List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
SeokHwan Oh, Myeong-Gee Kim, Youngmin Kim, Guil Jung, Hyuksool Kwon, Hyeon-Min Bae
Abstract
Recent improvements in deep learning have brought great progress in ultrasonic lesion quantification. However, the learning-based scheme performs properly only when a certain level of similarity between train and test condition is ensured. However, real-world test condition expects diverse untrained probe geometry from various manufacturers, which undermines the credibility of learning-based ultrasonic approaches. In this paper, we present a meta-learned deformable sensor generalization network that generates consistent attenuation coefficient (AC) image regardless of the probe condition. The proposed method was assessed through numerical simulation and in-vivo breast patient measurements. The numerical simulation shows that the proposed network outperforms existing state-of-the-art domain generalization methods for the AC reconstruction under unseen probe conditions. In in-vivo studies, the proposed network provides consistent AC images irrespective of various probe conditions and demonstrates great clinical potential in differential breast cancer diagnosis.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_74
SharedIt: https://rdcu.be/cVRUi
Link to the code repository
https://github.com/joseph9337/DSG-net
Link to the dataset(s)
https://github.com/joseph9337/DSG-net
Reviews
Review #1
- Please describe the contribution of the paper
This paper describes a framework to generate attenuation coefficient (AC) images from ultrasound (US) that is able to generalize to different probe geometries, in the context of quantitative US. The main contribution of the proposed method is a Deformable Sensor Adaptation (DSA) module, which takes the radiofrequency signal and the B-mode image as input, and learns to adjust the sensory data to a common sensory representation. The DSA module is trained leveraging on a meta-learning scheme, which improves generalization to unseen probe conditions. Both numerical and in-vivo validation using probe configurations not seen at training time is performed. Comparison with other domain randomization approaches is present.
- 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 framework presents an original approach to synthetize AC images starting from sensory data, even when acquired with US probe geometries different from the ones considered at training time. The proposed approach provides an interesting contribution in the context of US-guided procedures in general. In fact, learning-based methods are generally trained on images or methods acquired with a single fixed probe geometry, and they thus fail to generalize to images acquired with different US probes or systems. This paper presents a promising idea to improve the generalization capabilities of such methods. Experimental evaluation is performed to assess the performance of the method both on numeric examples and in in-vivo settings. An ablation study is also reported to show the contribution of each of the presented modules.
- 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 paper lies in the description of the conducted experiments and the obtained results. While the presented evaluations look reasonable, further details should be provided to better understand how the method assessment has been performed, why conducted experiments are relevant, and why the obtained results represent a proof of framework accuracy (more details below).
Moreover, description of the methods is sometimes difficult to follow due to the many details, parameters and acronyms. Used parameters and notation are not always properly defined.
- 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
In the reproducibility checklist, authors claim that they share a sample test dataset and their code. The current version of the manuscript has no links to open source repositories, but I expect authors would make data and code available upon acceptance. I see it very difficult to replicate the method and reproduce the results by only following the description provided in the paper, due to the complexity of the proposed architecture. Moreover, authors declare that ethical approval is “not applicable” to their work. Actually, since their evaluation include real patient data, ethical approval should be present.
- 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 description of the introduced symbols should be improved. Some notation is defined but never used (e.g., W_probe), other is never defined (e.g., R and omega in equation 1). Then, do y and I_s represent the same thing (i.e., output image)? In 2.4, authors say that y represents the ground truth, while in 2.1 y_i is defined as “organs and lesions”. Definitions should be more consistent.
The implementation details at the end of 2.4 are very difficult to follow, since the notation is quite heavy. In particular, it is not straightforward to me what the apices p, l and r associated to the inputs x indicate.
Authors claim that the method provides “accurate AC image”, but more details should be provided to support this claim. How do PSNR and MNAE allow to assess the method accuracy? Equations provided in the supplementary material B use a different notation than the one in the paper, thus are not really helping with the understanding. Moreover, measurement units should be reported.
In 3.1, authors refer to the “aggregated dataset” but they never define what it represents.
It is not clear how the Deep-All baseline method is exploited for evaluation. Are the proposed modules used together with Deep-All in the ablation study? If so, how are the different modules interfaced? When authors refer to “the NN” in section 3.1, do they mean the Deep-All model?
The description of in-vivo experiments should be improved. In particular: 1) How do authors define which is the AC related to lesions in in-vivo experiments? Do they extract lesion area from the generated images via segmentation? 2) To which approach do results related to the “ablated baseline” refer to (among those in Table 2)? 3) Since the paper reports an evaluation using in-vivo data, ethical approval must be mentioned. 4) I would suggest moving the details about the geometry of the probes used in in-vivo trials to section 3.2, to group there all the information related to such experiments.
As for the data-augmentation part, it would be interesting to report the data distribution after VSG-aug is used and compare it with the one in Table 1. This would allow to understand if the tested conditions fall within the data distribution obtained when using the augmented dataset.
Images can be improved for better clarity. Fig.1 should include the acronyms relative to each “module” to facilitate understanding of what is what. The font size in Fig.2 is too small, making it impossible to read, but it is important since it is the only way to know the dimensionality of the different parts of the architecture.
Typos:
- Euclidian -> Euclidean
- Each resolution subnetworks -> Each resolution subnetwork
- adjust -> adjusts
- 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?
Generalization to multiple probe geometries is a very interesting and novel contribution, with high potential impact to other US-guided applications. The paper is well-written and the evaluation looks robust, even if poorly described. I think the paper can be considered for publication, provided that authors improve the description of the conducted experimental evaluation and the obtained results.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- 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 #2
- Please describe the contribution of the paper
They propose a deformable sensor generalizable network to calibrate probe conditions. -They assessed the proposed method through numerical simulation and in-vivo studies.
- 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.
- They improved the accuracy of the AC image by 9.7%.
- In comparison with other methods, their method demonstrates 11.8% reconstruction improvement of the unseen data.
- Their proposed method demonstrates 26% enhancement for the reconstruction of out-of-distribution sensor geometry.
- 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 biomechanical properties used for simulation are not described.
- 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 authors agree to make public the, Training code, Evaluation code, (Pre-)trained model(s), Dataset or link to the dataset needed to run the code.
- 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 mentioned they used “biomechanical property that is set to cover general soft tissue characteristic” with a reference. It would help better clarification of the paper to mention these mechanical properties and how they were implemented in the simulations.
- 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?
- Network structure and the proposed method are explained well
- Number of papers in your stack
4
- 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
Review #3
- Please describe the contribution of the paper
The authors present a deep learning approach for AC reconstruction which is robust to the changes of the probe.
- 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 work has good novelties It is highly required for ultrasound imaging Very good abalation experiments convncing improvements
- 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.
Some of the parts should be refined so that the readers can have a better undrestanding of the approach (exp: Meta learning, DSA ) In simulation results, there is no experiment for CNN trained on signle imaging setting (I think Baseline without those modules still trained on multiple imaging settings?, if not, please clarify) There is no comparision with non-deep learning approches
- 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
good
- 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 a recent highly related paper that the authors have probably missed and I recommend to cite: A K Z Tehrani I Rosado-Mendez, H Rivaz, (2022). Robust Scatterer Number Density Segmentation of Ultrasound Images, IEEE Trans. UFFC (TUFFC)
The authors should simplify the paper. The DSA module should be explained in more details. Why using B-mode in DSA and why B-mode of one imaging setting is used. Overally, some part of this paper is not clear.
- 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?
The paper has the merit of acceptance in MICCIA. Some of the modules such as meta-learning and DSA should be explained clearer.
- Number of papers in your stack
3
- 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.
The authors present a deep learning QUS approach for attenuation coefficient (AC) reconstruction which can be generalized to different probe geometries. Learning- based methods are generally trained on images or methods acquired with a single fixed probe geometry, and they thus fail to generalize to images acquired with different US probes or systems. This paper presents a promising idea to improve the generalization capabilities of such methods. The ablation experiments are rigorous and the results are convincing.
However, the clarity of the presentation should be improved by addressing the questions and concerns that the reviewers have raised. There are also other issues that should be addressed in the revision, or at least acknowledged as shortcomings. There is no comparison with non-deep learning approaches.
- 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).
3
Author Feedback
N/A