Congenital Heart Defect Recognition Model Based on YOLOV5

Abstract

Congenital heart defect is an abnormality of the atrial ventricle or the large vascular structure connected to it. It is currently the most common fetal congenital defect, and the incidence accounts for about 30% of congenital defects. Fetal heart abnormalities ultrasound planes screening and the diagnosis of fetal heart defect is an important part of prenatal screening. In China, there is a large population base and obvious differences in medical resources in different regions. In this case, it is difficult for sonographers to diagnose congenital heart defect, and sonographers with rich experience and relevant qualifications are required to make the diagnosis, but the resources of sonographers are limited. This study proposes a deep learning method based on convolutional neural network (YOLOv5) to automatically identify and classify whether fetal-related cardiac ultrasound planes are abnormal. This study method can effectively identify and remind the sonographers of the possible abnormal fetal heart ultrasound section, improve the work efficiency of the sonographers, and reduce the burden of the sonographers. All the datasets used in this method are from university cooperative hospitals with a data volume of 1695, which can be divided into abnormal planes training set (595), normal planes training set (800) and anomalous planes test set (146), and normal planes test set (154). The Mean Average Precision (MAP) on the validation set reached 96.1%, the precision reached 85.2% and recall reached 96.5% in multiple repeated trials. We conduct some comparative experiments with different neural network methods and demonstrate that this method can not only improve the diagnostic efficacy of sonographers on congenital heart defect, but also hope to provide high-quality teaching tools to help low-qualified sonographers pay attention to and learn about fetal congenital heart defects.