Utilizing Weighted Spatio-Temporal Information for Automated Transthoracic Echocardiography View Classification.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Transthoracic echocardiography (TTE) is vital for assessing cardiac function and diagnosing cardiomyopathies.
• Accurate categorization of TTE video views is crucial for comprehensive diagnosis, necessitating automated solutions.

Purpose of the Study:

• To develop and evaluate models for classifying eleven standard TTE views from video input.
• To explore spatial, temporal, and fused feature extraction techniques for improved classification accuracy.

Main Methods:

• Utilized EfficientNet and Dilated Convolutional Networks for spatial and temporal feature extraction.
• Implemented a Dynamic Feature Fusion (DFF) technique to adaptively balance feature importance.
• Developed a custom cycle detector for partitioning TTE sequences into cardiac cycles for consistent input and ensemble prediction.

Main Results:

• Achieved a micro F1 Score of 0.951 with the proposed model architecture.
• Demonstrated the effectiveness of DFF in balancing spatial and temporal features.
• Showcased improved prediction reliability through cycle detection and ensemble techniques.

Conclusions:

• The developed model architecture represents a significant advancement in automated echocardiogram interpretation.
• This approach streamlines clinical workflows, reduces errors, and has potential applications in broader TTE data analysis, including disease classification.