List of Papers Browse by Subject Areas Author List
Abstract
Causal learning is crucial for echocardiography quality assessment (EQA), demonstrating strong potential for learning fundamental quality representations. However, confounded quality degradations in EQA lead to spurious correlations and feature entanglement, obscuring causal relationships in quality modeling. In this paper, we propose a Geometric Causal Aggregation Network (GeoCAN) that embeds nonlinear causal–geometric learning into EQA to explicitly model relational causality and improve robustness under complex degradations. Specifically, Causal-Geometric Learning (CGL) deconfounds quality factors from visual degradations, thereby enhancing the discriminability of causal representations and suppressing spurious correlations. Nonlinear Relational Aggregation (NRA) disentangles quality representations from relational structures, explicitly separating relational cues from local visual features and alleviating feature entanglement. Extensive experiments on two public datasets and a private clinical dataset demonstrate that GeoCAN consistently achieves state-of-the-art performance, validating nonlinear causal–geometric learning as a principled solution for robust and reliable EQA. The code is available at https://github.com/Yiran0325/Echo-GeoCAN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0202_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Yiran0325/Echo-GeoCAN
Link to the Dataset(s)
CACTUS dataset: https://users.encs.concordia.ca/~kadem/cactus/
BibTex
@InProceedings{LiYir_GeoCAN_MICCAI2026,
author = { Li, Yiran AND Zheng, Kai AND Li, Shuo},
title = { { GeoCAN: Nonlinear Causal-Geometric Learning for Echocardiography Quality Assessment } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This Paper introduces GeoCAN, a Geometric Causal Aggregation Network designed for echocardiographic quality assessment. Its core idea is to face the spurious correlation between quality factors and visual degradation factors. By introducing CGL embedding visual factors into a hyperbolic space and extracting causal interactions via spectral decomposition, while Nonlinear Relational Aggregation (NRA) decouples relational structures from quality representations based on KAN spline basis functions. GeoCAN was tested in 2 open source datasets and a private dataset. GeoCAN demonstrates that this method achieves improvements over baseline models.
- 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 overall design of the CGL and NRA modules is novel, and the authors have demonstrated that the discriminability enhancement of CGL by integrating latent hyperbolic learning into causal modeling to deconfound quality factors from visual degradations, and the quality representation under complex degradation conditions improving by NRA
2.The Experiment section is solidified and the evaluation has broad coverage. Three datasets constitute a relatively comprehensive multi-distribution test. The combination of multi-granularity metrics such as KRCC, SRCC, and PLCC is also more persuasive than a single metric.
- 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.Cij, as defined in Eq. (1), is merely a conditional ratio obtained by element-wise multiplication of feature maps followed by GLM normalization; it is not equivalent to the causal effect as defined in the Pearl framework or structural causal models (SCM). It’s fundamentally an attention reweighting mechanism. Without the constraint of an external explicit causal graph, the model may learn an “asymmetric matrix that minimizes the classification loss” rather than the true causal direction.
2.No ablation on hyperbolic space: If the author believes that hyperbolic space is the core mechanism that significantly outperforms traditional Eulerian embedding in deconfound quality factors from visual degradations, they should have designed an ablation study on hyperbolic space to validate its contribution.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
There’s a ? citing issue in the Introduction section, revise it.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.Novel Model Design 2.Detailed experimental setup
- Reviewer confidence
Somewhat confident (2)
- [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.
Accept
Review #2
- Please describe the contribution of the paper
This paper proposes GeoCAN (Geometric Causal Aggregation Network) for echocardiography quality assessment (EQA).
- 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 identifies two specific and well-motivated pain points in EQA — spurious correlations and feature entanglement — and designs CGL and NRA respectively to address them. The logical chain from problem to solution is coherent and easy to follow. 2.The attention visualizations in Fig. 3 show the model focusing on anatomically meaningful structures such as chamber boundaries and myocardial regions rather than artifacts, consistent with the paper’s claims about causally grounded representations.
- 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 paper repeatedly claims that CGL achieves “causal deconfounding,” but the quantity Cij=P(F(i)∣F(j))C_{ij} = P(F^{(i)} | F^{(j)})Cij=P(F(i)∣F(j)) in Eq. (1) is fundamentally a conditional dependency estimate, not an interventional causal effect under Pearl’s do-calculus framework. There is a conceptual gap between conditional probability and causal effect. The paper provides no interventional experiments, counterfactual analyses, or do-calculus-based justification to support genuine causal claims. This constitutes an overclaim.
2.Necessity of hyperbolic embedding not sufficiently validated. The paper embeds factor nodes into hyperbolic space and claims this preserves hierarchical causal topology, but 1) No ablation comparing hyperbolic vs. Euclidean embedding is provided; 2) No sensitivity analysis of the curvature parameter cc c is reported.
3.Eq. (2) introduces the GLM as a nonlinear aggregation operator, but no explanation is given for why GLM is preferable to standard softmax attention or other aggregation mechanisms. No ablation comparing GLM to alternatives is provided, and the learning stability of parameters α and β is not discussed.
4.The dimension-wise functional decomposition in Eq. (6) using learnable piecewise spline basis functions is formally very similar to Kolmogorov-Arnold Networks (KAN) [12]. While [12] is cited, the paper does not clarify whether NRA is an instantiation of KAN, nor does it articulate the specific improvements over directly applying KAN. This leaves the contribution boundary unclear.
5.The private dataset Echo-Q lacks description of annotation protocol, acquisition equipment, and institutional source, making it impossible to assess its representativeness.
6.The retraining of baselines does not describe hyperparameter search strategies, leaving open the risk of unfair tuning.
7.Several baseline entries in the result tables are missing (“-“) without explanation — it is unclear whether the method does not support that task or simply failed to converge.
8.A “[?]” placeholder reference appears on page 1, indicating an incomplete citation at submission time.
9.Several sentences contain redundant phrasing (e.g., “To sum up the above limitations” appears more than once).
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
GeoCAN presents an interesting framework combining hyperbolic geometry with causal modeling for EQA, and the experimental results consistently outperform baselines across three datasets. However, several concerns weaken the paper’s credibility: the core “causal” claims are not rigorously grounded (conditional dependency ≠ causal effect); the necessity of hyperbolic embedding is unvalidated by ablation; the relationship between NRA and KAN is ambiguous; the private dataset lacks an ethics statement; and there are notable writing oversights (the “[?]” citation, the suppressed title).
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
After reading the rebuttal, I appreciate the additional clarifications. However, the main methodological concerns remain unresolved.
Most importantly, the paper continues to make strong causal claims (causal deconfounding, causal interaction), while Eq. (1) fundamentally models directional conditional dependency rather than interventional causal effect. The rebuttal reframes this as modeling causal tendencies or causal disposition, but no intervention-based analysis, counterfactual validation, or causal-theoretic justification is provided.
Several key design choices also remain insufficiently validated. The necessity of hyperbolic embedding is still unsupported by formal ablation studies in the paper itself, and the claimed advantages of GLM over standard aggregation mechanisms are not validated through direct comparisons.
In addition, the relationship between NRA and KAN remains unclear. Although the rebuttal acknowledges KAN-style spline decomposition, it does not clearly establish the methodological novelty beyond existing KAN formulations.
Concerns regarding dataset transparency and reproducibility are also only partially addressed, particularly for the private Echo-Q dataset and baseline retraining details. Overall, although the work is interesting and experimentally promising, I do not believe the rebuttal sufficiently resolves the conceptual and methodological concerns necessary for acceptance.
Review #3
- Please describe the contribution of the paper
This paper proposes GeoCAN, a first nonlinear causal-geometric learning framework for echocardiography quality assessment. It features a Causal-Geometric Learning (CGL) module that integrates latent hyperbolic learning into causal modeling to de-confound quality factors from visual degradations, and a Nonlinear Relational Aggregation (NRA) mechanism that disentangles relational structures from quality representations.
- 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.This paper firstly introduce causal-geometric learning framework into echocardiography quality assessment, which is an interesting attempt. 2.The method proposed in this article has a good theoretical foundation. 3.The experimental results have demonstrated the effectiveness of the method and supported motivation.
- 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.Bad representation: i) Reference 18 is same as Reference 19.ii) On the first page, Introduction section, line 10, there is a reference error “[10,?]”. Please check the paper carefully to avoid these problems. 2.The paper claims the nonlinear relational aggregation mechanism can explicitly disentangles relational structures from quality representations, while the final performance of the model is still black box, without any demonstration of the effectiveness of decoupling in the middle layer.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This article proposes a novel non-linear causal-geometric learning framework with good theoretical support, although there are some shortcomings in other parts, it can be accepted by Weak overall.
- 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.
Figure 3 can only partially solve my problem. It is only a display of the final effect and cannot reflect the internal logic of the model’s operation. However, there is still a good novelty in this article, and the authors promise to open source. In summary, I maintain my decision to wear accept unchanged.
Author Feedback
Thanks to the reviewers for your valuable comments on our work. Especially, thanks to R1 and R3 for accepting the paper directly. We appreciate your recognition of: 1.Novelty (R1: “Novel model design”; R2: “an interesting framework”; R3: “an interesting attempt”) 2.Good theoretical support (R2: “specific and well-motivated pain points”; R3: “good theoretical foundation”) 3.Experimental effectiveness (R1: “detailed experimental setup” “the experiment section is solidified” “the evaluation has broad coverage”; R2: “consistently outperform baselines” R3: “demonstrated the effectiveness of the method and supported motivation”)
We will release the code, datasets, implementation details, and experimental details for reproducibility, while correcting typos/presentation issues to address all constructive comments (R1, R2, R3).
Q1: Novelty of CGL (R1/R2).
- CGL is a novel advance in causal-disposition modeling for EQA. Inspired by multiple studies without explicit causal graphs, e.g., [1,13, DOI: 10.52202/068431-2776], it infers causal tendencies between quality factors and visual degradations from directional conditional support under limited causal supervision.
- CGL infers causal direction from Cij/Cji asymmetric disposition built on directed conditional dependencies and organizes it into latent hyperbolic causal topology for causal-disposition-guided deconfounding, rather than learning a free classification-driven matrix.
Q2: Hyperbolic-embedding ablations (R1/R2).
- Hyperbolic-embedding ablations were left out due to space limitations. Euclidean replacement weakens quality-order consistency, while hyperbolic geometry yields average QA/KRCC gains of 3.94 percentage points/0.0801, supporting hierarchical asymmetric causal topology.
- Curvature analysis had indicated c=1.0 as a balanced hyperbolic scale for Lgeo: small c approaches Euclidean geometry, while overly large c may over-compress embeddings, causing an average 11.56% KRCC drop.
- Tables 3/4 further support CGL’s hyperbolic topology modeling. Removing CGL drops ranking-correlation metrics, showing its necessity for stable causal-geometric representations.
Q3: Novelty of GLM (R2).
- GLM has two advantages over softmax attention and other aggregation mechanisms: adaptive distribution sensitivity and stable order-aware pooling. Learnable α/β shift it from min- to mean- to max-like behavior, capturing weak-to-dominant activations for conditional-dependency estimation.
- The learning stability of α/β is theoretically supported by [14]. GLM is bounded within the input range, reduces to arithmetic mean when α→1, and is differentiable w.r.t. α/β.
Q4: Relation between NRA and KAN (R2).
- NRA advances KAN-style spline decomposition from generic nonlinear mapping to relation-aware aggregation, addressing feature entanglement by explicitly separating relational structures from quality representations.
Q5: Clarification about “-” entries, Echo-Q, and baseline retraining (R2).
- “-” denotes unsupported task/output settings rather than convergence failure. Specifically, single-task IQA baselines and w/o MTL variants lack the CA output head.
- Echo-Q was collected from routine echocardiography using Philips systems. Three expert sonographers annotated de-identified images under rigorous quality control.
- Baselines followed GeoCAN’s experimental protocol, with hyperparameters selected only on the validation set to avoid unfair tuning.
Q6: Clarification about disentanglement of NRA (R3).
- Tab. 3/4 support the disentanglement effect of NRA in stabilizing quality representations. Its removal causes drops in correlation metrics (PLCC/SRCC/KRCC), indicating weakened consistency between predicted and ground-truth quality scores.
- Fig. 3 supports the disentanglement effect by visualizing GeoCAN, which focuses on structured anatomical relations rather than isolated local activations or superficial degradations, consistent with key anatomical structures and quality-determining factors.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. 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 develops the novel GeoCAN framework that integrates hyperbolic geometry and causal modeling for EQA, with comprehensive experimental settings and consistent performance gains over baselines across three datasets. However, the reviewers are concernd with its core causal assertions, the necessity of hyperbolic embedding, missing ethical approval for its private dataset and noticeable writing errors. Overall, the proposed framework is innovative with decent theoretical backing but bears multiple unresolved flaws. Therefore, the recommendation is Invite for rebuttal.
- 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.
After the rebuttal, two reviewers choose Accept and one reviewer vote for Reject. The pros include good novelty and the authors plan to open-source codes. The cons include core methodological issues unaddressed, bold causal claims and the lack of adequate ablation tests and comparisons etc. Considering the pros and cons, I slightly incline to Accept.
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.
This paper proposes a framework (GeoCAN) that integrates causal learning and geometric feature extraction to address the challenges of false correlations and feature entanglement in echocardiogram quality assessment. Main concerns include the clarify of casual and dependency in this work, validation of hyperbolic embedding, motivation of GLM, novelty of NRA. Well most of them have been explained in the author’s response, an important concern raised by R2 is the overstated causal contribution. This should be tuned down to avoid misunderstanding .
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
I agree with Reviewer 2. The core issues remain unresolved: the strong causal claims lack interventional or counterfactual support, key components are not sufficiently validated, and dataset transparency concerns persist.