Automatic measurement of fetal head circumference using a novel GCN-assisted deep convolutional network

Abstract

The growth of the fetus can be effectively monitored by measuring the fetal head circumference (HC) in ultrasound images. Moreover, it is the key to assessing the fetus's health. Ultrasound fetal head image boundary is blurred. The ultrasound sound shadow results in a partial absence of the skull in the image. The amniotic fluid and uterine wall form a structure similar to the head texture and grayscale. All these factors result in challenges to ultrasound fetal head edge detection. The new convolutional neural network (CNN) named GAC Net was proposed in this paper, which can effectively solve the above problems. GAC Net is an end-to-end network model constructed by the encoder and decoder. In order to suppress the interference of ultrasound image quality defects on the HC measurement, the graph convolutional network (GCN) module was added to the connection channel between the encoder and the decoder. The new attention mechanism enhanced the network's ability to perceive border areas. Experiments were performed on the HC18 fetal head ultrasound image data set. The following objective evaluation indicators were calculated, including the Hausdorff distance (HD), the absolute difference (AD), the difference (DF), and the Dice similarity coefficient (DSC) of head circumference. Experimental results showed that GAC-Net had an HD of 1.22 ± 0.71 mm, an AD of 1.75 ± 1.71 mm, a DF of 0.19 ± 2.32 mm, and a DSC of 98.21 ± 1.16%. The overall performance of the proposed algorithm exceeded the state-of-the-art methods, which fully proved the effectiveness of the GAC Net presented in this paper.