Current Issue Cover


摘 要
目的 超声胎儿头部边缘检测是胎儿头围测量的关键步骤,因胎儿头部超声图像边界模糊、超声声影造成图像中胎儿颅骨部分缺失、羊水及子宫壁形成与胎儿头部纹理及灰度相似的结构等因素干扰,这给超声胎儿头部边缘检测及头围测量带来一定的挑战。在此,本文提出一种基于端到端的神经网络超声图像分割方法,用于胎儿头部边缘检测。方法 以UNet++神经网络结构为基础,结合UNet++最后一层特征,构成融合型UNet++网络。训练过程中,为缓解模型训练过拟合程度,在每一卷积层后接一个空间Dropout层。其具体思路:通过融合型UNet++深度神经网络提取超声胎儿头部图像特征,通过胎儿头部区域概率图预测,输出胎儿头部语义分割的感兴趣区域。进一步获取胎儿的头部边缘关键点信息,并采用边缘曲线拟合方法拟合边缘,最终测量出胎儿头围大小。结果 针对现有二维超声胎儿头围自动测量公开数据集HC18,以Dice系数(Dice Coefficient)、豪斯多夫距离(Hausdorff Distance,HD)、头围绝对差值(Abs Difference,AD)等指标评估本文所提出的模型性能。其结果为Dice系数98.06%,HD距离1.21 ± 0.69mm,头围测量AD 1.84 ± 1.73 mm。在妊娠中期测试数据中,Dice系数98.24%,HD距离1.15 ± 0.59mm,头围测量AD 1.76 ± 1.55mm。在生物医学图像分析平台Grand Challenge上HC18数据集已提交结果中,融合型UNet++的Dice系数排名第三,HD排名第二,AD排名第十。结论 与经典超声胎儿头围测量方法及已有的机器学习方法应用研究相比较,融合型UNet++能有效地克服超声边界模糊、边缘缺失等干扰,精准分割出胎儿头部感兴趣区域,获取边缘关键点信息。与现有神经网络框架比较,融合型UNet++充分利用上下文相关信息与局部定位功能,在妊娠中期的头围测量中,本文方法明显优于其他方法。
Fetal head edge detection in 2D ultrasound image using Fusion UNet++

Xing Yanyan,Yang Feng,Tang Yujiao,Zhang Liyun(School of Biomedical Engineering,Southern Medical University)

Abstract: Objective Ultrasound fetal head circumference measurement is very important to monitor the growth of fetus and estimate the gestational age. Computer aided measure fetal head circumference is valuable for sonographers who are short of experiments in ultrasound examinations. With the help of computer aided measurement they can detect fetal head edge more accurately and finish an examination more quickly. Fetal head edge detection is a necessary process for automatic measurement of fetal head circumference. Ultrasound fetal head image boundary is fuzzy, and the gray-scale of fetal head is similar to mother’s abdominal tissue especially in the first trimester. Ultrasound shadow leads to the loss of head edge and incomplete fetal head in the image, which brings certain difficulties to detect complete fetal head edge and fit head ellipse. The structures of amniotic fluid and uterine wall are similar to head texture and gray-scale, which often leads to misclassify this part as fetal head. All the above factors bring some challenges to ultrasound fetal head edge detection. Therefore, we propose a method for detecting the ultrasound fetal head edge by using convolutional neural network to segment the fetal head region end-to-end. Method In this paper, the model we proposed is based on UNet++. In deep supervised UNet++, every output can provide a predict result of the region of interest and each of them is different, but only the best predicted result will be used to predict the region of fetal head. Generally, the output results are more and more accurate from left to right. There are four feature blocks before four outputs of UNet++, the left feature contains more location information and the right contains more sematic information. In order to take full advantage of the feature map before outputs, we fuse them by concatenation and extract fused features further. The improved model is named Fusion UNet++. In order to prevent overfitting, we introduce spatial dropout after each convolutional layer instead of standard dropout, which extends the dropout value across an entire feature map. Here’s the specific idea of fetal head circumference measurement: Firstly, we use Fusion UNet++ to learn features of 2D ultrasound fetal head image, and then get the semantic segmentation result of fetal head by using fetal head probability map. Secondly, according to the result of image segmentation we extract the edge of the fetal head by using edge detection algorithm, then use the direct least square ellipse fitting method to fit the head contour. Finally, the fetal head circumference can be calculated according to the ellipse circumference formula. Result For the open data set of automated fetal head circumference measurement of two-dimensional ultrasound image named HC18 on Grand Challenges, which contains first, second and third trimester images of fetal heads. All of the fetal head images are the standard plane of measure fetal head circumference. In HC18 dataset, there are 999 2D ultrasound images with annotations of fetal head circumference in the train set and 335 2D ultrasound fetal head images without annotations in the test set. We use the train set to train convolutional neural network and submit the predicted results of the test set to participate in the model evaluation on HC18, Grand Challenges. In this paper, we use Dice coefficient, Hausdorff Distance (HD) and Absolute Difference (AD) as assessment indexes to quantitatively evaluate the proposed method. With proposed method, for all the three trimesters dataset of fetal head images, the fetal head segmentation Dice coefficient is 98.06%, HD is 1.21±0.69 mm and fetal head circumference measurement AD is 1.84±1.73 mm. The skull in second trimester is visible and shows as a bright structure, it is invisible in the first trimester and visible but incomplete in the third trimester. It is difficult to see complete skull in the first and third trimester, so second trimester fetal head circumference measure result is the best in all trimesters. And for most algorithms, they measure fetal head circumference only on second trimester or second and third trimesters fetal head ultrasound image. For the second trimester, the fetal head segmentation Dice coefficient is 98.24%, HD is 1.15±0.59 mm and fetal head circumference measurement AD is 1.76±1.55 mm. Among the results presented in the open test set, our Dice ranked the third, HD the second and AD the tenth. Conclusion Compared with the traditional method and machine learning methods, the proposed method can overcome the interference of fuzzy boundary and lack of edge effectively and can segment fetal head region accurately. Compared with the existed neural network method, in the second trimester of pregnancy, whether in fetal head segmentation or head circumference measurement the proposed method surpasses other methods. Proposed method achieves the state-of-the-art fetal head segmentation results.