面向微创脊柱手术经皮穿刺的图像引导技术研究
文献类型：学位论文
| 作者 | 李杨
|
| 学位类别 | 博士 |
| 答辩日期 | 2017-11-30 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 沈阳 |
| 导师 | 谈金东 ; 梁炜 |
| 关键词 | 图像引导手术，图像分割，图像识别，图像配准，增强现实 |
| 其他题名 | Research on Image-Guided Technology for Percutaneous Puncture of Minimally Invasive Spine Surgery |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 本文面向微创脊柱手术的实际应用需求，提出基于多模态数据融合的微创脊柱手术引导系统框架，对图像引导手术系统所涉及的腰椎CT图像分割、腰椎CT图像识别、腰椎图像2D-3D配准以及图像引导手术中的增强现实方法等关键技术进行了系统性研究，取得如下创新性成果： （1）为了解决腰椎CT图像分割中腰椎解剖结构个体差异问题，构建准确的术前腰椎模型，提出了一种新的基于全局信息水平集的腰椎图像分割方法。首先，为避免人工初始化对分割结果的影响，提出了一种自动初始化水平集函数，在目标边界附近生成光滑、简洁的水平集初始轮廓；其次，针对腰椎CT图像分割中广泛存在的弱边界泄露问题，提出了一种基于区域相关性的正则化水平集演化方程，通过计算水平集轮廓内外区域直方图马氏距离获得图像全局信息，增加了水平集轮廓对弱边界的捕获能力；最后，为了避免水平集函数过度分割，提出 指数形式的边缘停止函数，保障了水平集演化函数的快速收敛。实验结果显示，所提出的腰椎CT图像分割方法在不同噪声等级下的分割准确度和收敛速度均优于其他比较方法。 （2）针对任意角度和病理情况下CT图像腰椎特征复杂多变的问题，本文提出了一种基于特征融合深度学习模型的腰椎识别方法。为了解决深度学习模型训练数据不足的问题，提出了一种基于等值面的GPU加速数字影像重建方法，生成感兴趣区域的多角度二维图像，在降低DRR计算复杂度的前提下实现了对训练数据的增广。为了在病理情况下提取CT图像中普适的腰椎特征，提出包含两种不同卷积核的卷积神经网络模型，对腰椎图像中的形状信息和纹理信息进行融合，有效提取了CT图像中腰椎的本质特征。实验结果表明，本文方法在病理情况下的CT图像腰椎识别结果具有较高的准确率。 （3）为了根据术中二维图像将术前三维腰椎模型重构成术中患者的姿态，本文提出了一种基于曲率特征循环神经网络的2D-3D腰椎图像配准方法。首先，针对术中X线二维图像中感兴趣区域与背景特征混叠的问题，区别于传统的基于图像灰度的特征提取方法，提出了一种基于联合曲率的特征提取方法，获得二维图像和三维模型的共有特征；其次，为了解决术中X线二维图像在腰椎结构信息不完整情况下的腰椎识别问题，提出了一种基于层级循环神经网络的腰椎识别方法，通过融合腰椎上下文信息提高腰椎识别的准确性；最后，为确定三维腰椎模型相对于二维腰椎图像在3个自由度上的旋转角度，提出了一种基于双平面模型的腰椎姿态估计方法，通过融合正、侧位两个方向的特征点提高腰椎姿态估计准确度。实验结果显示，在正常和病理情况下，本文方法在腰椎识别精度和姿态估计准确性方面均优于其他比较方法。 （4）针对图像引导经皮穿刺实施过程中腰椎内部解剖结构不可视问题，本文提出了一种基于增强现实的腰椎经皮穿刺导航方法。为了解决传统增强现实方法在虚-实结合过程中依赖人工标记点的问题，本文提出了腰椎特征点自动识别算法，生成天然标记点，避免了由人工标记点的遮挡和偏移所引起的导航误差；为了整合腰椎模型和手术场景等不同空间坐标系，采用工业RGB相机增强移动C型臂，通过空间坐标系标定技术计算术前CT腰椎三维模型、术中X线二维图像和术中三维手术场景之间的坐标变换，实现虚-实结合的增强现实显示。实验结果表明，由于使用腰椎特征点作为天然标记点并且融合了虚拟模型和真实手术场景，本文方法在穿刺点定位精度方面优于使用人工标记点的导航方法。 |
| 英文摘要 | Oriented to the practical application of minimally invasive spine surgery needs, we improve the traditional IGS system and propose a framework of minimally invasive spinal surgery guidance system based on multi-modality data fusion. We conducted a systematic study on the key techniques of IGS including lumbar CT image segmentation, recognition, 2D-3D registration and augmented reality navigation. Therefore, we mainly achieved the following innovative results. (1) Concerning the individual difference issue of lumbar vertebra anatomy in CT image segmentation and in order to construct accurate preoperative lumbar model, we introduce a novel CT image segmentation method based on global information level set approach. First, to avoid the impact of manual initialization on segmentation results, an automatic initialization level set function is proposed to generate a smooth, concise level set initial contour around the target boundary. Second, aiming at the problem of weak boundary leakage existing widely in lumbar CT image segmentation, a regularization level set evolution equation based on regional correlation is proposed. The global information of the image is obtained by calculating the Mahalanobis distance between the inner and the outer regions of the level set contour, which increases the ability of level set function to capture weak boundaries. Finally, in order to avoid the over-segmentation of the level set function, the edge-stop function in the form of e-index is used to guarantee the rapid convergence of the level set evolution function. The experiment results show that the lumbar CT image segmentation method proposed in this paper has better segmentation accuracy and convergence speed under different noise levels than other comparison methods. (2) Aiming at the complicated and changeable lumbar features of CT images under any angle and pathological conditions, a lumbar vertebrae recognition method based on feature fusion deep learning is proposed. In order to solve the problem of insufficient training data of deep learning model, a GPU-accelerated digital image reconstruction method based on isosurface is proposed to generate a multi-angle 2D images of the area of interest. Under the premise of reducing the computational complexity of DRR, the augmentation of training data is realized. In order to extract the common lumbar features in CT images under pathological conditions, a convolutional neural network model containing two different convolution kernels is proposed to combine texture information and shape information of lumbar vertebrae CT images and extract of the essential features of lumbar vertebrae effectively. The experiment results show that the proposed method performs more accurate lumbar vertebrae recognition than other comparison methods in pathological conditions. (3) In order to reconstruct the preoperative lumbar model according to the intraoperative patients' pose, this paper proposes a 2D-3D registration method based on the recurrent neural network with curvature features. First, due to the feature overlap problem with the region of interest and the background in the X-ray projection images, the traditional feature extraction method based on image intensity is invalid. In this paper, a feature extraction method based on joint curvature is proposed, which improves the similarity of the features between 2D images and 3D model. Second, in order to solve the issue of partial occlusion in X-ray projection images, a hierarchical recurrent neural network with context information is exploited to improve the accuracy of lumbar recognition. Finally, in order to estimate the rotation angle of the 3D model relative to the 2D image in three degrees of freedom, a bi-planar model is applied to fuse the positive and lateral features of two directions to improve the accuracy of lumbar vertebrae pose estimation. The experiment results show that in both normal and pathological conditions, the proposed method is superior to other methods in lumbar recognition accuracy and pose estimation accuracy. (4) Aiming at invisible problems of internal anatomy of lumbar vertebrae during the process of image-guided percutaneous puncture, this paper presents an augmented reality based method for lumbar percutaneous puncture navigation. In order to solve the problem that the traditional augmented reality method is easy to be affected by artificial markers occlusion and drift during the process of virtual-real combination, this paper adopts the automatic recognition algorithm to extract the lumbar vertebral feature points as the natural markers to avoid these disadvantages. In order to integrate the different spatial coordinates system such as lumbar model and surgical scene, we enhanced the mobile C-arm with industrial RGB camera to achieve virtual-real combination, which is able to integrate preoperative 3D model, intraoperative 2D X-ray images and intraoperative surgical scenes through the calibration of space coordinate systems and achieve augmented reality display by virtual-actual combination. The experiment results show that due to the use of the lumbar vertebrae feature points as natural markers and the fusion of the virtual model and the actual surgical scene，the proposed method is superior to the other navigation method using artificial marker points in positioning accuracy of the puncture point. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 106页 |
| 源URL | [http://ir.sia.cn/handle/173321/21282]  |
| 专题 | 沈阳自动化研究所_工业控制网络与系统研究室 |
| 推荐引用方式 GB/T 7714 | 李杨. 面向微创脊柱手术经皮穿刺的图像引导技术研究[D]. 沈阳. 中国科学院沈阳自动化研究所. 2017. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。