Abstract

Info

Title:: Fetal facial ultrasound plane recognition based on real-time object detection network and its application

Author(s):: LIU Zhonghua1; 2; YU Weifeng1; WU Xiuming1; 2; XUE Hao2; L?Guorong3; 4; WANG Xiaoli5; LIU Peizhong2; 5; 1. Department of Ultrasound, Quanzhou First Hospital Affiliated to Fujian Medical University, Quanzhou 362000, China 2. College of Engineering, Huaqiao University, Quanzhou 362000, China 3. Collaborative Innovation Center for Maternal and Infant Health Service Application Technology, Quanzhou Medical College, Quanzhou 362000, China 4. Department of Ultrasound, the Second Affiliated Hospital of Fujian Medical University, Quanzhou 362000, China 5. School of Medicine, Huaqiao University, Quanzhou 362000, China

Keywords:: Keywords: ultrasound examination artificial intelligence real-time object detection network fetal face

Abstract:: Abstract: Objective To explore the role of an artificial intelligence (AI) model based on real-time object detection network in fetal facial ultrasound examination. Methods With the normal fetal facial ultrasound standard plane (FFUSP) at 20-24 weeks of gestation as the research object, a FFUSP recognition model based on real-time object detection network was constructed. The recognition accuracy of the model for FFUSP and the anatomical structures were analyzed, and the clinical value was evaluated by analyzing its performance in identifying FFUSP in 119 cases of fetal ultrasound images. Results The overall precision, recall rate, mAP@.5 and mAP@.5:.95 of the AI model were 97.8%, 98.5%, 98.1% and 61.0%, respectively. The clinical validation showed that the AI model had a sensitivity, specificity, positive predictive value, negative predictive value and accuracy of 100.0%, 98.5%, 87.4%, 100.0% and 98.7% for facial anatomy recognition, and the results were highly consistent with the classification of fetal ultrasound experts (k=0.925, P<0.001). The recognition accuracy of the model for 3 types of standard planes reached 100% and the average speed of dynamic video detection was 33.93 frames per second. Conclusion The FFUSP recognition model based on real-time object detection network exhibits excellent performance, and it can be applied to real-time ultrasound diagnosis, teaching and intelligent quality evaluation.