| 杨雨奇,罗嘉宁,杨永祥,邹东波,魏坤,夏永利,陈敏,马原.基于CT影像组学机器学习模型预测急性期创伤性脑损伤严重程度[J].中国医学影像技术,2024,40(7):992~996 |
| 基于CT影像组学机器学习模型预测急性期创伤性脑损伤严重程度 |
| Machine learning model based on CT radiomics for predicting severity of acute phase traumatic brain injury |
| 投稿时间：2023-09-20 修订日期：2024-04-03 |
| DOI：10.13929/j.issn.1003-3289.2024.07.008 |
| 中文关键词: 脑损伤 机器学习 影像组学 |
| 英文关键词:brain injuries machine learning radiomics |
| 基金项目:西部战区总医院院管课题(2021-XZYG-A13)。 |
| 作者 | 单位 | E-mail | | 杨雨奇 | 西南医科大学附属医院神经外科, 四川 泸州 646000
江油市人民医院重症医学科, 四川 江油 621700 | | | 罗嘉宁 | 中国人民解放军西部战区总医院神经外科, 四川 成都 610083 | | | 杨永祥 | 中国人民解放军西部战区总医院神经外科, 四川 成都 610083 | | | 邹东波 | 中国人民解放军西部战区总医院神经外科, 四川 成都 610083 | | | 魏坤 | 西南医科大学附属医院神经外科, 四川 泸州 646000 | | | 夏永利 | 西南医科大学附属医院神经外科, 四川 泸州 646000 | | | 陈敏 | 中国人民解放军西部战区总医院医疗保障中心信息科, 四川 成都 610083 | | | 马原 | 西南医科大学附属医院神经外科, 四川 泸州 646000
中国人民解放军西部战区总医院神经外科, 四川 成都 610083 | tianfu_47@163.com |
|
| 摘要点击次数: 611 |
| 全文下载次数: 292 |
| 中文摘要: |
| 目的 观察基于CT平扫(NCCT)影像组学特征建立的机器学习(ML)模型预测急性期创伤性脑损伤(TBI)严重程度的价值。方法 回顾性收集600例TBI为观察组,以另外65例TBI为外部验证集;另前瞻性纳入50例TBI为前瞻性验证集。根据出院时格拉斯哥预后评分(GOS)将观察组患者分为高危亚组(n=240)与低危亚组(n=360)。由医师A、B以相同标准分别评估观察组患者,基于首诊临床及NCCT资料以逻辑回归(LR)法建立人工模型,预测急性期TBI严重程度。按7 ∶ 3比例将观察组分为训练集(n=420,含168例高危、252例低危)与测试集(n=180,含72例高危、108例低危),基于训练集NCCT提取及筛选影像组学特征,采用LR、支持向量机(SVM)、随机森林(RF)、K近邻(KNN)4种ML法构建预测模型,分别于测试集、外部验证集(含34例高危、31例低危TBI)及前瞻性验证集(含21例高危、29例低危TBI)进行验证。结果 医师A、B判断观察组急性期TBI严重程度的曲线下面积(AUC)分别为0.606及0.771,人工模型的AUC为0.824。基于训练集NCCT筛选出的6个最佳影像组学特征构建的LR、SVM、RF和KNN ML 模型及人工模型在测试集的AUC分别为0.983、0.971、0.970、0.984及0.708,在外部验证集分别为0.879、0.881、0.984、0.863及0.733,而在前瞻性验证集分别为0.984、0.873、0.982、0.897及0.704。结论 基于CT影像组学建立的ML模型能有效预测急性期TBI严重程度。 |
| 英文摘要: |
| Objective To explore the value of machine learning (ML) models based on non-contrast CT (NCCT) radiomics features for predicting the severity of acute phase traumatic brain injury (TBI). Methods Totally 600 TBI patients were retrospectively collected as observation group, other 65 TBI patients were taken as external validation set, while 50 TBI patients were prospectively enrolled as prospective validation set. Patients in observation group were divided into high-risk subgroup (n=240) and low-risk subgroup (n=360) according to Glasgow outcome scale (GOS) at discharge. The severity of acute phase TBI in observation group was assessed by doctor A and B with the same criteria, then an artificial model was established based on clinical and NCCT data at the time of first diagnosis using logistic regression (LR) method for predicting the severity of acute phase TBI. Patients in observation group were divided into training set (n=420, including 168 in high-risk subgroup and 252 in low-risk subgroup) and test set (n=180, including 72 in high-risk subgroup and 108 in low-risk subgroup) at the ratio of 7∶3. Based on NCCT of training set, radiomics features were extracted and selected, and LR, support vector machine (SVM), random forest (RF) and K-nearest neighbor (KNN) were used to establish 4 ML models. The efficacies of the above models were validated in test set, external validation set (including 34 cases of high-risk and 31 cases of low-risk TBI) and prospective validation set (including 21 cases of high-risk and 29 cases of low-risk TBI), respectively. Results The area under the curve (AUC) of doctor A and B for evaluating the severity of acute phase TBI in observation group was 0.606 and 0.771, respectively, of artificial model was 0.824. Based on NCCT in training set, 6 optimal radiomics features were selected to construct LR, SVM, RF and KNN ML models, with AUC of 0.983, 0.971, 0.970 and 0.984 in test set, respectively, while the AUC of artificial model was 0.708. The AUC of LR, SVM, RF, KNN ML models and artificial model in external validation set was 0.879, 0.881, 0.984, 0.863 and 0.733, while in prospective validation set was 0.984, 0.873, 0.982, 0.897 and 0.704, respectively. Conclusion ML models based on CT radiomics could effectively predict the severity of acute phase TBI. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
