Prenatal 3D Ultrasound Image Segmentation And Automatic Positioning Of Anatomical Points

Medical ultrasound imaging because of being safe and non-destructive,plays an important role in reducing the congenital malformation rate of infants,and is the preferred imaging technique for prenatal diagnosis and birth defect screening.Two-dimensional ultrasound parameter measurement is widely used at present,but it is heavily dependent on the doctor's operation and experience.With the rapid development of three-dimensional ultrasound imaging technology in recent years,the volumetric biological parameters based on three-dimensional ultrasound and the localization of anatomical structures can reflect the development of the fetus and related tissues as a whole,on the one hand,the influence of fetal position or fetal posture is low.On the other hand,it can reduce the problem of two-dimensional ultrasound user differences.However,to obtain the biological parameters of volume and anatomical points,manual segmentation and manual locating of threedimensional ultrasound images are taken,which is difficult to be widely promoted in clinical practice.Therefore,an automatic segmentation of three-dimensional ultrasound images and automatic locating method of anatomical structures are desired.In this paper,for the calculation of fetal head volume in prenatal ultrasound diagnosis,as well as the calculation of thigh volume and the location of femoral structure points,fetal Head Segmentation in Volumetric Ultrasound in three-dimensional ultrasound images is proposed.The algorithm is used to obtain the fetal head volume parameter;Joint Image Segmentation and Landmark Localization in Thigh Volumetric Ultrasound algorithm for the joint segmentation and localization of the fetal thigh volume and anatomical points in the threedimensional ultrasound image is proposed to realize the automatic quantitative analysis of the threedimensional ultrasound image,providing efficient and accurate analytical tools for prenatal diagnosis.The segmentation of the fetal head volume image faces several challenges:(1)due to the strong echo of the skull,the segmentation of the fetal head volume image faces a large missing area;(2)the diversity of the fetus in the maternal body;(3)There is no obvious biological anatomical boundary between the fetal head and neck.Fetal thigh volume image segmentation and anatomical points locating have the following difficulties:(1)low contrast between tissues,blurred edge of the thigh;(2)occlusion of the femur,large area missing in the far field;(3)There are many similar structures in the volume data with the anatomical points,resulting in false positives in the location of the anatomical points.In order to overcome the challenges of fetal head volume image segmentation,this paper proposes a new model of three-dimensional ultrasound fetal head segmentation based on deep convolutional neural network,and attention mechanism used to make the network pay more attention to the target;Random data erasure strategies proposed to augment our data volume to prevent overfitting problems;automatic context iteration models is utilized to further refine the segmentation results.The quantitative results of the Dice coefficient obtained in 50 sets of data tests were 96.05% and were in good agreement with the results marked by the sonographer.In order to solve the difficulties in segmenting volume image and locating the fetal femur anatomical points,a deep neural network based on multi-task learning is proposed.This method makes full use of the feature sharing mechanism between tasks.In order to increase the similarity between the two anatomical points,we propose a distance constraint loss to optimize this problem.Finally,in order to utilize the relationship between thigh volume and anatomical points,GAN(Generative Adversarial Network)is used to optimized.Our results were finally obtained with a segmentation Dice Cofficient 91.05%.And the locating error of the anatomical points is an average of 3.97 pixels.Fetal Head Segmentation in Volumetric Ultrasound and Joint Image Segmentation and Landmark Localization in Thigh Volumetric Ultrasound algorithms proposed in this paper have good versatility and can be extended to other parts of the fetus for three-dimensional ultrasound volume image segmentation and anatomical structure point location,such as fetal abdominal volume image,fetal upper arm volume image,etc.,we can use then together in the future.Multiple locations of the fetus further optimize fetal development assessment in prenatal ultrasound.