基于深度学习的肝脏CT影像分割方法的研究与应用

发布时间：2018-02-27 00:00

本文关键词： 深度学习全卷积神经网络肝脏分割　出处：《吉林大学》2017年硕士论文　论文类型：学位论文

【摘要】：肝脏肿瘤的恶性度高,病情发展快,死亡率位居世界第二,而我国是肝癌的高发地区之一。随着医学成像技术的发展,计算机断层扫描技术(CT)在肝脏疾病的相关诊断中被广泛应用,并且已成为诊断肝脏疾病的首选方法。利用计算机图像处理技术,结合医学影像诊断技术,对肝脏疾病进行早期诊断、三维建模以及定量分析,能够使医生在术前掌握充足的数据,进行术前规划,提高手术的成功率并拟定合理有效的治疗方案。从腹部CT影像中准确可靠地分割肝脏轮廓,是肝脏疾病的早期诊断、肝脏大小及病情的估测和三维建模的第一步,也是非常关键的一步,其分割结果对后续工作有着直接的影响。在实际的临床应用中,一般通过拥有相关实践经验和专业知识的医师从CT影像中手工分割肝脏轮廓,但是,这个过程是非常耗费时间和精力的,而且受不同医师的主观因素、经验以及知识的差异的影响,往往会得到不同的分割结果。所以,为减轻医生的工作负担,提高工作效率,也为了获得更加客观、准确的分割结果,必然要引入计算机辅助诊断技术,帮助专业医师分割肝脏CT图像。传统的肝脏分割方法,是以图像处理方法为基础,主要依赖于图像的一些浅层特征,如:灰度、统计结构以及纹理等来分割肝脏轮廓。这种特征可以从图像中直接获得,或者通过人工设计的提取算子获得。这些浅层特征的鲁棒性较低,代表性不强,也易受到噪声的干扰。实践证明,往往是那些抽象的、深层的特征更具代表性。深度学习技术能够从大量数据中挖掘数据深层的抽象的特征,将其应用到肝脏分割任务中能够提高分割的精度和鲁棒性。本文提出了一种基于全卷积神经网络的CT影像肝脏分割的方法。我们以经典的卷积神经网络Alex Net为基本框架,对其网络布局做了一定的改动,使其成为全卷积网络结构。实验中使用大量的已标注好的腹部CT影像对网络进行训练,借由构造好的代价函数更新网络参数。为解决使用传统激活函数造成的梯度消失等问题,采用Re LU函数作为激活函数,同时为缓解过拟合问题,在网络中使用Dropout技术,增加网络的泛化能力。由于深度网络高层的输出缺乏足够的细节信息,导致获得的分割结果比较粗糙。为解决这个问题,本文提出了一种融合低层和高层特征的网络结构,通过将低层的局部细节信息与高层的抽象的语义信息融合,进而获得更加精确地分割结果。实验结果表明,本文提出的算法具有较好的鲁棒性和精度,同时与基于Patch的方法相比,本文方法具有更高的效率。
[Abstract]:Liver tumors have a high degree of malignancy, a rapid development of the disease, and a mortality rate of the second highest in the world, and China is one of the regions with a high incidence of liver cancer. With the development of medical imaging technology, Computed tomography (CT) has been widely used in the diagnosis of liver diseases, and has become the first choice in the diagnosis of liver diseases. Early diagnosis, 3D modeling and quantitative analysis of liver diseases enable doctors to master sufficient data and plan before operation. The accurate and reliable segmentation of liver contour from abdominal CT images is the first step in the early diagnosis of liver disease, the estimation of liver size and disease condition, and the three-dimensional modeling. It is also a very critical step. The segmentation results have a direct impact on the follow-up work. In practical clinical applications, the liver contour is usually segmented manually from CT images by doctors with relevant practical experience and professional knowledge, but, This process is very time-consuming and energy-intensive, and often results in different segmentation results due to different doctors' subjective factors, experience and knowledge. In order to obtain more objective and accurate segmentation results, it is necessary to introduce computer-aided diagnosis technology to help doctors to segment liver CT images. Traditional liver segmentation methods are based on image processing methods. It mainly depends on some shallow features of the image, such as grayscale, statistical structure and texture, to segment the liver contour. This feature can be obtained directly from the image. Or it can be obtained by artificial extraction operators. These shallow features are less robust, less representative, and easily disturbed by noise. Practice has proved that these shallow features are often abstract. The deep features are more representative. The deep learning technology can mine the abstract features of the deep layer of the data from a large amount of data. Applying it to liver segmentation can improve the accuracy and robustness of liver segmentation. In this paper, a method of liver segmentation in CT image based on full convolution neural network is proposed. We take the classical convolutional neural network Alex Net as the basic framework. The network layout is modified to make it a full convolution network structure. In the experiment, a large number of tagged abdominal CT images are used to train the network. In order to solve the problem of gradient disappearance caused by the traditional activation function, re LU function is used as the activation function. In order to alleviate the problem of over-fitting, Dropout technology is used in the network. In order to solve this problem, this paper proposes a network structure that combines the features of the lower layer and the higher level, because of the lack of sufficient detail information in the output of the high level network, which results in rough segmentation results. By merging the local details of the lower level with the high-level abstract semantic information, the segmentation results are obtained more accurately. The experimental results show that the proposed algorithm is robust and accurate. At the same time, compared with the method based on Patch, this method has higher efficiency.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R735.7;TP391.41

【参考文献】