【作者】 杜娟；

【导师】 韦岗；

【作者基本信息】 华南理工大学 ， 通信与信息系统， 2016， 博士

【摘要】 人类牙齿健康状况越来越受到重视。很多牙齿疾病和牙齿组织中成分损失和侵蚀有关,但是在适当的治疗下可以得到控制。因此牙齿早期损伤的检测和及时治疗对防止病灶恶化是非常关键的。超声已经被广泛应用于诸如诊断、无损检测和治疗等医学领域。因为高频超声具有波长短而分辨率高的特点,更适合用于牙科医学中。虽然早在上世纪60年代超声成像技术就开始应用于牙科,但是到现在还没有发展成为有效的诊断工具。已经有很多学者和组织对三维超声成像在牙科的应用进行了尝试和探索,但是从成像效果来看并不尽如人意,有的成像系统较为复杂,大部分实验都需要对牙齿样本进行预处理,改变了样本本来面貌,降低了成像结果的实用性。针对这些缺点,本文介绍了一种新的机械扫描系统,这个系统收集人牙齿的表面结构信息和牙釉质-牙本质的边界信息并进行三维成像。中心频率为50MHz的A超换能器由一个三维控制台控制,发射并接收超声信号。对一维回波信号进行一种基于频率的增益补偿算法,然后对二维图像信号进行基于Q学习的图像去噪算法之后,可以得到较清晰的牙釉质三维图像,并不需要对牙齿做任何改变形状的处理。釉质层的测量结果和CT扫描数据的测量结果进行了比较。通过对中山大学光华口腔医学院提供的4颗成人恒磨牙分别进行的三维重建及牙釉质厚度测量的实验结果可知,本文设计的成像系统及提出的成像方法可以很好地再现牙釉质的三维轮廓,为医生提供完整的三维视图。本文的主要研究内容可分为三个方面:设计开发一个基于A超探头的三维重建系统;对一维回波信号进行基于频率的增益补偿算法;对二维超声图像进行基于Q学习的融合滤波,提高成像质量。在三维超声重建系统方面,分析总结三维超声重建系统的模块组成、各模块的技术原理及功能,对比分析研究目前存在的几种扫描方法及各自优缺点。在此基础上,设计开发了基于机械扫描的A超三维重建系统。系统主要包括:中心频率为50MHz的单晶换能器、信号发生器、三维运动台、数字信号采集卡等硬件,图像数据的采集软件和一维、二维、三维图像的处理、生成显示软件开发。在一维信号处理算法上,参考已有的增益补偿算法,将线性自回归(AR)模型应用于A超回波信号的衰减参数估计,根据估计的衰减参数对信号进行补偿,得到了较好的图像质量。在线性自回归模型基础上,将自回归参数用少量基函数的线性加权和表示,把非平稳过程转换为线性过程,提高了参数估计的精度,从而提高了衰减系数估计的准确性,进一步提高了重建图像的质量。在二维图像去噪算法上,提出了一种基于强化学习中的Q学习算法的超声二维图像融合滤波算法,利用中值滤波器、小波变换滤波器和去斑各向异性扩散滤波器共同实现二维图像融合滤波算法,对不同滤波器处理结果进行加权和,Agent在学习过程中根据一定的图像评价指标选择行动(改变各个权重的大小),直到找到最优解。主观和客观评价显示此算法具有较好效果和普遍适应性。更多还原

【Abstract】 More and more attention has been paid to the health of human teeth.Many tooth diseases are related to the loss of tooth tissue and erosion,but can be controlled with proper treatment.Therefore,early detection and timely treatment is critical to prevent the deterioration of the disease.Ultrasound has been widely used in medical practice such as diagnosis,non-destructive testing and treatment.Due to its short wavelength and associated high resolution,high frequency ultrasound is more suitable for the applications concerning tooth.Although ultrasonography in dentistry can trace back to the early 1960’s,it has not been developed as an efficient diagnostic technique.Many exploration and research has been carried out by scholars and organizations on 3D ultrasound imaging in dental application,but the outcomes are not satisfactory so far.Some of the imaging systems are very complex.Most of them need to dedicatedly prepare the tooth samples,which changed the original stucture of the samples and in turn reduced the practicality of the results.To overcome these shortcomings,this dissertation introduces a new scanning system which collects structural information of tooth surface and enamel-dentine junction.A 3D motorized scanning stage was used to control the moving trajectories of an A-mode transducer so as to acquire the echoes from the surface of a tooth.A frequency-based compensation method was applied to 1-dimentional signals.Then a new denosing algorithm of 2-dimentional ultrasound images based on Q-learning was introduced.3-dimentional images of tooth enamel could be achieved and no special preparation was required before the scanning.The measurements of enamel thickness were compared with those of CT scanning data.Four permanent molars provided by from Guanghua College of Stomatology,Sun Yat-sen University were scanned and then 3D reconstructed.The 3D reconstruction and measurements of enamel thickness show that the imaging system and the imaging method can well reproduce the enamel layer and can provide the dentists with full 3D view of tooth enamel.The main research contents of this paper can be divided into three aspects: design and implementation of a 3D reconstruction system;a frequency-based gain compensation algorithm applied to 1D echo signals;a Q-Learning-based fusion filtering algorithm to improve the image quality applied to 2D ultrasound images.Design and implementation of a 3D reconstruction system.The components of a 3D ultrasound reconstruction system were studied 。 Several scanning methods and their advantages and disadvantages were compared and analyzed.A mechanical scanning system was developed based on the research which consisted of a single element focused transducer with a center frequency of 50 MHz,a general purpose ultrasonic pulser-receiver,a 3D translation device with its control box and control card,an acquisition card and a computer.Besides these hardware,the software which control the data acquisition,RF signal processing,2D image denosing and 3D visualization was developed.Introduce of a novel attenuation compensation algorithm.Linear autoregressive(AR)model was applied to the estimation of echo signal attenuation parameter based on existing gain compensation algorithms.The compensated signal can used to build images with better quality.Then a nonstationary AR algorithm was introduced which replaced the AR parameters with the combinations of small amount of base functions and provided even more improvement to image quality.Introduce of a new denosing algorithm of ultrasound images based on Q-learning.Three kinds of filters(median filter,wavelet transform filter and speckle reducing anisotropic diffusion filter)were fused to implement a two-dimensional image filtering algorithm.The results of these three filters were weighted and the agent chose next action(change of the weights)according to the reward(image quality evaluation)until the optimal solution was found.Subjective and objective evaluation shows that this algorithm has good effects and general adaptability.更多还原