| 宋若晨,褚相乐,黄勇华,刘海燕,张海深.基于颅脑T1WI对比增强图像构建卷积神经网络模型鉴别肺癌与乳腺癌脑转移[J].中国医学影像技术,2023,39(7):982~986 |
| 基于颅脑T1WI对比增强图像构建卷积神经网络模型鉴别肺癌与乳腺癌脑转移 |
| Convolutional neural network model based on contrast-enhanced cranial T1WI for differentiating brain metastases from lung cancer or breast cancer |
| 投稿时间：2023-02-28 修订日期：2023-04-24 |
| DOI：10.13929/j.issn.1003-3289.2023.07.006 |
| 中文关键词: 脑肿瘤 肺肿瘤 乳腺肿瘤 磁共振成像 神经网络,计算机 影像组学 |
| 英文关键词:brain neoplasms lung neoplasms breast neoplasms magnetic resonance imaging neural networks, computer radiomics |
| 基金项目:河南省医学科技攻关计划项目(LHGJ20210936)。 |
|
| 摘要点击次数: 1908 |
| 全文下载次数: 489 |
| 中文摘要: |
| 目的 分析基于对比增强颅脑T1WI(T1CE)构建的卷积神经网络(CNN)模型鉴别肺癌与乳腺癌脑转移的效能。方法 回顾性分析97例经手术病理证实的乳腺癌脑转移(39例)与肺癌脑转移(58例)患者,于颅脑T1CE中手动勾画病灶ROI并提取其影像组学特征,采用单因素分析及最小绝对收缩和选择(LASSO)算法进行特征降维,筛选最优特征;分别构建传统机器学习支持向量机(SVM)、随机梯度下降(SGD)、K邻近(KNN)、决策树(DT)、随机森林(RF)、极端随机树(ET)、逻辑回归(LR)模型及深度学习CNN模型,将按7:3数据分为训练集和验证集,采用受试者工作特征(ROC)曲线评估8种模型鉴别验证集中肺癌与乳腺癌脑转移的效能。结果 共纳入202个脑转移癌,含乳腺癌、肺癌脑转移各101个。基于颅脑T1CE提取1 050个特征,经单因素分析及LASSO算法降维后得到5个最优特征;以之构建的SVM、SGD、KNN、DT、RF、ET、LR及CNN模型鉴别验证集肺癌与乳腺癌脑转移的曲线下面积(AUC)分别为0.88、0.83、0.87、0.74、0.84、0.86、0.88及0.90,其中CNN模型的AUC最高。结论 基于颅脑T1CE构建的CNN模型可有效鉴别肺癌与乳腺癌脑转移。 |
| 英文摘要: |
| Objective To observe the efficacy of convolutional neural network (CNN) model based on cranial T1W contrast-enhanced images (T1CE) for differentiating brain metastases from lung cancer or breast cancer. Methods Data of 97 patients with surgically pathologically confirmed brain metastases from breast cancer (39 patients) or lung cancer (58 patients) were retrospectively analyzed. Lesion ROI on cranial T1CE was manually outlined, and the radiomics features were extracted. Then univariate analysis and the least absolute shrinkage and selection operator (LASSO) algorithm were used to reduce the feature dimension and screen the optimal features. Finally traditional machine learning support vector machine (SVM), stochastic gradient descent (SGD), K-nearest neighbor (KNN), decision tree (DT), random forest (RF), extremely randomized tree (ET), logistic regression (LR) model and deep learning CNN model were constructed, respectively. The data were divided into training and validation sets at the ratio of 7:3. Receiver operating characteristic (ROC) curves were drawn to evaluate the efficacy of the above 8 models for differentiating brain metastases from lung cancer or breast cancer in validation set. Results A total of 202 brain metastatic cancer lesions (101 from breast cancer and 101 from lung cancer) were detected, and 1 050 features were extracted from T1CE images. Five optimal features were obtained with univariate analysis and LASSO algorithm. Eight radiomics models were constructed with SVM, SGD, KNN, DT, RF, ET, LR and CNN, respectively, and the area under the curve (AUC) of each model for differentiating brain metastases from lung cancer and breast cancer in validation set was 0.88, 0.83, 0.87, 0.74, 0.84, 0.86, 0.88 and 0.90, respectively, among which CNN had the highest AUC. Conclusion CNN model based on cranial T1CE images could effectively differentiate brain metastases from lung cancer or breast cancer. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
