List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Kyungsu Lee, Jaeseung Yang, Moon Hwan Lee, Jin Ho Chang, Jun-Young Kim, Jae Youn Hwang
Abstract
Ultrasound (US) imaging has been widely utilized in medical fields. The acquisition of US images which contain pathological information is required for a better diagnosis. However, it is challenging to acquire informative US images due to their structural complexity, and it is significantly dependent on the expertise of a sonographer. Therefore, in this paper, we propose a fully automatic scanning-guide algorithm that assists unskilled sonographers to acquire informative US images. The proposed scanning-guide algorithm provides the proper directions of probe movement to search target regions and thus enables even unskilled sonographers to easily acquire ultrasound images of the target regions. The main contribution of this paper is to (1) propose a new scanning-guide task that aims to search a Rotator Cuff Tear (RCT) region using a deep learning-based algorithm (USG-Net) and (2) construct a dataset to optimize the corresponding deep learning algorithm. First, the multi-dimensional US images acquired from 80 patients with RCT are processed to optimize the scanning-guide algorithm. The optimized algorithm then classifies the existence of RCT successfully. Furthermore, if RCT is not in the current frame, the algorithm provides the proper direction toward RCT. The experimental results demonstrate that the fully optimized scanning-guide algorithm offers the proper directions to localize a probe to target regions with high accuracy and assists the acquisition of informative US images.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16449-1_3
SharedIt: https://rdcu.be/cVRUJ
Link to the code repository
https://github.com/kyungsu-lee-ksl/USG-Net
Link to the dataset(s)
https://github.com/kyungsu-lee-ksl/USG-Net
Reviews
Review #1
- Please describe the contribution of the paper
This work proposes USG-Net, a method for identifying RCT in US images. A generator portion of the method generates a 3D portion prior to the classification module which helps guide the direction of the probe motion.
In contrast to previous work, the proposed method does not rely on inertial measurement unit signals. While the authors don’t specifically say this (which I think they should - if I’m correct) this could mean a simpler setup, or in other words use of existing clinical pipelines without additional equipment (i.e. added cost/complexity)
- 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.
To attempt to remove the need for IMU signals simplifies the setup and allows use of this approach in existing clinical pipelines, with minimal changes.
- 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 main weakness of the work is lack of clarity in some areas. Some information needed for better understanding the method is given later in the paper, and at least one additional figure explaining some of the definitions (e.g. directions, tangential plane, etc) is needed.
Also, there is no comparison in this work (with similar methods or clinical evaluation)
- 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
Dataset acquisition method, system specifications and dataset are all anonymous, hence, it is not possible to comment on reproducibility of the work.
- 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
Abstract lacks quantitative information. E.g. results
2.1 pg 3: This section is quite confusing. What is the significance of the parallel lines? Assuming a cubic 3D volume, wouldn’t any cross section contain parallel lines? And is there a specific plane/direction we look at the RCT (axial, coronal)? Or does it not matter. E.g. In fig 1 if we tilt the probe around the yellow line, we would get a pink parallel line which is a bit further down the bottom face. Would that still be an acceptable plane? As for distance, are the authors measuring the distance between a plane and points on a 3D RCT volume? I think re-writing this section to highlight why this is done, along with a figure showing how this distance is measured, can clarify things.
When you talk about probe motion, would that be motion of the probe orthogonal to the surface only or could rotation/tilt of the probe (which could also eventually capture the RCT in the field of view) be a feasible solution?
In LU, L, LD, etc what are the up/down/left/right directions? What direction is S? (from later in the text it appears to mean “stop” but it is not clear here)
2.2. What do the authors mean by thick and thin volume?
2.2. If D is depth, I’m assuming W and H are width and height (again, relative to what? Please show in a figure). What is “p”? This is somehow shown in figure 2 but not clear from this section.
In section 2.2. Do the authors mean to say that the direction of movement is calculated from the prediction of the location of the RCT, which is computed by predicting the 3D volume around a 2D slice?
Do results depend on how close the starting probe position is to the RCT?
Please provide information regarding the dataset (size, train/test split, how the data was collected, an relevant clinical information etc)
Section 4: what is the significance of this method for dataset construction (e.g. why randomly slicing 3D US images lead to better results?)
The authors state that since their approach has never been studied in that the proposed model guides the probe to the target regions, they did not perform comparative studies. However, it is worth comparing the method to manual performance of experts. For example, how would using this method as a guide improve the performance of an expert compared to when they manually try to find the RCT. Would this improve the time needed? How would this help non-experts? The authors initially justify that this method would help non-experts. Would this improve the accuracy of them finding the RCT? By how much? Some clinical evaluation would strengthen the usefulness of this method. [The authors acknowledge this as a limitation, however “some” form of comparison is necessary in my opinion]
I wouldn’t consider limitation 2 a limitation, but a strength.
- 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?
Lack of clarity in many areas, lack of adequate comparison with other work or manual approaches.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
4
- 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
5
- [Post rebuttal] Please justify your decision
The author have addressed some of the concerns, particularly regarding improving clarity of the work within the text and figures.
Review #2
- Please describe the contribution of the paper
- A deep learning-based scanning-guide algorithm is proposed to guide to the exact target diseased region without external motion-tracking sensors.
- A new scanningguide task that aims to search target disease regions using a corresponding network.
- They develop an automatic dataset construction method to train the deep learning-based scanningguide network.
- 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 well-writen following the structure of Dataset construction, USG-Net Architecture, Loss function.
- The contributions are clear: A deep learning-based scanning-guide algorithm that guides to the exact target diseased region.
- The Dataset is constructed and the automatic dataset construction method to train the deep learning-based scanningguide network is summarized.
- 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.
- A formal mathematical formulation of the problem is recommended. The paper lacks the basic formulation, starting from the detailed dataset.
- More comparison with other methods is recommended, only the proposed method and its variants are compared.
- Training epoch in ablation studies can be more in the experiments parts.
- 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 work is solid and easy to follow and reproduce.
- training and testing code will be released.
- Clear implementation details of hyperparamters in appendix.
- Clear dataset construction pipeline.
- 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
There is some points to improve:
- More comparioson with other methods in Table I, as only the proposed method and its variants are compared.
- More training epoch is recommended in the experiments.
- More formal mathematical problem formulation in the method chapter.
- An overall introduction of the experiments to provide general picture.
- 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?
- Clear contributions in novel task,
- Well-written, but the problem formulation is lack and general design of experiments are missed.
- More methods should be compared, as only the proposed method and its variants are compared.
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
In this paper, the authors extract feasible 2D US images from clinical 3D US volumes and use these to create a deep-learning approach for simultaneously detecting rotator cuff tears (RCT) in 2D US images and providing feedback to sonographers on US probe movements if RCT is not detected in the plane.
- 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 Relevance: User dependence in the acquisition of 2D US images is a consistent and clinically relevant challenge. This also extends to sonographer training, which could be a potential application of this work.
Experiments and Results Section: The results presented are quite thorough and clearly described, including an ablation study.
- 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.
Clarity: Some sections of the work are unclear, particularly in the Methodology where some variables have not been defined and hyperparameters are not stated.
Novelty: As acknowledged in the literature review provided in the Introduction, the idea of providing US probe movement directions to sonographers is not novel and has been explored in other applications. Although this is acknowledged, the limitations of the existing work should be described further to justify the unmet need addressed by this contribution.
- 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
Although based on the reproducibility checklist, the authors demonstrate a commitment to publishing their code publicly, it is unclear whether the dataset will also me made available. The hyperparameters used appear to be missing 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/2022/en/REVIEWER-GUIDELINES.html
As noted in the prior sections, the two major revisions are as follows: 1) Ensure that all variables used in the manuscript have been defined and provide hyperparameters where relevant. 2) Expand or modify the literature review provided in the Introduction to justify the need for the proposed method to overcome existing limitations.
Minor revisions: Abstract: -Unclear what is meant by “structural complexity” -The tense used for the methodology is confusing. Please ensure that methods are written in past tense.
Introduction: -Paragraph 2: I would be careful with the definition of CAD as it is currently stated (“…the deep learning network diagnoses diseases in the acquired images”. The deep-learning algorithms currently aid in the diagnosis but are not making independent diagnoses. -Paragraph 2: Good job acknowledging existing work in this area. I would like to see more justification of “Despite the accurate guidance toward the standard scan plane by the proposed deep learning models, the unskilled sonographers have still difficulty in searching target disease regions” to explain the need for the proposed approach
Methodology: -The figures are very helpful for understand the approaches used. -2.1 Dataset Construction: In the last sentence, the meaning of a movement in the “S” direction is not defined until later in the experimental results. Please add it here. -2.2 Anatomical Representation Learning: Variables H, p, and W do not appear to be defined in the text.
Experiments and Results: -Clearly described
Discussion and Future Work:
- Limitations: For you second limitation, it sounds like this would be a straightforward transfer to another application but further work would need to be done for new applications to show the algorithm what to detect. Please clarify this in the text.
Conclusions: -Unclear how the high evaluation performance demonstrates “novelty”. Please rephrase.
- 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?
Overall, I think that the work presented will be of interest to the MICCAI community, providing a creative combination of deep-learning approaches to address a clinically relevant topic. The justification of the need for a new method and details of its implementation could be stronger.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
3
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
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.
This paper proposes a scanning-guide task that aims to search target disease regions using a corresponding network (USG-Net). USG-Net is capable of providing the direction of the probe movement toward the target region without external motion sensors. USG-Net utilizes 2D US images and 3D US images that contain anatomical representation. To this end, the authors adopt a generative model predicting 3D volume based on a 2D US image. In addition, they develop an automatic dataset construction method to train the deep learning-based scanning-guide network. The reviewers mention that this work has merit, but they point to several major issues. Some of the main issues are highlighted below. First, several instances of limited clarity are outlined in the reviews. I ask the authors to carefully address those in the rebuttal and improve the clarity of the paper and some variables that are not defined. There are also issues regarding better validation experiments. I encourage the authors to fully address all the comments in their rebuttal.
- 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).
9
Author Feedback
We appreciate that the key contributions of our work are affirmed by the reviewers: 1) a novel approach for a scanning-guide algorithm that does not rely on a motion-tracking sensor (R1, R2); 2) an automatic dataset construction method to train the scanning guide algorithm (R2); 3) clinical relevance for extending to sonographer training (R3). R1 and R2 suggest a comparison of ours and other methods. The output of our method is the direction of probe movement in nine directions. However, since [8] and [14] exhibit different environments and approaches compared to ours, there is a limitation in conducting a comparative experiment with the same environment. The output of the method in [8] is the angular signal of the IMU, and [14] exploits a reinforcement learning-based approach using a robot. Besides, R1 suggests comparing the manual performance of experts, but there is a limit to comparing the expert’s movement with our method because action cloning must be carried out using the IMU. Nevertheless, we will perform the associated works for further study. R1, R2, and R3 raised concerns about the ambiguity and clarity regarding mathematical formulations and definitions. For a better understanding of mathematical formulations and definitions we missed, we will add more detailed information to Fig. 1 and Fig. 2 (e.g., the meaning of tangent plane, direction, H, W, p (padding)). A detailed description of a dataset was not illustrated due to MICAAI anonymous policy (Data acquisition: pg 3, ln 4). We will include a detailed description in the final manuscript. Specification for the training is illustrated in Table 1 and Table 2 of the Appendix, and pg 6, ln 15. We missed information on hyperparameters and their detailed role since we used the same hyperparameters for the pre-training of the 3D volume generator as shown in the previous study (DEAR-3D) [21]. R1 may be confused about dataset construction. Intuitively, any cross-section contains parallel lines in a cubic 3D volume. At the implementation level, we used “parallel lines” to consider the sonographer’s behavior that acquired 2D US images using a probe. Therefore, we sliced 3D volume with randomly generated parallel lines and then obtained sliced images for constructing large datasets (key contribution in medical dataset construction). Moreover, in our study, the direction of slicing is not constrained with respect to axial, coronal, and sagittal planes, and the probe angle is also unconstrained. To calculate the distance between RCT and 2D plane, we find the minimum value by finding the distance of all voxels of RCT in 3D space and all voxels of the 2D plane. In addition, the meaning of thick and thin volume, commented by R1, is an expression of the volume size for large and small values of depth (D) and padding (p). In inferring the direction of movement, the shapes of the surrounding bones, muscles, and tendons are identified from the generated 3D volume, and thus the direction where the target disease exists is predicted. Therefore, since the target diseased region is predicted based on anatomical knowledge, the initial probe position is not dependent on the distance. R2 recommends more training epochs in the experiment. As a result, we trained more epochs than in Fig. 3, but there was no change in the tendency for loss in larger than 300 epochs. R3 suggests that we describe the limitations of existing work and justify our contribution. One of our contributions is to eliminate an external motion sensor in the existing deep learning-based scanning guide algorithm. The limitations of the previous study are described in paragraph 2 of the introduction. Since our study suggests a new paradigm of the scanning-guide algorithm, we refer to the high accuracy as a novelty. R3 raised concerns about the definition of CAD. We agree with that. We will correct the term (e.g., diagnose). Other minor points of clarification will be carefully addressed in the final paper.
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
I would like to thank the authors for addressing reviewers’ comments, especially those on clarity and validations. I encourage them to incorporate their response into the final paper.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
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).
6
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
I believe the authors have addressed the major comments. The idea of guiding the ultrasound probe without external tracking will be of great interest to the CAI community and will generate important discussions.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
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).
5
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The paper generally received positive reviews from all reviewers. Although similar approaches for modeling and providing the desired US probe movement directions have been explored in other contexts, all three reviewers found clinical novelty in the paper. As Reviewer 1 pointed out, the proposed method could greatly simplify the clinical workflow without the need for additional hardware components.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
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).
NR