acceptodds
Withdrawn from ICLR 2027

MUSE: Multi-View Echocardiography Segmentation via View-Adaptive Experts and Boundary-Aware Multi-Domain Fusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Echocardiography segmentation (ES) can substantially reduce radiologists' heavy clinical workload in image analysis. Despite recent advances, accurate multi-view segmentation remains challenging because echocardiograms are complex and ambiguous, whereas existing models mainly focus on a single echocardiographic view, limiting the clinical value of current models. To address these issues, we propose Multi-View Echocardiography Segmentation via View-Adaptive Experts and Boundary-Aware Multi-Domain Fusion (MUSE), a novel framework that aims to improve multi-view echocardiogram segmentation through view-adaptive expert paths. Specifically, MUSE adopts an encoder-decoder architecture. To handle ambiguous and complex echocardiographic images, MUSE constructs boundary priors from bottleneck features and performs multi-domain fusion through boundary-aware gating, which selectively enhances the wavelet, spatial-feature, and adaptive multiband Fourier branches. To further improve view-specific segmentation, we design a multi-expert routing mechanism that predicts view types from image features and employs shared and view-specific experts to extract common and distinctive features for segmentation. Additionally, to alleviate the limited view variety of existing datasets, we collect and annotate a multi-view, multi-structure echocardiography dataset, thereby improving clinical relevance through broader view coverage. Experimental results on two public datasets and one private dataset demonstrate the effectiveness of MUSE, which achieves state-of-the-art segmentation performance on both public and private datasets across diverse views and anatomical structures.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CV: Medical and Biological Imaging

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bFKpOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MUSE for multi-view echocardiography segmentation. The method combines a mixture of view experts (MoVE) and boundary-aware multi-domain fusion (BAMF) to improve the segmentation of different cardiac structures. Experiments on two public datasets and one private dataset show promising results.

Critical strengths
  1. The paper addresses an important problem in echocardiography segmentation. A unified model for different cardiac views is useful for clinical applications.

  2. The proposed framework is clearly presented, and the two main modules are reasonably designed.

  3. The experiments cover multiple views and cardiac structures. The results show improvements over several existing methods.

Critical weaknesses
  1. Limited technical novelty. The proposed MoVE and BAMF modules mainly combine existing techniques, including mixture of experts, wavelet transforms, Fourier features, and attention mechanisms. The authors should better explain the differences from previous methods and demonstrate the advantages of these designs.

  2. Insufficient analysis of view-specific learning. Although the method uses view-specific experts, it is unclear whether different experts really learn different anatomical information. The authors should visualize the expert routing results for different views and compare with a shared model and separately trained view-specific models.

  3. Limited ablation studies. The current ablation only evaluates the two main modules. More detailed experiments are needed to understand the contributions of the spatial, Fourier, wavelet, and boundary-aware components.

Questions for authors
  1. What are the main differences between the proposed modules and existing MoE and multi-frequency fusion methods?

  2. Can the authors provide expert routing visualizations and comparisons with independent models trained for each view?

  3. Can the authors provide more detailed ablations for BAMF?

Reviewer_MtEhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MUSE, a unified framework for multi-view and multi-structure echocardiography segmentation. MUSE combines a Mixture of View Experts (MoVE) module, which uses view-aware sparse routing over shared and view-specific experts, with a Boundary-Aware Multi-Domain Fusion (BAMF) module that integrates spatial, Fourier, and wavelet representations under an explicit boundary-guided gating mechanism. The method is evaluated on CAMUS, EchoNet-Dynamic, and an in-house five-view clinical echocardiography dataset with annotations for seven cardiac structures. The reported results show competitive or best segmentation performance on the evaluated datasets.

Critical strengths
  1. The paper addresses a clinically relevant segmentation setting beyond a single echocardiographic view. The attempt to build one model that handles heterogeneous views and multiple anatomical structures is practically meaningful, and the in-house dataset broadens the anatomical/view coverage beyond CAMUS and EchoNet-Dynamic.

  2. The method is clearly structured and empirically improves the baseline. The MoVE and BAMF modules target two distinct challenges—view heterogeneity and ambiguous boundaries—and the component-level ablation shows consistent improvements when either module is introduced.

Critical weaknesses
  1. The evaluation protocol for the in-house clinical dataset is insufficiently specified and may permit patient-level leakage. The manuscript describes the private dataset as 1,550 frames split into 1,240/155/155 training/validation/test images, but does not report the number of patients/examinations or state that the split is patient-disjoint. In echocardiography, multiple frames and views from the same examination can be highly correlated, so a frame-level split could substantially overestimate generalization. This is particularly important because the strongest empirical gains are reported on this dataset.

  2. The current ablations do not establish that the relatively complex MoVE/BAMF design is necessary. MoVE is trained with substantial ground-truth view-label teacher forcing, yet view-prediction accuracy and robustness to routing errors are not reported. Similarly, the appendix reports average BAMF fusion weights of 0.074, 0.038, and 0.888 for the spatial, Fourier, and wavelet branches, respectively, suggesting that the learned fusion is overwhelmingly dominated by the wavelet branch. The paper does not provide branch-level ablations or simpler controls that would establish the contribution of the Fourier/spatial branches, the boundary gate, and competitive fusion separately.

Questions for authors
  1. Is the in-house dataset split strictly patient-disjoint (or examination-disjoint)? Please report the number of patients/examinations in each split and clarify whether multiple frames or different views from the same patient can appear across training, validation, and test sets.

  2. Please clarify the ethics protocol for the in-house clinical dataset, including institutional/IRB approval or exemption, consent/consent waiver, and de-identification procedures. Who produced the segmentation annotations, and was inter-observer variability assessed?

  3. Since MoVE is trained with ground-truth view-label teacher forcing, what is the test-time view-prediction accuracy/confusion matrix when teacher forcing is disabled? How sensitive is segmentation performance to view-prediction or routing errors, and what is the performance without view supervision?

Reviewer_8PmgOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MUSE, a framework for multi-view echocardiography segmentation that integrates view-adaptive experts with boundary-aware multi-domain feature fusion. Specifically, it employs boundary-aware gating to fuse spatial, wavelet, and adaptive multiband Fourier features, aiming to improve the segmentation of ambiguous echocardiographic structures. It also introduces a multi-expert routing mechanism that combines shared and view-specific experts to capture common and view-dependent anatomical features. In addition, the authors collect and annotate a multi-view, multi-structure echocardiography dataset to address the limited view diversity of existing datasets. Experiments on two public datasets and one private dataset reportedly demonstrate state-of-the-art performance across multiple views and anatomical structures.

Critical strengths
  1. The paper addresses a clinically relevant problem, as multi-view echocardiography segmentation can potentially improve the applicability of automated segmentation systems in clinical workflows.
  2. The proposed framework considers both boundary ambiguity and view-specific feature variations, while the collection and annotation of a multi-view dataset may provide additional value for evaluating segmentation under diverse imaging conditions.
Critical weaknesses
  1. The methodological novelty appears limited. The two main concepts highlighted in the title—boundary-aware multi-domain feature fusion and view-adaptive expert routing—are related to established techniques in feature fusion, attention or gating mechanisms, and mixture-of-experts architectures. The paper needs to more clearly distinguish its contributions from existing methods and explain what fundamental methodological advances are introduced beyond combining these components.

  2. The rationale and effectiveness of the proposed components require stronger justification. It is unclear from the abstract why wavelet, spatial, and Fourier-domain features are jointly necessary, how boundary-aware gating specifically improves their integration, and whether view-adaptive experts provide benefits beyond conventional multi-view conditioning or mixture-of-experts approaches. More targeted ablation studies and comparisons with relevant alternatives would help establish the contribution of each design choice.

  3. The experimental evidence and dataset contribution need to be presented more clearly. Although the paper reports state-of-the-art results on two public datasets and one private dataset, the abstract does not clarify the diversity and scale of the newly collected dataset, the evaluation protocol, or the extent to which the improvements generalize across views and anatomical structures. The authors should provide sufficient evidence to support the claimed robustness and clinical relevance.

  4. The manuscript's presentation requires substantial improvement. The methodological illustrations, experimental figures and tables, and written descriptions need to be made clearer and more precise. In particular, the overall framework diagram should communicate the information flow and interactions among its components more intuitively, while the result figures and tables should facilitate direct comparison across methods, views, and anatomical structures. These presentation issues make it more difficult to understand and assess the proposed contributions.

Questions for authors
  1. How does MUSE fundamentally differ from existing boundary-aware feature fusion methods and mixture-of-experts architectures? What is the specific methodological novelty beyond integrating established techniques?

  2. What is the motivation for combining spatial, wavelet, and Fourier-domain features? Can the authors provide ablation studies or analyses demonstrating the individual and complementary contributions of these feature domains and the boundary-aware gating mechanism?

  3. How does the proposed view-adaptive expert routing compare with simpler alternatives, such as view-conditioned feature modulation or a conventional mixture-of-experts model? Does it consistently improve performance across different views and anatomical structures?

  4. Could the authors provide more details about the newly collected dataset, including its size, view and anatomical-structure distributions, annotation protocol, and train/test splitting strategy? How do the authors ensure that the reported improvements reflect genuine generalization rather than dataset-specific characteristics?

  5. Can the authors improve the framework illustrations and experimental visualizations to make the proposed mechanisms, performance differences, and failure cases easier to interpret? In addition, could they report the computational overhead introduced by multi-domain fusion and expert routing to clarify the practical trade-off between segmentation accuracy and inference efficiency?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.