摘要
目的基于超声影像组学特征、深度学习特征及临床资料构建联合模型,探讨其预测乳腺癌新辅助化疗(NAC)疗效的临床价值。方法选取接受NAC的乳腺癌患者264例,根据术后病理结果分为病理完全缓解(pCR)组73例和非pCR组191例。采用单因素分析和多因素Logistic回归分析筛选临床资料中预测乳腺癌NAC疗效的独立影响因素,构建临床模型。采用分层随机划分方法按照7∶3的比例将纳入患者分为训练集185例和测试集79例。基于训练集超声图像提取超声影像组学特征和深度学习特征,通过一致性评估、Z-score标准化、主成分分析降维及最小绝对收缩和选择算子(LASSO)回归筛选有效特征,构建超声影像组学特征联合深度学习特征的预测模型(Rad+DL模型)。采用多元Logistic回归方法将Rad+DL模型与临床模型联合,构建联合模型,并绘制列线图可视化。绘制受试者工作特征(ROC)曲线评估各模型预测乳腺癌NAC疗效的诊断效能;决策曲线分析各模型的临床适用性。结果pCR组人表皮生长因子受体-2(HER2)阳性和Ki-67高表达占比均高于非pCR组,雌激素受体、孕激素受体阳性占比均低于非pCR组,差异均有统计学意义(均P<0.05)。多因素Logistic回归分析显示,HER2为预测乳腺癌NAC疗效的独立影响因素,由此建立临床模型。经过Z-score标准化及主成分分析降维后,采用LASSO回归筛选出3个超声影像组学特征和6个深度学习特征,并通过随机森林算法构建Rad+DL模型。将Rad+DL模型预测值及HER2作为独立变量,构建联合模型。ROC曲线分析显示,训练集和测试集中临床模型、Rad+DL模型、联合模型预测乳腺癌NAC疗效的曲线下面积分别为0.719、0.886、0.914和0.736、0.805、0.861,以联合模型的曲线下面积最高。决策曲线分析显示,联合模型在训练集和测试集中均可获得较好的临床净收益。结论基于超声影像组学特征、深度学习特征及临床资料的联合模型在预测乳腺癌NAC疗效中具有较好的临床价值。
Objective To construct a combined model based on ultrasound radiomics,deep learning features,and clinical data,and to evaluate its clinical value in predicting neoadjuvant chemotherapy(NAC)treatment response in breast cancer.Methods A total of 264 breast cancer patients who underwent NAC were enrolled and divided into pathological complete response(pCR)group(73 cases)and non-pCR group(191 cases)based on postoperative pathological results.Univariate and multivariate Logistic regression analysis were used to screen the independent influencing factors for predicting NAC treatment response,and a clinical model was constructed.Patients were divided into training set(185 cases)and testing set(79 cases)by stratified random sampling at a ratio of 7∶3.Ultrasound radiomics features and deep learning features were extracted from training set images.After consistency evaluation,Z-score normalization,principal component analysis for dimensionality reduction,and least absolute shrinkage and selection operator(LASSO)regression,effective features were selected to construct a radiomics-deep learning(Rad+DL)model.Multiple Logistic regression was used to combine the Rad+DL model with the clinical model to establish a combined model,visualized by a nomogram.Receiver operating characteristic(ROC)curves were drawn to evaluate the diagnostic performance of each model in predicting NAC treatment response.Decision curve analysis was performed to assess clinical applicability.Results Compared with the non-pCR group,the proportion of positive human epidermal growth factor receptor-2(HER2)and Ki-67 high expression in pCR group were higher,and proportion of positive estrogen receptor and progesterone receptor were lower,with statistically significant differences(all P<0.05).Multivariate Logistic regression analysis showed that HER2 was an independent influencing factor for predicting NAC treatment response,and the clinical model was established accordingly.After Z-score normalization and principal component analysis of dimensionality reduction,LASSO regression selected 3 ultrasound radiomics features and 6 deep learning features,and the Rad+DL model was constructed by random forest algorithm.The combined model was established by Rad+DL model prediction values and HER2 status as independent variables.ROC curve analysis showed that the area under the curves of the clinical model,Rad+DL model,and combined model for predicting NAC treatment response were 0.719,0.886,0.914 in the training set and 0.736,0.805,0.861 in the testing set,respectively,with the combined model demonstrating the highest area under the curve.Decision curve analysis showed that the combined model provided better clinical net benefit in both sets.Conclusion The combined model based on ultrasound radiomics,deep learning features,and clinical data demonstrates good clinical value in predicting NAC treatment response in breast cancer.
作者
杨晨
贾晓宇
裴启华
胡可
黄忠江
晋建华
YANG Chen;JIA Xiaoyu;PEI Qihua;HU Ke;HUANG Zhongjiang;JIN Jianhua(School of Medical Imaging,Shanxi Medical University,Taiyuan 030001,China;Department of Ultrasound,Shanxi Provincial People’s Hospital,Taiyuan 030012,China;Department of Imaging,Shanxi Provincial Hospital of Traditional Chinese Medicine,Taiyuan 030012,China)
出处
《临床超声医学杂志》
2025年第10期820-826,共7页
Journal of Clinical Ultrasound in Medicine
基金
山西省卫生健康委科研课题(2024114)。
关键词
超声检查
影像组学
深度学习
乳腺癌
新辅助化疗
疗效预测
Ultrasonography
Radiomics
Deep learning
Breast cancer
Neoadjuvant chemotherapy
Treatment response prediction
