Development and Validation of a Deep-Learning Network for Detecting Congenital Heart Disease from Multi-View...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Pediatric Cardiology

Background:

• Early detection of congenital heart disease (CHD) is crucial for improving pediatric outcomes.
• Inexperienced sonographers often struggle with accurate CHD identification via transthoracic echocardiogram (TTE).
• Current diagnostic methods can be challenging, necessitating advanced tools for improved accuracy.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) framework for automated CHD detection in children.
• To integrate multi-view and multi-modal TTE data for enhanced diagnostic performance.
• To predict the probability of normal cardiac function, atrial septal defect (ASD), or ventricular septal defect (VSD).

Main Methods:

• Analysis of 2D and Doppler TTEs from 1,932 children (2018-2022).
• Development of a DL framework to identify cardiac views, integrate information, and predict CHD presence.
• Utilized five standard echocardiographic views: apical 4 chamber, subxiphoid, parasternal long-axis, parasternal short-axis, and suprasternal long-axis.

Main Results:

• The DL model achieved high accuracy in cardiac view classification (0.989).
• For CHD screening, the model demonstrated excellent performance with AUCs of 0.996 (within-center) and 0.990 (cross-center).
• Classification accuracy for healthy, ASD, and VSD was high, reaching 0.991 (within-center) and 0.986 (cross-center).

Conclusions:

• Deep learning models integrating multiple TTE views and modalities significantly enhance CHD detection accuracy in children.
• The developed DL framework shows potential to approximate the performance of experienced sonographers.
• This noninvasive approach can improve CHD screening efficiency and accuracy, aiding early diagnosis and treatment.