| 刘思腾,于湛,李哲人,刘译阳,翁思远,王洁洁.乳房X线片深度学习联合临床特征列线图预测乳腺癌表达人表皮生长因子受体2(HER-2)状态[J].中国医学影像技术,2023,39(11):1659~1664 |
| 乳房X线片深度学习联合临床特征列线图预测乳腺癌表达人表皮生长因子受体2(HER-2)状态 |
| Deep learning of mammograms combined with nomogram of clinical characteristics for predicting human epidermal growth factor receptor-2 (HER-2) expression status of breast cancer |
| 投稿时间：2023-07-13 修订日期：2023-09-10 |
| DOI：10.13929/j.issn.1003-3289.2023.11.014 |
| 中文关键词: 乳腺肿瘤 乳房X线摄影 ErbB受体 深度学习 |
| 英文关键词:breast neoplasms mammography ErbB receptors deep learning |
| 基金项目: |
|
| 摘要点击次数: 996 |
| 全文下载次数: 631 |
| 中文摘要: |
| 目的 观察深度学习(DL)术前乳房X线片(MG)联合临床特征列线图预测乳腺癌表达人表皮生长因子受体2(HER-2)状态的价值。方法 回顾性分析265例HER-2检测结果均为(++)的单发乳腺癌患者的MG和临床资料,包括93例HER-2阳性、172例HER-2阴性;按8 ∶ 2比例将其分为训练集(n=211,含74例HER-2阳性、137例HER-2阴性)和验证集(n=54,19例HER-2阳性、35例HER-2阴性)。以单因素及多因素logistic回归分析临床资料,筛选乳腺癌表达HER-2状态的独立预测因素,以之构建临床模型。将头足位和内外斜位MG输入孪生DL网络,获得肿瘤ROI图像,并输入至ResNet50网络提取DL特征,通过全连接层进行特征融合,共获得2 048个DL特征,经Softmax分类器输出二分类结果,得到DL评分(Deep-score),构建DL模型。以Deep-score和独立临床预测因素构建联合模型,并绘制列线图;以校准曲线评估其校准度。应用受试者工作特征(ROC)曲线评估各模型预测乳腺癌表达HER-2状态的效能,以决策曲线分析(DCA)评估其临床获益。结果 雌激素受体状态(OR=3.63)和Ki-67表达水平(OR=2.84)为乳腺癌表达HER-2状态的临床独立预测因子(P均＜0.05);基于此并联合Deep-score构建的联合模型的预测结果与实际结果的一致性良好,其在训练集的AUC(0.97)高于临床模型及DL模型(AUC=0.75、0.96,Z=7.15、2.03,P均＜0.05),在验证集的AUC(0.88)高于临床模型(AUC=0.70,Z=5.76,P＜0.01)而与DL模型差异无统计学意义(AUC=0.86,Z=1.50,P=0.33)。联合模型在训练集的临床净收益高于临床模型及DL模型,在验证集的临床净收益高于临床模型而与DL模型相当。结论 MG DL联合临床特征列线图可有效预测乳腺癌表达HER-2状态。 |
| 英文摘要: |
| Objective To explore the value of nomogram based on deep learning (DL) of preoperative mammograms (MG) combined with clinical features for predicting human epidermal growth factor receptor-2 (HER-2) expression status of breast cancer. Methods MG and clinical data of 265 patients with HER-2 (++) single breast cancer were retrospectively analyzed, including 93 cases of HER-2 positive (positive group) and 172 cases of HER-2 negative (negative group). The patients were divided into training set (n=211, including 74 cases of HER-2 positive and 137 cases of HER-2 negative) or validation set (n=54, including 19 cases of HER-2 positive and 35 cases of HER-2 negative) at the ratio of 8 ∶ 2. Univariate and multivariate logistic regression were used to analyze clinical data, and independent predictors of HER-2 expression status of breast cancer were screened to construct clinical models. Cranio-caudal position and mediolateral oblique position MG were put into a twin DL network to obtain tumor ROI images, which were input to ResNet50 network to extract DL features of tumors. Feature fusion was performed through the fully connected layer to obtain a total of 2 048 DL features, then the biclassification results were output through the Softmax classifier, and DL scores (Deep-score) was obtained to construct DL model. Then a combined model was constructed by combining Deep-score and clinical independent predictors, and its nomogram was drawn. Calibration curves were drawn to assess the calibration degree of the combined model. Receiver operating characteristic (ROC) curve was used to evaluate the efficacy of each model for predicting HER-2 expression status of breast cancer. Decision curve analysis (DCA) was used to evaluate the clinical benefit of the models. Results Estrogen receptor status (OR=3.63) and Ki-67 expression level (OR=2.84) were both independent predictors of HER-2 expression status of breast cancer (both P＜0.05), and clinical models were constructed. The predicted results of combined model based on Deep-score and clinical independent predictors were in good agreement with the actual results. AUC of combined model (0.97) in training set was higher than that of clinical model and DL model (AUC=0.75, 0.96, Z=7.15, 2.03, both P＜0.05), while in validation set (0.88) was higher than that of clinical model (AUC=0.70, Z=5.76, P＜0.01) but not significant different with DL model (AUC=0.86, Z=1.50, P=0.33). The clinical net benefit of combined model in training set was higher than that of clinical model and DL model, while in validation set was higher than that of clinical model and comparable to that of DL model. Conclusion DL of MG combined with nomogram of clinical features could effectively predict HER-2 expression status of breast cancer. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
