Abstract：
OBJECTIVE: To develop thrombus radiomics models based on dual-energy CT (DECT) for predicting etiologic cause of stroke.
METHODS: We retrospectively enrolled patients with occlusion of the middle cerebral artery who underwent computed tomography (NCCT) and DECT angiography (DECTA). 70 keV virtual monoenergetic images (simulate conventional 120kVp CTA images) and iodine overlay maps (IOM) were reconstructed for analysis. Five logistic regression radiomics models for predicting cardioembolism (CE) were built based on the features extracted from NCCT, CTA and IOM images. From these, the best one was selected to integrate with clinical information for further construction of the combined model. The performance of the different models was evaluated and compared using ROC curve analysis, clinical decision curves (DCA), calibration curves and Delong test.
RESULTS: Among all the radiomic models, model NCCT+IOM performed the best, with AUC = 0.95 significantly higher than model NCCT, model CTA, model IOM and model NCCT+CTA in the training set (AUC = 0.88, 0.78, 0.90,0.87, respectively, P < 0.05), and AUC = 0.92 in the testing set, significantly higher than model CTA (AUC = 0.71, P < 0.05). Smoking and NIHSS score were independent predictors of CE (P < 0.05). The combined model performed similarly to the model NCCT+IOM, with no statistically significant difference in AUC either in the training or test sets. (0.96 vs. 0.95; 0.94 vs. 0.92, both P > 0.05).
CONCLUSIONS: Radiomics models constructed based on NCCT and IOM images can effectively determine the source of thrombus in stroke without relying on clinical information.
METHODS: We retrospectively enrolled patients with occlusion of the middle cerebral artery who underwent computed tomography (NCCT) and DECT angiography (DECTA). 70 keV virtual monoenergetic images (simulate conventional 120kVp CTA images) and iodine overlay maps (IOM) were reconstructed for analysis. Five logistic regression radiomics models for predicting cardioembolism (CE) were built based on the features extracted from NCCT, CTA and IOM images. From these, the best one was selected to integrate with clinical information for further construction of the combined model. The performance of the different models was evaluated and compared using ROC curve analysis, clinical decision curves (DCA), calibration curves and Delong test.
RESULTS: Among all the radiomic models, model NCCT+IOM performed the best, with AUC = 0.95 significantly higher than model NCCT, model CTA, model IOM and model NCCT+CTA in the training set (AUC = 0.88, 0.78, 0.90,0.87, respectively, P < 0.05), and AUC = 0.92 in the testing set, significantly higher than model CTA (AUC = 0.71, P < 0.05). Smoking and NIHSS score were independent predictors of CE (P < 0.05). The combined model performed similarly to the model NCCT+IOM, with no statistically significant difference in AUC either in the training or test sets. (0.96 vs. 0.95; 0.94 vs. 0.92, both P > 0.05).
CONCLUSIONS: Radiomics models constructed based on NCCT and IOM images can effectively determine the source of thrombus in stroke without relying on clinical information.
摘要：
目的：建立基于双能CT（DECT）的血栓影像组学模型，以预测卒中的病因。
方法：我们回顾性地纳入了大脑中动脉闭塞患者，这些患者接受了计算机断层扫描（NCCT）和DECT血管造影（DECTA）。重建70keV虚拟单能量图像（模拟常规120kVpCTA图像）和碘叠加图（IOM）进行分析。基于从NCCT中提取的特征，建立了五个预测心栓塞（CE）的逻辑回归影像组学模型，CTA和IOM图像。从这些,选择最佳模型与临床信息整合,进一步构建联合模型.使用ROC曲线分析评估和比较不同模型的性能，临床决策曲线（DCA），校正曲线和德隆试验。
结果：在所有的放射学模型中，NCCT+IOM模型表现最好，AUC=0.95显著高于模型NCCT，CTA模型，训练集中的模型IOM和模型NCCT+CTA(AUC分别为0.88、0.78、0.90、0.87，P<0.05),测试集中的AUC=0.92，CTA显著高于模型组(AUC=0.71,P<0.05)。吸烟和NIHSS评分是CE的独立预测因子(P<0.05)。组合模型执行类似于NCCT+IOM模型，在训练或测试集中，AUC没有统计学上的显著差异。（0.96vs.0.95;0.94vs.0.92,均P>0.05)。
结论：基于NCCT和IOM图像构建的Radiomics模型可以有效地确定卒中血栓的来源，而无需依赖临床信息。
方法：我们回顾性地纳入了大脑中动脉闭塞患者，这些患者接受了计算机断层扫描（NCCT）和DECT血管造影（DECTA）。重建70keV虚拟单能量图像（模拟常规120kVpCTA图像）和碘叠加图（IOM）进行分析。基于从NCCT中提取的特征，建立了五个预测心栓塞（CE）的逻辑回归影像组学模型，CTA和IOM图像。从这些,选择最佳模型与临床信息整合,进一步构建联合模型.使用ROC曲线分析评估和比较不同模型的性能，临床决策曲线（DCA），校正曲线和德隆试验。
结果：在所有的放射学模型中，NCCT+IOM模型表现最好，AUC=0.95显著高于模型NCCT，CTA模型，训练集中的模型IOM和模型NCCT+CTA(AUC分别为0.88、0.78、0.90、0.87，P<0.05),测试集中的AUC=0.92，CTA显著高于模型组(AUC=0.71,P<0.05)。吸烟和NIHSS评分是CE的独立预测因子(P<0.05)。组合模型执行类似于NCCT+IOM模型，在训练或测试集中，AUC没有统计学上的显著差异。（0.96vs.0.95;0.94vs.0.92,均P>0.05)。
结论：基于NCCT和IOM图像构建的Radiomics模型可以有效地确定卒中血栓的来源，而无需依赖临床信息。
