目的 探讨对原始乳腺钼靶图像进行变换和采用机器学习算法融合不同类型的图像特征,以提高乳腺癌近期发病风险预测精度的价值.方法 自匹兹堡大学医学中心的临床数据库下载185例女性受检者头足(CC)位全数字化乳腺X线摄影(FFDM)图像.首先对原始灰度图像进行乳腺区域分割并将其变换为虚拟光学密度图像,而后从原始灰度图像和虚拟光学密度图像中分别提取不对称特征.基于此不对称特征分别训练第1阶段的2个决策树分类器,再以这2个分类器输出的得分值作为输入,训练第2阶段的1个决策树分类器.对乳腺癌近期发病风险预测性能采用留一法进行验证.结果 采用两阶段决策树融合方法预测乳腺癌的ROC曲线下面积(AUC)为0.9612±0.0132,敏感度为96.63％(86/89),特异度为91.67％(88/96),准确率为94.05％(174/185).结论 从虚拟光学密度图像中可提取出对乳腺癌具有较高预测力的特征,采用两阶段决策树方法对两类图像特征进行二次融合有助于提高乳腺癌近期发病风险预测精度.%Objective To investigate the value of improving the prediction accuracy of near-term risk for developing breast cancer by transforming the original mammography image and fusing the different types of image features using the algorithm of machine learning.Methods The craniocaudal (CC) full-field digital mammography (FFDM) of 185 women were downloaded from the clinical database at the university of Pittsburgh medical center.Firstly,the original gray images were segmented and transformed into virtual optical density images.Then the asymmetry features were separately extracted from original gray images and virtual optical density images.Two decision tree classifiers of the first stage were trained based on the features extracted from two types of image.And the scores output from the two classifiers were used as input to train the second stage of one decision tree classifier.Leave-one-case-out method was used to validate the prediction performance of near-term risk of breast cancer.Results Using two-stage decision tree fusion method to predict breast cancer,the area under the ROC curve (AUC) was 0.9612±0.0132.And the sensitivity,specificity and prediction accuracy were 96.63％(86/89),91.67％(88/96) and 94.05％(174/185).Conclusion The features extracted from virtual optical density image have higher discriminatory power of predicting breast cancer.Fusing the two kinds of image features twice by two-stage decision tree method can help to improve the prediction accuracy of near-term risk of breast cancer.
展开▼
