基于超声深度学习影像组学的乳腺癌新辅助化疗疗效预测  

Prediction of Efficacy of Neoadjuvant Chemotherapy for Breast Cancer Based on Ultrasound-Based Deep Learning Radiomics

作　　者：张恒 赵彤[5] 张赛 孙佳伟 李晓琴 倪昕晔 ZHANG Heng;ZHAO Tong;ZHANG Sai;SUN Jiawei;LI Xiaoqin;NI Xinye(Department of Radiotherapy,Changzhou Second People’s Hospital,Nanjing Medical University,Changzhou Jiangsu 213003,China;Jiangsu Province Engineering Research Center of Medical Physics,Changzhou Jiangsu 213003,China;Medical Physics Research Center,Nanjing Medical University,Changzhou Jiangsu 213003,China;Key Laboratory of Medical Physics in Changzhou,Changzhou Jiangsu 213003,China;Department of Ultrasound,Changzhou Second People’s Hospital,Nanjing Medical University,Changzhou Jiangsu 213003,China)

机构地区：[1]南京医科大学附属常州第二人民医院放疗科,江苏常州213003 [2]江苏省医学物理工程研究中心,江苏常州213003 [3]南京医科大学医学物理研究中心,江苏常州213003 [4]江苏省常州市医学物理重点实验室,江苏常州213003 [5]南京医科大学附属常州第二人民医院超声科,江苏常州213003

出　　处：《中国医疗设备》2024年第4期122-129,共8页China Medical Devices

基　　金：国家自然科学基金面上项目(62371243);江苏省重点研发计划社会发展项目(BE2022720);江苏省卫健委医学科研立项面上项目(M2020006);江苏省医学重点学科建设单位[肿瘤治疗学(放射治疗)](JSDW202237);江苏省自然基金面上项目(BK20231190);常州市社会发展项目(CE20235063)。

摘　　要：目的开发一种结合超声影像组学、深度学习和临床特征的综合模型,以预测乳腺癌新辅助化疗(Neoadjuvant Chemotherapy,NAC)后的病理完全缓解(Pathological Complete Response,pCR)。方法共纳入117例乳腺癌患者,按照7∶3的比例随机划分为训练集和验证集。采用Mann-Whitney U检验、随机森林递归消除算法和最小绝对收缩和选择算子进行特征筛选及影像组学/深度学习标签构建。对患者的临床参数进行单/多因素分析,以选择有效特征构建临床模型。然后利用逻辑回归算法将临床特征与影像组学、深度学习标签相结合,构建临床-影像组学-深度学习综合模型。从预测效果、校准能力和临床实用性方面评估模型性能。结果临床-影像组学-深度学习综合模型相比于单独的临床、影像组学和深度学习模型在训练集和验证集中均显示出最高的曲线下面积(训练集:0.949 vs.0.788 vs.0.815 vs.0.928;验证集:0.931 vs.0.643 vs.0.778 vs.0.901)。校准曲线和决策曲线证实综合模型具有良好的预测性能。结论与单一模型比较,综合模型对术前预测乳腺癌患者NAC后的pCR状态具有更高价值。Objective To develop a comprehensive model combining ultrasound radiomics,deep learning,and clinical features to predict the pathological complete response(pCR)after neoadjuvant chemotherapy(NAC)for breast cancer.Methods A total of 117 patients with breast cancer were included,and the training set and validation set were randomly divided according to a ratio of 7∶3.Mann-Whitney U test,random forest recursive feature elimination,and least absolute shrinkage and selection operators were used for feature screening and radiomics/deep learning signature construction.Single/multi-factor analysis of patients’clinical parameters were performed to select effective features to construct clinical models.Then Logistic regression algorithm was used to combine clinical features with radiomics and deep learning signatures to construct a clinical-radiomics-deep learning comprehensive model.Model performance was evaluated in terms of predictive efficacy,calibration ability,and clinical utility.Results The clinical-radiomics-deep learning comprehensive model showed the highest areas under curve compared to the separate clinical,radiomics,and deep learning models in both the training and validation sets(0.949 vs.0.788 vs.0.815 vs.0.928 for the training set;0.931 vs.0.643 vs.0.778 vs.0.901 for the validation set).The calibration curves and decision curves confirmed the good predictive performance of the comprehensive model.Conclusion Compared with the single model,the comprehensive model has a higher value in predicting the pCR status of breast cancer patients after NAC before surgery.

关 键 词：超声 影像组学 深度学习 乳腺癌 新辅助化疗 病理完全缓解 

分 类 号：R737.9[医药卫生—肿瘤]