Multi-Center Adversarial Bi-Phase Cross-Attention Network for Right Ventricular Segmentation and Functional...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated right ventricular (RV) analysis in 2D echocardiography faces challenges due to RV segmentation complexity and deep learning model fragility across ultrasound vendors.
• Existing frameworks lack a unified approach to address both RV segmentation and vendor-specific domain adaptation.

Purpose of the Study:

• Develop and validate TACA-Net (Bi-Phase Adversarial Cross-Attention Network), a multi-center framework for simultaneous RV endocardial segmentation and functional severity classification.
• Achieve vendor-agnostic performance for RV analysis in 2D apical four-chamber echocardiography.

Main Methods:

• TACA-Net integrates a domain discriminator for vendor-agnostic learning and a bi-phase cross-attention module for enhanced feature encoding.
• A dual-head decoder jointly optimizes segmentation and classification, utilizing an auxiliary bi-phase consistency loss.
• Prospective data from three clinical sites with different ultrasound vendors were used for training and external validation.

Main Results:

• TACA-Net achieved superior RV segmentation (DSC 0.903, HD95 7.1 mm) and functional classification (AUC 0.911) on an external test set, outperforming all baselines (p < 0.01).
• Ablation studies confirmed the independent contributions of domain alignment, bi-phase cross-attention, and multi-task learning.
• The model showed consistent performance across subgroups and the lowest expected calibration error (0.041) among classification models.

Conclusions:

• TACA-Net provides vendor-agnostic RV segmentation and functional classification from 2D echocardiography with strong prospective external validation.
• The framework offers a clinically interpretable and transparent foundation for scalable AI-assisted right heart assessment.