【作者】 于洋；

【导师】 赵宏；

【作者基本信息】 东北大学 ， 计算机应用技术， 2009， 博士

【摘要】 肺癌是导致当今世界男女死亡的首要因素。临床研究表明,肺癌的准确分期对疾病的预防和预后起至关重要作用。对分期为T1的恶性肺结节的早期检测和手术,能显著提高肿瘤的预后。计算机辅助诊断(Computer Aided Diagnosis, CAD)方法能帮助减少放射科医生阅片时间,并提高诊断的准确率。通常,肺癌CAD系统包括预处理、疑似结节检测、假阳(False Positive, FP)减少和分类过程。预处理阶段的任务是将肺结节检测的区域局限在肺内的ROI区域,同时,减少噪声和图像的伪迹。为了提高肺结节检测的敏感性,本文研究了作为肺结节检测预处理阶段的肺结节增强滤波算法。首先,本文提出了基于形态元分析(Morphological Component Analysis, MCA)的多尺度增强滤波算法。该方法源自稀疏表示和逼近理论,通过构造描述不同特征的字典稀疏表示各向同性和各向异性的特征。本文提出的基于MCA的增强滤波利用wavelet变换表示肺结节结构,curvelet变换描述主要可能构成肺结节检测的假阳的血管结构,并增强wavelet部分以达到增强肺结节的目的。实验结果表明该算法可以成功地根据形状特征将图像中的不同目标分离。其次,本文研究基于Hessian矩阵特征值分析方法,提出了基于平移不变冗余小波变换的增强滤波算法。同时,通过利用MPR (Multi-Planar Reconstruction)技术在VOI的不同视角上分析改进Hessian矩阵特征值,相对于仅在二维横断CT层面上的分析,该方法很大程度上减少了如血管横截或末端造成的肺结节检测的假阳数量。同时,本文也研究了如何利用形态学和纹理等多重医学影像特征减少肺结节检测的搜寻范围,并提出了相应的有效增强滤波算法。除了医学影像的增强技术研究外,本文也对医学影像分割和理解作了深入的研究。图像分割,通常作为计算机图像理解的第一个必要步骤。本文提出了一个基于Sobolev梯度改进的光流计算分割算法,用于提取肺CT影像的VOI (Volume of Interest)。该算法通过利用Sobolev空间下的梯度下降算法取代欧氏空间的梯度,使改进的光流计算对噪声鲁棒,且可以收敛到全局最优。本文也提出了基于肺CT影像的血管分割算法。血管分割以及血管的形态及拓扑的分析对于相关疾病的诊断、治疗、预测和手术计划都是至关重要的。本文在总结前人血管分割研究工作的基础上,提出了一个新的水平集分割算法,利用博弈论融合了基于区域的水平集(如CV)和基于边缘的水平集(如GAC)的优势,并通过实验证明该算法确实优于仅利用单一模型的水平集方法。因为通过分析随访前后肺结节的生长情况,通过计算倍增时间即能有效判断肺结节的良、恶性。因此,本文提出了基于冗余小波变换和分水岭的肺结节识别算法。通过冗余小波变换,可以增强肺结节的边缘信息,并同时确定疑似肺结节的ROI。利用冗余小波变换多尺度分析结果作为标记,利用快速准确的欧氏距离函数的分水岭变换将肺结节准确分割。对肺CT影像的实验证明,该算法可以获得比较满意的肺结节识别结果。更多还原

【Abstract】 Lung cancer is the most common cause of cancer death for both men and women around the whole world. Clinical studies show that the stage of the disease is the most important factor in preventing and prognosing lung cancer. Early detection and surgical resection of malignant lung nodules (T1) can improve the prognosis significantly. Computer Aided Diagnosis (CAD) methods can assist radiologists in reducing the reading time as well as in improving the diagnostic accuracy. Usually, CAD schemes for lung cancer consist of preprocessing, candidate detection, false positive (FP) reduction and classification. The preprocessing stage restricts the search space to the lung regions and reduces noise and image artifacts. To improve the sensitivity of candidate detection, a nodule enhancement filter can be applied as a preprocessing step prior to the initial nodule detection.First, we developed one enhancement filter to enhance the blob like nodules and suppress the curvilinear structures like vessels which are one of the primary sources for false positives. In the research, the dissertation proposed an enhancement filter to enhance the blob like nodules based on Morphological Component Analysis (MCA) theory, which can represent isotropic and anisotropic features as sparse combinations of atoms of predetermined dictionaries. According to its results, the enhancement can be achieved by separating and analyziong objects with different shapes.Second, we developed another enhancement filter through employing the shift-invariant undecimated wavelet transform and analyzing the eigenvalues of the Hessian matrix. In contrast with the results obtained when using the blob enhancement filter on 2D slices, most crossings and end points of blood vessels as FPs have been substantially eliminated, by means of analyzing the connectivity between their context slices with MPR techniques.Meanwhile, we investigated how to exploit multiple medical image features to limit the search space to the regions of interests.Besides medical image enhancement, the dissertation also worked on medical image segmentation and analysis. Segmentation is often a necessary first step to computer analysis. This dissertation proposed a segmentation framework to extract the volumes of interest (VOIs) using optical flow constraint refined with Sobolev gradient. In this work, to avoid using normal gradient that could be sensitive to noise and make the flow converge to local minima, optical flow method was exploited in Sobolev space instead of Euclidean space.The dissertation proposed a segmentation framework on analysis of blood vessels in thoracic CT scans. Determination and assessment of vessel morphology and topology is crucial for diagnosing, treatment, outcome prediction, and surgical planning. The dissertation presented a new level set functional combined with both region-based models (e.g. CV) and edge-based models (e.g. GAC) through mutual information sharing based on game theory.Finally, the dissertation presented a method to recognize the lung nodules based on undecimated wavelet and watershed transforms, since doubling time of pulmonary nodules is important for estimating malignancy versus benignity. The thoracic CT images are enhanced using translation invariant redundant wavelet transform. The lung nodules can be segmented by exploiting watershed in enhanced volume. Experiments on the thoracic CT images showed that the proposed method obtained satisfactory results of nodule recognition.更多还原