深度学习在CT血管造影中检测脑动脉瘤的应用
本研究基于两家医院四台扫描仪2015年1月至2019年6月的头CT血管造影图像构建深度学习算法,并在2019年7月至12月的400例独立CT血管造影上进行外部验证。算法检测脑动脉瘤的敏感度为97.5%(633/649;95%CI:96.0,98.6),并检出8个初始报告遗漏的动脉瘤(1.2%,8/649);在算法辅助下,放射科医师的加权AFROC曲线下面积提高0.01(95%CI:0.00,0.03)。
要点
- 算法检测脑动脉瘤敏感度97.5%(633/649;95%CI:96.0,98.6)。
- 算法检出8个初始报告遗漏的动脉瘤(1.2%,8/649)。
- 放射科医师加权AFROC曲线下面积提高0.01(95%CI:0.00,0.03)。
诊断标准与关键数值
- 算法检测脑动脉瘤的敏感度为97.5%(633/649;95%CI:96.0,98.6)。
- 算法检出8个初始报告遗漏的动脉瘤(1.2%,8/649)。
- 借助算法,放射科医师的加权AFROC曲线下面积提高0.01(95%CI:0.00,0.03)。
局限性
- 原文未报告局限性。
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
Background Cerebral aneurysm detection is a challenging task. Deep learning may become a supportive tool for more accurate interpretation. Purpose To develop a highly sensitive deep learning-based algorithm that assists in the detection of cerebral aneurysms on CT angiography images. Materials and Methods Head CT angiography images were retrospectively retrieved from two hospital databases acquired across four different scanners between January 2015 and June 2019. The data were divided into training and validation sets; 400 additional independent CT angiograms acquired between July and December 2019 were used for external validation. A deep learning-based algorithm was constructed and assessed. Both internal and external validation were performed. Jackknife alternative free-response receiver operating characteristic analysis was performed. Results A total of 1068 patients (mean age, 57 years ± 11 [standard deviation]; 660 women) were evaluated for a total of 1068 CT angiograms encompassing 1337 cerebral aneurysms. Of these, 534 CT angiograms (688 aneurysms) were assigned to the training set, and the remaining 534 CT angiograms (649 aneurysms) constituted the validation set. The sensitivity of the proposed algorithm for detecting cerebral aneurysms was 97.5% (633 of 649; 95% CI: 96.0, 98.6). Moreover, eight new aneurysms that had been overlooked in the initial reports were detected (1.2%, eight of 649). With the aid of the algorithm, the overall performance of radiologists in terms of area under the weighted alternative free-response receiver operating characteristic curve was higher by 0.01 (95% CI: 0.00, 0.03). Conclusion The proposed deep learning algorithm assisted radiologists in detecting cerebral aneurysms on CT angiography images, resulting in a higher detection rate. © RSNA, 2020 Online supplemental material is available for this article. See also the editorial by Kallmes and Erickson in this issue.
引用
- Yang J, Xie M, Hu C, et al. Deep Learning for Detecting Cerebral Aneurysms with CT Angiography. Radiology. 2020. PMID: 33141003. doi:10.1148/radiol.2020192154
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。