人工智能对外伤性脑出血的诊断及定量临床应用研究  

Study on Diagnosis and Quantitative Clinical Application of Artificial Intelligence in Traumatic Intracerebral Hemorrhage

作　　者：窦末未 刘远健 陈凯 成官迅 DOU Mo-wei;LIU Yuan-jian;CHEN Kai;CHENG Guan-xun(Medical College of Shantou University,Shantou 515000,Guangdong,China;Department of Medical Imaging,Shenzhen Samii Medical Center/the Fourth People's Hospital of Shenzhen,Shenzhen 518118,Guangdong,China;Medical Image Artificial Intelligence,Hunan key Laboratory/Xiangnan University,Chenzhou 423000,Hunan,China;Department of Medical Imaging,Shenzhen Hospital,Peking University,Shenzhen 518000,Guangdong,China)

机构地区：[1]汕头大学医学院,广东汕头515000 [2]深圳市萨米医疗中心/深圳市第四人民医院医学影像科,广东深圳518118 [3]医学影像人工智能湖南省重点实验室/湘南学院,湖南郴州423000 [4]北京大学深圳医院医学影像科,广东深圳518000

出　　处：《医学信息》2023年第15期57-63,共7页Journal of Medical Information

基　　金：深圳市萨米医疗中心院内课题基金(编号:SSMC-2023-A6);医学影像人工智能湖南省重点实验室开放课题(编号:YXZN2022001)。

摘　　要：目的探讨人工智能在外伤性脑出血(TICH)定量评估方面的准确性、一致性及其临床价值。方法对2021年1月-2022年9月深圳市萨米医疗中心/深圳市第四人民医院就诊的87例TICH患者和90例外伤性非脑出血患者的X线计算机断层扫描(CT)影像进行回顾性研究。以2名工作10年以上的神经放射科专家利用3Dslicer软件手动测量平均值为“金标准”,uAI算法、1名未参加AI算法的神经放射科医师(R1)用3Dslicer软件手动分割与R1在uAI软件上(uAI+R1)手动分割,并进行一致性检验。结果177例患者中uAI检测TICH、蛛网膜下腔出血(SAH)、脑实质内出血(IPH)、硬膜外血肿(EDH)、硬膜下血肿(SDH)的灵敏度、特异度分别为79.31%、92.20%;74.00%、75.28%;50.00%、96.55%;66.67%、98.72%;75.00%、95.04%;R1检测上述指标分别为97.70%、100.00%;96.00%、100.00%;90.63%、100.00%;90.48%、100.00%;69.44%、100.00%;uAI+R1检测上述指标分别为98.85%、100.00%;98.00%、95.28%;90.63%、100.00%;90.48%、94.87%;83.33%、95.74%。uAI算法与金标准及R1得到的TICH及各亚型的出血体积量化结果显示TICH组、SAH组及EDH组中uAI算法与金标准、uAI算法与R1、金标准与R1比较,差异无统计学意义(P>0.05);IPH组中uAI算法与金标准、uAI算法与R1比较,差异有统计学意义(P<0.05),而金标准与R1比较,差异无统计学意义(P>0.05);SDH组中uAI算法与金标准比较,差异无统计学意义(P>0.05),而uAI算法与R1、金标准与R1比较,差异有统计学意义(P<0.05)。ICC分析显示,金标准与uAI算法及R1具有很强的一致性,且金标准与uAI算法具有很强的正相关性(P<0.05)。结论AI对TICH及各亚型检测的灵敏度低,但可快速准确量化血肿体积,提供了AI在TICH中的实践经验,诊断结果可供参考,但放射诊断医生仍然需要注意观察并在AI辅助下量化TICH体积。Objective To explore the accuracy,consistency and clinical value of artificial intelligence in quantitative assessment of traumatic intracerebral hemorrhage(TICH).Methods A retrospective study was conducted on the X-ray computed tomography(CT)images of 87 patients with TICH and 90 patients with traumatic non-cerebral hemorrhage who visited Shenzhen Sami Medical Center/The Fourth People's Hospital of Shenzhen from January 2021 to September 2022.Based on the"gold standard"for manual measurement of average by 2 neuroradiologists who had worked for more than 10 years using 3Dslicer software,the uAI algorithm and one neuroradiologist(R1)who did not participate in the AI algorithm were manually segmented with 3Dslicer software and R1 was manually segmented on uAI software(uAI+R1),and consistency was checked.Results The sensitivity and specificity of uAI for TICH,subarachnoid hemorrhage(SAH),intraparenchymal hemorrhage(IPH),epidural hematoma(EDH),subdural hematoma(SDH)were 79.31%,92.20%;74.00%,90.00%;50.00%,93.75%;66.67%,100.00%;75.00%,100.00%;R1 detected that the above indicators were 97.70%,100.00%;96.00%,100.00%;90.63%,100.00%;90.48%,100.00%;72.22%,100.00%;The above indicators detected by uAI+R1 were 98.85%,100.00%;98.00%,100.00%;90.63%,100.00%;90.48%,100.00%;83.33%,97.22%.The quantitative results of TICH and various subtypes obtained by uAI algorithm and gold standard and R1 showed that there was no significant difference between uAI algorithm and gold standard,uAI algorithm and R1,gold standard and R1 in TICH group,SAH group and EDH group(P>0.05).In IPH group,there were significant differences between uAI algorithm and gold standard,uAI algorithm and R1(P<0.05),but there was no significant difference between gold standard and R1(P>0.05).In SDH group,there was no significant difference between uAI algorithm and gold standard(P>0.05),but there was significant difference between uAI algorithm and R1,gold standard and R1(P<0.05).The correlation coefficients of ICC analysis showed that the gold standard had strong consi

关 键 词：人工智能 外伤性脑出血 X线计算机断层扫描 

分 类 号：R651.15[医药卫生—外科学]