| 王博,谭红娜,周燕,孙慧芳,耿亚媛,高剑波.乳腺X线摄影影像组学特征预测三阴性乳腺癌[J].中国医学影像技术,2021,37(6):904~908 |
| 乳腺X线摄影影像组学特征预测三阴性乳腺癌 |
| Radiomics features of mammography in predicting triple negative breast carcinoma |
| 投稿时间：2020-05-08 修订日期：2021-03-28 |
| DOI：10.13929/j.issn.1003-3289.2021.06.026 |
| 中文关键词: 三阴性乳腺肿瘤 乳房X线摄影术 影像组学 |
| 英文关键词:triple negative breast neoplasms mammography radiomics |
| 基金项目:国家重点研发计划项目（2017YFC0109404）。 |
|
| 摘要点击次数: 2718 |
| 全文下载次数: 655 |
| 中文摘要: |
| 目的 评估乳腺X线摄影影像组学特征预测三阴性乳腺癌（TNBC）的价值。方法 回顾性分析177例非特殊类型浸润性乳腺癌，包括77例TNBC和100例非TNBC，按7∶3比例将其随机分为训练集（n=123）和验证集（n=54）。提取乳腺X线摄影内外侧斜位（MLO）和头足位（CC）图像的影像组学特征，并采用Select-K-Best方法、最小绝对值收敛和选择算子（Lasso）回归分别筛选MLO、CC及MLO+CC图像中与TNBC相关性最强的影像组学特征，以支持向量机方法分别构建TNBC预测模型；采用受试者工作特征（ROC）曲线分析模型及影像组学特征的预测效能。结果 自CC位、MLO位及MLO+CC图像各提取1 033个影像组学特征。经Select-K-Best方法和Lasso回归分析，分别于训练集CC位图像、MLO位图像和CC+MLO位图像各筛选出7个与TNBC相关性最强的影像组学特征，以之构建TNBC预测模型。ROC曲线结果显示，训练集中CC位、MLO位、CC+MLO位预测模型预测TNBC的ROC曲线下面积（AUC）分别为0.82、0.73、0.84，验证集分别为0.82、0.70、0.78；CC位预测模型在训练集和验证集中预测TNBC的准确率较高，分别为81.30%和81.48%。基于CC位图像筛选出的7个与TNBC相关性最强的影像组学特征中，灰度级依赖矩阵中的Maximum Probability及经过滤波变换的Gray Level NonUniformity及Large Dependence Low Gray Level Emphasis分类预测TNBC的AUC为0.587、0.599及0.615（P均<0.05）。结论 乳腺X线摄影图像、尤其乳腺CC位图像影像组学特征对术前预测TNBC有一定临床应用价值，。 |
| 英文摘要: |
| Objective To observe the value of radiomics features of mammography in predicting triple negative breast carcinoma (TNBC). Methods Data of 177 patients with non-special type of invasive breast carcinoma, including 77 cases of TNBC and 100 non-TNBC patients, were retrospectively analyzed. The patients were divided randomly into training set (n=123) and validation set (n=54) in a ratio of 7 ∶3. The radiomics features of mediolateral oblique (MLO) and cranial caudal (CC) images were extracted, and the Select-K-Best methods, least absolute shrinkage and selection operator (Lasso) regression were used to screen the radiomics features with the most correlation with TNBC in MLO, CC and MLO+CC images, respectively, and the support vector machine(SVM)method was used to construct TNBC prediction models. The predictive efficacy of these models and the radiomics features were evaluated with receiver operating characteristic (ROC) curves. Results Each 1 033 radiomics features were obtained from CC, MLO and MLO+CC images, respectively. In training set, 7 radiomics features with strongest correlation with TNBC were selected from CC, MLO and MLO+CC images using Select-K-Best methods and Lasso regression, respectively, which were used to construct the prediction model of TNBC. ROC curves showed that the area under curve (AUC) of the prediction models based on CC, MLO, CC+MLO images was 0.82, 0.73, 0.84 in training set and 0.82, 0.70, 0.78 in validation set, respectively; and the accuracy of prediction model based on CC image in training set and validation set was 81.30% and 81.48%, respectively. Among 7 radiomics features with strongest correlation with TNBC selected based on CC images, AUC of Maximum Probability in the Gray Level Dependent Matrix, the Gray Level Nonuniformity and Large Dependence Low Gray Level Emphasis of wavelet features for predicting TNBC was 0.587, 0.599 and 0.615, respectively (all P<0.05). Conclusion The radiomics features based on mammography, esp.CC images, had certain clinical application value for preoperative prediction of TNBC. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
