The Automatic Quality Control Of Fetal Ultrasound Cardiac Four-chamber Plane

Ultrasound is widely used in prenatal screening due to its advantages of low cost,good realtime performance and no radiation.Fetal ultrasound cardiac four-chamber plane is one of the most commonly used cardiac planes.It has been used in the early 1980 s to diagnose fetal heart abnormalities,and most cardiac abnormalities can be detected through this section.In prenatal ultrasound screening,ultrasound image quality control is an important part of clinical diagnosis.The process is usually done manually,with some drawbacks,such as requiring experienced operators,time-consuming and laborious operations,and discriminating between different physicians.Therefore,in order to improve the efficiency and accuracy of screening,reduce the workload of doctors,and assist doctors in diagnosis,this paper proposes a new automatic quality control framework for fetal ultrasound cardiac four-chamber plane.In this study,with the assistance of expert doctors,the quality control protocol for the fourchamber plane of the heart in fetal ultrasound was firstly formulated,including the cardiac fourchamber plane,gain,zoom,pulmonary vein angle,apex cordis and moderator band,and multiple ribs,and different evaluation scores were assigned according to the importance of the protocols.The four-chamber view of the heart is the target view,which needs to be screened from the hospital database first.Gain and zoom are important imaging factors.Excessive zoom often results in incomplete structure display,while too small zoom may result in some anatomical structure display too small and not clear enough.Whether the gain is too high or too low,it may lead to poor imaging quality and even misdiagnosis.Some key anatomical structures must be examined for each target section,including pulmonary vein angles,apex cordis and moderator band,and multiple ribs,because these structures are critical for determining whether the section is abnormal or not.In this study,the new quality control framework of fetal ultrasound cardiac four-chamber plane mainly consists of three networks.(1)the Basic Convolutional Neural Network(B-CNN)has the recognition speed of 125 FPS(Frame Per Second).We adopted a lightweight SqueezeNet network structure and improved it with modules in the ResNeXt network to form the final B-CNN.The main purpose was to roughly extract the target four-chamber plane,including the apical,bottom and parasternal,from the hospital database.(2)deeper CNN(DCNN),with 33 FPS recognition speed.On the basis of B-CNN extraction into the four-chamber cardiac tangent plane,the imaging factors of the four-chamber cardiac tangent plane,namely the gain and zoom,were further identified by using a deeper D-CNN network model combined with multi-label learning.(3)Aggregated Residual Visual Block Network(ARVBNet),the detection speed with 101 FPS.By imitating human visual mechanism,we proposed a ARVB with stronger feature extraction ability,and combined it with the advanced Single Shot MultiBox Detector(SSD)to form the final ARVBNet.The network is mainly used to detect the anatomical structure of the four-chamber plane,including the clear pulmonary vein angle,apex cordis and moderator band,as well as the multiple ribs due to irregular operation.Finally,according to the output of the three networks,the overall quantitative evaluation score of each four-chamber target plane is obtained,so as to realize automatic quality control.The overall quality control framework has the detection speed of 26 FPS,which meets the requirements of real-time quality assessment of the section.Experiments on the fetal ultrasound dataset show that compared with some other advanced detection models,our ARVBNet detection network not only has the highest mAP(mean Average Precision),but also maintains a fast rate of 101 FPS.ARVBNet is the most innovative network model in this paper.In order to verify the universality and generalization of the model,this paper tested the model on the PASCAL VOC data set and achieved better performance than the traditional detection network.The quality control framework in this paper is universal,and it can be extended to other ultrasonic sections,such as fetal abdomen,head,face,etc.,for quality control in the future.