Abstract

Dual-screen contrast-enhanced ultrasound (CEUS) has been the first-line imaging techniques for the differential diagnosis of primary liver cancer (PLC), since the imaging of tumor micro-circulation perfusion as well as anatomic features of B-mode ultrasound (BUS) view. Although previous multi-view learning methods have shown their potential to boost diagnostic efficacy, correlation variances of different views among subjects are largely underestimated, arising from the varying imaging quality of different views and the presence of valuable findings or not. In this paper, we propose a correlation-adaptive multi-view fusion method (CAMVF) for dual-screen CEUS based PLC diagnosis. Towards a reliable fusion of multi-view CEUS findings (i.e., BUS, CEUS and its parametric imaging), our method dynamically assesses the correlation of each view based on the prediction confidence itself and prediction consistency among views. Specifically, we first obtain the confidence of each view with evidence-based uncertainty estimation, then divide them into credible and incredible views based on cross-view consistency, and finally ensemble views with weights adaptive to their credibility. In this retrospective study, we collected CEUS imaging from 238 liver cancer patients in total, and our method achieves the superior diagnostic accuracy and specificity of 88.33% and 92.48%, respectively, demonstrating its efficacy for PLC differential diagnosis. Our code is available at https://github.com/shukangzh/CAMVF.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0746_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

https://github.com/shukangzh/CAMVF

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Wan_Correlationadaptive_MICCAI2024,
        author = { Wan, Peng and Zhang, Shukang and Shao, Wei and Zhao, Junyong and Yang, Yinkai and Kong, Wentao and Xue, Haiyan and Zhang, Daoqiang},
        title = { { Correlation-adaptive Multi-view CEUS Fusion for Liver Cancer Diagnosis } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15005},
        month = {October},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a novel approach in the field of hepatocellular carcinoma diagnosis for multi-view CEUS fusion. The core of this approach is to dynamically measure the weights of the views by considering both the uncertainty of the views and the consistency among them. View consistency is considered compared to other multi-view fusion methods, thus reducing the risk of being biased by a view with very low relevance.

  • 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) The paper effectively mitigates the problem that DST may produce counter-intuitive results when dealing with highly contradictory evidence by calculating the semantic consistency between views and categorizing them into credible and non-credible views. (2) This paper is well-written and well-organized.

  • 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 novelty of this paper may be limited. Although using DST theory to measure view uncertainty seems to a novel point, it is not the first work to do so. It is only the first application of this method in this field. The use of DST theory to measure uncertainty in medical image analysis first appears in this article [1], and I have not seen any relevant references and discussions compared to this paper. (2) Lack further considerations in terms of view consistency. When only one view is “good” and all other views are perfusion features, and they are not valuable for PLC diagnosis. In this case, a “good” view being assigned a lower weight because of its low consistency with other views. The author might have ignored this situation. (3) The rational of why DST was used to measure view uncertainty is unclear. There are numerous ways to measure view uncertainty, and it is not stated in this paper why DST was selected rather than other methods.

    [1] Zou, Ke, Xuedong Yuan, Xiaojing Shen, Meng Wang, and Huazhu Fu. “Tbrats: Trusted brain tumor segmentation.” In International Conference on Medical Image Computing an d Computer-Assisted Intervention, pp. 503-513. Cham: Springer Nature Switzerland, 2022.

  • 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?

    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

    (1) Some additional details were needed. For example, “why the DST metric of view uncertainty was used rather than other methods.” Since it is about the integrity of the paper, the methods mentioned in the paper should say why and what the benefits are. (2) Please add the experiment of using as hyperparameters the thresholds for whether a view is considered to be a credible view or a non-credible view. (3) Existing experiments are conducted on the in-house dataset, it would be better to verify the effectiveness on publicly available dataset to further prove the generalization performance of the proposed method.

  • 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 Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper chooses to use DST to measure view uncertainty, but fails to state why this method is used and not others. It is the foundation of the thesis to state the rationale for the use of each method used in the thesis.Moreover, the extreme case where there is only one “good” view and all other views are “bad” views is not considered, in which case the “good” view is assigned a lower weight.

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    The author has addressed most of my concerns, and i am glad to raise my score.



Review #2

  • Please describe the contribution of the paper

    The authors propose a correlation-adaptive multiview fusion method (CAMVF) for dual-screen CEUS based PLC differential diagnosis. Towards a reliable fusion of multi-view CEUS findings (i.e., BUS, CEUS and its parametric imaging), their method dynamically assesses the correlation of each view for individual subject. Methodologically, view correlation is jointly determined by the prediction confidence of view itself and prediction consistency among views. Specifically, they first obtain the confidence of each view with evidence-based uncertainty estimation, then divide them into credible and incredible views based on cross-view consistency, and finally ensemble view at an evidence level with weights adaptive to their credibility. In this retrospective study, they collected CEUS imaging from 238 liver cancer patients in total, and their method achieves the superior diagnostic accuracy and specificity, demonstrating its efficacy for PLC differential diagnosis.

  • 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’s key strength lies in its inventive approach to correlation-adaptive multi-view fusion for liver cancer diagnosis using CEUS imaging. By dynamically combining CEUS views and weighting them based on credibility, determined through evidence-based uncertainty estimation and cross-view consistency assessment, the method effectively addresses the challenge of integrating multiple views, resulting in improved diagnostic accuracy and specificity. Additionally, the proposed method is evaluated through a retrospective study involving CEUS imaging from 238 liver cancer patients, demonstrating superior diagnostic accuracy and specificity.

  • 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) Figure 1 should be more detailed. 2) Lack of clarity in defining the problem statement and research objectives. 3) Insufficient explanation of the methodology used for correlation assessment and view fusion. 4) Limited discussion on the dataset used and its representativeness. 5) Inadequate comparison with existing state-of-the-art methods in the field. 6) Absence of detailed information on the model architecture and parameters. 7) Limited explanation of how the proposed method addresses potential biases in the data. 8) Insufficient exploration of potential limitations or constraints of the proposed method.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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?

    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

    The authors propose a correlation-adaptive multiview fusion method (CAMVF) for dual-screen CEUS based PLC differential diagnosis. Towards a reliable fusion of multi-view CEUS findings (i.e., BUS, CEUS and its parametric imaging), their method dynamically assesses the correlation of each view for individual subject. Methodologically, view correlation is jointly determined by the prediction confidence of view itself and prediction consistency among views. Specifically, they first obtain the confidence of each view with evidence-based uncertainty estimation, then divide them into credible and incredible views based on cross-view consistency, and finally ensemble view at an evidence level with weights adaptive to their credibility. In this retrospective study, they collected CEUS imaging from 238 liver cancer patients in total, and their method achieves the superior diagnostic accuracy and specificity of 88.33% and 92.48%, respectively, demonstrating its efficacy for PLC differential diagnosis. Many shortcomings must be treated to ameliorate the manuscript: 1) Figure 1 should be more detailed. 2) Lack of clarity in defining the problem statement and research objectives. 3) Insufficient explanation of the methodology used for correlation assessment and view fusion. 4) Limited discussion on the dataset used and its representativeness. 5) Inadequate comparison with existing state-of-the-art methods in the field. 6) Absence of detailed information on the model architecture and parameters. 7) Limited explanation of how the proposed method addresses potential biases in the data. 8) Insufficient exploration of potential limitations or constraints of the proposed method.

  • 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 recommendation of weak accept is based on several factors that contribute to the paper’s overall quality and potential impact:

    Innovative Methodology: The paper introduces a novel correlation-adaptive multiview fusion method (CAMVF) for PLC (Primary Liver Cancer) differential diagnosis using CEUS (Contrast-Enhanced Ultrasound) imaging. This innovative approach demonstrates the authors’ contribution to advancing diagnostic techniques in the field. Clear Description of Methodology: Despite some shortcomings in methodology explanation, the paper provides a clear overview of the CAMVF method’s principles, including how it dynamically assesses the correlation of each view and how it combines CEUS views based on credibility weights. Clinical Relevance: The study’s clinical viability is supported by a retrospective analysis involving CEUS imaging from 238 liver cancer patients. The reported diagnostic accuracy and specificity results (88.33% and 92.48%, respectively) indicate the potenti

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    The authors have addressed most of my comments, I maintain the weakl accept decision.



Review #3

  • Please describe the contribution of the paper

    This paper proposed a multi-view fusion method for Dual-screen contrast-enhanced ultrasound (CEUS) to differential diagnosis of HCC and ICC.

  • 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. Correlation-adaptive fusion of multi-view CEUS findings.
    2. Dynamically assesses the correlation of each view with evidence-based uncertainty estimation.
  • 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 effectiveness of the proposed method may have certain limitations for different ultrasound scanner.

  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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

    The abstract section is too long and does not clearly describe the main ideas and innovative points.

  • 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 proposed correlation-adaptive multi-view fusion method for CEUS is interesting and novel.

  • 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

    Accept — should be accepted, independent of rebuttal (5)

  • [Post rebuttal] Please justify your decision

    The authors have addressed concerns.




Author Feedback

We thank all reviewers for their comments. Due to space limit, we only respond to major comments here and will address all comments in final paper. For R#1 and R#4 “addresses potential biases in the data…” “potential limitations or constraints…” “When only one view is “good” and all other views are perfusion features…” == The potential biases mainly refer to view correlation inferred from collected data. From Fig. 2, we find it difficult to determine a definitive ranking of the views in terms of their diagnostic value. Additionally, we have set a hyper-parameter, temperature factor, to adjust the sensitivity to inter-view consistency for weight calculation. By setting a relatively large temperature factor, inter-view differences could be narrowed, avoiding an excessive emphasis on the single or certain modalities and allowing the model to receive more gradient information from various modalities. Furthermore, the optimal value of temperature factor is determined by cross-validation on the training set.
== We admit the limited discussion on the extreme case where only one view is “good” and all others are “bad” views with invaluable feature. In this case, the weight of the only“good”view could be reduced, and the final prediction could be biased. Nonetheless, we can identify the abnormal condition in a simple way. Specifically, the average distance among views, theta, which functions as a threshold for credible/incredible view split, is likely larger than those of more typical scenarios. As expected, the only“good”view is distant from the remaining“bad”views, and the distances among“bad”views tend not to be small with nearly random predictions. Thus, we can perform a statistical analysis of the average distance among views (theta) and detect the abnormal condition when its value exceeds a predefined limit, such as the third quartile. Furthermore, clinicians are required to re-evaluate the model predictions for such extreme cases. “why the DST metric of view uncertainty was used.” “Insufficient explanation of the methodology” == Towards trustworthy prediction, uncertainty quantification in deep learning can be roughly divided into Bayesian neural networks (BNN), deep ensemble learning, and evidential deep learning. The first line of method, based on BNN, replaces deterministic parameters with distributions over weights, allowing the model to output the distribution of predictions and their uncertainty. However, explicit modeling of distributions over weights is computationally expensive and multi-view uncertainty quantification tends to be hard to converge. Ensemble-based methods integrate multiple deep models to assess the uncertainty of predictions, resulting in an increased number of trainable parameters, which is particularly problematic in multi-view learning scenarios. Hence, we resort to evidential deep learning, which introduces Dempster-Shafer Evidence Theory (DST) to directly model prediction uncertainty using beliefs from different views. Subjective logic theory in DST allows for a much more efficient way to model the prediction with Dirichlet distribution. The parameters of the Dirichlet distribution (beliefs) can be obtained by replacing the softmax operator with a non-negative activation (ReLU) and optimized by a modified cross-entropy loss. For the application of liver cancer aided diagnosis, DST metric of view uncertainty is significantly more feasible due to its substantially reduced computational complexity.
“add the experiment of using as hyper-parameters the thresholds.” “verify the effectiveness on publicly available dataset.” “Inadequate comparison with existing state-of-the-art methods…” == We acknowledge suggestions for further experiments and consider them for future work. For R#1 “Absence of detailed information…” ==To present the process of feature extraction, model construction and training, we will clarify more details in the revision and publish the project code.




Meta-Review

Meta-review #1

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A

  • 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).

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    All reviewers agreed on acceptance.

  • 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).

    All reviewers agreed on acceptance.



back to top