中国科学院自动化研究所机构知识库

心血管疾病已经成为了威胁人类健康的“头号杀手”。作为治疗心血管疾病的主要方式，冠状动脉介入手术以其创伤小、痛苦小、术后恢复快等优势而被越来越多的患者所接受。然而，现有的冠脉介入手术仍然存在一些突出问题：手术过程中介入器械运动识别和跟踪主要依赖于影像系统提供的低质量二维图像，缺少重要的三维解剖结构信息；X射线的累积辐射对医生的身体伤害非常大；大剂量的造影剂注射容易导致其它并发症的发生。此外，介入操作技能评价主要基于专家观察的手术操作核查评价表或总体评价量表，主观性非常强，而且涉及到大量的评价内容，评价过程极为耗时。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)和国家自然科学基金项目“基于专家医生操作技能的血管介入机器人的自适应学习控制”(61611130217)等课题的支持下，针对当前冠脉介入手术存在的问题，围绕手术过程中复杂自然操作行为的智能分析与应用展开研究。

论文的主要内容和创新点如下：

（1）针对冠脉介入手术自然操作中医生操作行为的获取问题，设计了一款新型的复杂自然操作行为采集系统。该系统通过有效整合多种运动捕捉传感器(肌电信号传感器、电磁跟踪传感器、三轴加速度计以及光纤弯曲传感器)，可同步获取医生的肌肉活动、手部运动、近端操作力以及手指运动等四类自然操作行为。深入分析了传统冠脉介入手术中医生的自然操作，并将复杂操作有效分解成轴向平移操作、周向搓捻操作和复合操作。详细研究了介入器械运动的影响因素，分析了自然操作行为在冠脉介入手术中的作用。

（2）针对冠脉介入手术中介入器械运动的识别问题，提出了一种基于操作行为融合的介入器械运动模式精准识别方法。首先，基于冠脉介入手术操作仿真平台，在导引导丝的六种子运动模式下同步采集了10名受试医生的四类自然操作行为。建立了基于隐马尔可夫模型的行为相似度矩阵，分析了操作行为之间的冗余性。在此基础上，进一步设计了相关操作行为的加权前馈选择算法。基于相关操作行为并利用运动模式解耦设计了介入器械运动的分层识别方法。 数据分析结果表明，分层识别方法的识别准确率高达96.39%，显著优于未解耦的基本识别方法，实现了冠脉介入器械多种运动模式的精准识别。

（3）针对冠脉介入手术中介入器械运动的跟踪问题，提出了一种基于操作行为融合的介入器械运动精确跟踪方法。首先，10名受试医生在仿真平台上分别进行了连续的导引导丝平移和搓捻操作实验，获取了手术过程中医生的四类自然操作行为，以及导引导丝的轴向平移距离和周向旋转角度。基于回归分析与集成融合理论，建立了自然操作行为与器械运动之间的映射关系。统计结果表明，所提出的跟踪方法可以有效利用不同操作行为之间的互补性，提高对介入器械运动的跟踪精度，其中基于高斯混合回归的三模态集成融合获得了1.07±0.17 (mm)，20.05±3.36 (度)的跟踪精度，显著优于基于较少模态或单一初级融合的跟踪方法。

（4）针对冠脉介入手术中介入操作技能的评价问题，提出了一种基于操作行为融合的介入操作技能综合评价方法。首先开展了活体动物实验，由10名不同技能水平的介入医生将导引导丝递送至左冠回旋支的两条靶血管中，并同步采集了手术过程中的四类自然操作行为。然后，利用非参数检验方法对不同技能组别的行为特征进行差异性分析，获得了用于技能评价的有效操作行为。基于机器学习理论设计了定性和定量评价方法，并进行了大量的数据分析实验。实验结果表明，基于高斯混合模型聚类的定性评价方法可以有效区分初级医生和高级医生的操作，评价准确率高达92%；基于马氏距离度量的定量评价得分与临床应用的总体评价量表评分之间存在显著相关性，相关系数高达0.9225，从而验证了所提出方法的临床有效性。

Cardiovascular diseases have become the No.1 killer to human. As the main treatment of cardiovascular diseases, percutaneous coronary intervention (PCI) has been accepted by more and more patients due to the advantages such as less trauma, less pain and shorter recovery time. However, there are still some crucial problems existing in the current practice of PCI: Recognizing and tracking endovascular tool motions still mainly depend on the DSA, which only provides low-quality 2D images and lacks significant 3D anatomical information. In addition, X-ray fluoroscopy exposes medical staff to significant doses of X-ray radiation and large dose of contrast agent can also lead to nephrotoxicity and complications. Moreover, technical skill assessment is mainly based on structured checklists or global rating scales through expert observation, which is highly subjective and involves a large number of evaluation items, resulting in the scoring process being quite laborious, burdensome and time consuming. Supported by the project of the National Natural Science Foundation of China (Grants 61533016, 61611130217), this paper focuses on the intelligent analysis and application of complex natural behaviors, aiming to solve the problems in the current percutaneous coronary intervention. The main contributions and innovations of this research are as follows:

(1) To acquire multimodal behaviors in natural manipulations of PCI, a novel acquisition system is developed by integrating multiple motion-sensing devices (i.e., electromyography, electromagnetic, three-axis accelerometer and fiber-optic bend sensors), which can simultaneously collect interventionalists' natural behaviors (i.e., muscle activity, hand motion, proximal force and finger motion, respectively). Then complex natural manipulations are deeply analyzed and decomposed to axial translation, radial twisting and combined manipulations. Next, the influence of endovascular tool motions is further studied by considering the function of four types of natural behaviors. 

(2) To accurately recognize multi-patterns of endovascular tool motions, a recognition method is proposed by fusing natural behaviors of interventionalists in PCI. To start with, four types of natural behavior are simultaneously acquired from ten interventionalists while manipulating a medical guidewire on a simulation platform with six motion subpatterns. To analyze the redundancy between behaviors, hidden Markov model-based similarity matrices are established, and a weighted feedforward selection algorithm is designed to obtain relevant behaviors for each motion subpattern. Based on relevant behaviors and motion decoupling, a hierarchical recognition method (HRM) is developed. Extensive analysis demonstrates that the HRM has implemented accurate multi-pattern recognition of endovascular tool motions with an accuracy of 96.39%, significantly outperforming that of the basic recognition method without motion decoupling.

(3) To precisely track endovascular tool motions, a tracking approach is proposed by fusing natural behaviors of interventionalists in PCI. Firstly, simulation experiments are performed to collect natural behaviors of 10 interventionalists while continuously translating and twisting a guidewire, as well as the translation distance and rotation angle of the guidewire. Based on regression analysis and ensemble fusion, the mapping relationship between relevant behaviors and endovascular tool motions is established. Statistical results show that the proposed method can effectively make use of the complementary between different behaviors to improve tracking performance. The three-modal fusion scheme based on Gaussian mixture regression can achieve the root mean square error of 1.07±0.17 (mm)，20.05±3.36 (degree), significantly better than the cases using less modalities and single preliminary fusion.

(4) To effectively evaluate technical skills, a comprehensive assessment technique is proposed by fusing natural behaviors of interventionalists in PCI. To begin with, in vivo experiments are conducted to acquire natural behaviors of 10 different skill-level interventionalists while delivering a guidewire to two target vessels of left circumflex arteries. Next, a nonparametric test is used to evaluate the difference between expert and novice attempts to select effective behaviors for technical skill assessment. By using machine learning-based techniques, qualitative and quantitative assessment methods are designed, and extensive experiments of data analysis are conducted. Experimental results indicate the Gaussian-mixture-model-based qualitative assessment can effectively distinguish novice and expert manipulations with a clustering accuracy of 92%; the Mahalanobis-distance-based quantitative assessment scores strongly correlate to those obtained from the modified global rating scales used in clinical practice with a high correlation coefficient of 0.9225, verifying the effectiveness of the proposed assessment technique.