Abstract

Carotid atherosclerosis represents a significant health risk, with its early diagnosis primarily dependent on ultrasound-based assessments of carotid intima-media thickening. However, during carotid ultrasound screening, significant view variations cause style shifts, impairing content cues related to thickening, such as lumen anatomy, which introduces spurious correlations that hinder assessment. Therefore, we propose a novel causal-inspired method for assessing carotid intima-media thickening in frame-wise ultrasound videos, which focuses on two aspects: eliminating spurious correlations caused by style and enhancing causal content correlations. Specifically, we introduce a novel Spurious Correlation Elimination (SCE) module to remove non-causal style effects by enforcing prediction invariance with style perturbations. Simultaneously, we propose a Causal Equivalence Consolidation (CEC) module to strengthen causal content correlation through adversarial optimization during content randomization. Furthermore, we design a Causal Transition Augmentation (CTA) module to ensure smooth causal flow by integrating an auxiliary pathway with text prompts and connecting it through contrastive learning. The experimental results on our in-house carotid ultrasound video dataset achieved an accuracy of 86.93%, demonstrating the superior performance of the proposed method. Code is available at https://github.com/xielaobanyy/causal-imt.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1017_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/xielaobanyy/causal-imt

Link to the Dataset(s)

N/A

BibTex

@InProceedings{GaoShu_ACausalityInspired_MICCAI2025,
        author = { Gao, Shuo and Yang, Meng and Xue, Jun and Chen, Yang and Zhang, Jingyang and Zhou, Guangquan},
        title = { { A Causality-Inspired Model for Intima-Media Thickening Assessment in Ultrasound Videos } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},
        page = {13 -- 22}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a novel approach to disentangle view variations in the assessment of intima-media thickness using ultrasound. The authors introduce the SCE, CEC, and CTA modules to effectively address style variations, leveraging causal techniques.

  • Please list the major strengths of the paper: you should highlight 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 presents a novel contribution through a creative use of causal graphs to disentangle view and style variations, which is both original and well-aligned with the problem at hand.

    2. Overall, it is clearly written, with a strong motivation that highlights the clinical relevance and technical challenges of intima-media thickness assessment.

    3. The methods and results are well presented, and the accompanying figures are informative.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. The approach to modeling style perturbations using the mean (μ) and standard deviation (σ) terms is interesting. However, it would be helpful for the authors to clarify how well these perturbations reflect realistic style shifts encountered in actual ultrasound imaging. Some discussion or empirical justification of their realism would strengthen the manuscript.

    2. The CEC module would benefit from additional explanation. Specifically, it is unclear how randomizing the feature vectors leads to causal consolidation without compromising important style-related information. A clearer description of the underlying intuition and empirical validation would be helpful.

    3. In causal analysis, directed acyclic graphs (DAGs) are typically preferred to represent causal relationships. The CTA module introduces an auxiliary pathway with a double-arrowed connection for text prompts. A justification for this design choice would enhance clarity.

    4. In Table 2, the sensitivity scores for the SCE and CEC modules appear to be relatively low. It would be valuable for the authors to comment on this and discuss whether there are trade-offs between sensitivity and other performance metrics in their design.

    5. A brief discussion of how style variations display in the dataset used for training and evaluation would be appreciated. This would help contextualize the challenges being addressed and validate the choice of techniques.

    6. Finally, it would be interesting to see how the proposed method performs on other publicly available datasets, such as the one referenced here: https://doi.org/10.17671/gazibtd.804617. A comment on the generalizability of the approach to other ultrasound datasets would strengthen the paper’s impact.

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    Minor: Section 3 Experiments, Datasets ‘sensitivity’ is mentioned twice. Fig.5 caption needs more explanation.

  • 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.

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

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

    While the paper presents a novel and promising approach that creatively leverages causal graphs to address view and style variations in ultrasound-based intima-media thickness assessment, there were concerns regarding the clarity of some modules (e.g., CEC), the realism of style perturbations, and the justification for certain causal modeling choices. The rating will be changed depending on the rebuttal.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal clarifies certain unclear points that was unclear to me. I believe it meets the standards for acceptance at MICCAI.



Review #2

  • Please describe the contribution of the paper

    The paper proposes an innovative causality-inspired model that explains how to assess carotid intima-media thickening even in the presence of style interference.

    The design includes SCE, CEC, and CTA modules, which reduce irrelevant noise information, enhance the sensitivity to causal content, and improve the ability to capture fine-grained pathological features.

    Experiments demonstrate that this method shows its highest accuracy and clear advantages in IMT assessment on an internal dataset of carotid artery ultrasound videos.

  • Please list the major strengths of the paper: you should highlight 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 proposes a framework based on causal inference that analyzes the relationship between style changes and real pathological signals from the perspective of structured causal models, and subsequently designs targeted modules to eliminate spurious correlations.
    2. Designs Spurious Correlation Elimination, Causal Equivalence Consolidation, and Causal Transition Augmentation to achieve spurious correlation elimination and causal enhancement.
    3. The paper implements complementary multi-modal information. With weaker image features, this method uses textual information to provide more background knowledge support, especially for clinical applications.
    4. The proposed method achieved an accuracy of 86.93% in experiments on the internal carotid artery ultrasound video dataset, surpassing competing methods.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. The writing logic in the introduction and abstract needs to be improved.
    2. The introduction has many logical jumps. The introduction of spurious correlation and causal correlation lacks background and connection.
    3. The paper does not sufficiently discuss the limitations of previous methods for similar problems, nor does it provide enough evidence to prove the advantages of causal modeling. It is hard to understand the reasons for adopting causal methods compared with the existing methods.
    4. Although the paper mentions causal analysis, it does not explain why traditional methods cannot handle causal interference issues.
    5. While the paper introduces Structural Causal Models and describes causal chains, it doesn’t discuss the rationality of these causal assumptions in actual ultrasound images.
    6. The introduction of causal flow in the CTA module seems to introduce a new concept not mentioned in the problem description.
    7. This paper does not sufficiently explain how these three modules directly correspond to the motivation of extracting causal relationships and suppressing non-causal factors, making the necessity unconvincable.
    8. The impact of the weights between multiple loss terms on final performance has no analysis, affecting the evaluation of each module’s contribution.
    9. The CTA module introduces text prompts generated by chain thinking as assistance, but the paper lacks necessary ablation study to demonstrate the specific contribution of this part and the differences when using traditional text prompts or not using text prompts at all.
    10. The lack of experiments using only auxiliary text without contrastive loss, or only using contrastive loss without auxiliary text prompts, makes it hard to determine the contributions of auxiliary text prompts and contrastive loss to overall performance.
    11. Not displaying AUC affects the assessment of the model’s relative advantages and the overall persuasiveness of the results.
    12. Although the paper conducts a binary classification task, the number of positive and negative samples is missing.
    13. There is a typo in the learning rate: 6×105.
    14. Equation (9) should define the symbols (such as L_overall = ) on the left side, otherwise it’s incomplete. Also, equations (5) and (8) have typos where the periods should be commas.
  • Please rate the clarity and organization of this paper

    Poor

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • 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.

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

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

    This paper proposes a relatively novel framework in structural causal modeling and multimodal information fusion. However, although the paper introduces a causal model, it fails to clearly articulate differences from traditional methods and why causal modeling is necessary for this task. The three modules (SCE, CEC, CTA) do not establish a clear corresponding relationship with the motivation. The experimental section lacks necessary ablation studies and details. Data descriptions and important evaluation metric AUC are incomplete. Also, unclear writing logic, typos, and formula punctuation errors reduce the paper’s readability.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    Thank you to the authors for addressing some of the concerns I raised.

    However, several critical issues remain insufficiently resolved: the writing logic and organizational structure still lack clarity; multiple formatting and notation errors in the equations were not promised to be corrected; the motivation remains weak, with no clear justification provided for adopting a causal framework; the authors did not explain the absence of the key evaluation metric AUC, which is especially important given the highly imbalanced dataset; furthermore, the connection between the proposed modules and the stated causal motivation remains unclear.



Review #3

  • Please describe the contribution of the paper

    This paper uses ideas from causality to isolate content from style in Carotid Intima-Media Thickening estimation. They develop modules to make prediction invariant to spurious features and strongly correlated with causal features. They also use additional text inputs to further enhance this correlation.

  • Please list the major strengths of the paper: you should highlight 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.
    • Nice SCM in Fig 1
    • Excellent diagram in Fig 2
    • Very clearly written and maths seems correct
    • Good implementation details
    • Impressive empirical results and ablation study to demonstrate necessity of each component
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Pg 1 “ultrasound prob”
    • Very small point but in Eq. 6 I would put mu(f) at the beginning to make it clearer
    • Pg 6 “we adpot”
  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • 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.

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

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

    This is a very well written paper that integrates a range of different ideas to form an interesting and effective model. I cannot give it a 6, however, since I am not familiar with work on ultrasound.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A




Author Feedback

We thank the reviewers for their valuable comments. They highlight that our method is with good novelty (R1&R2&R3), very clearly written with a strong motivation (R1&R3), well-presented results with informative figure (R1&R3).

Q1. Description on style shifts (R1) During carotid US scanning, the probe is maneuvered along the carotid artery with different scanning orientations and pressures, which affect the US propagation path and thus cause changes in echo intensity. These echo discrepancies alter the overall characteristics of each frame, referred to as style shifts, such as brightness variations with uneven illumination varying across frames and also contrast variations blending the intima-media layer with the surrounding carotid muscle.

Q2. Justification for modeling style perturbation via mean and standard deviation in SCE (R1) The mean term captures the overall brightness of the image by representing the average activation across spatial locations, while the standard deviation reflects bright-dark contrast by measuring intensity dispersion. Since brightness and contrast constitute the primary style shifts in carotid US images (see Q1), these statistics are well-suited for style manipulation.

Q3. Clarification on content randomization in CEC (R1) The main idea for enhancing content causal correlation is to enforce the equivalence between the content factor and prediction result. Specifically, we perturb the content factor by shuffling the feature channels, while preserving style information by re-injecting the original style-related statistics (i.e., mean and standard deviation) via instance normalization. Furthermore, we adopt adversarial optimization to generate uninformative predictions from the content-randomized features, consolidating content-prediction equivalence and thus enhancing causal correlation.

Q4. Justification for bidirectional pathway between text prompts and image features in CTA (R1) for enhancing causal impact flow (R2) Text prompts provide fine-grained symptom descriptions to guide image features (T->F), facilitating causal flow from content to prediction (C->F->Y). Conversely, image features calibrate the embedding process of text prompts (F->T), ensuring the textual cues are well-aligned with the causal reasoning required for the task. Overall, this bidirectional connection complements the strengths of image feature enhancement and text embedding modulation, outperforming unidirectional methods.

Q5. Limitations of traditional methods and motivation of our causality-inspired model (R2) Traditional IMT assessment methods often assume minimal probe view variation, which is inconsistent with the significant view changes in carotid US scanning. This leads to style shifts (see Q1) that should be theoretically independent of diagnostic inference but are, in practice, intertwined with diagnosis-relevant content cues (e.g., lumen anatomy). In this regard, these traditional methods are often biased by spurious correlations with style shifts, while neglecting causal correlations with content cues. This motivates us to design a causal model that mitigates spurious correlations through SCE and strengthens causal correlations via CEC and CTA. For more details, please refer to the first paragraph on page 2 in the manuscript.

Q6. Data description on positive/negative samples (R2) Our dataset contains 30 thickened (positive) and 90 non-thickened (negative) samples.

Q7. Trade-off between accuracy and sensitivity (R1) In Table 2, combining SCE and CEC yields high accuracy but low sensitivity, reflecting the trade-off in our class-imbalanced dataset (see Q6). The high accuracy compared to existing methods in Table 1 shows the advantage of SCE and CEC, while the reduced sensitivity emphasizes the importance of the CTA module.

Q8. Ablation study on CTA (R2) We conducted the CTA ablation study in Table 2 by removing the contrastive loss, as CTA is incorporated via contrastive learning to integrate text prompts.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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



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’

    Despite one skeptical reviewer, I believe this paper will be of interest to miccai audience. I encourage authors to pay attention to the post-rebuttal comments, especially to the remaining issues with formatting and notation.



Meta-review #3

  • 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



back to top