A Deep Convolutional Neural Network based Hybrid Framework for Fetal Head Standard Plane Identification
Jingyu Ye1*, Ruizhi Liu2*, Bing Zou1*, Hongyang Zhang2, Nianji Zhan1, Cong Han2, Ying Yang1,3, Hongguo Zhang2, Jian Guo1, Fang Chen1,3, Shida Zhu1#, Shucheng Hua2#
1 BGI-Shenzhen, Shenzhen 518083, China;2 Reproductive Medicine & Prenatal Diagnosis Center, The First Hospital, Jilin University, Changchun, China.;3 Shenzhen Engineering Laboratory for Birth Defects Screening, BGI-Shenzhen, Shenzhen 518083, China;*These authors contributed equally;#Correspondence author: Shucheng Hua, The First Hospital, Jilin University, No.71 Xinmin Street, Changchun 130021, China. Tel. +86043188782707. Email: shuchenghua@126.com; Shida Zhu, BGI-Shenzhen, Building 11, Beishan Industrial Zone, Yantian District, Shenzhen 518083, China. Tel. +8615814003823. Email: zhushida@bgi.com
Abstract
Objective To develop more effective deep convolutional neural network based technology for automatic identify five types of fetal head standard view planes, i.e., Transventricular plane (TV), Transthalamic plane (TT), Transcerebellar plane (TC), Coronal view of eyes (Eyes), Coronal view of nose (Nose), and other non-standard fetal ultrasound images (Background) in ultrasound images.
Design Randomly select samples for the research including the abnormal fetuses.
Setting Historical records of prenatal examinations from The First Hospital of Jilin University.
Samples 19928 2-D fetal ultrasound images of fetuses with the gestational age ranging from 18 weeks to 31 weeks.
Methods A novel convolutional neural network based hybrid framework is designed for automatic identification of standard fetal head view planes in ultrasound images which is an important topic of fetal screening and diagnosis. In the proposed framework, YOLO-V3 is applied for locating possible fetal head region in the candidate image. Then, a group of object classification networks which includes ResNet50, ResNeXt50, InceptionResNet-V2, and SonoNet64 are employed to make predictions on the located fetal head region. The predictions of the classification models are stacked so as to generate the final prediction for the suspected fetal head region, and the original image as well.
Main Outcome Measures The average precision, average recall, average F-1 score, and the average accuracy on identifying six categories of fetal ultrasound images for each method.
Results The proposed deep convolution neural network based framework achieves the average precision of 89.67%, the average recall of 89.61%, the average F-1 score of 89.61%, and the average accuracy of 89.61% which demonstrates the state-of-the-art performance. The average AUC is 0.9893.
Conclusions The experimental results indicate that the proposed framework is effective on identifying fetal head standard view planes in ultrasound images. Since the experiments are designed to reproduce the scenario happened in the real life, the proposed method could be potentially applied to the automatic fetal screening and diagnosis.
Keywords : fetal ultrasound screening, deep convolutional neural network, fetal head standard plane, model stacking.