动态增强影像映射图的深度学习方法预测乳腺癌新辅助化疗疗效  

Deep Learning Model Using DCE-MRI Mapping for Prediction of Response to Neoadjuvant Chemotherapy in Breast Cancer

作　　者：刘鑫 范明 厉力华[1] Liu Xin;Fan Ming;Li Lihua(Institute of Biomedical Engineering and Instrumentation,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学生物医学工程与仪器研究所,杭州310018

出　　处：《中国生物医学工程学报》2023年第6期710-719,共10页Chinese Journal of Biomedical Engineering

基　　金：浙江省自然科学基金(LR23F010002);国家自然科学基金(62271178,U21A20521)。

摘　　要：新辅助化疗有助于提高乳腺癌患者的后期存活率,但疗效评估具有一定滞后性,准确的新辅助化疗疗效评估可以给医师更有效的临床建议,实施更优化的治疗方案。为了更好利用影像的空间信息和增强影像的时间序列信息,本研究提出一种基于动态增强影像映射模式图的乳腺癌新辅助化疗疗效预测方法。回顾性收集208例接受新辅助化疗前患者乳腺癌影像,依据Miller&Payne分级系统将数据标签为有反应组和无反应组,并随机划分为训练集126例,测试集82例。对影像做预处理及分割感兴趣区域后挑选肿瘤最大径及其邻近共7张切片构建映射模式图,结合增强时间序列构建原始切片影像、不同映射模式下多序列影像及融合两种映射模式多序列影像。使用深度学习网络对映射模式图做预测,绘制预测结果的ROC曲线并计算AUC、敏感性、特异性等评价指标对模型类型性能做评估。其中融合两种映射模式的多序列影像的预测模型效果最佳,AUC为0.832。实验表明,相较于使用原始切片影像,结合时间序列影像和切片间空间特征的方法对新辅助化疗疗效预测分类效果提升有效。Neoadjuvant chemotherapy is helpful to improve the later survival rate of breast cancer patients,but the efficacy evaluation has a certain lag.Accurate evaluation of the efficacy of neoadjuvant chemotherapy can give medical doctors more effective clinical suggestions and implement more optimized treatment plans.In order to make better use of the spatial information of the image and the time series information of the enhanced image,a dynamic enhanced image mapping pattern map was proposed to predict the efficacy of neoadjuvant chemotherapy for breast cancer.The images of 208 patients with breast cancer before neoadjuvant chemotherapy were retrospectively collected.According to the Miller&Payne grading system,the data were labeled as response group and non-response group,and randomly divided into training set(126 cases)and test set(82 cases).After image preprocessing and segmentation of the region of interest,the maximum diameter of the tumor and its adjacent 7 slices were selected to construct the mapping mode map.The original slice image,multi-sequence images under different mapping modes and multi-sequence images under fusion two mapping modes were constructed by combining the enhanced time series.The deep learning network was used to predict the mapping pattern graph,the ROC curve of the prediction results was drawn,and the evaluation indicators such as AUC,sensitivity,specificity were calculated to evaluate the performance of the model type.Among them,the prediction model of multi-sequence images fused with two mapping modes achieved the best result,with an AUC of 0.832.Experimental results showed that compared with the original slice images,the method combining longitudinal time series images and spatial features between slices effectively improved the classification effect of neoadjuvant chemotherapy response prediction.

关 键 词：乳腺癌 新辅助化疗疗效 深度学习 映射模式 

分 类 号：R318[医药卫生—生物医学工程]