List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Hexin Dong, Jiawen Yao, Yuxing Tang, Mingze Yuan, Yingda Xia, Jian Zhou, Hong Lu, Jingren Zhou, Bin Dong, Le Lu, Zaiyi Liu, Li Zhang, Yu Shi, Ling Zhang
Abstract
Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal cancer in which the tumor-vascular involvement greatly affects the resectability and, thus, overall survival of patients. However, current prognostic prediction methods fail to explicitly and accurately investigate relationships between the tumor and nearby important vessels. This paper proposes a novel learnable neural distance that describes the precise relationship between the tumor and vessels in CT images of different patients, adopting it as a major feature for prognosis prediction. Besides, different from existing models that used CNNs or LSTMs to exploit tumor enhancement patterns on dynamic contrast-enhanced CT imaging, we improved the extraction of dynamic tumor-related texture features in multi-phase contrast-enhanced CT by fusing local and global features using CNN and transformer modules, further enhancing the features extracted across multi-phase CT images. We extensively evaluated and compared the proposed method with existing methods in the multi-center (n=4) dataset with 1,070 patients with PDAC, and statistical analysis confirmed its clinical effectiveness in the external test set consisting of three centers. The developed risk marker was the strongest predictor of overall survival among preoperative factors and it has the potential to be combined with established clinical factors to select patients at higher risk who might benefit from neoadjuvant therapy.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43904-9_24
SharedIt: https://rdcu.be/dnwG2
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #3
- Please describe the contribution of the paper
This paper proposed a novel method using a learnable neural distance to measure the distance between the tumor and its surrounding vessels for predicting Pancreatic ductal adenocarcinoma (PDAC) survival.
- 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.
(1) This paper further investigates the relative position relationship between tumor and vessel to improve the PDACs survival prediction.
(2) This work integrates CNN and Transformer for better capture of local and global information.
- 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.
(1) The parameter “K” is important for neural distance. But the authors didn’t clarify why choose 32 in this study.
(2) The paper should clarify which marker in Table 3 is not a pre-operative marker. I don’t know if PT is not a pre-operative marker.
(3) Table 1 should also provide the results of testing dataset.
- 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
As the data used in the paper is not publicly available, reproducing the results may be challenging. However, if the work is accepted, the code could be released to allow others to test the method.
- 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
Please see the weaknesses.
- 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
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Based on the strengths and weaknesses.
- 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 #4
- Please describe the contribution of the paper
This study proposed a multi-branch transformer-based framework for predicting cancer survival, with the following contributions: (1) Prosing a novel approach for aiding survival prediction in PDAC by introducing a learnable neural distance that explicitly evaluates the degree of vascular invasion between the tumor and its surrounding vessels. (2) Introducing texture-aware transformer block to enhance the feature extraction approach, combining local and global information for comprehensive texture information, and validating that the cross-attention is utilized to capture cross-modality information and integrate it with in-modality information, resulting in a more accurate and robust prognostic prediction model for PDAC. (3) Through extensive evaluation and statistical analysis, demonstrating the effectiveness of our proposed method.
- 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.
This an interesting and Innovative work, by introducing the learnable neural distance, structure information and the texture-aware transformer. The experimental results and statistical analysis demonstrate the effectiveness of the proposed method. Furthermore, the proposed model can be combined with established high-risk features to aid in the patient selections who might benefit from neoadjuvant therapy before surgery, which demonstrates the clinical application.
- 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 methods seem to be complex and not to be introduced clearly, especially for the most novelty point: neural distance. Additionally, the patient cohorts are confused, even I don’t know that what is the label for supervising the model, survival or not at 36 months for neoadjuvant therapy or surgery?
- Please rate the clarity and organization of this paper
Satisfactory
- 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
There is not a good reproducibility, because of these weaknesses as listed above.
- 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
1) Why did this study use multi-phase contrast-enhanced CT but no other medical imaging, such multi-sequence MRI?
2) What is difference between PDACs and previous studies [1, 11, 2, 5], for “PDACs differ significantly from the tumors in these studies”. 3) The learnable neural distance should be the most novelty, but it not was introduced clearly, such as what is the CrossAttention in formula (5). Which phase of CT was it constructed based on? Moreover, it just used the space coordinate to learn the neural distance. Why not add the imaging information?
4) What does the proposed model predict, i.e. what is label for model training? It study seems to predict the survival at 36 months, I don’t sure that. - 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
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed methods have better novelty, clinical significance, clinical feasibility. However, the reproducibility need to be improved further.
- 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 #1
- Please describe the contribution of the paper
This paper proposes a new method for predicting the prognosis of pancreatic ductal adenocarcinoma (PDAC). The main contributions are: • Accurately assessed the relationship between the tumor and nearby blood vessels using a learnable neural distance. • Improved the extraction of dynamic tumor-related texture features in multi-phase contrast-enhanced CT using a combination of CNN and transformer modules. • Leveraging the ability of contrast-enhanced CT scans to provide extensive contextual information, experiments are performed on a multicentre dataset with 1070 PDAC patients. • This neural distance between the tumour and the vessels can serve as a biomarker that, combined with other clinical factors, can help to select the patients for risk stratification and treatment decisions for patients with PDAC.
- 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.
• This paper introduces a novel biomarker in terms of neural distance extracted using a texture-aware transformer to improve the prognostic prediction of pancreatic cancer using Multi-Phase CT. • A brief clinical analysis is provided in terms of Hazard Ratios. • The proposed biomarker combined with other strong clinical features, can be used to select patients who might benefit from neoadjuvant therapy before surgery.
- 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 proposed method in this research is intriguing but has the following weaknesses: • In section 2.1, the authors have mentioned texture-aware convolution block and texture-aware transformer blocks consisting of 3x3x3 and 1x1x1 convolution layers. However, it is not defined how these layers are different from conventional convolution operations and how precisely they extract “texture-aware feature.”. • In Fig. 2, the difference between the convolution block (blue blocks) and 3x3x3 convolution layer (yellow block) is not clearly defined. • Cross-attention block is used to extract cross-modality features, and results in Table 1 and Table 2 show its effectiveness; however, the structure of cross-attention block is not defined. Moreover, the results must elaborate the effectiveness of proposed neural distance without cross-attention block. • Before calculating neural distance, nnUnet is used in a semi-supervised manner to extract structural information. However, the details for this step are missing. • Moreover, the ablation study can be extended to signify the effectiveness of structural information • This work has achieved satisfactory outcomes for a small dataset of 1072 patients, among which 892 are used for nested cross-validation. However, for independent testing, only 178 samples were used. This small number of samples in the test set raises concerns about the generalizability of the proposed framework. • The font size in Fig. 1 and Fig. 2 is tiny and hard to read. • In Fig. 2, an arrow is missing after structural aware feature extraction.
- 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 uses an in-house dataset, which introduces concern about reproducibility. Code availability is not mentioned in the manuscript
- 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
Please see para 6
- 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
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Due to the weaknesses mentioned above, I recommend a weak acceptance for this paper.
- 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
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.
- Presents integrated transformer/CNN for PDAC prognosis with a learnable neural distance parameter
- Introduces novel biomarker in terms of neural distance extracted using a texture-aware transformer
- Details of some of the transformer/CNN blocks are unclear. CrossAttention block should be better explained.
- More details on nnUnet should be provided, empirically set parameters should be justified better
- Good comparisons to SOTA methods + survival/KM analysis
Author Feedback
We would like to express our gratitude to all reviewers and meta-reviewer detailed review of our paper. We appreciate the time and effort they have invested in providing us with valuable feedback. We have carefully considered the comments and suggestions and have prepared this rebuttal to address the issues they have raised. Q1.Details of the Texture-aware blocks (R1\&Meta-Reviewer). We would like to clarify that the Texture-aware blocks consist of 3x3x3 and 1x1x1 convolution layers, with the former capturing local spatial information and the latter mapping the input tensor to a higher dimensional space. In Fig. 2, the convolution layer highlighted in blue is a single layer, while the yellow block represents the entire CNN block, composed of a convolution layer, a batch normalization layer, and a LeakyReLU activation layer. Q2.Details of CrossAttention block in Nerual distance (R1\&R3\&R4\&Meta-Reviewer). We implemented this block based on the Self-Attention block widely used in the Transformer model, with an additional mask defined in Eq.1. Specifically, we employed a 2-way CrossAttention block to calculate the surface distance between the tumor set (P_c) and the vessel set (V_c). We then obtained a K-dim distance and concatenated the features extracted from other components to predict the survival outcome using a fully connected (FC) layer. Furthermore, the parameter “K” plays a critical role in the Neural distance method as it selects the closest points to the opposite surfaces. We will perform an ablation on this parameter in the camera-ready submission to further investigate its influence. Q3.More details on nnUnet (R1\&Meta-Reviewer). We apologize for the lack of information on the nnUNet training parameters. The multi-phase CT images were directly concatenated as input channels and the pseudo-annotations were generated by combining two teacher models. The student model was trained on both manual and pseudo-annotated multi-phase images, with the teacher models trained for 200 epochs (250 batches per epoch) and the student models trained for 33 epochs (1500 batches per epoch). Once again, we would like to express our gratitude to all reviewers for their insightful comments and suggestions. We will rewrite the unclear descriptions and conduct additional experiments to improve the quality of our paper in the camera-ready submission.