Echofusion: Multi-View Transformer Architecture With Anatomical Constraints for Comprehensive Echocardiographic Assessment Publisher

Summary: Can AI read heart ultrasound videos more accurately? A proposed model improved 12 of 15 clinical tasks and cut ejection-fraction error from 4.2% to 3.1% in testing. #CardiacAI #Echocardiography

Avesta M ; Baboukani A Y ; Najafabadian B ; Behjati M
Source: 11th International Conference on Signal Processing and Intelligent Systems, ICSPIS 2025 Published:2025

Abstract

This paper presents a rigorous theoretical and practical analysis of multi-view transformer architectures for echocardiographic video interpretation. We develop a complete mathematical framework describing the information flow from raw pixel inputs to clinical predictions, establishing formal connections between convolutional feature extraction, temporal modeling, and multi-task learning. Our contribution includes: (1) a comprehensive mathematical formalization of the architecture with complete derivation of forward and backward propagation; (2) theoretical analysis of representational capacity, optimization landscape, and generalization bounds; (3) novel multi-view cross-attention mechanism with formal guarantees on information integration efficiency; (4) auxiliary anatomical segmentation tasks with joint optimization framework; (5) detailed experimental validation protocol with statistical significance t e sting across 15 clinical metrics; (6) comprehensive ablation studies isolating architectural components; and (7) clinical impact analysis with cardiologist validation. The mathematical framework establishes precise relationships between architectural hyperparameters and clinical performance metrics, providing design principles for future work. Our proposed extensions demonstrate statistically significant improvements a cross 12 of 15 clinical tasks (p<0.01), with largest gains in measurements requiring spatial-temporal integration (EF MAE reduced from 4.2% to 3.1%). This work bridges theoretical deep learning with clinical echocardiography interpretation, establishing rigorous mathematical foundations for video-based cardiac assessment. ©2025 IEEE.