Automatic Recognition of Fetal Heart Standard Section Based on Fast-RCNN

Abstract

Congenital heart defect is one of the most common fetal congenital defects. Every year, about 1% of newborns in the world suffer from this disease, and the proportion in developing countries reaches up to 4%-5%. Automatic identification of standard fetal heart sections from 2D ultrasound scanning video is an important prerequisite for examining a fetus with congenital heart disease. In China, most areas belong to rural areas. In this difficult environment, it is extremely difficult to diagnose fetuses with congenital heart disease by prenatal sonographers, which requires sonographers with rich qualifications to make the diagnosis, but few sonographers meet this qualification in rural areas. In this study, a new method based on Fast-RCNN deep learning with Mobilenet as the backbone network is proposed to automatically identify the standard section of fetal heart. This model can not only help sonographers collect fetal ultrasound images in practice, but also provide a reliable basis for later analysis of the fetal images, save more time and enhance efficiency. And this method can not only help new ultrasound physicians, but also provide high-qualified sonographers enough auxiliary diagnosis effects. All the data sets used in this method collected from the cooperative hospitals of colleges and universities, and the data volume is 1839, which can be split into training set(1479) and test set(360). In three centuries of repeated trials, the Mean Average Precision (MAP) on the validation set reaches 92.49%, and the accuracy rate reaches 90%. In the later period, some comparative experiments of different neural networks have been carried out, which proves that the method in this study is superior to other neural networks and can bring enough benefits to ultrasound physicians.