基于机器学习的阿尔伯塔卒中项目早期CT评分在临床常规中检测早期梗死征象
研究在119例急性大脑中动脉缺血患者的CT扫描中验证e-ASPECTS软件自动评估ASPECTS的性能,并与三名神经放射科医师独立评估2380个脑区进行比较,以明确梗死核心为参考标准。软件在正常脑组织中可能提供有效数据,但在存在脑白质病变、梗死或非典型脑实质缺损的患者组中相关性不显著,最终仍需人工合理性核查。
要点
- e-ASPECTS软件在正常脑组织中可能提供有效数据,但在脑白质病变、梗死或非典型脑实质缺损患者中相关性不显著。
- 软件ASPECTS评估与明确梗死核心的相关系数为0.59,低于三位神经放射科医师的0.71、0.76和0.80。
- 软件敏感度83%、特异度57%、PPV 55%、NPV 82%、准确率67%,均低于三位医师的相应指标。
核心结论
- 对于ASPECTS评估,所检验的软件在正常脑组织情况下可能提供有效数据,可能增强神经放射科医师临床决策的工作,尤其对已有脑实质改变的患者组需要最终人工合理性核查。
研究设计与样本
- 研究纳入119例急性大脑中动脉缺血患者的CT扫描,患者集按正常脑、脑白质病变、梗死和非典型脑实质缺损(如多发性硬化等)区分;ASPECTS由e-ASPECTS软件包(Brainomix®,英国)自动提供,并由三名神经放射科医师独立检查2380个脑区,以最佳内科治疗(溶栓和/或取栓)后的明确梗死核心为参考标准进行评估者间比较。
诊断标准与关键数值
- 以明确梗死核心为参考标准,软件(A)与梗死核心的相关系数为0.59,医师(B)0.71、(C)0.76、(D)0.80;软件(A)敏感度83%、特异度57%、PPV 55%、NPV 82%、准确率67%,医师(B)分别为74%、76%、69%、83%、77%,医师(C)分别为80.8%、85.2%、76%、84%、80%,医师(D)分别为63%、90.7%、82%、79%、80%。
临床应用与阅片要点
- 作者认为该软件在正常脑组织情况下可能提供有效数据,可能增强神经放射科医师临床决策的工作,但尤其对已有脑实质改变的患者组需要最终人工合理性核查。
局限性
- 软件在存在脑白质病变、梗死和非典型脑实质缺损的患者组中相关性不显著。
- 作者指出尤其对已有脑实质改变的患者组需要最终人工合理性核查。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
PURPOSE: New software solutions emerged to support radiologists in image interpretation in acute ischemic stroke. This study aimed to validate the performance of computer-aided assessment of the Alberta Stroke Program Early CT score (ASPECTS) for detecting signs of early infarction.
METHODS: ASPECT scores were assessed in 119 CT scans of patients with acute middle cerebral artery ischemia. Patient collective was differentiated according to (I) normal brain, (II) leukoencephalopathic changes, (III) infarcts, and (IV) atypical parenchymal defects (multiple sclerosis, etc.). ASPECTS assessments were automatically provided by the software package e-ASPECTS (Brainomix®, UK) (A). Subsequently, three neuroradiologists (B), (C), and (D) examined independently 2380 brain regions. Interrater comparison was performed with the definite infarct core as reference standard after best medical care (thrombolysis and/or thrombectomy).
RESULTS: Interrater comparison revealed higher correlation coefficient of (B) 0.71, (C) 0.76, and of (D) 0.80 with definite infarct core compared to (A) 0.59 for ASPECTS assessment in the acute ischemic stroke setting. While (B), (C), and (D) showed a significant correlation for individual patient groups (I), (II), (III), and (IV), except for (D) (II), (A) was not significant in patient groups with pre-existing changes (II), (III), and (IV). The following sensitivities, specificities, PPV, NPV, and accuracies given in percent were achieved: (A) 83, 57, 55, 82, and 67; (B) 74, 76, 69, 83, and 77; (C) 80.8, 85.2, 76, 84, and 80; (D) 63, 90.7, 82, 79, and 80, respectively.
CONCLUSION: For ASPECTS assessment, the examined software may provide valid data in case of normal brain. It may enhance the work of neuroradiologists in clinical decision making. A final human check for plausibility is needed, particularly in patient groups with pre-existing cerebral changes.
引用
- Guberina N, Dietrich U, Radbruch A, et al. Detection of early infarction signs with machine learning-based diagnosis by means of the Alberta Stroke Program Early CT score (ASPECTS) in the clinical routine. Neuroradiology. 2018. PMID: 30066278. doi:10.1007/s00234-018-2066-5
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。