【作者】 姜娈；

【导师】 宋恩民；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2009， 博士

【摘要】 乳腺癌是50岁以上妇女的主要死因之一，它是一种严重危害女性身心健康的常见恶性肿瘤。在乳腺癌早期诊断的多种方法中，乳腺X线摄片被认为是最可靠和最有效的方法。基于乳腺X线摄片的计算机辅助检测乳腺癌方法，可以有效地辅助放射科医师提高乳腺癌检测的精准度、一致性和效率。然而目前的计算机辅助检测乳腺癌肿块病灶方法的性能没有被放射科医师认可和信赖，这严重影响了该类型系统在实际临床中的应用。课题研究以提高基于单幅乳腺X线摄片的计算机辅助检测肿块方法的性能，和得到放射科医师容易接受且有信心的计算机辅助检测结果为目标，对基于乳腺X线摄片的计算机辅助检测肿块方法进行了深入研究。主要研究内容包括：基于单幅乳腺X线摄片的计算机辅助检测肿块方法中可疑肿块病灶的准确分割和相关特征的有效量化；基于多幅乳腺X线摄片的计算机辅助检测肿块方法中同侧乳腺的轴位(CC)视图和斜位(MLO)视图中可疑肿块区域的匹配和匹配区域对相关多视图特征的提取；基于内容图像检索的计算机辅助检测肿块方法中相似确诊参考病例图像的准确和有效查找。首先，针对基于单幅乳腺X线摄片的计算机辅助检测肿块方法中肿块分割和特征提取两个难点问题，提出了一种基于最大熵原则和活动轮廓模型的自动分割方法，并基于分割的肿块边界，实现了病理学相关毛刺组织的检测与量化。其次，根据同一病患同侧乳腺组织拍摄的CC视图和MLO视图中相应肿块病灶区域在中轴线上的投影点与乳头之间的距离基本保持不变的性质，匹配了多视图中相应的可疑病灶区域，并基于匹配区域对提取了视图不变性特征和相似性度量特征。分别对单幅乳腺X线摄片中的每个感兴趣区域提取的39个图像学特征和多幅乳腺X线摄片中的每个匹配区域对提取的23个图像学特征形成的初始特征集合，采用了基于逐步判别法的特征选择方法进行优化选择，并基于FISHER线性分类器得到了可疑肿块病灶区域的检测分数，实现了基于单幅乳腺X线摄片的计算机辅助检测肿块方法和基于多幅乳腺X线摄片的计算机辅助检测肿块方法。经过对基于单幅乳腺X线摄片的计算机辅助检测肿块方法和基于多幅乳腺X线摄片的计算机辅助检测肿块方法整体性能的评估和比较分析，显示了基于多幅乳腺X线摄片的计算机辅助检测肿块方法在提高基于单幅乳腺X线摄片的计算机辅助检测肿块方法灵敏度的同时降低了假阳性率，提高了基于单幅乳腺X线摄片的计算机辅助检测肿块方法的性能。最后，通过基于最大熵原则和活动轮廓模型的可疑肿块病灶分割、60个可疑肿块病灶相关的特征提取、基于遗传算法的特征优化选择、基于粒子群优化算法的特征权重学习、基于加权特征向量间欧氏距离的相似性度量准则和基于K近邻分类的决策值计算，实现了基于内容图像检索的计算机辅助检测肿块方法。基于可疑肿块病灶的分割结果，将参考病例数据库包含的全部感兴趣区域进行划分，分别对每个数据集合包含的可疑病灶相关的特征向量集合进行了优化选择和权重学习，并通过基于特征向量间的加权欧氏距离的相似性度量准则与基于图像像素值分布的Pearson相关性的有效结合，调节了可疑肿块病灶分割质量对感兴趣区域间相似性度量的影响，改进了基于内容图像检索的交互式计算机辅助检测肿块方法的性能，得到了放射科医师容易接受且有信心的计算机辅助检测结果。通过对基于单幅乳腺X线摄片计算机辅助检测肿块方法、基于多幅乳腺X线摄片计算机辅助检测肿块方法和基于内容图像检索的计算机辅助检测肿块方法的研究，为实现计算机辅助检测系统在临床中的广泛应用奠定了理论基础。更多还原

【Abstract】 Breast cancer is one of the leading causes of death in women over the age of 50. For the early detection and diagnosis of breast cancer, mammography is currently considered the most reliable and cost-effective method. Computer-aided detection (CAD) based on mammography could eventually provide a valuable "second opinion" for improving accuracy, efficiency and consistency of detecting breast cancer in the clinical evironment. However, the performance of current CAD in mass detection remains relatively low, and the radiologists could not have confidence in and accept this type of schemes.To improve the performance in the computer-aided detection of mass based on single mammography, as well as to increase radiologists’ confidence in and reliance on CAD-prompted mass detection results, computer-aided detection methods of breast mass based on mammography are studied in detailed in this dissertation. The main contents are as follows: accurate mass segmentation and related characteristic quantification in CAD based on single mammography, matched regions identification and related feature extraction in CAD based on multiple mammographies (ipsilateral views), accurate and efficient search of similar reference images in CAD based on content-based image retrieval (CBIR).First, due to the two challenging problems, mass segmentation and related feature extraction, an automated segmentation method based on maximum entropy principle and active contour model is proposed in this dissertation. With the segmented mass contour, spiculated tissues surrounding the mass are detected, and a quantitative spiculation index is computed to assess the degree of spiculation.Second, based on the projected distance to the nipple along the centerline, matched regions identification and related feature extraction are carried out on both ipsilateral views (cranio-caudal view and mediolateral oblique view). With the 39 features extracted from single mammography and the 23 features extracted from multiple mammographies, stepwise feature selection method and linear discriminant analysis (Fisher) are sequentially used to obtain the detection scores of the suspicious regions. The performances in mass detection of CAD based on single mammography and CAD based on multiple mammographies are evaluated and compared. The experimental results show that CAD based on multiple mammographies could improve the sensitivities, while reducing false-positive detection rates in the CAD based on single mammography. CAD based on multiple mammographies could improve the performance of CAD based on single mammography.Third, with the mass segmentation based on maximum entropy principle and active contour model, 60 related feature extraction, GA-based (Genetic Algorithm) feature selection, PSO-based (Particle Swarm Optimization) feature weights study, similarity measure based on weighted Euclidean distance and KNN-based (K Nearest Neighbor) decision index computation, CAD based on CBIR is implemented in this dissertation.Based on the mass segmentation, the reference databased is divided. Feature selection and weight study are respectively applied to each dataset. With the combination of weighted multi-image feature-based similary measure and pixel value-based Pearson’s correlation, the effect of the segmentation qualify on the similariy measure is adjusted. This method improves the performance of ICAD (Interactive CAD) based on CBIR, and obtains CAD-prompted mass detection results which will increase the radiologists’ confidence and reliance.The researches on CAD method based on single mammography, CAD method based on multiple mammographies and CAD method based on content-based image retrieval in this dissertation establish theoretical basis for widely use of CAD systems in clinics.更多还原