CASIA OpenIR  > 复杂系统管理与控制国家重点实验室  > 先进机器人
介入手术机器人导航中导丝跟踪与造影剂检测方法研究
吴玉东
Subtype硕士
Thesis Advisor侯增广
2019-05-31
Degree Grantor中国科学院自动化研究所
Place of Conferral智能化大厦17层第七会议室
Degree Discipline控制理论与控制工程
Keyword三维影像导航 二维x射线图像 导丝跟踪 造影剂检测
Abstract

心血管疾病已成为全人类死亡头号原因,我国居民的患病率持续攀升。冠状动脉粥样硬化,俗称冠心病,是心血管疾病的主要类型,其治疗方式主要包括药物治疗、血管介入手术和冠脉搭桥手术。血管介入手术因其疗效确定、手术创伤小、术后恢复快等优点在临床上得到广泛应用。然而该手术在减少病人痛苦的同时也因辐射问题和缺乏有效导航手段问题给医生身体健康和操作技能带来挑战。集成有三维影像导航功能的主从式介入机器人系统是应对上述问题的有效解决方案,已成为工业界和学术界研究热点。目前,三维影像导航系统仍面临诸多挑战,如:术中导丝跟踪、术中造影剂检测、2D/3D图像配准等。

因此,本文在国家自然科学基金重点项目“血管微创介入手术机器人的基础问题研究”(61533016)支持下,在已有主从式血管介入机器人平台基础上,根据实际临床需求,以术中二维X射线图像为切入点,围绕术中导丝跟踪、造影剂检测、导丝头部末端定位、多任务集成等问题展开研究。本文的主要研究内容和贡献如下:

1. 基于编码器—解码器结构的导丝头部跟踪算法研究:首先,针对现有导丝跟踪算法运行时大多需手动初始化,且对相邻帧间变化不鲁棒的问题,本文将导丝跟踪任务建模为序列内单帧图像的语义分割问题,并采用编码器—解码器结构的全卷积网络实现导丝头部自动分割。为增加样本多样性,本文基于导丝头部为柔性结构、术中易形变的特性,提出了改良的局部弹性形变数据增强算法。此外,采用Dice Loss缓解训练时因背景像素占比较多导致的样本类别不平衡问题。最后,实验分析了不同损失函数及造影剂对分割结果的影响。量化实验显示,该算法在临床测试集上达到了导丝头部跟踪误差中值为0.4885像素,F1分数为0.9519的跟踪精度。

2. 造影剂检测算法研究:鉴于造影剂的自动检测是影像导航系统自动运行的前提,本文分别实现了基于SVM的多特征融合造影剂检测算法和基于卷积神经网络的造影剂检测算法。在基于SVM的多特征融合方法中,使用Frangi多尺度滤波器进行血管结构增强,并采用灰度均值、Hu不变矩和Haralick三种特征对图像进行描述。在实验中,我们探究了不同特征组合及不同融合方式对检测精度的影响,该算法在测试集上的最优$F_1$分数为0.9219。

基于卷积神经网络的方法中采用深度残差网络,在测试集上的F1分数为0.9568。

此外,为进一步展示神经网络所学习到的经验知识,本文通过CAM算法将模型对不同输入图像的判别区域进行可视化。可视化结果表明,卷积神经网络在造影剂检测时主要关注冠脉区域,与人类经验知识较为吻合。

3. 导丝头部末端定位算法研究及多任务系统集成:针对导丝头部分割算法无法有效定位导丝头部末端的问题,本文提出了基于关键点检测的导丝头部末端定位算法。该算法易于和前文导丝头部分割算法集成,且几乎不增加额外计算量。该算法在测试集上的定位误差为3.1617像素。此外,本文提出了一种结合导丝头部分割信息的后处理方案,以进一步提升末端定位精度。实验表明使用后处理可将定位误差减少至2.2414像素。最后,我们基于多任务学习理论,将导丝头部分割、导丝头部端点定位和造影剂检测三个任务进行集成形成了高效、统一的术中二维X射线处理框架。实验表明,集成后在不损失精度的情况下,减少了69.8GFLOPs的计算量,约占总计算量的30%。 该多任务系统在NVIDIA TITAN Xp GPU平台运行速度达20帧每秒,满足低辐射模式下导航系统的实时性要求。

Other Abstract

Cardiovascular disease has become the first cause of death worldwide, and the prevalence of Chinese residents continues to rise. Coronary atherosclerosis, commonly known as coronary heart disease, is a major type of cardiovascular disease, and its treatment methods mainly include drug therapy, vascular intervention surgery and coronary artery bypass surgery. Among them, vascular interventional surgery has been widely used in clinical practice because of its curative effect, small surgical trauma and rapid postoperative recovery. Vascular interventional surgery reduces the patient's pain but also poses challenges to the doctor's health and surgical operations due to radiation problems and lack of effective navigation. The master-slave interventional robot integrated with the 3D image-guided navigation system is an effective solution to the above problems, and has become a focused area both in industry and academia. At present, 3D image-guided navigation systems still face many challenges, such as intraoperative guidewire tracking, intraoperative contrast detection and 2D/3D image registration.

Therefore, supported by the key project of the National Natural Science Foundation of China (Grant 61533016), this paper focus on the problems such as guidewire tracking, contrast agent detection, guidewire endpoint localziation and integration of multiple tasks based on the requirements of vascular interventional surgery 3D image-guided navigation system and actual clinical. The main contributions of this paper are as follows:

1.  Guidewire tip segmentation based on encoder-decoder structure: At present, existing guidewire tracking algorithms are mostly based on spline fitting or hand-crafted feature learning, which need to be manually initialized, and are not robust to sharp changes between adjacent frames. In this paper, the guidewire tracking is modeled as a segmentation problem, and the fully convolution network based on the encoder-decoder structure is used to realize automatic segmentation of the guidewire tip. In addition, based on the flexible structure of the guidewire tip, the local elastic deformation data augmentation algorithm is proposed. And dice loss is used to mitigate the problem of class imbalance caused by the background pixel accounted for a large number. Finally, the effects of different loss functions and the effect of contrast agent on the segmentation results were analyzed experimentally. The algorithm achieves the median of tip precision of 0.4885 pixel and $F_1$ of 0.9519 on the clinical test set.

2. Contrast agent detection: The detection of contrast agent helps the image-guided navigation system to automatically apply 2D/3D image registration and breathing, heartbeat motion compensation algorithms. In this paper, contrast agent detection is modeled as a classification problem. SVM-based multi-feature fusion classification method and convolutional neural network-based (CNN-based) method are adopted. We use multi-scale Frangi filters to enhance vessels in 2D X-ray images and use mean of pixel value, Hu moment and Haralick to extract features, subsequently. In the experiment, we explored the effect of different feature combinations and fusion methods on the classification accuracy. The final F1 score of the SVM algorithm is 0.9219. For CNN-based method, the $F_1$ score of the CNN-based is 0.9568. To further demonstrate the empirical knowledge of CNN, we visualized the activation maps and discriminant regions of the model for different images using the CAM algorithm. The results show that the CNN determines the image containing the contrast agent through the vessels containing the contrast agent.

3. Guidewire endpoint localization and multitask intergated system: We proposed an algorithm for the endpoint localization based on keypoints detection algorithm. The algorithm can be easily integrated with the guidewire tip segmentation algorithm with little extra computational effort. The error of algorithm on the test set is 3.1617 pixels. Then, we proposed a post-processing method to make use of tip segmentation results to improve the accuracy of endpoint localizations. Experiments show that using this post-processing can reduce the localization error to 2.2414 pixels. Finally, we try to integrate the three tasks of guidewire tip segmentation, guidewire tip endpoint localization and contrast agent detection based on multi-task learning theory to achieve an efficient and unified intraoperative 2D X-ray processing framework. Experiments demonstrate that after integration, the computational cost is reduced by approximately 30% while without loss of accuracy of each task. The algorithm runs at 20 frames per second on the NVIDIA TITAN Xp GPU platform.

Pages82
Funding ProjectNational Natural Science Foundation of China[61533016]
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23834
Collection复杂系统管理与控制国家重点实验室_先进机器人
Recommended Citation
GB/T 7714
吴玉东. 介入手术机器人导航中导丝跟踪与造影剂检测方法研究[D]. 智能化大厦17层第七会议室. 中国科学院自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
Thesis.pdf(18414KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[吴玉东]'s Articles
Baidu academic
Similar articles in Baidu academic
[吴玉东]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[吴玉东]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.