On the YOLOv4 Architecture for Fast and Real Time Congenital Heart Disease Detection Via Ultrasound Videos

Abstract

Congenital Heart Disease (CHD) is one of the most frequent cardiac defects in infants, and it is becoming more common. Various research studies have been conducted for CHD identification based on clinical and non-clinical data. This study conducts an artificial  intelligence system for real-time congenital heart disease (CHD) detection using ultrasound videos. The YOLOv4 (You Only Look Once) is employed for localizing the congenital defect through ultrasound videos. The performance is evaluated by mean Average Precision (mAP) which compares the classification results with the medical ground truth. The model has great performance in CHD detection with mAP values of training data of 98.36% for YOLOv4 and 87.24% for YOLOv4 tiny. This is useful for doctors and radiologists who require a simple, fast, yet accurate model for the detection of CHD.