（阅读笔记）利用CTA自动确定急性缺血核心

• Background
• Methods
• • Study Pupulation
  • Study Design
  • DeepSymNet Description
  • Statistical Analysis
• Results
• 其他

by HPC_ZY

期刊文献《Machine Learning–Enabled Automated Determination of Acute Ischemic Core From Computed Tomography Angiography [1] 》阅读笔记，它主要讲了如何通过CTA推测梗死体积，以及效果如何。

Background

Specifcally, patients enrolled in these latest studies were identifed using emergent magnetic resonance imaging or computed tomography perfusion (CTP). However, the majority of community hospitals and primary stroke centers do not have these capabilities 24/7 nor the expertise to interpret these studies emergently. Also, prior studies have shown that head movement in CTP acquisitions limits its interpretability in up to 25% of patients.3,4 Moreover, the capacity of inpatient Neurologists and Emergency Medicine physicians to evaluate all stroke patients up to 24 hours from LKWT with the same level of urgency as 0- to 6-hour presentations is limited.

关于CTP诊断的问题：

1.大多数医院和中风中心可能没法全天候，或不能快速及时的解释CTP检查结果；

2.采集过程中的头部运动使得25%的结果无法反应真实情况；

3.住院神经科医生和急诊内科医师对所有中风患者进行评估的能力是有限的，从最后一次已知的健康时间起24小时内，其紧急程度与0到6小时相同

Therefore, should all stroke patients up to 24 hours from LKWT be routed directly to comprehensive stroke centers? Doing so would result in a large number of non-EST or thrombolysis candidates bypassing centers fully qualifed to care for them and overstretch the capacity of comprehensive stroke center hospitals and staff. As such, there is a pressing need to screen patients that may be candidates for treatment using imaging data and expertise that are available reliably at primary stroke centers.

是否应该将所有24小时内的中风患者都应该直接送到综合性卒中中心？大量的非溶栓治疗中心的工作人员对他们进行全面的治疗，会导致他们的治疗能力过剩。因此，迫切需要利用原发性脑卒中中心可靠的影像学数据和专业知识筛选可能需要治疗的患者。

Methods

Study Pupulation

All patients presenting to our Emergency Department for an acute stroke evaluation undergo the same neuroimaging protocol, which consists of a noncontrast head CT (NCHCT) followed by CTP and CTA. As such, an identical imaging protocol was performed in both patients who were ultimately diagnosed with LVO AIS, non-LVO AIS, as well as those ultimately diagnosed as not having AIS or TIA.

所有到急诊科接受急性中风评估的患者都要接受相同的神经影像学检查，包括非接触式头部CT（NCHCT），然后是CTP和CTA。因此，对最终诊断为LVO-AIS、非LVO-AIS的患者以及最终诊断为没有AIS或TIA的患者均采用相同的成像方案。

Study Design

Two primary end points were chosen to reflect the questions that the AIS neuroimaging evaluation must address. The frst end point was accuracy in detection of LVO. The second primary end point was accuracy in detection of ischemic core.

实验的评判主要有两点：

1.LVO检测的准确性

2.检测缺血核心的准确性

DeepSymNet Description

DeepSymNet, which leverages brain symmetry information to learn an outcome variable.

DeepSymNet[2]，利用大脑的对称性信息来学习结果变量。

an initial preprocessing pipeline automatically removes the skull, registers the CTA to a common template, and normalizes the voxel intensity.

先预处理去除颅骨，将CTA配准到通用模板，最后归一化体素灰度值。

Then, the DeepSymNet architecture learns the convolutional flters common to the 2 brain hemispheres using 3-dimensional Inception modules.

The DeepSymNet was trained and tested using a 10-fold cross-validation on 2 binary variables: presence/absence of large vessel occlusion and a dichotomized ischemic core size with a threshold of 30 or 50 mL as described in Figure I in the online-only Data Supplement. The DeepSymNet probabilities for the 2 binary variables were then used for the statistical analysis. The contribution of each voxel toward reaching the output classifer probability was evaluated using ϵ-Layerwise Relevance Propagation (online-only Data Supplement).

这部分没理解透彻啊，啥跟啥哦……

Statistical Analysis

Univariate comparisons between continuous variables were performed using Student t test or Wilcoxon rank-sum testing for nonnormal data, and univariate comparisons of categorical variables were performed using Fisher exact test.

连续变量间的单变量比较采用Student t检验或Wilcoxon秩和检验，分类变量的单变量比较采用Fisher精确检验。

From the DeepSymNet probabilities described above for the 3 comparisons (LVO detection, ischemic core ≤30 mL and ischemic core ≤50 mL), receiver-operator curve analysis was performed, and DeepSymNet determinations were evaluated using area under the curve (AUC) measurements with 95% CIs.

对缺血面积小于等于30%的CIs进行了深度分析（95%可信区间），并对受试者进行了缺血面积小于等于50%的测定。

Correlation between DeepSymnet probabilities and CTP-RAPID ischemic core volume predictions (as a continuous variable) was performed using Pearson correlation.

DeepSymnet概率与CTP快速缺血核心体积预测（作为连续变量）之间的相关性采用皮尔逊相关法进行。

Analyses were performed using the Scikit-Learn/Statsmodels Python libraries and confrmed with StataMP 14 (StataCorp LLC, College Station, TX), Prism 7 (GraphPad, La Jolla, CA) statistical software.

使用Scikit Learn/Statsmodels Python库进行分析，并与StataMP 14（StataCorp LLC，College Station，TX）和Prism 7（GraphPad，La Jolla，CA）统计软件一致。

Results

其他

[1] Sheth S A, Lopez-Rivera V, Barman A, et al. Machine Learning–Enabled Automated Determination of Acute Ischemic Core From Computed Tomography Angiography[J]. Stroke, 2019, 50(11): 3093-3100.

[2] Barman A, Inam M E, Lee S, et al. Determining ischemic stroke from CT-angiography imaging using symmetry-sensitive convolutional networks[C]//2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). IEEE, 2019: 1873-1877.