基于FFDM图像的乳腺癌肿块与结构紊乱病症检测方法研究

The incidence of breast cancer is steadily increasing in recent decades, and breast cancer is a serious threat to women’s health in China. Early diagnosis can greatly improve the cure rate and the life quality of patients. Full Field Digital Mammography (FFDM) is one of the most effective means of early detection of breast cancer, and this research project involving early diagnosis of breast cancer on FFDM is of great significance. Existing researches have several disadvantages: most of them were based on mammograms from European and American women whose breast are not dense generally, while Chinese women generally have dense breast, the mass detection in dense mammogram has not be investigated. Besides, most mass detection methods were performed on a single image, ignoring the information embedded in related images, thus, their accuracies were not high. The research of architectural distortion (another common abnormality) was lacking. The diagnosis of benign/malignant of mass ignored useful information embedded in easy-to-get unlabeled images, the accuracy was not high. In this project, we try to overcome the above shortcomings. We separate dense breast from non-dense breast, and deal with them separately to improve the usage of mass detection procedure. We will integrate the two-view comparison and bilateral comparison based on an anatomically oriented breast coordinate system, and achieve mass detection with multiple kernel support vector machine. Based on the characteristics of architectural distortion in images, we will investigate the detection problem with orientation field analysis, multiple-scale multiple-direction Gabor and Curvelet transformations. We will integrated semi-supervised level set segmentation and semi-supervised multiple kernel support vector machine to improve the diagnosis of benign/malignant mass. We will optimize the methods based on dataset of Chinese women’s mammogram. Related research results may improve the technology of detection and diagnosis of breast cancer, promote the development of medical image processing, pattern recognition and other displines.

乳腺癌严重威胁我国女性健康，早期诊断可大幅度提高患者的治愈率与生活质量。全视野数字乳腺X线摄影(FFDM)是当前乳腺癌早期检测最有效的手段之一。本项目研究基于FFDM的乳腺癌检测，具有重要意义。现有研究大都针对欧美非致密型乳腺，缺乏对我国女性常见致密型腺体的考虑，肿块检测大多仅处理单幅图像，未充分利用多幅图像对照信息；缺乏对常见的结构紊乱异常检测研究；肿块良恶性诊断准确率不高。针对这些问题，本项目拟将致密型乳腺图像与非致密型分开处理，研究基于乳腺内在坐标系、融合双视图与左右侧对比、结合多核支持向量机的肿块检测方法；根据结构紊乱的图像特点，提出基于方向场分析、多尺度多方向Gabor与Curvelet变换等的检测方法；结合水平集分割与半监督多核支持向量机方法提高肿块良恶性识别准确度，并在中国女性数据集上优化算法。相关研究成果有望提高乳腺癌检测和诊断水平，促进医学图像处理、模式识别等学科的发展。

乳腺癌是女性最常见的恶性肿瘤之一。我国每年约有18万妇女患上乳腺癌，1.3万妇女死于乳腺癌。早期诊断可大幅度提高患者的治愈率与生活质量。钼靶X线摄影是当前乳腺癌早期检测最有效的手段之一，图像上主要包括钙化、肿块和结构紊乱异常，医生一般通过经验来判断是否有以上的异常发生以及决定后续检查或手术。虽然医生双重阅片可以提高诊断准确率，但由于待阅图像数据量巨大，双重阅片实际难以推广。本项目研究基于钼靶图像的乳腺癌辅助检测与诊断。. 对肿块的检测提出了一种检测方法，首先用一种称为多同心层MCL的自适应区域增长方法分割可疑区域，然后，使用窄带活动轮廓模型方法细化初始分割的可疑区域。对可疑区域与分割边界提取几何与纹理特征，可疑区域最终用支持向量机进行分类。提出的方法在包含429张头尾位CC图像的数据集上进行了实验，灵敏度为78.2%时伪正率为1.48 FPsI。此外，研究了融合多张图像信息的肿块检测算法，对肿块检测提出了一种结合MLO和CC视图的方法，并且引入了twin支持向量机（TWSVM）分类器，提出了一种迭代的特征选择方法，在DDSM数据集上进行了实验。对肿块的良恶性识别方面，提出了一种结合TWSVM分类器与L21范式稀疏特征选择的良恶性识别方法。. 在结构紊乱检测问题中，提出了一种结合TWSVM与多次递归特征消除特征选择的结构紊乱检测方法。对结构紊乱与正常组织的识别问题中，提出了一种基于迁移学习A-SVM的方法。. 在钙化检测中，提出了一种集成可能性模糊c均值PFCM和加权支持向量机WSVM的钙化簇检测方法。在410张钼靶图像上进行了实验，并与标准非加权SVM进行了比较。结果显示提出的方法是有效的。. 项目成果将有助于钼靶图像的乳腺疾病自动检测与诊断。