目的 开发一种名为上下文注意力网络(Context Attention Network,CA-Net)的新方法,基于深度学习利用超声(Ultrasound,US)图像来早期预测局部晚期乳腺癌患者的新辅助化疗(Neoadjuvant Chemotherapy,NAC)疗效.方法 选取101例女性为研究对象,在NAC前和NAC的6～8个周期后接受US监测.CA-Net中的嵌入式特征集成模块与多尺度的空洞卷积和通道注意力机制相结合,利用空间注意力模块特征图的空间关系生成空间注意力图.这种组合架构可以增加感受野,也能够更加准确地关注重要的目标区域,提高网络性能.通过消融实验和基本网络对比实验确定CA-Net的最优模型结构.将CA-Net预测病理完全缓解(Pathologic Complete Response,pCR)的性能与最新研究的3种方法进行比较,并通过受试者工作特征曲线下面积(Area Under the Curve,AUC)评估模型的预测性能.结果 CA-Net训练集的AUC值为0.950(95%CI:0.870～0.960),准确度为92.3%,敏感度为89.5%,特异性为94.7%;测试集的AUC值为0.920(95%CI:0.850～0.930),准确度为90.7%,敏感度为88.2%,特异性为94.4%,表明模型具有较高的性能和潜在的临床应用价值,且比其他现存模型性能有所提高.结论 基于超声的乳腺癌NAC疗效预测的深度学习模型预测效能较好,可作为预测NAC能否达到pCR的新方法.
Prediction of Efficacy of Neoadjuvant Chemotherapy for Breast Cancer Based on Ultrasound
Objective To develop a new method called context attention network(CA-Net)based on deep learning using ultrasound(US)images for early prediction of neoadjuvant chemotherapy(NAC)efficacy in patients with locally advanced breast cancer.Methods A total of 101 women were selected as research objects to receive US surveillance before NAC and 6 to 8 cycles after NAC.The embedded feature ensemble in CA-Net took a multi-scale null convolution and channel attention mechanism was combined and then spatial attention maps were generated using the spatial relations of the spatial attention feature maps.The sensory field was increased by this combined architecture and this combined architecture was also able to focus more accurately on important target regions,improving network performance.The optimal model structure of CA-Net was determined by ablation experiments and basic network comparison experiments.The diagnostic performance of CA-Net for predicting pathologic complete response(pCR)was compared with three recently studied methods,and the predictive performance of the models was assessed by the area under the curve(AUC)of the receiver operating characteristic.Results The AUC value of CA-Net training set was 0.950(95%CI:0.870-0.960),with an accuracy of 92.3%,a sensitivity of 89.5%,and a specificity of 94.7%.The AUC value of the test set was 0.920(95%CI:0.850-0.930),with an accuracy of 90.7%,a sensitivity of 88.2%,and a specificity of 94.4%,which indicated that the the model had high performance and potential clinical applications,and improved performance over other extant models.Conclusion The deep learning model for prediction of the efficacy of neoadjuvant chemotherapy for breast cancer based on ultrasound has a better prediction efficacy,and can be used as a method to determine whether NAC reaches pCR.
deep learningbreast cancerneoadjuvant chemotherapypathologic complete responseultrasound imagingattention mechanismablation experiment
国家科技期刊平台
NSTL
万方数据
