急性胸痛患者阻塞性冠状动脉疾病检测深度学习模型的开发与验证:一项多中心研究
研究开发并验证了基于CCTA的深度学习模型,用于急诊急性胸痛患者阻塞性CAD(≥50%狭窄)检测。训练集为单中心378例(10,060幅曲面MPR图像),外部验证集为3个急诊中心298例。模型在外部验证集上每支动脉敏感度92.7%、特异度89.9%、AUC 0.919,每例患者敏感度93.3%、特异度80.7%、AUC 0.871。
要点
- 外部验证集阻塞性CAD患病率27.5%(82/298)。
- 每支动脉敏感度92.7%、特异度89.9%、AUC 0.919。
- 每例患者敏感度93.3%、特异度80.7%、AUC 0.871。
核心结论
- 该深度学习模型在因急性胸痛行CCTA的患者中识别阻塞性CAD表现出高敏感度和高阴性预测值,提示其可能有助于急诊医师检测CAD。
研究设计与样本
- 研究为深度学习模型的开发与验证研究;训练数据集为2015年1月至2022年12月单中心急诊378例急性胸痛患者的CCTA曲面MPR图像10,060幅;外部验证数据集为2021年1月至2022年12月3个急诊中心298例患者;参考标准为阻塞性CAD(≥50%狭窄)。
诊断标准与关键数值
- 模型基于You Only Look Once v4,需手动预处理提取曲面MPR,每支主要冠状动脉使用15幅手动预处理MPR图像进行开发;外部验证集每支动脉敏感度92.7%、特异度89.9%、阳性预测值62.6%、阴性预测值98.5%、AUC 0.919;每例患者敏感度93.3%、特异度80.7%、阳性预测值67.7%、阴性预测值96.6%、AUC 0.871。
临床应用与阅片要点
- 作者认为该模型对急性胸痛CCTA患者阻塞性CAD检测具有高敏感度和高阴性预测值,可能有助于急诊医师检测CAD。
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
PURPOSE: This study aimed to develop and validate a deep learning (DL) model to detect obstructive coronary artery disease (CAD, ≥ 50% stenosis) in coronary CT angiography (CCTA) among patients presenting to the emergency department (ED) with acute chest pain.
MATERIALS AND METHODS: The training dataset included 378 patients with acute chest pain who underwent CCTA (10,060 curved multiplanar reconstruction [MPR] images) from a single-center ED between January 2015 and December 2022. The external validation dataset included 298 patients from 3 ED centers between January 2021 and December 2022. A DL model based on You Only Look Once v4, requires manual preprocessing for curved MPR extraction and was developed using 15 manually preprocessed MPR images per major coronary artery. Model performance was evaluated per artery and per patient.
RESULTS: The training dataset included 378 patients (mean age 61.3 ± 12.2 years, 58.2% men); the external dataset included 298 patients (mean age 58.3 ± 13.8 years, 54.6% men). Obstructive CAD prevalence in the external dataset was 27.5% (82/298). The DL model achieved per-artery sensitivity, specificity, positive predictive value, negative predictive value (NPV), and area under the curve (AUC) of 92.7%, 89.9%, 62.6%, 98.5%, and 0.919, respectively; and per-patient values of 93.3%, 80.7%, 67.7%, 96.6%, and 0.871, respectively.
CONCLUSIONS: The DL model demonstrated high sensitivity and NPV for identifying obstructive CAD in patients with acute chest pain undergoing CCTA, indicating its potential utility in aiding ED physicians in CAD detection.
引用
- Kim JY, Park J, Lee KH, et al. Development and validation of deep learning model for detection of obstructive coronary artery disease in patients with acute chest pain: a multi-center study. Radiol Med. 2025. PMID: 40810889. doi:10.1007/s11547-025-02064-1
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。