【作者】 王瑞平；

【导师】 万柏坤；

【作者基本信息】 天津大学 ， 生物医学工程， 2003， 博士

【摘要】 乳腺癌是妇女常见的恶性肿瘤之一，早期发现、早期诊断、早期治疗是降低乳腺癌死亡率的关键。乳腺钼靶 X 线摄影是目前临床诊断乳腺癌的有力工具。但钼靶 X 线影像的信息只有很少部分能为人眼识别，即使富有经验的医生也很难及时发现钼靶 X 线影像上早期乳腺癌的微小钙化点，以致延误病人的治疗时机。可以说，实现乳癌早期诊断的关键技术之一是及时发现乳癌 X 影像中的微小钙化并判断其是否有恶化倾向。随着计算机技术的飞速发展，基于传统乳腺钼钯 X 线影像的计算机辅助检测微小钙化点已成为乳癌早期诊断的研究热点。这主要是因为细小、颗粒状的成簇微钙化点是乳癌的一个重要早期表现。国外统计资料表明占 30%～50%的乳腺恶性肿瘤伴有微钙化。因此，不断提高微小钙化点的检出率和准确判别其恶性度成为众多学者孜孜以求的目标。 本文建立了一个基于模块化设计思路的计算机辅助诊断系统借以对乳腺钼靶图像上的微钙化点进行检测和模式识别。该系统分为四个模块：①预处理模块-对乳腺钼靶 X 片图像进行数字化和归一化处理，得到具有相同空间分辨率和灰度分辨率的规格化图像，以便于计算机作进一步后续处理。②感兴趣区域(ROI)提取模块-自动寻找并分割含有微钙化点的区域，以节省后续处理的工作量。本研究将独立分量分析(ICA)用于乳腺 ROI 的特征提取，在此基础上用人工神经网络(ANN)分类器进行模式识别。③微钙化点自动检测模块-实现乳腺 ROI 上微钙化点的自动检测与定位。本文将差值图像去噪、阈值化分类技术和小波去噪、ANN 分类技术分别用于乳腺 ROI，得到含高频信号和极高频噪声位置信息的二值化图像及含高频信号和低频背景位置信息的二值化图像。将两者进行与操作得到含微钙化点位置信息的二值化图像。④微钙化点病变类型识别模块-实现微钙化点特征提取和优化及病变类型模式识别，给出初步诊断结果。本文建立了一套表征微钙化点形态、纹理等特性的 33 维特征矢量。然后，用遗传算法进行特征选择得到 17 维优化特征矢量，优化特征矢量与 ANN 组成判别模型完成微钙化点病变类型的判定。上述各模块相互独立，可以单独改进和优化而不影响其它模块。此外，本文将适用于小样本的支持矢量机(SVM)分类器应用于上述分类模型中，并用接受者操作特征曲线(ROC)对分别由 ANN 和 SVM 分类器组成的判别模型分类性能进行评价。 I<WP=4>运用该系统对临床病例和标准乳腺库中乳腺图像进行分析，得到 87.5%(ANN)和 90.0%(SVM)的乳腺 ROI 检出率、96.3%(ANN)和 97.0%(SVM)的微钙化点检出率及 88.7%(ANN)和 93.0%(SVM)的恶性微钙化点识别率。结果证明了本文建立的计算机辅助诊断系统具有较高的微钙化点检出率和较准确的恶性度判别性能，为乳癌早诊研究提供了一套新方法。 本研究中的创新性主要体现在：①提出基于模块化的早期乳腺癌辅助诊断系统设计思路，建立了基于微钙化点检测的早期乳腺癌计算机辅助诊断的完整模型；②首次将独立分量分析成功地应用于乳腺图像 ROI 自动提取；③首次将集合论思想应用到乳腺图像微钙化点的检测中，提出一种能发挥差值图像技术、阈值化分类和小波变换技术、人工神经网络分类等多种技术优势的综合处理检测方法；④建立了一套比较完整的可以表征乳腺图像微钙化点各类特异性的特征参数矢量；⑤首次将现代统计学习理论引入到本研究中，并且成功地建立了基于 SVM 的判别模型；⑥提出了基于 ROC 曲线的阈值选择方法和系统诊断价值评价方法。更多还原

【Abstract】 Breast cancer is one of the most common malignant diseases among women.Clear evidence shows that early discovery, early diagnosis and early treatment ofbreast cancer can significantly increase the chance of survival for patients.Mammography is the most effective method for the early detection of breast cancer.However normally, viewed mammograms display only a very small part of the totalinformation they contain. It is very hard to find the microcalcifications (MCCs) ofearly breast cancer in mammograms even for an experienced radiologist. Therefore,any increase in the detection and classification of MCCs will lead to furtherimprovement in its efficacy in the detection of early breast cancer. With the rapidprogress of computer technology, computer aided detection and identification ofMCCs have been a hot research field since clustered MCCs in mammograms are animportant sign for early detection of breast cancer. It is estimated that about 30% to50% of breast carcinomas detected radiographically demonstrates MCCs inmammograms. So the increase in the detection and classification of MCCs inmammograms has been of interest to many researchers. This paper presents a prototype of a computer-aided diagnostic system (CAD) formammography screening to automatically detect and classify MCCs in mammograms.It comprises four modules. The first module, called the mammogram preprocessingmodule, digitizes and normalizes the original mammogram, and makes it to be fit forcomputer processing. Since the region of interest (ROI) covers only a small part ofthe whole mammograms, the second module, called the ROI finder module, finds andlocates suspicious areas of MCCs. Independent component analysis is implemented toextract the features of ROI, and artificial neural network(ANN) classifier is used tolabel the region as either true or false ROI. Since only MCCs are of interest inproviding a sign of breast cancer, the third module, called the MCCs detectionmodule, is a computer automated MCCs detection system that takes as inputs theROIs provided by the ROI finder module. Two methods are used to detect the MCCs.The first one based on difference-image technique is used to remove thelow-frequency background, while the second one based on wavelet denoising andneural network classifying technique is used to remove the very high-frequency noise.Signals coming out from two methods are combined through a logical AND operationto get the final detected result that contains the position information of MCCs. Finally, III<WP=6>the fourth module, called the MCCs classification module, includes featuresextraction, feature optimization and pattern recognition. A pool of many features (33)with the information about shape, texture and so on of MCCs is computed. Geneticalgorithm is introduced to get the optimal features (17). The ANN classifier is used tolabel the MCC as either malignant or benign. Moreover, support vector machinerelying on the statistical learning theory is applied to the patter recognition in theresearch. One advantage of the designed system is that each module is a separatecomponent that can be individually upgraded to improve the whole system. Moreover,receiver operation curve is used to evaluate the performance of each decision modelin this research. The above methods are adopted in the processing of the test sampleswith a true positive rate (TPR) of 87.5% (ANN) and 90.0% (SVM) in the ROIautomated finder module, a TPR of 96.3% (ANN) and 97.0% (SVM) in the MCCsdetection module, a TPR of 88.7% (ANN) and 93.0% (SVM) in the MCCsclassification module. The SVM classifiers get slightly better results than the ANNclassifiers. The results show that the method has a high performance on the detectionand classification of MCCs, and gives a new method for the research on the diagnosisof early breast cancer. The originalities of this thesis are the followings: 1. One modularization design thought is int更多还原