List of Papers Browse by Subject Areas Author List
Abstract
Transesophageal echocardiography (TEE) plays a pivotal role in cardiology for diagnostic and interventional procedures. However, using it effectively requires extensive training due to the intricate nature of image acquisition and interpretation. To enhance the efficiency of novice sonographers and reduce variability in scan acquisitions, we propose a novel ultrasound (US) navigation assistance method based on contrastive learning as goal-conditioned reinforcement learning (GCRL). We augment the previous framework using a novel contrastive patient batching method (CPB) and a data-augmented contrastive loss, both of which we demonstrate are essential to ensure generalization to anatomical variations across patients. The proposed framework enables navigation to both standard diagnostic as well as intricate interventional views with a single model. Our method was developed with a large dataset of 789 patients and obtained an average error of 6.56 mm in position and 9.36 degrees in angle on a testing dataset of 140 patients, which is competitive or superior to models trained on individual views. Furthermore, we quantitatively validate our method’s ability to navigate to interventional views such as the Left Atrial Appendage (LAA) view used in LAA closure. Our approach holds promise in providing valuable guidance during transesophageal ultrasound examinations, contributing to the advancement of skill acquisition for cardiac ultrasound practitioners.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1889_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1889_supp.pdf
Link to the Code Repository
N/A
Link to the Dataset(s)
https://www.cancerimagingarchive.net/collection/lidc-idri/
https://www.kaggle.com/datasets/andrewmvd/pulmonary-embolism-in-ct-images
BibTex
@InProceedings{Ama_Goalconditioned_MICCAI2024,
author = { Amadou, Abdoul Aziz and Singh, Vivek and Ghesu, Florin C. and Kim, Young-Ho and Stanciulescu, Laura and Sai, Harshitha P. and Sharma, Puneet and Young, Alistair and Rajani, Ronak and Rhode, Kawal},
title = { { Goal-conditioned reinforcement learning for ultrasound navigation guidance } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15011},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper studies ultrasound navigation for transesophageal echocardiography imaging using reinforcement learning (RL). It is based on Contrastive RL and proposes two improvements, contrastive patient batching and contrastive data augmentation loss. Experiments in built simulation environments show the advantage of the proposed method over existing RL methods and baselines, including novel view navigation.
- 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 techniques, including contrastive patient batching and data augmentation loss, improve the basslines.
- Experiments show that the proposed existing off-policy RL method and an RL algorithm for ultrasound navigation on two datasets with 140 and 5 patients.
- Arbitrary views can be achieved besides standard views.
- 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 technical novelty compared to the vanilla contrastive RL is relatively limited. The main difference is the contrastive patient batching. The paper should give a more detailed analysis of the proposed strategy.
- For the results, given such large standard deviations, the paper should provide a statistical analysis.
- Given predefined trajectories, other imitation learning algorithms should be included.
- From the provided video results, it seems the system struggles to stop the navigation.
- Please rate the clarity and organization of this paper
Very 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?
Some of the training details are missing. The release of the code and simulation environment is expected to reproduce the experiments.
- 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
- For the data augmentation, what if image augmentation is used during training, but without constructing the K^2 matrices? The paper should also discuss the number of K. The paper should also provide more details on data augmentation.
- The paper should elaborate on whether the initial poses for different methods are the same.
- For the video, the paper should include comparisons with other methods.
- 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?
The paper works on an important ultrasound navigation task and uses the RL formulation to solve it. It proposes two techniques to improve the RL baseline, which is examined on two datasets with varied patient numbers and view settings. The technical novelty and the experiments should be further discussed and strengthened.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 authors are proposing a new reinforcement learning strategy aiming to perform ultrasound navigation guidance. For that, the authors claim that are generating realistic ultrasound images from CT volumes, using a triplet of (observation, action, goal) to generate the US view/position. The presented state-of-the-art is well organized and recent. The authors claim that they are the first to attempt to develop an ultrasound navigation model capable of navigation to arbitrary views given a specific goal.
- 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 written, though its organization could use some refinement. The chosen topic aligns well with the conference’s objectives. In my view, the authors correctly highlight the contributions of their work, including: (i) introducing a novel method for simulated TEE guidance; (ii) extending the CRL framework; and (iii) conducting experiments and validation.
From my perspective, the primary contribution lies in the design of the simulated TEE guidance, offering a valuable technical resource to enhance current simulation models.
- 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.
Although I agree with the authors’ assertion, I also believe that this work represents an initial attemp. Several factors were ignored,in particular the dynamic information. Evaluating US images inherently involves dynamics, and without incorporating this aspect, it’s impossible to simulate all relevant clinical analyses accurately. Additionally, I have reservations regarding the image quality of the generated US. Upon reviewing the supplementary materials, I found these representations to be unrealistic, despite their anatomical accuracy.
The paper’s organization requires improvement. While I understood the intention behind the last paragraph of the introduction, many of the terms mentioned therein were not introduced earlier. Perhaps relocating this paragraph to the discussion section would be beneficial.
Regarding the results, more image examples are necessary. In the section on “Interventional view navigation,” the authors mention utilizing the FUMPE dataset and reference additional LAA segmentations. It’s unclear to me whether the FUMPE dataset includes LAA segmentation or if these were generated by the team.
In the concluding statement, the authors propose that “Using this method as a guidance system could help train sonographers, improve the acquisition quality, and reduce variability among experienced users.” However, the approach to achieving this is unclear. How do they propose to conduct guidance training solely using 2D TEE, considering its limited field of view? The inclusion of fluoroscopy seems mandatory for such simulations. Furthermore, their strategy for reducing variability among users warrants clarification.
- Please rate the clarity and organization of this paper
Very 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
In a future work, I suggest that the authors consider expanding this framework to incorporate 3D volumes, as it appears entirely achievable based on the existing model. Moreover, I’m curious if this model is also compatible with external probes (not TEE). Please include this point in your discussion.
- 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 is focused on a relevant topic. The paper technically sound good. The validation is relevant for a proof of concept. I see novelty on this work.
- 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 #3
- Please describe the contribution of the paper
The paper introduces a novel ultrasound navigation assistance technique that incorporates contrastive learning within a goal-conditioned reinforcement learning (GCRL) framework. The new framework enables navigation to arbitrary views. It integrates a new contrastive patient batching method (CPB) and a data-augmented contrastive loss. These components are crucial in enhancing the generalization of the model to account for anatomical variations among different patients. The proposed method outperforms competing methods in terms of pose accuracy.
- 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.
Innovative Approach: The paper introduces a novel ultrasound navigation assistance technique that incorporates contrastive learning within a GCRL framework to navigate sonographers to arbitrary planes they want.
Novel Mechanisms: It integrates a new CPB method and a data-augmented contrastive loss. These components are crucial in enhancing the generalization of the model to account for anatomical variations among different patients.
- 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 is that only synthetic data are used for evaluation. It might be acceptable because of the difficulty of obtaining real 3D TEE data along with the ground truth of the poses. Another weakness may be the absence of subjective evaluation from the clinical side. While their method has demonstrated superior performance compared to competing methods, it is uncertain whether it is accurate enough for use in real-world scenarios. What range of accuracy would be considered acceptable by sonographers?
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Do you have any additional comments regarding the paper’s reproducibility?
No
- 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
Add subjective evaluation from the clinical side. I hypothesise that clinicians would not particularly care about certain nuances of pose estimation accuracy. This could indicate what is “good enough” in terms of an output.
- 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 contibutes to an important problem in TEE or even ultrasound navigation. According to the performance reported in the paper, the proposed method can be useful. The paper is well written and structured.
- Reviewer confidence
Somewhat confident (2)
- [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
Author Feedback
We would like to thank the reviewers for their comments and constructive feedback. All the reviewers have acknowledged the novelty of this work for ultrasound acquisition guidance. For brevity, we have grouped reviewers’ questions and provided a response. We will subsequently update the manuscript by including these clarifications.
Dataset and simulation:
We recognise that the lack of dynamic information such as cardiac motion is a limitation of this study, and the ultrasound image quality can be further improved by making it appear more realistic. We indeed plan to address these topics in our future work. In this work, we emphasized the simulator’s correctness and consistency w.r.t anatomical structures and studied whether an agent can be trained to learn this context and navigate to desired targets. To facilitate this, all segmentations were obtained using an algorithm, which is referenced in the supplementary material.
We agree with the reviewers that this framework could be extended to work with 3D volumes as well as handle other ultrasound modalities such as TTE or ICE.
Evaluation:
When testing the model, the initial poses are randomly sampled, hence they are different for every method. For every view in each patient dataset, we sample 5 initial poses. While a high pose error indicates that the desired view is not reached, a lower error indeed does not imply clinical usability. Further analysis is required to correlate the error values to clinically acceptable ranges. We believe that our simulation pipeline can be used for this purpose.
Regarding the number of data augmentations K, we’ve added results for K = 1 and K = 4 in the supplementary material. Our choice of K=2 yields better results, showing that the regularisation induced by the data augmentation loss is beneficial. At higher values of K, the performance gain saturates as the variability in the distribution caused by the additional random shifts is limited.
Applications of the proposed method: Several works [2, 3] have demonstrated the benefits of using AI to guide ultrasound image acquisition in reducing the variability among users and helping novices acquire diagnostic-quality images. In the manuscript, we used the term guidance to refer to this process, rather than interventional guidance, which is typically coupled with fluoroscopy. We will update the manuscript to clarify this.
Meta-Review
Meta-review not available, early accepted paper.