Research On Automatic Measurement Of Fetal Abdominal Circumference By Ultrasound Based On Deep Learning

Fetal ultrasound is a prenatal examination performed in almost all pregnant women.This is not only because of its low price,convenience and speed,but more importantly,there is no radiation,which will not cause damage to pregnant women and fetuses.Regular examination of the health status of the fetus and screening for fetal malformation can reduce the birth rate of children with severe malformations and achieve eugenism.The assessment of fetal growth and development mainly relies on the measurement of fetal growth parameters.For example,the measurement of fetal abdominal circumference can help determine whether it is macrosomia,and the prevention of macrosomia can reduce the proportion of dystocia caused by excessive fetal size,which has important practical significance.In the current clinical examination,the measurement of fetal abdominal circumference is estimated by the doctor’s manual delineation of the fitting ellipse,which is highly subjective and inefficient,and more dependent on the doctor’s clinical experience due to the poor quality of ultrasound images and high noise.Therefore,this thesis proposes a method for automatic segmentation of fetal abdominal ultrasound images and automatic measurement of abdominal circumference.This thesis is divided into two parts:(1)to establish an automatic segmentation system for fetal abdominal ultrasound images;(2)Abdominal circumference is automatically measured according to the results of automatic segmentation.For the problems of edge blurring,missing,and poor contrast in abdominal ultrasound images,based on deep learning method,we improve it under the U-Net network architecture.U-Net3+ model is used to change the jump connection mode,reduce the semantic gap between the shallow and deep features of the image,and add deep supervision.The hybrid loss function is introduced to fully extract and utilize the original features of the image,so as to realize the automatic and accurate segmentation of the abdominal circumference.In the automatic measurement of abdominal circumference,the results of automatic segmentation are first processed and extracted to obtain the abdominal circumference edge,and then the abdominal circumference is fitted based on the least squares ellipse fitting.The ellipse parameters are estimated and the ellipse circumference is calculated to obtain the abdominal circumference.Finally,we use the test set images to validate the above model.In the automatic segmentation system,the Dice coefficient and IOU are used to evaluate the effect of automatic segmentation.Among them,the Dice coefficient and IOU of U-Net3+ model for automatic abdominal segmentation reach 94.73% and 88.96%,respectively.Compared with the U-Net++model,the Dice and IOU values obtained by automatic segmentation are improved by 4.59% and4.36%,respectively.For the automatic measurement of abdominal circumference,the mean absolute error and mean relative error are used to evaluate the results of automatic measurement.The mean absolute error of abdominal circumference measured by U-Net3+ model is 2.37 mm,and the mean relative error is 1.68%.It can be seen that the automatic segmentation system of fetal abdominal ultrasound image and the automatic measurement system of abdominal circumference established in this thesis have good effects.