A hybrid spatial and temporal attention driven network for left ventricular function assessment using echocardiography

Comprehensive Summary

The authors developed LV-STANet, a segmentation-free deep learning model that estimates left ventricular ejection fraction (EF), global longitudinal strain (GLS), and fractional shortening (FS) directly from apical four‑chamber echocardiography videos. The model used a ResNet50-based spatial encoder plus a temporal attention module, combining frame-wise features with learned weights across the cardiac cycle. After being trained on over 10,000 studies from EchoNet-Dynamic and tested both internally and on the CAMUS dataset, LV-STANet achieved an EF mean absolute error of about 5% on EchoNet-Dynamic and about 7% on CAMUS, with correlations around 0.8 and 0.7 respectively. Performance for GLS and FS was slightly less stable than for EF. Ablation experiments showed that removing spatial attention, temporal attention, or their weighted fusion each led to worse EF accuracy.

Outcomes and Implications

This work suggests that reliable EF can be obtained automatically from routine A4C clips without manual LV tracing or explicit ES/ED frame selection. At the bedside, such a tool could shorten echo reporting time, reduce inter-reader variability, and provide more consistent EF estimates in labs with limited staffing or expertise. Because performance was best on the development dataset and may degrade with different scanners or image quality, multi-center validation and workflow integration are needed before widespread adoption.