Research On Angle Of Progression Measurement Based On Intrapartum Ultrasound Image

The Angle of Progression(Ao P)indicates the relative position of the pubic symphysis and fetal head during labor.It is used to locate the fetal head and evaluate the labor progress.Nowadays,the obstetricians can measure Ao P with the help of ultrasound equipments.This paper presents an automatic Ao P measurement approach based on semantic segmentation of Transperineal Ultrasound(TPU)images.First,a convolutional neural network is used to segment the area of pubic symphysis and fetal head in a TPU image.Then the equation of an ellipse that best fits the area is obtained,based on which the Ao P is measured and annotated.Typical TPU images contains noise,blurred fetal head area,and inconspicuous area boundaries,making it difficult for semantic segmentation.This paper proposes a U-Net based Double Branch Segmentation Network(DBSN)tailor-made for TPU images.It consists of two parts: the encoder receives the image input,and the decoder is composed of both deformable convolution blocks and ordinary convolution blocks.The decoder is further divided into upper and lower branches.The feature map of the lower branch is used as the input of the upper branch,and after constrained by the Attention Gate(AG),assists the upper branch in its decoding.Experimental results show that DBSN yields an average segmentation accuracy of 93.38%(in terms of Dice coefficient),and the average angle error of Ao P between DBSN and an obstetrician is 5.993°.It is worth noting that,given a parameter size of 39.21 MB,DBSN can process 8 TPU images per second in a single GTX2080 Ti environment.To further deploy DBSN in scenarios with limited resources,this paper proposes a Grouped Atrous Convolutional Network(GACN),which significantly reduces memory usage and speedup inference through a reasonable combination of grouped,atrous,and decomposed convolutions.Experimental results show that it yields an average segmentation accuracy of92.94%(in terms of Dice),and the average angle error of Ao P between GACN and an obstetrician is 7.315°.With a modest parameter size of 3.05 MB,GACN can process 98 TPU images in a second in a single GTX2080 Ti environment.Three post-segmentation region processing algorithms are proposed,namely Center of Gravity-Standard Circle(CGSC)algorithm,Segmentation Region Cropping(SRC)algorithm,and Protective Region Clipping(PRC)algorithm.Given the fact that a fetal head is morphologically similar to a standard circle,CGSC fits the fetal head area into a standard circle based on its geometric characteristics.It reduces the average measurement error of Ao P by0.73°,as the experimental results suggest.SRC continuously cuts the deformation area of the pubic symphysis and fetal head to make the remaining area closer to an ellipse.Experimental results show that it reduces the average measurement error of Ao P by 0.96°.PRC is an SRC enhancement which introduces a protection mechanism to the cutting process,and thus retains the normal fetal head area.Experimental results show that a combination of PRC and SRC reduces the average measurement error of Ao P by 1.12°.