基于TI-RADS的甲状腺结节超声图像特征提取与可视化技术研究

发布时间：2018-03-14 14:23

本文选题：结节　切入点：TI-RADS　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：在医学领域中,结节是指未经诊断不确定良恶性的肿块,甲状腺结节包含良性结节和恶性结节。其中甲状腺癌就是恶性结节,良性结节多是炎症性结节或者囊肿。由于甲状腺结节无明显的病症表现,因此病症极易被人忽视。目前,甲状腺疾病的常用诊断方法是超声检测,但是人工进行超声诊断仍然存在一些主观缺陷。随着医学影像设备的广泛应用和数字图像处理技术的飞速发展,利用图像处理进行计算机辅助诊断的研究越来越多。计算机辅助诊断的主要目的是通过计算机的识别处理把超声图像准确地分类,为医生和病人提供可参考的诊断结果。本文主要研究的是基于TI-RADS表的甲状腺结节超声图像的特征提取和可视化。目的是研究TI-RADS等级不同的甲状腺结节超声图像,把结节的分级结果和不同等级特征的差异利用图像表现出来。主要研究内容包括三部分:甲状腺结节超声图像的预处理,甲状腺结节超声图像的特征提取以及甲状腺结节超声图像的可视化设计。超声图像预处理包含图像去噪和图像分割。针对甲状腺图像中的斑点噪声,应用了基于边缘增强的各向异性扩散模型(EEAD),在保留超声图像质量的情况下去除了超声图像中的斑点噪声。针对超声图像的结节分割,提出了基于边缘梯度算子和形状约束的图割算法(Graph Cut),主要通过最小化能量函数得到结节区域。分割算法优化了超声图像分割结果形状不准确以及边缘毛躁的现象。甲状腺结节的特征提取提出了基于TI-RADS表的超声图像特征量化方法。把结节的特征分成形态、边界、回声、纵横比和钙化5类。通过形态学特征提取、灰度特征提取等多种方法,获得5类共计34个数据特征,并用相关性、T检验和聚类等方法对特征数据进行了效果验证。在可视化研究阶段,主要工作包括聚类分析和可视化设计。聚类算法根据特征的样本规律将其划分成不同等级,数据可视化把分级结果利用可视化布局展示。实验中针对单一类别的特征聚类和多类特征聚类分别应用了基于遗传学的蚁群算法聚类(GACO)和多视图加权聚类(TW-Kmeans)。可视化实验针对不同的数据结构设计可视化布局,实现针对基本信息的可视化和结节特征的可视化设计。应用圆形分区图和矩形树状图表示基本信息之间的关系。应用雷达图、平行坐标图和星形散点图表现结节的分级结果和不同级别的特征差异。
[Abstract]:In the field of medicine, nodule is an undiagnosed benign and malignant mass. Thyroid nodule contains benign and malignant nodule. Thyroid carcinoma is a malignant nodule. Benign nodules are mostly inflammatory nodules or cysts. Because thyroid nodule has no obvious symptom, it is easy to be ignored. At present, ultrasound is commonly used to diagnose thyroid diseases. However, there are still some subjective defects in artificial ultrasound diagnosis. With the wide application of medical imaging equipment and the rapid development of digital image processing technology, The main purpose of computer-aided diagnosis is to classify ultrasonic images accurately by computer recognition. This paper mainly studies the feature extraction and visualization of thyroid nodules based on TI-RADS table. The purpose of this paper is to study the ultrasound images of thyroid nodules with different TI-RADS grades. The difference between the classification results of the nodules and the characteristics of different grades is represented by the image. The main contents of the study include three parts: the preprocessing of the ultrasonic images of the thyroid nodules, The feature extraction of thyroid nodule ultrasound image and the visualization design of thyroid nodule ultrasonic image. Ultrasonic image preprocessing includes image denoising and image segmentation. An anisotropic diffusion model based on edge enhancement was applied to remove speckle noise in ultrasonic images without preserving the quality of ultrasound images. A graph cutting algorithm based on edge gradient operator and shape constraint is proposed, which is mainly used to minimize the energy function to obtain the nodule region. The segmentation algorithm optimizes the phenomena of inaccurate shape and hairy edge of ultrasonic image segmentation results. Feature extraction of thyroid nodule A method of ultrasonic image feature quantization based on TI-RADS table is proposed. The feature of thyroid nodule is divided into shape. The boundary, echo, aspect ratio and calcification are classified into five categories. By means of morphological feature extraction and gray feature extraction, a total of 34 data features of 5 categories are obtained. In the visualization research stage, the main work includes clustering analysis and visual design. The clustering algorithm divides the feature data into different grades according to the law of the samples. Data visualization shows the hierarchical results by visual layout. In the experiment, the genetic ant colony algorithm (ACO) and the multi-view weighted clustering (TW-Kmeansan) are applied to single class and multi-class feature clustering, respectively. Design visual layout for different data structures, The basic information is visualized and the nodule feature is visualized. The relationship between the basic information is expressed by using the circular partition map and the rectangular tree chart, and the radar image is used to show the relationship between the basic information and the basic information. Parallel coordinate map and star scatter plot show nodule classification results and characteristics of different grades.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R581;TP391.41

【参考文献】