平扫CT深度学习模型对自发性脑出血早期血肿扩大的预测价值OA
Predictive Value of a Non-Contrast CT-Based Deep Learning Model for Early Hematoma Expansion in Spontaneous Intracerebral Hemorrhage
目的 探讨基于平扫CT(non-contrast CT,NCCT)的深度学习模型与CTA点征对自发性脑出血早期血肿扩大(hematoma expansion,HE)的预测价值. 方法 回顾性纳入2021年9月—2022年4月于首都医科大学附属北京天坛医院就诊,发病6 h内且完成基线NCCT及CTA检查的自发性脑出血患者,所有患者均于神经功能恶化时或发病24～48 h内复查NCCT.以发病48 h内复查NCCT血肿体积-基线血肿体积≥6 mL为HE的诊断依据,将纳入患者分为HE组和非HE组.对本研究团队前期构建的基于NCCT的深度学习模型在本队列中进行外部验证,并比较深度学习模型与CTA点征对自发性脑出血早期HE的预测效能. 结果 共纳入185例自发性脑出血患者,男性139例(75.1％),中位年龄为55(46～65)岁.其中HE组51例(27.6％),非HE组134例(72.4％).Hosmer-Lemeshow拟合优度检验显示,基于NCCT的深度学习模型的拟合优度良好(P=0.167).CTA点征预测自发性脑出血早期HE的敏感度为0.392,特异度为0.978,AUC为0.685(95％CI 0.589～0.781).深度学习模型预测自发性脑出血早期HE的敏感度为0.647,特异度为0.769,AUC为0.713(95％CI 0.625～0.800).DeLong检验结果显示,深度学习模型与CTA点征的AUC差异无统计学意义.与CTA点征相比,深度学习模型预测自发性脑出血早期HE的净重新分类指数为0.232(95％CI-0.088～0.553,P=0.155),综合判别改善指数为0.038(95％CI-0.091～0.168,P=0.561). 结论 基于NCCT的深度学习模型可有效预测自发性脑出血早期HE风险,其预测效能与CTA点征相当.在无法进行CTA检查或不具备开展CTA检查条件的临床场景中,该深度学习模型可作为自发性脑出血早期HE风险评估的替代工具.
Objective To investigate the predictive value of a non-contrast CT(NCCT)-based deep learning model and the CTA spot sign for early hematoma expansion(HE)in spontaneous intracerebral hemorrhage(ICH). Methods Patients with spontaneous intracerebral hemorrhage who presented to Beijing Tiantan Hospital,Capital Medical University,between September 2021 and April 2022 were retrospectively enrolled.All patients presented within 6 h of symptom onset and underwent baseline NCCT and CTA.Follow-up NCCT was performed upon neurological deterioration or within 24-48 h of symptom onset.Patients were divided into an HE group and a non-HE group,with HE was defined as an increase in hematoma volume of≥6 mL from baseline to follow-up NCCT within 48 h of onset.External validation of the NCCT-based deep learning model previously developed by our research team was performed in this cohort.The predictive performance of the deep learning model and the CTA spot sign for early HE in spontaneous intracerebral hemorrhage was compared. Results A total of 185 patients with spontaneous intracerebral hemorrhage were enrolled,including 139 men(75.1％),with a median age of 55(46-65)years.There were 51 patients(27.6％)in the HE group and 134 patients(72.4％)in the non-HE group.The Hosmer-Lemeshow goodness-of-fit test indicated good fit for the NCCT-based deep learning model(P=0.167).The CTA spot sign predicted early HE in spontaneous intracerebral hemorrhage with a sensitivity of 0.392,a specificity of 0.978,and an AUC of 0.685(95％CI 0.589-0.781).The deep learning model predicted early HE in spontaneous intracerebral hemorrhage with a sensitivity of 0.647,a specificity of 0.769,and an AUC of 0.713(95％CI 0.625-0.800).The DeLong test showed no statistically significant difference in AUC between the deep learning model and the CTA spot sign.Compared with the CTA spot sign,the deep learning model yielded a net reclassification improvement of 0.232(95％CI-0.088-0.553,P=0.155)and an integrated discrimination improvement of 0.038(95％CI-0.091-0.168,P=0.561)for predicting early HE in spontaneous intracerebral hemorrhage. Conclusions The NCCT-based deep learning model can effectively predict the risk of early HE in spontaneous intracerebral hemorrhage,with predictive performance comparable to that of the CTA spot sign.In clinical settings where CTA examination is unavailable or cannot be performed,this deep learning model may serve as an alternative tool for assessing the risk of early HE in spontaneous intracerebral hemorrhage.
单诗宇;丁少东;姬泽强;刘丹;郝允逸;王安心;李娜
北京 100070 首都医科大学附属北京天坛医院神经病学中心北京 100191 北京航空航天大学生物与医学工程学院北京 100070 首都医科大学附属北京天坛医院神经病学中心||北京 100070 神经系统疾病国家临床医学研究中心,首都医科大学附属北京天坛医院北京 100070 首都医科大学附属北京天坛医院神经病学中心北京 100070 神经系统疾病国家临床医学研究中心,首都医科大学附属北京天坛医院北京 100070 神经系统疾病国家临床医学研究中心,首都医科大学附属北京天坛医院北京 100070 首都医科大学附属北京天坛医院神经病学中心||北京 100070 神经系统疾病国家临床医学研究中心,首都医科大学附属北京天坛医院
医药卫生
自发性脑出血血肿扩大深度学习计算机体层扫描血管造影点征
Spontaneous intracerebral hemorrhageHematoma expansionDeep learningCTASpot sign
《中国卒中杂志》 2026 (9)
1134-1139,6
科技部"十四五"国家重点研发计划(2022YFC2504903)
评论