Li Xiao-Yang, Xiong Jun-Feng, Jia Tian-Ying, Shen Tian-Le, Hou Run-Ping, Zhao Jun, Fu Xiao-Long
Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University, Shanghai 200000, China.
School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200000, China.
J Thorac Dis. 2018 Dec;10(12):6624-6635. doi: 10.21037/jtd.2018.11.03.
We aim to analyze the ability to detect epithelial growth factor receptor () mutations on chest CT images of patients with lung adenocarcinoma using radiomics and/or multi-level residual convolutionary neural networks (MCNNs).
We retrospectively collected 1,010 consecutive patients in Shanghai Chest Hospital from 2013 to 2017, among which 510 patients were -mutated and 500 patients were wild-type. The patients were randomly divided into a training set (810 patients) and a validation set (200 patients) according to a balanced distribution of clinical features. The CT images and the corresponding EGFR status measured by Amplification Refractory Mutation System (ARMS) method of the patients in the training set were utilized to construct both a radiomics-based model (M) and MCNNs-based model (M). The M and M were combined to build the Model (M). Clinical data of gender and smoking history constructed the clinical features-based model (M). M was then added into M, M, and M to establish the Model (M), the Model (M) and the Model (M). All the seven models were tested in the validation set to ascertain whether they were competent to detect mutations. The detection efficiency of each model was also compared in terms of area under the curve (AUC), sensitivity and specificity.
The AUC of the M, M and M to predict mutations was 0.740, 0.810 and 0.811 respectively. The performance of M was better than that of M (P=0.0225). The addition of clinical features did not improve the AUC of the M (P=0.623), the M (P=0.114) and the M (P=0.058). The M demonstrated the highest AUC value of 0.834. The M did not demonstrate any inferiority when compared with the M (P=0.742) and the M (P=0.056).
Both of the M and the M could predict mutations on CT images of patients with lung adenocarcinoma. The M outperformed the M in the detection of mutations. The combination of these two models, even added with clinical features, is not significantly more efficient than M alone.
我们旨在分析使用放射组学和/或多级残差卷积神经网络（MCNN）在肺腺癌患者胸部CT图像上检测表皮生长因子受体（EGFR）突变的能力。
我们回顾性收集了2013年至2017年在上海胸科医院连续就诊的1010例患者，其中510例患者为EGFR突变型，500例患者为野生型。根据临床特征的均衡分布，将患者随机分为训练集（810例患者）和验证集（200例患者）。利用训练集中患者的CT图像以及通过扩增阻滞突变系统（ARMS）方法测得的相应EGFR状态，构建基于放射组学的模型（M1）和基于MCNN的模型（M2）。将M1和M2合并构建模型（M3）。性别和吸烟史的临床数据构建基于临床特征的模型（M4）。然后将M4分别加入M1、M2和M3中，建立模型（M5）、模型（M6）和模型（M7）。在验证集中对所有七个模型进行测试，以确定它们是否能够检测EGFR突变。还根据曲线下面积（AUC）、敏感性和特异性比较了每个模型的检测效率。
M1、M2和M3预测EGFR突变的AUC分别为0.740、0.810和0.811。M2的性能优于M1（P = 0.0225）。添加临床特征并未提高M1（P = 0.623）、M2（P = 0.114）和M3（P = 0.058）的AUC。M6表现出最高的AUC值，为0.834。与M2（P = 0.742）和M3（P = 0.056）相比，M6没有表现出任何劣势。
M1和M2都可以在肺腺癌患者的CT图像上预测EGFR突变。在检测EGFR突变方面，M2优于M1。这两种模型的组合，即使添加了临床特征，也并不比单独使用M6更有效。
