摘要
目的:探讨商用人工智能(AI)软件(联影AI与数坤AI)在临床实际场景中测量自发性颅内出血破入脑室系统时的准确性及一致性。方法:回顾性纳入89例自发性颅内出血破入脑室的患者,采用联影uCT 960+、西门子SOMATOM Definition Flash及飞利浦Ingenuity Core 16 CT扫描仪获取图像。通过两款AI软件自动分割计算脑室内出血(IVH)及脑实质出血(IPH)量,由两位资深神经影像医师手动修正后取平均值作为金标准。采用Friedman检验、组内相关系数(ICC)及Bland-Altman分析评估测量差异与一致性,临床阈值设定为IVH±1 mL、IPH±6 mL。结果:联影AI与数坤AI对IVH的测量均值分别为(8.68±8.12)mL和(9.39±7.87)mL,经Wilcoxon符号秩检验,两者差异无统计学意义(Z=-0.92,P=0.37)。在IPH测量中,联影AI[(30.6±27.5)mL]略低于数坤AI[(30.7±28.4)mL],两者差异亦无统计学意义(P=0.89)。与人工修正值[IVH:(9.03±8.85)mL;IPH:(32.7±27.8)mL]相比,两者虽数值相近,但Bland-Altman分析结果显示IVH测量的95%一致性界限(LoA)超出临床阈值(如联影AI与人工修正的LoA为-6.25~5.54 mL)。IPH测量中,两款AI软件的LoA亦超出阈值。ICC分析结果显示,IVH测量中联影AI与人工修正的一致性(ICC=0.94)高于数坤AI(ICC=0.82),而在IPH测量中两者与人工修正的ICC均≥0.97。结论:商用AI在颅内出血破入脑室的测量中与人工修正存在趋势一致性,但随机误差超出临床可接受阈值。临床应用需结合人工修订以提升准确性。
Objective:To investigate the accuracy and consistency of commercial artificial intelligence(AI)software(United Imaging AI and Shukun AI)in measuring spontaneous intracranial hemorrhage(ICH)with ventricular extension in real-world clinical scenarios.Methods:A total of 89 patients with spontaneous intracranial hemorrhage breaking into ventricles were retrospectively included.Images were acquired using United Imaging uCT 960+,Siemens SOMATOM Definition Flash,and Philips Ingenuity Core 16 CT scanners.Two AI software systems automatically segmented and calculated the volumes of intraventricular hemorrhage(IVH)and intraparenchymal hemorrhage(IPH).The gold standard was established as the mean value of manual corrections by two senior neuroradiologists.Friedman test,intraclass correlation coefficient(ICC),and Bland-Altman analysis were used to evaluate measurement differences and consistency,with clinical thresholds set as±1mL for IVH and±6mL for IPH.Results:The mean IVH measurements by United Imaging AI and Shukun AI were(8.68±8.12)mL and(9.39±7.87)mL,respectively,with no significant difference(Z=-0.92,P=0.37)by Wilcoxon signed-rank test.For IPH measurements,United Imaging AI(30.6±27.5)mL showed a slightly lower value than Shukun AI(30.7±28.4)mL,without statistical significance(P=0.89).Compared with manual corrections[IVH:(9.03±8.85)mL;IPH:(32.7±27.8)mL],both AI systems showed similar values,but Bland-Altman analysis revealed that the 95%limits of agreement(LoA)for IVH measurement exceeded clinical thresholds(e.g.,LoA of United Imaging AI vs.manual correction:-6.25 to 5.54 mL).The LoA of both AI systems for IPH measurement also exceeded the thresholds.ICC analysis showed that the consistency of United Imaging AI with manual correction in IVH measurement(ICC=0.94)was higher than that of Shukun AI(ICC=0.82),while both showed ICC≥0.97 with manual correction in IPH measurement.Conclusion:Commercial AI demonstrates trend consistency with manual correction in measuring ICH with ventricular extension,but random errors exceed clinical acceptable thresholds.Clinical application needs to be combined with manual revision to improve accuracy.
作者
邓喜青
殷灿
申跃明
罗正卯
DENG Xi-qing;YIN Can;SHEN Yue-ming(Department of Radiology,Zhuzhou Central Hospital,Hunan 412000,China)
出处
《放射学实践》
北大核心
2026年第1期12-16,共5页
Radiologic Practice
关键词
颅内出血
脑室内出血
脑实质出血
人工智能
体层摄影术
X线计算机
测量一致性
Intracranial hemorrhage
Intraventricular hemorrhage
Intraparenchymal hemorrhage
Artificial intelligence
Tomography,X-ray computed
Measurement consistency
