中国科学院自动化研究所机构知识库

心血管疾病是威胁人类健康的头号杀手。冠心病作为一种常见的心血管疾病，其死亡人数占据了所有心血管疾病死亡人数的主要部分。目前，冠脉介入手术已经成为治疗冠心病的主要手术方式。随着冠脉介入手术例数的迅速增长，手术医师短缺的问题暴露了出来，而且这一问题在未来还会更加严重。想要解决手术医师短缺的问题，一个有效的解决方法是开发血管介入手术机器人，将手术医师从冠脉介入手术中解放出来。血管介入手术机器人需要拥有与手术医师相同的视觉感知能力，具体来说，就是从数字减影血管造影（Digital Subtraction Angiography， DSA）影像中获取手术所需关键信息的能力。本文在国家自然基金重点项目``血管微创介入手术机器人的基础问题研究''（61533016）等项目的支持下，针对赋予血管介入手术机器人视觉感知能力这一目标，结合手术医师实施冠脉介入手术的具体需求，围绕DSA影像中导丝端点定位、冠脉血管分割以及冠脉血管宽度估计展开研究。论文的主要内容和创新点如下：

（1）针对DSA影像中多导丝端点定位任务，提出了一个两阶段的多导丝端点定位方法。该方法将多导丝端点定位任务分解为导丝检测与单导丝端点定位两个子任务。在导丝检测子任务上，通过利用连续帧中的时序信息，设计了一个检测结果的后处理算法，该算法可以提升连续帧中导丝检测结果的正确率。在单导丝端点定位子任务上，基于多任务学习思想设计了一个分割注意力模块，并将其与堆叠沙漏模型相结合，形成了SA-hourglass模型。实验结果显示，SA-hourglass模型在单导丝端点定位任务上的平均定位误差仅为2.20±0.15像素，显著优于目前已有方法。在实验中。SA-hourglass模型还被应用到了视网膜手术的手术器械关键点定位任务中，平均定位误差仅为5.30±0.18像素。

（2）在DSA影像中多导丝端点定位任务上，进一步提出了单模型的多导丝端点定位方法KL R-CNN。KL R-CNN模型以实例分割模型Mask R-CNN为基础，该模型的创新点在于两点：第一，对模型参数进行了诸多修改，解决了模型输出分辨率过低的问题；第二，设计了一个新的关键点定位分支，解决了现有关键点定位分支容易出现的过拟合问题。实验结果显示，在多导丝端点定位任务上，KL R-CNN模型的多个评价指标显著优于目前已有方法。当KL R-CNN模型与本文提出的两阶段多导丝端点定位方法进行对比时，两种方法在不同的评价指标上各有优势。

（3）针对DSA影像中冠脉血管分割任务，提出了一个基于多尺度特征融合的冠脉血管分割模型CAU-net。鉴于目前缺乏任务相关的公开数据集，建立了一个用于DSA影像中冠脉血管分割任务的数据集。为解决现有血管分割算法没有充分利用多尺度特征的问题，设计了一个多尺度特征融合模块，并将该模块与经典医学图像分割模型U-net相结合，形成了CAU-net模型。实验结果表明，多尺度特征融合模块的引入可以将分割模型在冠脉血管分割任务上的Dice分数提高2.18%。同时，CAU-net模型的冠脉血管分割Dice分数达到了90.35%，显著优于目前已有算法。

（4）针对二维图像的血管宽度估计任务，提出将血管宽度估计任务定义为一个像素级的数值回归任务，并在该任务定义下，提出了一个基于卷积神经网络的血管宽度估计模型。鉴于目前缺乏用于训练血管宽度估计模型的训练标签，设计了一个血管宽度标签生成算法，该算法可以利用血管分割标签生成血管宽度值标签。在提出的血管宽度估计模型中采用了分治思想，用于缓解因训练样本分布不均衡带来的长尾效应。实验结果显示，在冠脉血管宽度估计任务上，在血管宽度估计模型中引入分治思想可以有效缓解由训练样本分布不均衡带来的不利影响，平均误差标准差仅为1.51个像素。在视网膜血管宽度估计任务上，基于卷积神经网络的血管宽度估计模型在宽度估计精度和运行速度上都优于目前已有方法。

Cardiovascular diseases have been the number one killer to humans. Coronary heart disease, as a common cardiovascular disease, accounts for the majority of all cardiovascular deaths. At present, Percutaneous Coronary Intervention (PCI) has become the main surgical treatment for coronary heart disease. With the rapid increase in the number of PCIs, the shortage of surgeons has been exposed, and the shortage problem will become more serious in the future. To solve the problem of surgeons' shortage, an effective solution is to develop the PCI robot, which can liberate surgeons from PCIs. The PCI robot needs to have the same visual perception as the surgeon, specifically, the ability to obtain image information from Digital Subtraction Angiography (DSA) images. Supported by the National Natural Science Foundation of China (61533016), this dissertation aims at endowing the PCI robot with visual perception ability. Combining with the specific needs of surgeons when performing PCIs, this dissertation focuses on the problems of guidewire endpoint localization, coronary artery segmentation, and coronary artery width estimation in DSA imaging. The main contributions and innovations of this dissertation are as follows:

（1）A two-stage multi-guidewire endpoint localization method is proposed for the task of multi-guidewire endpoint localization in DSA images. This method decomposes the multi-guidewire endpoint localization task into two sub-tasks: guidewire detection and single guidewire endpoint localization. In the sub-task of guidewire detection, a post-processing algorithm of detection results is designed by using the time sequence information in continuous frames, which can improve the accuracy of guidewire detection results in continuous frames. In the sub-task of single guidewire endpoint localization, a segmentation attention module is designed based on the multi-task learning idea, and this module is combined with the stacked hourglass model to form the SA-hourglass model. Experimental results show that the mean error of the SA-hourglass model is only 2.20±0.15 pixels on single guidewire endpoint localization, which is significantly better than the existing methods. In the experiment, the SA-hourglass model is applied to the keypoint localization task of surgical instruments in retinal surgery, and the mean error is only 5.30±0.18 pixels.

（2）Aiming at the multi-guidewire endpoint localization in DSA images, a unified multi-guidewire endpoint localization model named KL R-CNN is further proposed. KL R-CNN model is based on the instance segmentation model Mask R-CNN. The innovation of the KL R-CNN model lies in two points: Firstly, some model settings are modified to solve the problem of low output resolution of localization results. Secondly, a novel keypoint localization branch is designed to solve the overfitting problem of the existing keypoint localization branch. Experimental results show that compared with the existing methods, the proposed KL R-CNN achieves the state-of-the-art performance in many evaluation metrics of the multi-guidewire endpoint localization task. When the KL R-CNN model is compared with the two-stage multi-guidewire endpoint localization method proposed in this dissertation, the two methods have their advantages in different evaluation metrics.

（3）Aiming at the coronary artery segmentation task in DSA images, a coronary artery segmentation model CAU-net based on multi-scale feature fusion is proposed. Due to the lack of available task-related datasets, a dataset is established for the coronary artery segmentation task in DSA images. In order to solve the problem that existing vessel segmentation algorithms do not make full use of multi-scale features, a multi-scale feature fusion module is designed, and this module is combined with the classical medical image segmentation model U-net to form the CAU-net model. The experimental results show that the introduction of the multi-scale feature fusion module can improve the Dice score of coronary vessel segmentation by 2.18%. Meanwhile, the Dice score of the CAU-net model is 90.35%, which is significantly better than the existing methods.

（4）For the vessel width estimation task in two-dimensional images, the vessel width estimation task is defined as a pixel-level numerical regression task, and a convolutional neural network-based vessel width estimation model is proposed under this task definition. In view of the lack of training labels for training vessel width estimation models, a vessel width label generation algorithm is designed, which could generate vessel width labels by using vessel segmentation labels. The divide-and-conquer idea is adopted in the proposed vessel width estimation model to alleviate the long tail effect caused by the uneven distribution of training samples. In the coronary artery width estimation task, the experimental results show that the introduction of the divide-and-conquer idea can effectively alleviate the negative effects caused by the uneven distribution of training samples, with the mean standard deviation of error is only 1.51 pixels. In the task of retinal vessel width estimation, the convolutional neural network-based vessel width estimation model performs better than the existing methods in both estimation accuracy and running speed.