Cross-View Constrained Encoder and Cross-Modal Semantic-Hierarchy Fusion Network for Congenital Heart Disease Identification.
Authors
Affiliations (4)
Affiliations (4)
- Southwest Petroleum University, Xindu District, Chengdu, Sichuan, China.
- Sichuan Provincial Maternity and Child Health Care Hospital, Chengdu, Sichuan, China.
- Southwest Petroleum University, Xindu District, Chengdu, Sichuan, China. [email protected].
- Southwest Petroleum University, Xindu District, Chengdu, Sichuan, China. [email protected].
Abstract
Congenital heart disease (CHD) is a common structural birth defect, and early screening is important for improving prognosis. Clinical cardiac ultrasound combines brightness-mode (B-mode) and color Doppler images from multiple standard views to assess cardiac morphology and hemodynamic abnormalities. Existing case-level models remain limited in modeling information exchange among standard views within each modality and cross-modal fusion between B-mode morphology and Doppler hemodynamics. The proposed Cross-View Constrained Encoder and Cross-Modal Semantic-Hierarchy Fusion Network (CVCE-CMSHFNet) addresses these limitations through two complementary components: a Multi-View Constrained Encoder (MVCE) and a Bidirectional Progressive Semantic-Hierarchy Fusion (BiPSHF) module. MVCE models relationships among the eight standard views within each modality. Sinkhorn-constrained transport and a parallel gated transformation update the view representations. Training-time Intra-Modal View Covariance Diversity Regularization (IVCDR) separately penalizes covariance similarity among B-mode view representations and among Doppler view representations. BiPSHF performs bidirectional cross-modal interaction at the view, view-group, and case levels. The view-group level organizes the eight standard views into four study-specific groups based on anatomical relationships and clinical assessment roles. CVCE-CMSHFNet aims to support clinical diagnosis by providing a case-level CHD probability for physician reference. Evaluation included 384 complete eight-view dual-modality cases from a 506-subject single-center echocardiography database. CVCE-CMSHFNet achieved 95.64% accuracy, 95.23% F1 score, 97.05% AUC, 94.44% sensitivity, and 96.67% specificity.