III. Methodology
The process of fetal abnormality screening normally consists of two steps: (1) looking for standard planes, (2) evaluation and diagnosis based on standard planes. When the sonographer sweeping the ultrasound tube and looking for the standard planes of fetal bodies, it is actually object detection discussed in computer vision. In order to discriminate fetal standard planes, a deep neural network based hybrid framework is proposed. The idea behind the proposed framework is to use an object detector to locate the Region of Interests (ROI), i.e., fetal head region, then another network is applied to make further decision on the extracted ROI. In this work, YOLO-V3(14) is applied as the object detector, and ResNeXt(15) is employed as the fine-grained classifier.