Methods: 2378 frames from 51 normal fetal cardiac screening movies at 18-20 weeks, were used as a training dataset for machine learning. 701 frames from 28 normal fetal cardiac screening movie used as test data. The movie included the hole images of fetal heart from stomach to base of heart. Correct positions of the heart and peripheral organs were annotated. The positions were following crux, ventricular septum, right atrium, tricuspid valve, right ventricle, left atrium, mitral valve, left ventricle, pulmonary artery, ascending aorta, superior vena cava, descending aorta, stomach spine, umbilical vein, inferior vena cava, pulmonary vein, ductus arteriosus. We calculated the accuracy of each parts of normal fetal heart ultrasound screening movie. Results: The recall ratio of crux, ventricular septum, left ventricle and atrium, right ventricle and atrium, ascending aorta, pulmonary artery, stmach and spine were 97.1%, 69.3%, 96.6% and 90.6%, 84.8% and 96.9%, 61.9%, 100%, 100% and 100%, respectively. Conclusions: A novel system of machine learning automatically detected parts of normal fetal heart at high level. This system can provide us useful information of segmental diagnosis and be a great power for prenatal detection of CHD.