Echocardiographic image segmentation via multi-expert consensus diffusion model
- VernacularTitle:基于多专家共识扩散模型的心脏超声图像分割
- Author:
Han ZHANG
1
;
Kang LI
1
Author Information
1. West China Biomedical Big Data Center, West China Hospital, Sichuan University, Chengdu, 610041, P. R. China
- Publication Type:Journal Article
- Keywords:
Echocardiographic image segmentation;
diffusion models;
annotation diversity;
annotation consensus;
left ventricle segmentation;
uncertainty-aware segmentation;
inter-observer variability;
probabilistic consensus strategy
- From:
Chinese Journal of Clinical Thoracic and Cardiovascular Surgery
2026;33(07):1057-1063
- CountryChina
- Language:Chinese
-
Abstract:
Objective To address the inter-observer annotation variability in echocardiographic image segmentation caused by image boundary ambiguity and differences in experts’ clinical backgrounds. Methods A framework was proposed that leveraged a conditional diffusion model, guided by cardiac ultrasound images, to learn the probability distribution of multi-expert annotations. A probabilistic consensus strategy was further introduced to dynamically fuse these annotations. This approach was designed to simultaneously aggregate clinical consensus among experts and preserve the diversity of the annotation distribution, thereby jointly modeling both the stable morphological features of the target anatomy and the reasonable variations inherent in expert annotations. Given the scarcity of authentic multi-expert annotations in public datasets like CAMUS, a synthetic multi-expert annotation system was developed using morphological operations. This system was designed to simulate three typical clinical annotation style (conservative, optimal, and aggressive), to provide a reliable data foundation for validating the proposed method. Results The proposed model significantly outperformed state-of-the-art methods in both left ventricular endocardium (LVEndo) and left ventricular epicardium (LVEpi) segmentation tasks. For LVEndo, the generalized energy distance () at the end-diastolic (ED) phase reached 0.073 1, representing a 43.9% reduction compared to the D-Persona model. For LVEpi, the decreased by 42.0% and 39.0% at the ED and end-systolic phases, respectively, relative to D-Persona. Furthermore, the structural fidelity improved by 2.7-3.5 percentage points for LVEndo and 1.4-2.2 percentage points for LVEpi compared to single-expert models, indicating a superior ability to capture diverse expert preferences. Conclusion By jointly modeling population-level consensus and annotation diversity, this work enables a unified characterization of anatomical structures and their associated annotation uncertainty in cardiac ultrasound images, offering a novel approach toward robust and interpretable segmentation in clinical settings.
