甲状腺结节超声图像纹理特征提取及半监督分级方法研究

发布时间：2018-10-10 06:45

【摘要】：甲状腺结节是临床上常见的疾病,由多种病因引起。B型超声影像技术因其检查费用低、无创伤、成像速度快,可实时诊断、可重复性强等特点已成为最常用检查的方法。本文以超声图像为研究对象,以提取超声子图的纹理特征和量化结节感兴趣区域(Region of Interest,ROI)中的超声征象为研究目标,利用已有病例的超声影像和临床资料,重点分析研究超声图像纹理特征和甲状腺超声影像报告和数据系统(thyroid imaging reporting and data system,TI-RADS)表现的超声征象,为鉴别模型提供特征集合。论文首先分析和整理了 449个甲状腺结节超声图像病例。从超声仪视频中截取超声图像,标注结节ROI的边界。按照TI-RADS的标准,整理了多种结节征象以及每一征象的表现状态和结节的最终诊断结果。介绍了两种常见的图像分割技术NCut和Snakes,并将其应用于甲状腺结节ROI提取。接着,以常用的纹理特征提取方法双树复小波变换(Dua1 Tree Complex Wavelet Transform,DT-CWT)和Gabor滤波器为基础,提出一种多尺度融合的甲状腺结节纹理特征提取方法。该方法首先将含有甲状腺结节ROI的超声子图进行DT-CWT和Gabor变换得到纹理图像,然后计算纹理图像的均值和方差,采用首尾相接的方法实现特征融合。在分类判别方法下,最终实现甲状腺结节的良恶性判别。最后,论文提出了一种基于TI-RADS的甲状腺结节超声图像半监督分级方法。TI-RADS作为诊断甲状腺结节的标准,能够规范、准确地指导超声工作者进行超声检查。作为计算机辅助诊断系统,首先定量分析TI-RADS表现出的超声征象,将这些征象作为判别甲状腺结节的特征向量,然后应用半监督模糊C-均值聚类集成模型得到的聚类结果。实验结果表明,该方法可以区分不同级别的甲状腺结节。本文从甲状腺结节纹理特征和分级方法两个方面研究了甲状腺超声图像,并取得了较好的效果。作为一种计算机辅助诊断方法,在临床上辅助医生诊断甲状腺结节病症,减小医生主观判断,提供有效的诊断建议,进一步促进机器学习方法在医学中的应用。
[Abstract]:Thyroid nodule is a common disease in clinic. Because of its low examination cost, no trauma, fast imaging speed, real-time diagnosis and strong repeatability, B-mode ultrasound imaging technology has become the most commonly used examination method. In this paper, we take ultrasound image as the research object, take the extraction of the texture feature of ultrasound subimage and the ultrasonic sign of region of interest (Region of Interest,ROI) of quantized nodule as the research object, and use the ultrasonic image and clinical data of the existing cases as the research object. In order to provide the feature set for the identification model, the ultrasonic features of ultrasound image texture and thyroid ultrasound image report and data system (thyroid imaging reporting and data system,TI-RADS) were analyzed and studied. First of all, 449 cases of thyroid nodules were analyzed and sorted. Ultrasound images were intercepted from ultrasonic video and the boundary of nodular ROI was labeled. According to the standard of TI-RADS, various nodular signs, the manifestation of each sign and the final diagnosis of nodule were arranged. Two common image segmentation techniques, NCut and Snakes, are introduced and applied to ROI extraction from thyroid nodules. Then, based on two tree complex wavelet transform (Dua1 Tree Complex Wavelet Transform,DT-CWT) and Gabor filter, a multi-scale fusion method is proposed to extract the texture features of thyroid nodules. In this method, first of all, the ultrasonic subimage containing thyroid nodule ROI is transformed by DT-CWT and Gabor to obtain the texture image, then the mean value and variance of the texture image are calculated, and the feature fusion is realized by combining the head and tail. Finally, the benign and malignant thyroid nodules can be distinguished by classifying and discriminating. Finally, this paper presents a semi-supervised classification method of thyroid nodules based on TI-RADS. TI-RADS is the standard for the diagnosis of thyroid nodules. As a computer-aided diagnosis system, the ultrasonic features of TI-RADS are quantitatively analyzed, and these signs are used as characteristic vectors to distinguish thyroid nodules, and then the clustering results obtained by semi-supervised fuzzy C-means clustering model are applied. The experimental results show that the method can distinguish different thyroid nodules. In this paper, we studied the thyroid ultrasound images from the texture features and classification methods of thyroid nodules, and achieved good results. As a kind of computer-aided diagnosis method, it can help doctors diagnose thyroid nodules in clinic, reduce doctors' subjective judgment, provide effective diagnostic advice, and further promote the application of machine learning in medicine.
【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

【相似文献】