摘要
目的利用宫颈MRI图像提取影像组学特征,建立随机森林模型识别ⅠA期宫颈癌与高级别鳞状上皮内病变(high-grade squamous intraepithelial lesion,HSIL)。材料与方法回顾性分析经手术病理证实的43例ⅠA期宫颈癌患者与51例HSIL患者,按照4∶1的比例设置训练集(ⅠA=34,HSIL=41)与测试集(ⅠA=9,HSIL=10)。收集其术前MRI图像,经预处理后上传至影像组学云平台,分别在OSag-T2WI、OAx-T1WI以及OAx-T2FS上逐层手动勾画宫颈,获得宫颈三维容积感兴趣区(volume of interest,VOI),提取组学特征。采用方差阈值分析法(Variance Threshold)、单变量特征选择法(SelectKBest)以及最小绝对值收缩和选择法(least absolute shrinkage and selection operator,LASSO)进行数据降维、特征选择。采用随机森林模型进行机器学习,绘制ROC曲线,分析不同序列组学模型的诊断效能。结果基于OSag-T2WI、OAx-T1WI、OAx-T2FS以及OSag-T2WI联合OAx-T2FS分别得到8个、10个、6个以及9个有效特征。以OSag-T2WI联合OAx-T2FS的组学特征值建立的随机森林模型诊断效能最高,AUC为0.89[95%CI(0.741.00)];基于OAx-T1WI的模型诊断效能最低,AUC为0.51[95%CI(0.230.78)]。结论基于MRI的影像组学随机森林模型可以较好地在没有明确病灶的情况下区分ⅠA期宫颈癌与HSIL,对于术前减少侵入性检查与指导术式有着重大的意义。
Objective:Using cervical MRI images to extract radiomic features and establish machine learning models to identify stageⅠA cervical cancer and high-grade squamous intraepithelial lesions(HSIL).Materials and Methods:A retrospective analysis of 43 patients with stageⅠA cervical cancer and 51 patients with HSIL confirmed by surgery and pathology was performed,and 20%(n=19)of the samples were selected as the test set.The preoperative MRI images were collected to upload to the radiomics cloud platform.sagittal T2WI,axial T2WI and T2FS were manually segmented layer by layer to obtain the three-dimensional volume of interest(VOI)of the cervix,and extract the omics features.Variance threshold analysis method,univariate feature selection method(SelectKBest)and least absolute shrinkage and selection operator(LASSO)were used for data dimensionality reduction and feature selection.Random forest model was used for machine learning,ROC curve was drawn to analyze the diagnostic efficiency of different sequence radiomics models.Results:Based on OSag-T2WI,OAx-T1WI,OAx-T2FS and OSag-T2WI combined with OAx-T2FS,8,10,6 and 9 effective features were obtained.The random forest model based on OSag-T2WI combined with OAx-T2FS has the highest diagnostic performance,with AUC of 0.89[95%CI(0.74—1.00)];the model based on OAx-T1WI has the lowest diagnostic performance,with AUC of 0.51[95%CI(0.23—0.78)].Conclusions:The random forest model of radiomics based on MRI can better distinguish stageⅠA cervical cancer from HSIL without clear focus,which is of great significance for reducing invasive examination and guiding surgical procedures before surgery.
作者
樊知昌
夏雨薇
甄俊平
周宇堃
靳波
边文瑾
杨洁
FAN Zhichang;XIA Yuwei;ZHEN Junping;ZHOU Yukun;JIN Bo;BIAN Wenjin;YANG Jie(Shanxi Medical University,Taiyuan 030001,China;Huiying Medical Technology,Beijing 100192,China;Department of Imaging,the Second Hospital of Shanxi Medical University,Taiyuan 030001,China;Department of Imaging,the First Hospital of Shanxi Medical University,Taiyuan 030001,China;Department of Imaging,the Shanxi Children's Hospital,Taiyuan 030001,China)
出处
《磁共振成像》
CAS
CSCD
北大核心
2021年第6期38-43,共6页
Chinese Journal of Magnetic Resonance Imaging
基金
山西省回国留学人员科研资助项目(编号:2014-077)。
关键词
宫颈癌
高级别鳞状上皮内病变
磁共振成像
影像组学
机器学习
cervical cancer
high-grade squamous intraepithelial lesion
magnetic resonance imaging
radiomics
machine-learning
