Deep learning-based video-level view classification of two-dimensional transthoracic echocardiography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Deep learning (DL) shows promise for 2D transthoracic echocardiography (TTE) view classification.
• Existing DL models face challenges in view coverage, accuracy, inference speed, and real-world validation.

Purpose of the Study:

• To develop a clinical requirement-driven DL framework (TTESlowFast) for accurate and efficient TTE view classification.
• To address class imbalance and limited labeled data in TTE videos.

Main Methods:

• Utilized the SlowFast architecture for TTE video analysis.
• Implemented a sampling balance strategy to handle class imbalance.
• Employed a data augmentation strategy to overcome limited labeled TTE data.

Main Results:

• Achieved high performance on the test set: 0.9881 accuracy, 0.9870 precision, 0.9867 recall, and 0.9867 F1 score.
• Demonstrated strong field performance: 0.9607 accuracy, 0.9586 precision, 0.9499 recall, and 0.9530 F1 score.
• Achieved low inference times (105.0 ± 50.1 ms on desktop GPU, 186.0 ± 5.2 ms on edge device), meeting clinical demands.

Conclusions:

• The TTESlowFast framework offers effective and efficient TTE view classification.
• Low inference delay makes TTESlowFast suitable for real-time medical applications.
• The framework shows significant potential for practical clinical use in echocardiography.