王琳,曹艳,邓子微,等.基于3D-nnUnet的三维超声胎肺容积自动测量[J].深圳大学学报理工版,2022,39(3):324-333.[doi:10.3724/SP.J.1249.2022.03324]
WANG Lin,CAO Yan,DENG Ziwei,et al.Automatic measurement of fetal lung volume by 3D ultrasound based on 3D-nnUnet[J].Journal of Shenzhen University Science and Engineering,2022,39(3):324-333.[doi:10.3724/SP.J.1249.2022.03324]
点击复制
基于3D-nnUnet的三维超声胎肺容积自动测量
王琳1，曹艳2，邓子微2，胡炯通2，梁嘉敏2，曹晓焱1，潘文雄1，严玉玲1，孙志伟1，杨鑫2，倪东2

1.深圳市福田妇幼保健院医学影像科，广东深圳518016;2.深圳大学医学部生物医学工程学院，广东深圳518071

人工智能;胎肺成熟度;三维超声;语义分割;深度学习;网络自适应;医学影像

Automatic measurement of fetal lung volume by 3D ultrasound based on 3D-nnUnet
WANG Lin1，CAO Yan2，DENG Ziwei2，HU Jiongtong2，LIANG Jiamin2，CAO Xiaoyan1，PAN Wenxiong1，YAN Yuling1，SUN Zhiwei1，YANG Xin2，and NI Dong2

1.Medical Image Department, Futian Women and Children Hospital, Shenzhen 518016, Guangdong Province, P. R. China;2.School of Biomedical Engineering, Health Science Center, Shenzhen University, Shenzhen 518071, Guangdong Province, P. R. China

artificial intelligence; fetal lung maturity; three-dimensional ultrasound; semantic segmentation; deep learning; network adaptation; medical image

DOI: 10.3724/SP.J.1249.2022.03324

备注

胎肺发育不良常会令胎儿在出生时引发严重的呼吸窘迫，甚至造成新生儿死亡．胎肺容积测量是临床上无创评估肺成熟度的一项重要手段，但现有的胎肺容积测量方法误差大、过程繁琐、耗时长且临床实用性差．本研究基于3D-nnUnet提出一种高效稳定的胎肺自动分割和测量方法，利用网络对胎肺数据的自适应，有效克服图像组织对比度低和边缘模糊问题，实现了三维超声胎肺的精确分割．针对胎肺超声图像在不同孕周差异大，以及样本数分布极不均衡的问题，提出利用困难样本聚焦和简单样本惩罚（hard-miningandeasy-penalized，HMEP）损失来提升模型的泛化能力和稳定性．与二维最优分割网络DeepLabV3+和3D-Unet分割结果相比，基于3D-nnUnet的分割网络性能最佳，分割准确率高达85.7%；HMEP损失能够使3D-nnUnet模型专注地学习少数困难样本，将分割准确率提升近2%；分割模型在不同孕周的数据上所测得胎肺容积和医生手动勾画的胎肺容积经一致性检验无显著的统计学差异．实验结果表明，该方法可高效实现三维超声胎肺的自动精确分割和容积测量，具有良好的稳定性和泛化能力，可避免以往胎肺容积测量方法繁琐耗时、误差较大的问题，在诊断胎肺发育状况及评估肺成熟度方面有较好的应用前景．
Fetal lung dysplasia often causes the severe respiratory distress and even neonatal death at birth. Fetal lung volume measurement is an important method for clinical non-invasive assessment of lung maturity. However, the existing methods for measuring fetal lung volume suffer from huge errors, tedious and time-consuming operations, and poor clinical practicability. For this reason, we first propose an efficient and stable automatic segmentation and measurement method for fetal lung volume based on 3D-nnUnet, which uses the network adaptation to effectively overcome the problems of low tissue contrast and blurred edges and achieves accurate segmentation of 3-dimensional (3D) fetal lung ultrasound images. In addition, to solve the problem of large difference and extremely uneven distribution of sample numbers at different gestational weeks, we innovatively propose a hard-mining and easy-penalized (HMEP) loss to improve the model generalization ability. Compared with the segmentation results of 2-dimension optimal DeepLab V3+and 3D-Unet, our network model has the best segmentation performance, and the accuracy is as high as 85.7%. In addition, HMEP loss enables the 3D-nnUnet model to focus on learning a few difficult samples, which increases the segmentation accuracy by nearly 2%. There is no significant statistical difference between the fetal lung volumes measured by our network and gold standard using the consistency test in different gestational weeks. The experimental results show that our method can efficiently realize the automatic and accurate segmentation and volume measurement of 3D ultrasound fetal lung, and has the good stability and generalization ability, which can avoid the problems of cumbersome, time-consuming operations and large errors in the previous fetal lung volume measurement methods. Our method has the promising application prospects in diagnosing fetal lung dysplasia and assessing fetal lung maturity.
·