Research On Measurement Method Of Fetal Head Circumference Based On Convolutional Neural Network

Measuring the head circumference of the fetus can estimate the gestational age and monitor the growth of the fetus.Medical ultrasound imaging technology has become the preferred method for measuring fetal head circumference due to its cheapness,non-radiation and real-time nature.However,the current main measurement method is manual measurement by experienced physicians,and the measurement accuracy depends on the physician’s clinical experience.This method easily leads to time-consuming ultrasound image measurement and increases the workload of the physician.Existing resources of experienced physicians are not sufficient,and it is more obvious in some backward areas,which affects the accuracy of prenatal examinations for pregnant women.In order to improve the efficiency and accuracy of ultrasound measurement,this thesis proposes a measurement solution to help automatically measure fetal head circumference parameters.The research of this subject has a positive effect on alleviating the work pressure of medical prenatal examination and improving the efficiency of prenatal ultrasound diagnosis.In this thesis,firstly,the ultrasound image of the fetal head is segmented based on the U-Net model of the medical image segmentation network.Based on the U-Net network model,the network structure is adjusted,and the H-Unet fetal head segmentation network model is proposed.H-Unet uses the Non-local module to capture long-range dependencies,integrates the spatial pyramid module to obtain multi-scale feature information,and proposes an improved loss function for network training;then,it detects the edge of the head through the Canny operator,and uses the minimum The ellipse fitting method of the square method automatically calculates the fetal head circumference value.The background area of the ultrasound image inevitably affects the accuracy of fetal segmentation.In order to improve the segmentation network’s attention to the fetal head region,this thesis further uses the Mask R-CNN network model as the framework,combined with the H-Unet network as the segmentation network module.First,perform target detection on the ultrasound image of the fetal head to find the fetal head area of interest.Then,the H-Unet network proposed in this thesis s used to segment the fetal head area in small areas.Because ultrasound imaging may make the contours on both sides of the fetus’ s head missing,resulting in inaccurate contours detected on both sides of the fetus’ s head.This thesis proposes an anomaly detection method based on robust statistics to remove the abnormal pixels of the segmentation result and improve the accuracy of fetal head circumference measurement.This thesis conducts experiments on the HC18 fetal head ultrasound image data set to calculate and analyze the absolute head circumference error(AD),Hausdorff distance(HD)and Dice coefficient of this method.The results show that the method in this thesis is superior to other methods in these evaluation indicators,which verifies the effectiveness and superiority of the method in the measurement of fetal head circumference.