摘要
目的基于头颅CT放射组学探究高血压性脑出血早期血肿扩大预测的最佳机器学习方法。方法回顾性分析130例脑出血患者CT图像,提取头颅CT平扫纹理特征,通过选定特征训练分类器,用六种经典的机器学习方法进行交叉验证,评估预测脑出血血肿扩大的稳定性和性能。结果支持向量机(SVM-Radial)的预测性能(AUC为0.714,准确性为0.723),广义线性模型(GLM)的预测性能(AUC为0.643,准确性为0.587),随机森林(RF)的预测性能(AUC为0.686,准确性为0.680),k近邻(kNN)的预测性能(AUC为0.657,准确性为0.639),梯度提升树算法(GBM)的预测性能(AUC为0.718,准确性为0.670),神经网络(NNet)的预测性能(AUC为0.659,准确性为0.680),其中SVM-Radial表现出较高的预测性能,GLM表现出较低的预测性能。结论基于头颅CT放射组学方法预测高血压性脑出血早期血肿扩大的6种机器学习方法中,SVM-Radial预测性能最好,具有潜在的临床应用价值。
Objective To study the best machine learning method for early prediction of hematoma expansion in hypertensive intracerebral hemorrhage based on head CT plain scan.Methods The CT images of 130 patients with cerebral hemorrhage were retrospectively analyzed,and the texture features of the head CT plain scan were extracted.The classifier was trained by selecting the features,and the six classic machine learning methods were cross-validated to evaluate the stability and performanceof predicting cerebral hemorrhage hematoma expansion.Results The prediction performance of support vector machine(SVM-Radial)(AUC 0.714±0.144,accuracy 0.723±0.109),generalized linear model(GLM)prediction performance(AUC 0.643±0.125,accuracy 0.587±0.136),random forest(RF)prediction performance(AUC 0.686±0.128,accuracy 0.680±0.130),k-nearest neighbor(kNN)prediction performance(AUC 0.657±7C 15,accuracy 0.639±39 performance 19),gradient boosting tree algorithm(GBM)Prediction performance(AUC 0.718±0.141,accuracy 0.670±0.126),neural network(NNet)prediction performance(AUC 0.659±0.162,accuracy 0.680±0.130),in which support vector machines showed high prediction performance,generalized linear model showed low predictive performance.Conclusion Among the six machine learning methods based on cranial CT radiomics to predict early hematoma expansion in hypertensive intracerebral hemorrhage,support vector machine(SVM-Radial)has the best predictive performance and has potential clinical application value.
作者
丁川
李小虎
王俊
李红文
王玉萍
余长亮
葛亚琼
王海宝
刘斌
Ding Chuan;Li Xiaohu;Wang Jun;Li Hongwen;Wang Yuping;Yu Changliang;Ge Yaqiong;Wang Haibao;Liu Bin(Dept of Radiology,The First Affiliated Hospital of Anhui Medical University,Hefei 230022;GE Healthcare(China),Shanghai 210000)
出处
《安徽医科大学学报》
CAS
北大核心
2022年第1期161-164,共4页
Acta Universitatis Medicinalis Anhui
基金
安徽省科技攻关项目(编号:201904a07020060)。