List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yunkui Pang, Xu Chen, Yunzhi Huang, Pew-Thian Yap, Jun Lian
Abstract
Prostate magnetic resonance imaging (MRI) offers accurate details of structures and tumors for prostate cancer brachytherapy. However, it is unsuitable for routine treatment since MR images differ significantly from trans-rectal ultrasound (TRUS) images conventionally used for radioactive seed implants in brachytherapy. TRUS imaging is fast, convenient, and widely available in the operation room but is known for its low soft-tissue contrast and tumor visualization capability in the prostate area. Conventionally, practitioners usually rely on prostate segmentation to fuse the two imaging modalities with non-rigid registration. However, prostate delineation is often not available on diagnostic MR images. Besides, the high non-linear intensity relationship between two imaging modalities poses a challenge to non-rigid registration. Hence, we propose a method to generate a TRUS-styled image from a prostate MR image to replace the role of the TRUS image in radiation therapy dose pre-planning. We propose a structural constraint to handle non-linear projections of anatomical structures between MR and TRUS images. We further include an adversarial mechanism to enforce the model to preserve anatomical features in an MR image (such as prostate boundary and dominant intraprostatic lesion (DIL)) while synthesizing the TRUS-styled counterpart image. The proposed method is compared with other state-of-art methods with real TRUS images as the reference. The results demonstrate that the TRUS images synthesized by our method can be used for brachytherapy treatment planning for prostate cancer.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_46
SharedIt: https://rdcu.be/cVRTF
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a novel weakly supervised learning strategy to train DCLGAN, thus achieving the transition from input MRI images to output TRUS images. Prostate Contour Segmentation and MRI Pattern Exaction are proposed to realize the weakly supervised learning for DCLGAN.
- 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 author collected a dataset and carried out detailed experiments on this basis. Compared with the baselines, the performance of the proposed method is much improved.
- Prostate Contour Segmentation and MRI Pattern Exaction contribute to realize the weakly supervised learning for DCLGAN.
- 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.
- Qualitative results show that the proposed method still has a lot of room for improvement.
- Weakly supervised learning is emphasized in the title and contribution of the paper, but there is too little description about weakly supervised learning in the main body of the paper.
- The authors did not conduct ablation experiments. This is not a big problem, but the author uses too many modules, I am not sure whether all these modules could work.
- Please rate the clarity and organization of this paper
Poor
- 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
This paper is clear enough that an expert could confidently reproduce.
- 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/2022/en/REVIEWER-GUIDELINES.html
The authors did not conduct ablation experiments. This is not a big problem, but the author uses too many modules, I am not sure whether all these modules could work. For example, Contour, IDT loss and BCE loss are all used in Prostate Contour Segmentation, but I am not sure whether these settings are all necessary.
- 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?
I think this thesis is probably below the threshold for acceptance. The main problem lies in the structure of the paper, the lack of necessary ablation experiment, and the qualitative effect is not particularly good.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
3
- Reviewer confidence
Somewhat 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 #3
- Please describe the contribution of the paper
To cope with low soft tissue contrast and tumor visualization in the prostate area of TRUS imaging, this paper proposes a method that synthesizes TRUS images from unpaired MR images. Anatomical structural constraints with weakly supervised modules are considered to emphasize the structural content of the synthesized TRUS image.
- 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 method for synthesizing US images from MRI images is novel, the application has significant clinical value. Comparison and evaluation of the results are provided.
- 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.
-
- 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
I believe it could be reproduced from the paper.
- 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/2022/en/REVIEWER-GUIDELINES.html
It is very helpful that the authors can provide additional explanation on the optimization objective for the generators G and H.
- 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?
Coping with low soft tissue contrast and visualization of the tumor in the prostate region is a difficult task. The paper presents a method that synthesizes TRUS images from unpaired MR images. No major revision is required but improvement could be on clarifying the MR image pattern extraction which is very technical.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Somewhat 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 #4
- Please describe the contribution of the paper
The paper employs DCLGAN to generate TRUS-styled MR images using weakly supervised learning. This can be applied to LDR prostate brachytherapy planning and treatment. The results show both image quality enhancement and precise segmentation.
- 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 proposed method creatively constructs the style transfer architecture upon a generative adversarial network based on cycle consistency. Particularly, the discriminator classifies the real and synthetic MR images, TRUS images, MR prostate contours, segmentation of tiny anatomical structures of prostate.
- Adopt weakly supervised learning to alleviate the requirement of paired data for training, which is usually too challenging to have segmentation masks on MR Images.
- 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.
Need to specify how to select the weights in the loss function.
- Please rate the clarity and organization of this paper
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 paper has included enough details to facilitate reproducibility of this work, except for the weights in the loss function. In addition, it would be a large contribution to our community if the authors consider opening the training data.
- 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/2022/en/REVIEWER-GUIDELINES.html
I have no suggestions here.
- 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?
The proposed method is novel with most of the specification provided. Using weakly supervised learning for prostate segmentation on MRI resolves the issue of lacking prostate labels.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- 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.
Interesting paper and application, although limited validation which probably should be acknowledged in the revised version.
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and 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).
nr
Author Feedback
Reviewer #1
- “The method has room to be improved.” We agree. The contrast of the prostate and internal structures can be further improved.
- “Too little description of weakly supervised learning in the main body of the paper.” Fully supervised learning means all the networks or sub-modules are trained using labels, while weakly supervised learning means only some sub-modules need labels and other modules don’t. In our method, we only create labels for the CTR module that segments prostate contour from US images. Other modules are not supervised because we do not create labels to train them. Therefore, the proposed method is “weakly supervised”.
- “The authors did not conduct ablation experiments.” We actually performed the ablation study and both the quantitative and qualitative results were presented in the original submission. However, we probably did not mention it clearly in the paper. In Table. 1, the “Ours (w/o CTR)” row is the ablation that removes the CTR module from our proposed method. The “Ours (w/o Edge)” row is the ablation that removes the Edge module. And the “DCLGAN” row shows the ablation that removes both the CTR module and Edge Module. Similarly, in Figure 4, the three columns: “DCLGAN”, “Ours (w/o CTR)”, and “Ours (w/o Edge)” are qualitative results of the ablation study.
Reviewer #3
- “Additional explanation on the optimization objective for the generators G and H.” In Appendix C, we described how the loss functions are combined to train generators G and H (the weight selection is described in the answer to reviewer 4). The objective for generator H is simpler because it only serves as a cycle-consistency constraint, similar to the Cycle GAN. The first loss in L_H helps generator H to generate the MR-styled image. And the second loss forces generator H to preserve structures from the input TRUS image. Similar to generator H, generator G also has these two loss functions. But in practice, we find the second loss is a weak constraint for structure preservation. Therefore, we add loss between the prostate contour of TRUS and the estimated prostate contour of the synthesized image (need to extract it using the CTR module). We also add loss between the estimated prostate contours of MR and the synthesized image. (since there is no prostate label on MR images). Lastly, the edge loss is added to ensure the internal structures of the MR image are preserved in the synthesized image.
Reviewer #4
- “Need to specify how to select the weights in the loss function.” In Appendix C, we described how the loss functions are combined. In practice, to train Generator H, lambda_A=0.5 and lambda_B=0.5. To train Generator G, in the first 30k epochs, lambda_A=0.5, lambda_S=0.5, lambda_C=lambda_E=0. (first train a stable base network) In the next 70k epochs, lambda_A=lambda_S=0.45, lambda_C=0.05, lambda_E=0.05.