Multi-Task Deep Learning Framework for Real-Time Quality Assessment and Probe Guidance in Echocardiography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Transthoracic echocardiography quality is operator-dependent.
• Current AI systems need extensive labeled data and offer limited real-time feedback.
• Existing methods provide only global quality scores, hindering clinical application.

Purpose of the Study:

• To develop a label-efficient, multi-task AI framework for real-time echocardiography image quality assessment.
• To enable structure-specific quality scoring without additional annotations.
• To provide real-time corrective probe maneuver suggestions.

Main Methods:

• A multi-task framework performing view classification and structure segmentation.
• Entropy-based metrics for label-efficient, structure-specific quality scoring.
• A lightweight maneuver predictor for real-time corrective action suggestions.
• Training and validation on a 43-case maneuver-tagged dataset.

Main Results:

• The AI quality metric achieved high AUC (0.901-0.987) in distinguishing standard from nonstandard views.
• The maneuver predictor demonstrated 85.1% top-1 accuracy (mAP: 0.904).
• The system achieved a fast inference time of 17 ms per frame, suitable for real-time use.

Conclusions:

• The proposed AI system assists users in acquiring acceptable echocardiographic views consistently.
• It reduces the need for extensive annotations and operator dependency.
• The framework enhances clinical efficiency and reliability in cardiac imaging.