【作者】 翟永平；

【导师】 刘云辉；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2013， 博士

【摘要】 结核病是由结核杆菌感染引起的慢性传染疾病，传统诊断方法存在漏检率高、劳动强度大、严重依赖专家的知识和经验等问题。为了克服诸多缺陷，本文将图像处理及模式识别技术引入结核病的诊断并实现全自动化。论文围绕结核杆菌自动检验中的显微镜自动聚焦及目标检测两大关键技术展开研究，取得了若干重要成果。显微镜自动聚焦方面：（1）针对聚焦函数无定量评估指标的问题，设计了陡峭区宽度、平缓区波动量等6个定量评价指标，为聚焦函数的设计及选取提供了定量标准；（2）为了解决稀疏图像内容情况下传统聚焦方法聚焦成功率低的问题，提出一种基于图像内容重要度的聚焦算法。实验结果表明：当图像内容非常稀疏时，本算法聚焦成功率高达90%，而此时传统算法成功率仅为24%；（3）提出一种基于视频连续反馈的显微镜自动聚焦算法。算法首次建立了显微镜连续聚焦模型以及相关参量之间的函数约束关系，并给出了一种多视野扫描中聚焦初始方向的启发式选取方法。实验结果表明：本算法聚焦速度约为传统算法的2.2倍，聚焦成功率高达96.7%，表明本算法性能远优于传统算法。结核杆菌识别方面：（1）建立了结核杆菌目标的颜色及形状模型，并提出一种基于目标骨架的关键点自动标注及单分支结构目标的形状简化建模方法。该方法的提出解决了经典变形模型无法自动标注关键点、无法自动获取目标形状参数的问题，从而使得大样本训练及基于变形模型的实时目标识别成为可能；（2）建立了面向实际工程应用的由粗到精、逐步细化、层层筛选的多阶段、多层次图像分割框架，并提出一种基于形状几何统计特征及决策树的结核杆菌目标识别算法。为了验证算法性能，分别进行了典型样本实验和大样本统计实验。典型样本实验选取7组分割及识别难度比较大的典型结核杆菌显微图像，结果表明本文算法能很好地适应结核杆菌图像的背景及杂质的复杂变化，分割及识别的效果比较好。大样本实验中，从大量痰涂片样本中随机选取22836幅图像进行实验，结果为：算法灵敏度（也即真阳性率）为95.2%，特异度（也即真阴性率）为91.1%，表明本算法性能较好，但对一些极端情况还需要进一步研究。更多还原

【Abstract】 Tuberculosis (TB) is a serious communicable chronic disease which is casued bythe Mycobacterium Tuberculosis. The conventional method employed for TB diagnosisinvolves a labor-intensive task with poor sensitivity and requires highly trained experts.To overcome these shortcomings, this dissertation employed image process and patternrecognition techniques for automatic TB diagnosis. Two key technologies employed forautomatic TB indentification, including the microscopy autofocusing technique and theobject detection technique, were studied. The technical contribution of this dissertationis summarized as follows:First, in microscopy autofocusing the following contribution is made:(1) Wedeveloped six quantative performance metrics of focus function, including the width ofsloped part of focus curve, the steepness, the variance of flat part of the focus curve,et.al. These metrics provide the quantative standards for focus function design andperformance evaluation;(2) To overcome the shortcoming that the traditionalautofocusing algorithms may fail to find the optimal focal plane under thecircumstances with low image content densities, this dissertation proposed a contentbased focus measure for guiding automatic search of the optimal focal plane. Theexperimental results show that performance of the proposed method is far superior tothe traditional methods: the autofocusing success rate of the proposed method is largerthan90%under the circumstances with low image content density while the traditionalmethod only gains a success rate of24%;(3) A continuous autofocusing method, inwhich the z-stage keeps moving during the search process and the image exposure andthe subsequent focus measure calculation are carried out in parallel with the z-stagemovement, was proposed. This dissertation established a continuous autofocusingmodel and the relationship of the motor speed, the frame rate, and the depth of field ofthe optical system, which is essentially important for achieving continuous autofocusing.Moreover, a heuristic initial search direction selection method was proposed, whichdramatically improve the speed of autofocusing in muli-fileds scanning system.Experimental investigation has been conducted based on our automatic microscopysystem. The experimental results show that the proposed method gains a success rate of96.7%and the autofocusing speed is2.2times faster than the tranditional “stop-and-go”method. These results confirm the superior performance of the proposed method overthe traditional methods.In automatic detection of Mycobacterium Tuberculosis (TB) in microscopicimages, we have made the following contribution:(1) A color model and a deformableshape model of the TB object were established. Moreover, a landmark automaticannotation method which is based on the morphologyical skelecton was proposed and a skelecton based simplified deformable shape model was established. The skelectonbased shape model outperforms the tranditional coutour based shape model not only forits low complexity but also for its capability that makes the landmark be auto-annotated,which makes the shape parameter auto-extraction and recognition possible;(2) A coarseto fine, multi-stage, multi-level image segmentation framework was established to dealwith the complicated situation in practical image segmentaion task. Moreover, aclassification and recognition algorithm based on the shape feature descriptors and thedecision tree classifier was proposed.In order to demonstrate the performance of the proposed algorithms, two types ofexperiments had been condoncted, including the typical samples experiment using7typical TB images and the statistical experiments using22836TB images randomlyselected from larger number of TB specimens. Experimental results show that theproposed algorithm can accommodate the complex variety of specimens and the imagebackground. The statistical experiments result show that the sensitivity (true posistiverate) and the specificity (true negative rate) are95.2%and91%respectively. This is agood score considering that the samples used in the experiments are randomly selectedfrom TB specimens whose image quality may be influenced by many factors. In futurework, we will do futher effort to improve the performance especially the specificity ofthe algorithm.更多还原