Multi-task Learning For Automated Quality Control Of Fetal Head Ultrasound Images

During prenatal ultrasound,sonographers often need to measure related parameters to estimate the growth and development of the fetus.Obtaining standard planes is the prerequisite for doctors to make accurate measurements.The biparietal diameter and head circumference parameters of the transthalamic plane studied in this paper are used to predict the weight of the fetus.It is a key reference parameter for pregnant women to choose the final pregnancy method.However,the acquisition of the standard surface depends on the clinical experience of the ultrasound doctor.There is a great subjectivity and it is easy to obtain a low-quality ultrasound plane.Clinically,an experienced expert ultrasound doctor evaluates the ultrasound image quality of a novice doctor,but it is very time-consuming and labor-intensive and takes up clinical expert resources.In order to alleviate the work pressure of expert doctors,we have proposed a method for automated quality control of fetal head ultrasound images.This paper proposes a new multi-task learning framework that uses a multi-task faster region convolutional neural network(MF R-CNN)architecture for automated quality assessment.MF R-CNN can identify the key anatomical structure of the fetal head,analyze whether the ultrasound image magnification is appropriate,and then evaluate the quality of the ultrasound image according to the clinical protocol.First of all,the first five convolutional blocks of MF R-CNN learn features shared in the input data,which can be associated with detection and classification tasks and then extended to task-specific output streams.In training,in order to speed up different convergence speeds for different tasks,this paper designs a segmented training method based on transfer learning.Secondly,this paper adds a clinical prior knowledge module to improve the accuracy of the test results.The position of each anatomical structure is fixed,the coverage between anatomical structures is statistically calculated,and the results of abnormal coverage in the prediction result are eliminated to avoid interference from other structures and tissues in the ultrasound image.At last,a multi-scale attention module is added in the shared residual feature layer.The attention feature layer is calculated from the two dimensions of channel and space,and then the attention feature layer is multiplied with the input feature layer to adaptively learn the features.The multi-scale attention module trains the network to pay attention to the details of the target,and at the same time,it can improve the feature expression ability of the network model.The automated ultrasound fetal head quality control system proposed in this paper evaluates the ultrasound plane of the fetal head by identifying the key anatomical structure and magnification of the ultrasound image to determine whether it is a standard image.Finally,this paper also uses Py Qt5 to design a simple and easy-to-operate automated quality control interface for fetal head to facilitate the operation of clinical ultrasound doctors.Experimental results show that the method can accurately evaluate the quality of the ultrasound plane in half a second.Compared with the current most advanced methods,our method has achieved good detection results,improved detection efficiency,and reduced measurement errors caused by improper ultrasound scanning.