viewer (like Slicer).
-
- Section 3.1 - 1400 paired scans is a lot for 40 patients - I think you mean 2D slices not scans?
- Section 3.1 last paragraph - what manual adjustments were made ( eg. translation/rotation only)?
- Section 4.2 “overly blurry” isn’t scientific language - more important information are HU differences and impact on dose.
- 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.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
see above
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
There are some weaknesses with the paper, but the application is of interest and I think if the authors can correct the issues mentioned above it would be an interesting MICCAI paper.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
Thanks to the authors for the rebuttal. I will retain my original score.
Review #3
- Please describe the contribution of the paper
The paper introduces a new method to improve planning CT (pCT) generation from cone-beam CT (CBCT) scans in radiation therapy, overcoming challenges such as anatomical discrepancies and Hounsfield Unit (HU) inaccuracies in CBCT datasets. By integrating diffusion models and utilizing supervised and unsupervised learning techniques, it outperforms existing methods. Through thorough evaluation and comparison, the paper highlights its potential to enhance treatment plans in adaptive radiation therapy (ART). Overall, the innovative approach proposed in the paper shows promise for improving patient treatment outcomes in clinical practice.
- 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 paper discusses the potential implications of the proposed approach for improving patient treatment outcomes in clinical practice, highlighting its relevance and significance in the field of radiation therapy. -The paper conducts rigorous evaluation using multiple metrics, including Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM), providing comprehensive evidence of the effectiveness of the proposed method. -By integrating diffusion models into the approach, the paper offers a sophisticated solution to improve the synthesis of CT-like images from CBCT data, potentially enhancing the precision of treatment planning.
- 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.
The paper may assume homogeneity in the anatomical structures and imaging characteristics across different patients or treatment scenarios. However, if the method’s performance varies significantly in diverse patient populations or clinical settings, it could limit its applicability and effectiveness in practice.
- 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.
- Do you have any additional comments regarding the paper’s reproducibility?
N/A
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
-While your study demonstrates superior performance in image quality metrics, it would be valuable to validate the clinical utility of your proposed method through direct evaluation of its impact on patient outcomes. Consider conducting retrospective or prospective clinical studies to assess the effectiveness of your approach in improving treatment planning accuracy, reducing treatment-related toxicities, or enhancing overall patient care. -Address the generalizability of your findings to diverse patient populations and clinical scenarios. Discuss any potential limitations or challenges associated with applying your method in real-world clinical practice, particularly in cases where anatomical variations or imaging artefacts may affect the performance of your approach. -Enhance the quality of the images in the paper.
- 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
Accept — should be accepted, independent of rebuttal (5)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a novel approach to improving the accuracy of planning CT (pCT) generation from cone-beam CT (CBCT) scans in radiation therapy. By integrating diffusion models and employing a combination of supervised and unsupervised learning techniques, the proposed method offers a unique and innovative solution to address challenges in the field.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Accept — should be accepted, independent of rebuttal (5)
- [Post rebuttal] Please justify your decision
The authors have thoroughly addressed all the concerns raised in the initial review respected to the proposed Brownian Bridge Diffusion Models to synthesis diagnostic-CT-like image based on daily check up CT acquired in radiation therapy.
Author Feedback
Dear Reviewers,
We greatly appreciate your efforts and constructive feedback on our manuscript. We have carefully considered each of your comments and would like to address the key concerns as follows:
Two reviewers (1 and 5) raised concerns about the clinical feasibility and anatomical correctness of the generated images, noting potential issues with generalizability across different CT scanners and diverse patient populations and clinical scenarios. We fully agree that demonstrating such generalizability is indeed a critical step for clinical deployment. However, we believe this is beyond the scope of the current study, given its initial phase in the development process. We will focus on addressing generalizability in the next phase, accommodating scanner variability, patient population diversity, and practical clinical scenarios. Advanced data normalization techniques, including multi-frequency processing, are currently under investigation in conjunction with the diffusion-based approaches.
We would like to emphasize that the superior performance of the proposed model compared to existing models has been clearly demonstrated in this study. Although the proposed method consistently showed improved performance over existing models and techniques, we acknowledge the concern raised by one reviewer regarding the potential mismatch of actual Hounsfield units (HU) and anatomical structures with the ground truth, which could cause issues if used for treatment planning. However, our goal in developing this model was not to completely replace planning CT in clinical practice. Instead, it aims to enhance the quality of daily cone-beam CT (CBCT) images used for patient positioning verification, allowing for more accurate alignment. This improvement is intended to better align the patient and monitor overall changes during the treatment process, such as tumor size reduction and patient weight loss. For these purposes, the emphasis on accurate HU is less critical than for planning and dose calculation. Therefore, we believe that this model can be effectively used in the decision-making process for treatment progress monitoring and determining the need for adaptive planning, providing clinical utility without direct dose calculation.
We also acknowledge the suggestion to collaborate more closely with clinical experts to gain a deeper understanding of clinical requirements.
Additionally, Reviewer 4 pointed out the absence of a discussion on the MICCAI SynthRad challenge and validation against its data. We appreciate this comment. Unfortunately, we became aware of the challenge dataset only recently and did not have enough time to incorporate it into our study. However, we fully expect that the proposed method will consistently outperform other approaches.
Kind regards, Author
Meta-Review
Meta-review #1
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The authors have adequately addressed all the concerns.
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
The authors have adequately addressed all the concerns.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper has two accept recommendations and the rebuttal explains the possible clinical applicability. I thus recommend accept.
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
The paper has two accept recommendations and the rebuttal explains the possible clinical applicability. I thus recommend accept.