【作者】 张宏；

【导师】 顾力栩；

【作者基本信息】 上海交通大学 ， 生物医学工程（专业学位）， 2021， 硕士

【摘要】 经皮冠状动脉介入手术是治疗冠心病的常见手段之一。在手术前医生会拍摄病人的CT血管成像(CTA,CT Angiography),了解病人冠状动脉的三维空间结构,为接下来的手术路径规划提供丰富的冠状动脉结构信息。在术中,医生会拍摄X光冠状动脉造影(XA,X-ray Angiography),在术中XA的引导下将导管送至病变处,然后通过放置球囊或对堵塞段切割等方式进行治疗。由于心脏的跳动和人体的呼吸作用,冠状动脉的拓扑结构会发生复杂的形变,临床医生需要丰富的经验才能根据二维的术中XA图像来判断导管在冠状动脉中正确的行进方向。而且,术中XA图像本身的低分辨率和二维信息的不确定性也给医生带来了很大的困扰。如果能将术前CTA与术中XA实时准确地融合起来,将很大程度上降低术中XA图像的不确定性,帮助医生更好的做出路径规划,进一步提高手术的成功率。但是,由于二者模态和维度上的不同,冠状动脉配准面临着巨大的挑战。本研究充分利用深度学习方法实时准确等优势,创新性地提出了基于深度学习的两阶段术前CTA与术中XA的配准方法。首先,冠状动脉配准本质上是三维图像与二维图像配准。为了解决二者模态和维度上的不同,本文分别训练了3D U-Net和2D U-Net对术前CTA和术中XA的冠状动脉进行提取,得到对应的三维冠状动脉和二维冠状动脉分割结果。为了消除分割结果的噪声和背景伪影导致的假阳性,本文采用最大连通区域算法保留了主要的冠状动脉部分。接着,本文利用Zhang-Suen骨架提取算法对分割结果进行中心线提取,为接下来冠状动脉配准提供基本数据特征。然后,针对于三维CTA冠状动脉中心线和二维XA冠状动脉中心线,本文采取基于特征的配准方法,创新性地提出了基于深度学习的两阶段配准方法。第一阶段寻找冠状动脉配准的初始位姿,对齐拍摄角度,弥补由于拍摄机器导致的几何空间上的不一致。本文采用基于深度自编码器的配准方法,深度自编码器能提取出冠状动脉中心线的高维特征,然后计算特征之间的误差,以迭代优化的方式去寻找一个最优变换矩阵。第二阶段在获取到了初始位姿的基础上,对三维CTA进行投影,寻找投影中心线和术中XA中心线的非刚性形变,弥补心脏运动和人体呼吸导致的冠状动脉解剖结构的变化。本文提出了一个点云配准网络来预测CTA投影中心线和XA之间的形变场。为了让形变场更加合理,本文提出了全局拓扑学约束和局部拓扑学约束来约束形变场满足一定的解剖学上的意义,进一步提高配准的精度。综上所述,本文对冠状动脉配准任务中的冠状动脉分割和中心线配准两个关键技术进行了探究,并提出了基于深度学习的两阶段配准方法。实验结果表明,相对于之前的方法,本文提出的方法是实时准确的,具有良好的临床应用前景。更多还原

【Abstract】 Percutaneous coronary intervention is one of the common methods to treat coronary heart disease.Before the operation,the preoperative CT Angiography(CTA)allows the doctor to understand the three-dimensional structure of the patient’s coronary artery,so as to provide abundant information about the structure of the coronary artery for the following surgical path planning.During the operation,the catheter will be delivered to the stenosis under the guidance of the intra-operative X-ray Angiography(XA),and then treat the stenosis by placing a balloon or cutting the blockage.Due to the heartbeat and respiration,the topology of the coronary arteries will undergo complex deformation,and the doctor needs rich experience to judge the correct direction of the catheter in the coronary artery based on the intra-operative XA images.Moreover,the low resolution and uncertainty of the XA also present great challenges to doctors.If the preoperative CTA and intra-operative XA can be accurately fused in real time,the uncertainty of the XA will be greatly reduced,which will help doctors make better path planning and further improve the success rate of surgery.However,due to their differences in mode and dimension,it challenges the registration of the coronary arteries.In this thesis,a two-stage registration method of preoperative CTA and intra-operative XA based on deep learning is proposed by taking full advantage of the real time and accuracy of deep learning method.First,coronary artery registration is essentially 3D-2D registration.In order to solve the differences in mode and dimension,3D U-Net and 2D U-Net were trained to extract the coronary arteries of preoperative CTA and intra-operative XA respectively.In order to eliminate the noise of segmentation and false positives caused by background artifacts,the maximum connected region algorithm was used to preserve the main coronary artery.Then,we use Zhang-Suen skeleton extraction algorithm to extract the centerline of coronary arteries from the segmentation,which provides basic data features for the subsequent step.Then,for 3D CTA centerline and 2D XA centerline,we adopt a features-based registration method and proposes a two-stage registration method based on deep learning.The first stage aims to find the initial posture to make up of the inconsistency in the geometric space caused by the shooting machine.We employ an autoencoder to extract the high-dimensional features of the centerline of the coronary artery and calculate the error between the features to find an optimal transformation matrix by iterative optimization.In the second stage,based on the initial position,the CTA is projected to find the non-rigid deformation of the projection centerline and the intra-operative XA centerline to compensate for changes in the anatomical structure caused by the heartbeat and respiration.This paper proposes a point cloud registration network to predict the deformation field between the centerline of the projection and XA.In order to make the deformation field more reasonable,global and local topology constraints are proposed to constrain the deformation field to meet certain anatomical significance and further improve the accuracy of registration.To sum up,we explore two key techniques of coronary artery segmentation and centerline registration and carry out a two-part experiment through experimental design.The results show that,compared with the previous methods,our method is real-time and accurate,and has a good prospect of clinical application.更多还原