肿瘤手术导航中图像分割与配准方法研究

发布时间：2018-03-03 01:32

本文选题：手术导航　切入点：肿瘤图像　出处：《北京工业大学》2016年博士论文　论文类型：学位论文

【摘要】：计算机辅助手术(Computer Aid/Assisted Surgery,CAS)是依靠图像引导的介入性手术,是医学研究领域目前的热点之一。它通过术前手术规划,术中配准对手术进行引导,术后评估等一系列过程,对病灶进行定位、诊断,引导医生对病灶进行相关的专业处理,解决常规手术中难定位的问题,同时减少手术并发症。在肿瘤手术中,由于肿瘤附着器官的复杂结构如脑肿瘤周围血管神经密布,以及肿瘤组织本身的浸润性,使得临床对高精度的计算机辅助肿瘤手术有着迫切的需求。本文针对肿瘤手术导航中的关键问题——手术引导的精确度进行研究,针对术前肿瘤病灶定位诊断及术中配准这两大影响手术精确度的主要因素均提出了新的解决方案。目的是为临床医生完成高精度的肿瘤外科手术提供新的契机,确保手术安全,减少病人痛苦,并在最短时间内到达靶点病灶完成手术。本论文主要研究内容与成果如下:肿瘤自动分割与诊断算法研究肿瘤的精确自动分割及早期诊断,可以提供靶区病灶位置,规避重要器官及血管神经,辅助术前手术规划,解决传统手工勾勒费时及精确度低的问题。然而长期以来,由于问题的复杂性,肿瘤分割一直存在精度低及误诊断的问题,而手工勾勒费时,对勾勒者专业要求高且缺乏自动性。针对上述问题,本文提出了新的方法:1.从综合多种模式肿瘤图像的角度出发,提出一种基于传统卷积神经网络(Convolutional Neural Networks,CNNs)的脑肿瘤自动分割与诊断算法。该算法针对传统脑肿瘤分割算法使用单一图像模式及分割精度低的问题,设计了新的CNNs架构模型。该模型可以自动学习多模态图像中的有用特征,综合利用多模态图像的信息。实验结果表明,本文所提出方法的分割与诊断精确度优于传统算法,可为医生的临床诊断及术前手术规划提供可靠信息。2.在传统CNNs基础上,提出多通道CNNs算法。首先,该算法摒弃了传统机器学习类算法手动设计特征的特征提取方式,采用自学习的方法提取多模态图像的明显特征。其次,由于肿瘤组织边界的浸润性及低对比性,该算法克服了传统CNNs方法仅利用图像边界即局部信息的弊端,综合利用肿瘤图像的全局及局部信息。实验结果表明,该算法分割与诊断的精确度优于目前最流行的肿瘤分割算法,也优于传统的CNNs算法。肿瘤图像配准算法研究术中实时采集的医学图像与术前手术规划图像之间的配准,是影响手术导航的关键指标。而肿瘤的易变形、持续生长等特征,对术中配准提出了新的挑战。本文提出了一种用于存在大变形情况的配准框架,解决了传统方法在此情况下失效的弊端。同时改进传统的相似性测度,在保证配准精度的前提下,提高算法的收敛速度。1.提出一种深度迭代配准框架及基于CNNs分类器的初始预配准算法。针对传统配准框架,仅调用一次初始预配准,固定次数非刚性配准的弊端,提出深度迭代配准框架。即在后续每次非刚性配准迭代中,再次调用初始预配准,以充分利用两次配准达到提高配准精度的目的。针对传统初始预配准方法无法处理存在大变形的情况,本论文提出分别求解初始预配准仿射变换中的旋转、平移、尺度参数。其中旋转参数,首先离线训练CNNs分类器,使其能识别多达360类旋转角度;尺度参数,使用图像大小信息将固定及移动图像尺寸达到一致;平移参数,首先通过统计学方法计算每张图像的形心,通过形心位置信息使两张图像达到一致。实验结果证明,本文方法能够替代传统框架处理存在大的变形下的配准问题。2.提出一种高效的相似性测度算法。主成分分析(Principle Component Analysis,PCA)用来提取配准中固定以及移动图像最主要的特征点,避免由额外噪声带来的误差。将PCA与传统相似性测度诸如Spearman及Pearson等相关系数相结合组成新的相似性测度。实验结果证明该算法能够在保证配准精度的同时进一步提高算法的收敛速度。
[Abstract]:Computer assisted surgery (Computer Aid/Assisted, Surgery, CAS) by interventional image-guided surgery, is one of the current hot spots in the field of medical research. It is through the preoperative planning, intraoperative registration of surgical guidance, evaluation and a series of process after operation, location of lesions, diagnosis, treatment of the lesions to guide doctors the relevant professional, solve the routine operation difficult to locate the problem at the same time, reduce the postoperative complications. In tumor operation, due to the complex structure of the tumor attachment organs such as brain tumor blood vessels around the nerve and tumor tissue with infiltration of itself, making the clinical on the high precision of computer aided surgery have urgent needs in this paper. Aiming at the key problems in navigation surgery - surgical guidance accuracy study for the preoperative diagnosis and intraoperative tumor lesion location registration of the two operation The main factors are the accuracy of the new solution is put forward. The purpose is to provide a new opportunity for clinicians to complete tumor surgery with high precision, to ensure the safety of the operation, reduce the pain of patients, and in the shortest time to reach the target point of focus to complete the operation. The main research contents and results are as follows: automatic segmentation of tumor and diagnosis the precise automatic segmentation algorithm of tumor and early diagnosis, can provide the target location, avoid the important organs and blood vessels and nerves, assist in surgery planning, to solve the traditional problems of low accuracy and time-consuming hand sketched. However, due to the complexity of the problem, there has been a tumor segmentation accuracy is low and the problem of false diagnosis, and a hand sketched outline of their time-consuming, high professional requirements and lack of initiative. Aiming at the above problems, this paper puts forward a new method: 1. from the mixed pattern of tumor image angle Of a traditional convolution based on neural network (Convolutional Neural Networks, CNNs) automatic segmentation of brain tumors and the diagnosis algorithm. The algorithm for brain tumor segmentation of traditional single image segmentation model and the low accuracy of the algorithm, design a new CNNs frame model. The model can automatically learn the useful features of multimodal in the image, the comprehensive utilization of multi modality image information. The experimental results show that the proposed segmentation algorithm is superior to the traditional method and the diagnostic accuracy of the surgical planning, to provide reliable information for clinical diagnosis and surgery for.2. before the doctor on the basis of traditional CNNs, the multichannel CNNs algorithm. Firstly, the algorithm discards the traditional feature machine learning algorithm design manual feature extraction method, the extraction characteristic of multi modality images by using the method of self-learning. Secondly, due to the infiltration of tumor tissue boundary circle And low contrast, the algorithm overcomes the disadvantages of traditional CNNs method only uses the image boundary disadvantages of local information, comprehensive utilization of global and local tumor image information. The experimental results show that the segmentation and the diagnostic accuracy of the algorithm is superior to the most popular tumor segmentation algorithm is better than the traditional CNNs algorithm. The registration between surgical planning the image of the medical image and the real-time image registration algorithm of tumor surgery in the study before, is the key index of surgical navigation. While tumor deformation, sustained growth characteristics, put forward a new challenge on intraoperative registration. This paper proposes a framework for the registration of large deformation situation, solve the traditional method of failure problems. At the same time improved the traditional similarity measure, under the premise of ensuring the accuracy of registration, improve the speed of convergence of the algorithm.1. proposed an iterative depth registration box Frame and initial registration algorithm based on CNNs classifier. The traditional registration framework, called only once the initial pre registration, the number of drawbacks of fixed non rigid registration, advanced iterative registration framework. In each subsequent iteration in the non rigid registration, call the initial registration, in order to make full use of the two to improve the accuracy of registration registration the purpose of the initial registration. The traditional methods cannot deal with the existence of large deformation, this paper respectively solve the initial pre registration of affine transformation, rotation, translation and scale parameters. The rotation parameters from the first line to train the CNNs classifier, which can identify up to 360 kinds of rotation angle; scale parameter, fixed and mobile the size of the image to use image size information; the translation parameters, first through statistical methods to calculate the image centroid, through the centroid of the two image information to achieve Consistent. Experimental results show that this method can replace the traditional.2. registration framework for dealing with large deformation under an efficient similarity measure algorithm. Principal component analysis (Principle Component, Analysis, PCA) is used to extract the feature points and the main fixed mobile image registration, to avoid the error caused by the extra noise. The combination of new similarity measure PCA with traditional similarity measures such as Spearman and Pearson correlation coefficient. The experimental results show that the algorithm can guarantee the accuracy of registration and further improve the convergence speed of the algorithm.

【学位授予单位】：北京工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R730.56;TP391.41

【相似文献】