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

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate quantification of cardiac left ventricle (LV) function is crucial for diagnosing cardiovascular diseases.
• Traditional methods rely on segmentation-based models, which demand large annotated datasets, often unavailable in medical settings.
• Ultrasound image characteristics like low inter-class variability and high noise pose significant challenges for model training.

Purpose of the Study:

• To develop a segmentation-free model, LV-STANet, that minimizes reliance on ground truth annotations for LV function assessment.
• To maintain accuracy and computational efficiency in quantifying key LV functional parameters.
• To offer a viable solution for clinical deployment in resource-constrained environments.

Main Methods:

• LV-STANet directly estimates LV function from 2D echocardiogram videos, integrating spatial and temporal features.
• A spatial encoder captures anatomical details, while a temporal attention module models frame-to-frame dynamics.
• A weighted aggregation strategy combines these components to predict ejection fraction (EF), global longitudinal strain (GLS), and fractional shortening (FS).

Main Results:

• LV-STANet achieved competitive performance on the EchoNet-Dynamic dataset.
• Mean absolute errors (MAE) were 5.1% for EF, 3.35% for GLS, and 4.95% for FS.
• The model demonstrates accurate and reliable cardiac function assessment without requiring segmentation.

Conclusions:

• LV-STANet offers a promising, segmentation-free approach for cardiac function quantification.
• The model's efficiency and reduced annotation requirement make it suitable for clinical settings with limited resources.
• This method advances automated cardiac assessment from echocardiograms.