【作者】 丁建睿；

【导师】 承恒达；

【作者基本信息】 哈尔滨工业大学 ， 人工智能与信息处理， 2012， 博士

【摘要】 癌症严重威胁着人类的健康，其成功治愈依赖于病变的早期发现与治疗。超声成像以其无创、经济、便捷、无辐射等特点，已经成为浅表器官病变早期检测的首选方式。但是对超声图像的正确判读依赖于医生的知识水平和临床经验，具有较强的主观性。计算机辅助诊断技术的引入，为临床医生提供了客观、量化的病变信息和辅助决策信息，有助于消除由于医生主观因素造成的误诊，是临床医学研究的热点问题，对提高疾病的诊断准确率具有重要意义。目前基于超声图像的计算机辅助诊断系统主要存在两个问题：一是由于图像质量导致病变的精确定位相当困难，定位误差会对特征的准确提取与分类性能造成影响；二是缺乏在定位不精确前提下的分类方法研究。本文针对以上问题，对超声图像计算机辅助诊断系统中不依赖于病变精确定位的分类方法进行了研究，提出将超声图像分类问题转化为多示例学习问题。本文考虑样本分布情况，提出了局部加权的Citation-kNN算法；考虑到超声图像存在良恶性特征交叉现象，提出了示例空间向概念空间映射的多示例学习方法；提出了评价弹性超声图像的量化指标，针对弹性图像和B超图像的特点，提出全局特征与局部特征相结合的多示例学习方法。主要工作包括以下三方面内容：一、提出局部加权的Citation-kNN算法。对超声图像进行划分，其整体作为示例包，子区域作为示例，将超声图像分类问题转化为多示例学习问题。在传统Citation-kNN算法投票集合构造的基础上，进一步考虑样本的空间分布，对投票者采用距离加权和离散度加权，并对不同加权方式进行组合，在乳腺超声图像库和标准测试数据集上进行了实验并取得了良好的效果。二、提出局部特征与多示例学习相结合的超声图像分类方法。超声图像中非病变组织会对病变的准确分类造成一定的干扰，对病变的定位将有助于提高分类的准确性。在对病变初步定位的基础上，对其进行划分，采用局部纹理特征对病变进行描述，将该定位区域作为示例包，其子区域作为示例。考虑到超声图像存在良恶性特征混叠的现象，传统的多示例学习定义不再适用。采用聚类方法获取概念空间，将示例包向概念空间投影，在概念空间构造分类器。在乳腺超声图像库上的实验表明，相比于其他方法，本文所提出的方法具有更高的分类准确性。三、提出评价甲状腺弹性超声图像的定量指标，提出全局特征与局部特征相结合的多示例学习方法。通过对弹性超声图像的分析，提出可用于评价弹性图像的量化指标，与目前临床上评价弹性图像的方法相比，具有更高的准确性和可靠性。通过对弹性图像特征的分析，提出弹性图像的整体特征相比局部特征更有利于良恶性病变的区分。在对病变初步定位的基础上，将弹性图像的全局特征与B超图像的局部特征相结合，采用多示例学习方法进行分类，取得了良好的效果。本文在超声图像分类方面的研究工作，较好地解决了病变定位不精确情况下的特征提取与分类问题，对超声图像计算机辅助诊断系统的研究具有积极的意义。更多还原

【Abstract】 Cancer is a serious threat to human health. Its successful cure is dependent onearly detection and treatment. Ultrasound imaging is noninvasive, economic,convenient, no radiation and so on. It has become the preferred method for earlydetection of superficial organ diseases. However, correct interpretation for theultrasound images is dependent on the physician’s knowledge level and clinicalexperience, has strong subjectivity. Computer aided diagnosis technology canprovides objective, quantification and decision making information about lesion.It is helpful to eliminate misdiagnosis in clinical practice due to subjective factors.It has become a hot issue in the medical domain. Moreover, it has importantsignificance for improving diagnosis accuracy. At present, there are twoproblems in the computer aided diagnosis system for ultrasound images. One isvery difficult to position lesions precisely due to the image quality. The error canaffect feature extraction and classification performance. The other is the lack ofstudy for classification method under the inaccurate positioning condition.To solve above problems, this dissertation study the classification method ofcomputer-aided diagnosis system for ultrasound images, which do not rely onlesions precisely positioning. It proposes to convert the classification problem ofultrasound image to multiple-instance learning problem. It proposes locallyweighted Citation-kNN algorithm based on sample distribution. There existscrossover of benign and malignant lesions on ultrasound images. Thephenomenon is taking into account and the multiple-instance learning method,which based on mapping sample space to concept space, is proposed. Thedissertation proposes a quantitative criterion for evaluating elastogram. Based onthe characteristics of the B-mode images and elastogram, it proposes amultiple-instance learning method which combining global and local features.The main work includes the following three aspects:1. Locally weighted Citation-kNN algorithm is proposed. By dividing theultrasound image, the whole image is a bag and sub-regions are instances. Thenthe classification problem of ultrasound image can convert to a multiple-instancelearning problem. By further considering spatial distribution of samples, the voter weighted according to distance and dispersion based on traditionalCitation-kNN’s voting set. The different weighting methods are combined andtested. The good results are achieved when experimented on breast ultrasoundimage database and benchmark for multiple-instance learning.2. The classification method for ultrasound images which combining localfeatures and multiple-instance learning is proposed. The non-diseased tissue caninterfere with accurate classification of lesions. To position lesion will help toimprove the classification accuracy. The lesion is roughly located as region ofinterest (ROI) and the ROI is divided. Local texture features are used to describethe ROI. The ROI is a bag and sub-regions are instances. The traditionaldefinition of multiple-instance learning is no longer applicable due to thefeatures’ overlapping between benign and malignant lesion. Clustering method isused to construct the concept space. The bag is projected to the concept space.The classifier is trained on the concept space. The experiments show that theproposed method has higher classification accuracy on the ultrasound imagedatabase.3. The quantitative criterion to evaluate thyroid elastogram is proposed. Amultiple-instance learning method is proposed which combining global and localfeatures. A quantitative indicator for evaluating elastogram is presented byanalyzing elastogram features. It has higher accuracy and reliability comparedwith current clinical evaluation methods. For the elastogram, the global featuresare more discriminable for benign and malignant lesions than local features. Afterlocated the lesion roughly, the global features of elastogram and the local featuresof B-mode images are combined. The proposed multiple-instance learningmethod has achieved better classification accuracy than others.The research work is mainly on classification of ultrasound image. It solves thefeatures extraction and classification problem when imprecisely positioning thelesion well. In addition, it has positive significance to the computer-aideddiagnosis system.更多还原