基于深度学习三维计算超声的早孕期胎儿发育精准评估新方法研究

Evaluation of fetal development based on ultrasound volumetric parameters is a hot topic in the development of prenatal ultrasound diagnosis. However, the calculation of ultrasound volumetric parameters requires clinical experts to manually segment volumetric images, which is extremely time-consuming and labor-intensive. In addition, the segmentation process is highly user-dependent. Therefore, it is very difficult to apply such volumetry based evaluation in clinical diagnosis. In this project, we propose a new method for the evaluation of fetal development based on deep learning driving computerized volumetric ultrasound. The method combines the power of medical ultrasound diagnosis and computer science. Based on our customized 3D Fully Convolutional Network, we propose to inject a Recurrent Neural Network (RNN) to flexibly explore 3D semantic knowledge from a novel, sequential perspective, and therefore significantly refine the local segmentation result which is initially corrupted by the ubiquitous boundary inconsistency in ultrasound volumes. Considering sequence hierarchy, we introduce a hierarchical deep supervision mechanism to effectively boost the information flow within RNN and further improve the semantic segmentation results. We then precisely segment the fetal, placenta and chorionic sac from volumetric ultrasound, and then calculate the growth table of fetal volumetric parameters and obtain the gestational age prediction model. A series of research methods proposed in this project are primitive and can provide practical methods and tools for early diagnosis of ultrasound, as well as to accumulate knowledge for a number of scientific problems in the study of medical ultrasound.

基于超声容积生物学参数评估胎儿发育是产前超声诊断发展的前沿热点课题。但是，容积生物学参数计算需临床专家手动分割容积图像，极其耗时耗力，且存在高用户差异，难以在临床诊断中应用。本项目深度融合产前超声诊断和计算机科学，提出研究基于深度学习三维计算超声的早孕期胎儿发育评估新方法，主要包括：研究基于全卷积神经网络的深度学习模型，进行端到端学习和预测，并通过有效融合不同尺度信息获得准确的分割结果；研究基于循环神经网络的空间上下文信息编码方法，提高容积超声中解剖结构模糊边界和缺失处的预测准确性；研究多层次递进误差协同驱动的深度监督策略，提升深度神经网络的训练效率和泛化能力；实施容积数据中胎儿、胎盘和绒毛膜腔的精准自动分割，进而建立胎儿容积生物学参数生长表和胎龄精准预测模型。本项目提出的一系列研究方案具有原始初创性，可为早孕期超声诊断提供实用方法与工具，也可为计算医学超声研究的若干科学问题提供知识积累。

本项目是融合产前三维超声诊断和计算机科学的研究，提出并研究了基于深度学习三维计算超声的早孕期胎儿发育评估新方法。在数据收集上，已采集早孕期胎儿容积数据不少于9000个，其中137套由产前超声医师标注并由专家修改，用于深度学习模型研究。深度学习三维计算超声模型研究方面：定制了三维容积下的全卷积神经网络模型；研究了基于循环神经网络的空间上下文信息编码方法；研究了多层次递进误差协同驱动的深度监督策略。最终，本研究的模型对三维超声中胎儿和羊膜腔的分割精度均达到85%以上。本研究亦通过模拟羊膜腔实验了解了胎儿实际容积与模型所计算的容积差异，并利用该模型的自动和半自动分割功能计算了妊娠早期胎儿各孕周容积的参考值。本项目基本完成了项目目标：共发表学术论文24篇，其中SCI收录8篇。已经及正在培养研究生13人。出版专著2本，申请专利1个。本研究的深度学习三维计算超声模型可在容积超声中准确自动获取容积生物学参数，有望推动容积生物学参数在产前超声诊断中的广泛应用。