Chapter 6 - Towards building an efficient deep neural network based on YOLO detector for fetal head localization from ultrasound images

Abstract

High-risk pregnancy complications include ectopic pregnancy, spontaneous abortion, fetal chromosomal abnormalities, fetal growth restriction, congenital anomalies, placenta previa and abruption, gestational diabetes, preeclampsia, and cesarean delivery. All these problems require increased monitoring or surveillance. Particularly, fetal growth restriction is one of the major risk concerns of high-risk pregnancy problems. In the field of Obstetrics, measuring fetal biometry through ultrasound has become a regular practice [21]. To ascertain the well-being of the fetal and safe pregnancy, measuring the head circumference (HC) of fetal at different gestational age is of primary importance. HC measurement is one of the basic biometric parameters used to ascertain the fetal growth and identify cerebral anomalies. This, in turn, is extrapolated to approximate the fetal age and estimated due date. But, ultrasound images are idealistic and naïve for most pregnant women as it requires trained radiologists or sonographers to interpret the images. Considering the scarcity of trained sonographers, there is an absolute necessity for automated HC estimation from ultrasound images. Of late, the arena of computer vision (CV) has monumental success in visual object detection and recognition. CV applications can render a trustworthy help for physicians and radiologists by improving the accuracy of screening. Inspired by the tremendous advancements in the area of CV, automatic HC detection is proposed in this chapter employing the efficient state-of-the-art object detection algorithm “You Only Look Once” (YOLO). YOLO is a very adept convolutional neural network that runs in real time for object detection. Object detection is rather more intricate than classification as it not only identifies the objects but also indicates where the location of the object in the image. YOLO detectors are giving a stunning performance in terms of speed, accuracy, and learning capabilities. YOLO applies a single feed forward propagation through the neural network speeding up the process of object recognition. Ultimately, the aim of this study revolves on actively pushing the envelope of this technology and to build a model that precisely locates the head of the fetal from US images and consequently measures the fetal HC.