Fetal Congenital Heart Disease Echocardiogram Screening Based on DGACNN: Adversarial One-Class Classification...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Fetal congenital heart disease (FHD) is a significant cause of child mortality, with high prevalence in Asia.
• Current screening relies on echocardiography, but obtaining clear fetal heart images, especially four-chamber views, is challenging.
• Interpreting dynamic fetal heart structures and hemodynamics requires specialized expertise, highlighting the need for automated screening.

Purpose of the Study:

• To develop an automated system for accurate and robust screening of fetal congenital heart disease (FHD).
• To address the challenge of limited annotated training data for deep learning models in FHD detection.
• To improve the recognition accuracy and robustness of FHD screening using unlabeled echocardiogram data.

Main Methods:

• Proposed a novel deep learning model, DGACNN, integrating DANomaly and GACNN (WGAN-GP and CNN) components.
• DANomaly utilizes cycle adversarial learning for robust one-class classification of echocardiogram slices.
• GACNN employs WGAN-GP with end-systolic four-chamber heart (FCH) views for feature extraction and data augmentation.

Main Results:

• The DGACNN model achieved an 85% accuracy rate in recognizing FHD, outperforming existing state-of-the-art networks by 1%-20%.
• Experimental validation showed DGACNN's performance reached 84% in a test set, exceeding that of expert cardiologists.
• The model effectively leveraged unlabeled video slices to enhance training dataset size and improve recognition capabilities.

Conclusions:

• The DGACNN architecture offers a promising solution for automated early screening of fetal congenital heart disease.
• This AI-driven approach has the potential to significantly assist cardiologists in timely and accurate FHD detection.
• The method's ability to utilize unlabeled data makes it valuable for improving diagnostic tools in resource-limited settings.