基于临床-影像组学特征的机器学习模型预测颅内动脉瘤的破裂风险

Prediction of rupture risk of intracranial aneurysms using machine learning models based on clinico-radiomic features

目的探讨基于临床-影像组学特征的机器学习模型预测颅内动脉瘤破裂风险的效能。方法回顾性分析2019年1月至2022年12月收治的153例颅内动脉瘤的临床资料。利用单因素分析筛选出影响动脉瘤破裂的临床特征;收集DICOM格式影像数据,利用3D Slicer软件对载瘤动脉进行三维重建与分割,利用radiomics插件提取动脉瘤的形态学特征,并利用LASSO算法筛选影响动脉瘤破裂的重要形态学特征;构建基于临床和形态学特征的机器学习模型,计算各模型预测颅内动脉瘤破裂风险的曲线下面积、准确度、精准度、灵敏度、特异度。结果153例中,破裂动脉瘤43例(破裂组),未破裂动脉瘤110例(未破裂组)。与未破裂组相比,破裂组高血压病人占比较高(P<0.05)。与未破裂组相比较,破裂组动脉瘤形态学特征的伸长率、球形度明显缩小(P<0.05),而动脉瘤最小轴径、最大轴径、冠状面最大直径、最大三维直径、网格体积、表面积、表面积体积比和体素体积明显增大(P<0.05)。纳入分析的14个形态学特征经过LASSO回归与十折交叉验证选择最优lambda值0.023,最终筛选出6个最优形态学特征,根据这六大形态学特征构建SVM、KNN、LR机器学习模型,其曲线下面积(AUC)分别为0.73(95%CI 0.47~0.98)、0.80(95%CI 0.72~0.87)、0.75(95%CI 0.61~0.89)。基于高血压这一临床特征和形态学特征构建的SVM1、KNN1和LR1模型的AUC分别为0.79(95%CI 0.67~0.93)、0.85(95%CI 0.79~0.92)、0.83(95%CI 0.72~0.95)。结论本研究基于影像组学技术自动提取颅内动脉瘤的形态学特征,并构建了六种机器学习模型,能准确识别动脉瘤的状态,从而可以尽早对具有高破裂风险的颅内动脉瘤进行干预,具有重要的临床指导意义。

Objective To explore the effectiveness of machine learning models based on clinico-radiomics features in predicting the risk of intracranial aneurysm rupture.Methods The clinical data of 153 patients with intracranial aneurysms admitted to our hospital from January 2019 to December 2022 were retrospectively analyzed.Univariate analysis was used to screen the clinical features affecting aneurysm rupture.DICOM format image data were collected,and the 3D Slicer software was used to reconstruct and segment the parent artery.The radiomics plug-in was used to extract the morphological features of the aneurysm,and the LASSO algorithm was used to screen the important morphological features affecting aneurysm rupture.Machine learning models based on clinical and morphological features were constructed,and the area under the curve(AUC),accuracy,precision,sensitivity,and specificity of each model in predicting the risk of intracranial aneurysm rupture were calculated.Results Of these 153 patients,43 patients had ruptured intracranial aneurysms(ruptured group)and 110 had unruptured intracranial aneurysms(unruptured group).Compared with the unruptured group,the proportion of patients with hypertension was significantly higher(P<0.05),the elongation and sphericity of the aneurysm morphological features were significantly reduced(P<0.05),while the minimum axial diameter,maximum axial diameter,maximum coronal diameter,maximum three-dimensional diameter,grid volume,surface area,surface area volume ratio,and voxel volume of the aneurysm were significantly increased in the ruptured group(P<0.05).The AUC of SVM,KNN and LR models based on the six morphological features were 0.73(95%CI 0.47~0.98),0.80(95%CI 0.72~0.87),0.75(95%CI 0.61~0.89),respectively.The AUC of SVM1,KNN1 and LR1 models based on the clinical feature of hypertension and morphological features were 0.79(95%CI 0.67~0.93),0.85(95%CI 0.79~0.92),0.83(95%CI 0.72~0.95),respectively.Conclusions The morphological features of intracranial aneurysms were automatically extracted based on radiomics technology,and six machine learning models were constructed,which can accurately identify the status of aneurysms,so as to intervene as early as possible for intracranial aneurysms with high risk of rupture,which has important clinical significance.