基于D-UNET的胎儿产前超声检测

胎儿的产前超声检测对判断或预测孕妇孕期、估计胎儿尺寸和质量具有重要意义.超声检测主要是基于超声图像对胎儿的腹围(abdominal circumference)、股骨长(femur length)以及头臀径(crown-rump length)等参数进行测量.这些生物参数的测量可以用来判断胎儿的生长状况是否良好以及胎儿是否畸形.当前,这些参数需要依靠医生进行手动测量,该方法效率低下且严重依赖医生经验,从而导致检测准确率下降.对此,本文提出了一种基于深度学习的算法来对腹围、头围、股骨等部位进行自动分割,并对这些生物参数进行自动测量.该算法基于UNET结构,并采用扩张卷积(dilation convolution)以提高网络提取上下文信息的能力,因此本文将其命名为D-UNET.本文基于腹围、股骨以及头臀径三个临床数据集对所提D-UNET进行了实验验证,并与一些其他先进的深度学习分割算法进行了比较.实验结果表明,本文算法在三个数据集上都表现了与专家医师手动标注接近的测量结果.由此说明,本文算法能辅助医师对生物参数进行自动测量.

Fetal prenatal ultrasound detection based on D-UNET

The ultrasound detection of the fetus is of great significante to judge or predict the pregnancy and estimating the fetal size and quality. Ultrasonic measurement is mainly conducted in abdominal circumference, femur length and crown-rump length of the fetus. Measurement of these biological parameters can be used to determine whether the fetal growth is good or deformed. In general, these parameters need to be measured by doctor manually, which is inefficient because of repeated operations and often misdiagnosed due to shortcomings such as relying on the doctor''s experience. In response to this kind of problem, in this paper, we propose a deep learning-based method to automatically measure the abdominal circumference, femur length and crown-rump length. The method is based on the UNET structure and combines the strategy of dilation convolution to improve the network''s ability to extract context information and named it D-UNET. Validation was done on three datasets: abdominal circumference, femur and crown-rump length and compared with some other advanced deep learning segmentation algorithms. Clinically, our method demonstrated measurements that were very close to the manual labeling by expert physicians on all three datasets. This shows that our method can assist physicians in the automatic measurement of these biological parameters.