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

This study addresses the challenge of accurately quantifying cardiac left ventricle (LV) function, critical for diagnosing cardiovascular diseases. Existing methods typically depend on segmentation-based models that require large annotated datasets, a resource often scarce in the medical field. Moreover, the low inter-class variability and high noise in ultrasound images further complicate the model training. To overcome these limitations, we propose LV-STANet, a segmentation-free model designed to minimize reliance on ground truth annotations while maintaining accuracy and computational efficiency. LV-STANet estimates LV function directly from 2D echocardiogram videos by integrating spatial and temporal features. A spatial encoder captures anatomical features, while a temporal attention module models the dynamic behavior across frames. These components are combined using a weighted aggregation strategy to predict key LV functional parameters: ejection fraction (EF), global longitudinal strain (GLS), and fractional shortening (FS). We evaluate our model on the publicly available EchoNet-Dynamic dataset. LV-STANet achieves a mean absolute error (MAE) of 5.1% for EF, 3.35% for GLS, and 4.95% for FS, demonstrating competitive performance. These results highlight the model' s ability to provide accurate and reliable cardiac function assessment without the need for segmentation, offering a promising direction for clinical deployment in resource-constrained settings.