A Clinically Interpretable AI System for Real-Time Quality Control of Transthoracic Echocardiography: Development,...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Echocardiography quality control (QC) is vital but traditionally subjective and time-consuming.
• Existing artificial intelligence (AI) systems often lack clinical interpretability and broad use.
• There is a need for objective, real-time QC in echocardiography.

Purpose of the Study:

• To create, validate, and implement an interpretable, rule-based AI system for real-time echocardiographic view quality assessment.
• To automate the evaluation of key echocardiographic domains: visualized structures, cardiac axis, depth, and gain.
• To provide immediate, objective feedback for improving echocardiographic image acquisition.

Main Methods:

• Developed a quantifiable scoring rubric for nine standard echocardiographic views.
• Automated the rubric using a modular AI pipeline with SlowFast-Echo for view classification and deep learning models (SSD, U-Net) for domain assessment.
• Trained on 2,966 videos, validated prospectively on 1,801 videos, and externally on 1,821 videos from public datasets.

Main Results:

• The AI system achieved 98.6% accuracy in view classification.
• Overall image quality assessment showed 95.0% accuracy against expert consensus on the prospective test set.
• The system demonstrated high accuracy across quality domains (94.7%-98.7%) and strong generalizability (91.7%-92.3%) with a mean inference time of 303 ms.

Conclusions:

• A comprehensive AI system for accurate, real-time, and interpretable echocardiographic quality assessment was successfully developed and deployed.
• The system translates expert criteria into an automated framework, standardizing image acquisition.
• This tool has the potential to improve diagnostic confidence and efficiency in clinical practice and sonographer training.