Automatic Standard Plane Localization In Fetal Ultrasound Image

During the routine ultrasound examination,the accurate localization of the standard plane from dynamic ultrasound images is quite vital for subsequent biometric measurement and diagnosis.Traditional manual localization of standard plane by doctors is labor-intensive and time-consuming,and relies heavily on the operator's clinical experience,resulting in large inter-observer variability.Therefore,the development of automatic method for localizing ultrasound standard plane is highly desirable.In this paper,we present deep convolutional network based approach to localize the ultrasound standard plane.Computerized automated localization of fetal ultrasound standard plane can greatly improve the efficiency in prenatal ultrasound diagnosis,relieve doctors' work pressure,and save time for ultrasound doctors to process and analyze fetal ultrasound images.This paper proposes a method based on deep convolutional neural network to achieve automatic localization of fetal ultrasound standard plane.We propose a fetal ultrasound standard aspect localizing system based on Ultrasonic Image Networks(UInet)and Region Search Networks(RSN).We compared the results of the above two automated methods with doctor's manual annotations,and achieved good agreement.To the best of our knowledge,we propose to automatically localize Fetal Thalamus Standard Plane,Fetal Femur Standard Plane and Fetal Abdominal Standard Plane simultaneously for the first time.In UInet's method,we regard the localizing problem as a supervised classification problem,and the supervision information comes from the doctor's manual annotation.First,we cut out the key anatomical regions of the ultrasound image manually;then,we introduce adjacent similar plane into the data set to solve the problems of similar anatomical structure interference and adjacent plane confusion;finally,we visualize the localizing basis of UInet by convolutional neural network visualization technology and the results obtained are consistent with the anatomical regions analyzed by the ultrasound doctor.In the RSN-based method,the RSN automatically searches for the key anatomical region of the plane on the high-order feature map obtained by the deep convolutional network,and then completes the localizing of the standard plane.Finally,based on the UInet's method,the accuracy of prediction results for Fetal Thalamus Standard Plane?Fetal Femur Standard Plane and Fetal Abdominal Standard Plane on the 18,736 ultrasound images reached 85%,96%,and 92%,respectively;the RSN-based method was used to predict 899 ultrasound videos.The accuracy of the results reached 87%,91%,and 90%,respectively.The experimental results show that the RSN can accurately search the key anatomical region of the plane and has good localizing performance.