CT图像中肝脏分割的数学建模方法

发布时间：2018-10-22 16:37

【摘要】：从腹部CT图像中分割出精准的肝脏轮廓是医学图像分割中的一个热点问题,同时也是一个极具挑战的问题。本文针对此问题,提出了一种半自动和一种全自动的数学分割模型。一方面,肝脏CT数据具有复杂的背景、低对比度、弱边界以及高噪声的特点;另一方面,肝脏肿瘤或者其它肝脏病变的出现往往会将肝脏分成两个表征完全不同的子区域。特别的,如果肝脏肿瘤位于肝脏表面,那么准确的提取肝脏边界将更为困难。此问题虽然可以通过多次分割得到解决,但却需要对体数据进行多次处理,这必然增加时间消耗。为此,在基于活动轮廓模型的框架下,我们提出了新的混合变分模型。此模型仅仅用一个水平集函数就可以一次分割出含两个子部分的整个肝脏组织。此模型创新性地引入了测地距离选择的技巧来自适应地选择最相似的特征去驱动曲线演化到准确的肝脏边界。另外,在定义局部区域表征项时,提出的加权直方图技巧对精准边界的提取也有一定的贡献。基于测地距离选择的肝脏分割方法需要人工给定初始化,这将降低方法的临床应用性。为此,我们借助于深度学习和图割给出了一种全自动的肝脏分割方法。此方法包含两个内容:(ⅰ)利用三维卷积神经网络学习出肝脏概率图,同时确定初始肝脏区域;(ⅱ)利用肝脏先验概率信息和图割的技巧精分割初始的肝脏轮廓。此模型完全不需要任何人工操作,因此具有潜在的临床应用价值。另外,此模型中三维卷积神经网络的应用在肝脏分割领域也属于较早的工作。本文所提出的所有分割模型均在公共数据集上进行了验证。大量定性和定量的实验结果显示了所提模型可以准确地从CT图像中提取肝脏轮廓。
[Abstract]:Segmenting accurate liver contour from abdominal CT image is a hot issue in medical image segmentation, and it is also a very challenging problem. In order to solve this problem, a semi-automatic and a fully automatic mathematical segmentation model is proposed in this paper. On the one hand, liver CT data have the characteristics of complex background, low contrast, weak boundary and high noise. On the other hand, liver tumors or other liver lesions often divide the liver into two subregions characterized by completely different characteristics. In particular, if liver tumors are located on the surface of the liver, accurate extraction of liver boundaries will be more difficult. Although this problem can be solved by multiple segmentation, it is necessary to process the volume data many times, which will inevitably increase the time consumption. Therefore, under the framework of active contour model, we propose a new hybrid variational model. In this model, the whole liver tissue with two subparts can be segmented by using only one level set function. This model innovatively introduces the geodesic distance selection technique to adaptively select the most similar features to drive the curve to the exact liver boundary. In addition, the proposed weighted histogram technique also contributes to accurate boundary extraction when defining local representation terms. The method of liver segmentation based on geodesic distance selection needs artificial initialization, which will reduce the clinical application of the method. For this reason, we give a fully automatic liver segmentation method by means of deep learning and graph cutting. This method consists of two parts: (I) learning the liver probability map by using three-dimensional convolution neural network and determining the initial liver region at the same time; (II) segmenting the initial liver contour precisely by using the prior probability information of the liver and the technique of graph cutting. This model does not need any manual operation, so it has potential clinical application value. In addition, the application of three-dimensional convolution neural network in this model is also an early work in the field of liver segmentation. All the segmentation models proposed in this paper are verified on the common data set. A large number of qualitative and quantitative experimental results show that the proposed model can accurately extract liver contours from CT images.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R816.5;TP391.41

【相似文献】