基于Adaboost-决策树算法的乳腺微钙化区域真假阳性检测

True-and false-positive detections of breast microcalcifications based on Adaboost-decision tree algorithm

乳腺癌的早期症状在乳腺钼靶图像中主要表现为微钙化点,微钙化区域的真假阳性检测对于乳腺癌早期筛查具有重要意义。本研究选取DDSM图像进行实验,手动截取了400个疑似钙化区域。首先提取全部区域的Haralick纹理特征和灰度游程矩阵特征建立特征集,然后使用Adaboost算法集成决策树,构建强分类器AB-DT,对400个疑似钙化区域进行分类。实验发现当集成462棵决策树时,模型分类性能最佳。最后进行10折交叉验证,AB-DT算法达到了91.75%的准确率,91.75%的敏感性,91.79%的特异性,F1指数为0.9187。该模型在微钙化真假阳性检测上性能优越,可用于辅助乳腺微钙化点检测,具有一定的临床应用价值。

The early manifestation of breast cancer is mainly characterized by microcalcifications in mammograms.The trueand false-positive detections of microcalcifications are of great significance for the early screening of breast cancer.DDSM images were selected for the experiment,and 400 suspected calcification regions were manually intercepted.The feature set was firstly established by extracting Haralick texture features and grey-level run length matrix features of all regions;and then,Adaboost algorithm was integrated with decision tree to construct a strong classifier AB-DT for classifying 400 suspected calcification regions.It was found that the model classification performance was the best when 462 decision trees were integrated.Finally,10-fold cross-validation was conducted,and the results revealed that the accuracy,sensitivity and specificity of AB-DT algorithm reached 91.75%,91.75%and 91.79%,respectively,and that F1 score was 0.9187.The proposed model has superior performance in the true-and false-positive detections of microcalcifications,and it can be used to assist the detection of breast microcalcifications,which has certain clinical application value.