【作者】 邹波；

【导师】 林春漪； 张岚；

【作者基本信息】 华南理工大学 ， 工程硕士（专业学位）， 2018， 硕士

【摘要】 乳腺癌是女性最常见的恶性肿瘤,严重危害女性的生命健康。穿刺活检是乳腺病灶良恶性诊断的“金标准”,然而穿刺活检有创,且临床上存在过多不必要的穿刺活检,因此需要一种具备高精度且无创的乳腺病灶良恶性分类方法来为临床医生提供有效的辅助诊断信息,减少不必要的穿刺活检。本文以超声射频(Radio Frequency,RF)时间序列为数据源,以乳腺病灶为研究样本,从时域、频域、时频域、非线性动力学四个维度进行特征提取,以支持向量机(Support Vector Machine,SVM)和随机森林作为分类器,提出两个具备高精度且无创的乳腺病灶良恶性分类解决方案,并开发了乳腺病灶良恶性分类的计算机辅助诊断(computer aided diagnosis,CAD)系统。解决方案一从时域、频域、非线性动力学三个维度对超声RF时间序列进行特征提取,SVM受试者工作特征曲线下的面积(area under receiver operating characteristics curve,AUC)为0.914,最好准确率为86.59%,平均准确率80.08%,随机森林AUC为0.937,最好准确率为95.12%,平均准确率87.33%,对比目前研究中最好结果,SVM和随机森林的AUC分别提升1.6%和10.2%。解决方案二先对超声RF时间序列进行三层Mallat小波分解,在时频域对低频系数和高频系数提取统计特征和非线性动力学特征,然后对解决方案一中频域特征进行改进提取,最后对特征采用Relief-F和主成分分析(Principal Component Analysis,PCA)进行特征降维。选择Relief-F作为降维方式,SVM的AUC为0.932,最好准确率为86.67%,平均准确率82.22%,随机森林AUC为0.997,最好准确率为98.67%,平均准确率、平均敏感性、平均特异性分别为94.46%、94.92%、95.25%,对比目前研究中最好结果,SVM和随机森林的AUC分别提升3.6%和17.3%;选择PCA作为降维方式,SVM的AUC为0.915,最好准确率为88.75%,平均准确率为82.22%,随机森林AUC为0.985,最好准确率为96.25%,平均准确率为92.5%,平均敏感性为94.58%,平均特异性为89.78%,对比目前研究中最好结果,SVM和随机森林的AUC分别提升1.7%和15.9%。同时在Visual Studio2013下开发的乳腺病灶良恶性分类CAD系统,可以为临床医生提供有效的辅助诊断信息,有效减少不必要的穿刺活检,为乳腺病灶良恶性识别提供无创的解决新思路。更多还原

【Abstract】 Breast cancer is the most common malignant tumor in women,which seriously endangers women’s life and health.Biopsy is the "gold standard" for the diagnosis of benign and malignant breast lesions.However,biopsy is invasive,and there are too many unnecessary biopsy in clinical practice,so it is necessary to have a high accuracy and noninvasive classification of breast lesions to provide effective diagnostic information for clinicians to reduce unnecessary biopsy.In this paper,the ultrasound frequency(RF)time series is used as the data source,and the breast lesions are studied as samples,feature extraction is carried out from four dimensions: time domain,frequency domain,time-frequency domain and nonlinear dynamics,with support vector machine(SVM)and random forest as classifiers,two high-precision and non-invasive methods for classifying benign and malignant breast lesions are proposed,and developed a computer aided diagnosis(CAD)system for the classification of benign and malignant breast lesions.Solution 1 extracted the features of ultrasonic RF time series from the three dimensions of time domain,frequency domain and nonlinear dynamics.the area under receiver operating characteristic curve(AUC)of the SVM was 0.914,the best accuracy rate was 86.59%,and the average accuracy rate was 80.08%,the AUC of random forest was 0.937,the best accuracy rate was 95.12%,the average accuracy was 87.33%.Compared with the best results in current study,the AUC of SVM and random forest were increased by 1.6% and 10.2%,respectively.Solution 2 firstly decomposes the ultrasonic RF time series with three layers of Mallat wavelet,then extracts the statistical and nonlinear dynamic features of the low frequency coefficients and high frequency coefficients in the time-frequency domain,then the frequency domain features of solution one are improved and extracted,and finally uses the Relief-F algorithm and Principal Component Analysis(PCA)to reduce the dimension of the features.Relief-F algorithm was chosen as the dimension reduction method.The AUC of SVM was 0.932,the best accuracy rate was 86.67%,and the average accuracy rate was 82.22%,the AUC of random forest was 0.997,the best accuracy rate was 98.67%,and the average accuracy,average sensitivity,and average specificity were 94.46%,94.92%,and 95.25%,respectively.compared with the best results in current study,the AUC of SVM and random forest increased by 3.6% and 17.3%,respectively;PCA was chosen as the dimension reduction method.The AUC of SVM was 0.915,the best accuracy rate was 88.75%,and the average accuracy rate was 82.22%,the AUC of random forest was 0.985,the best accuracy rate was 96.25%,the average accuracy rate was 92.5%,the average sensitivity was 94.58%,and the average specificity was 89.78%.compared with the best results in current study,the AUC of SVM and random forest increased by 1.7% and 15.9%,respectively.At the same time,the(CAD)system for the classification of benign and malignant breast lesions was developed under Visual Studio 2013,which can provide clinicians with effective auxiliary diagnostic information,effectively reduce unnecessary biopsy,and provide a noninvasive solution for the identification of benign and malignant breast lesions.更多还原