深度学习全自动定量吸气相与呼气相胸部CT功能小气道病变
本研究为回顾性研究,纳入196例患者共249次配对吸气相与呼气相胸部CT,评估基于深度学习的全自动工具定量功能小气道病变(fSAD)的准确性与时间效率。结果显示自动定量与半手动法一致性极好(Spearman r=0.93),对临床相关fSAD(≥28%肺受累)诊断敏感度100%、特异度96.8%、准确率97.2%,且半手动分析中位时间分别为3.47分钟和5.21分钟。
要点
- 深度学习工具可全自动定量fSAD,与半手动法一致性极好(r=0.93)。
- 自动工具对≥28%肺受累的fSAD诊断准确率达97.2%。
- 半手动法需3.47–5.21分钟,自动工具可提升工作效率。
核心结论
- 基于深度学习的fSAD自动定量与半手动评估一致性极好,诊断准确率高,可作为传统半手动方法的高效可靠替代。
研究设计与样本
- 回顾性研究,纳入2016年1月至2022年7月配对吸气相与呼气相胸部CT共249次(196例患者,中位年龄56岁,男性120例)。
- fSAD依据既定标准定义,半手动评估使用商业后处理软件并含多个手动步骤,自动评估使用深度学习工具全自动分析。
- 由两名心胸放射科医师在20例随机选取病例中测量半手动分析时间,采用Wilcoxon符号秩检验评估读者间系统差异,Spearman相关和Bland-Altman分析评估一致性。
诊断标准与关键数值
- 以半手动fSAD评估为参考,自动fSAD评估与之Spearman r=0.93(95%CI:0.89–0.96;P<.01),Bland-Altman偏倚-1.7%。
- 对临床相关fSAD(≥28%肺受累)分类,自动fSAD评估敏感度100%、特异度96.8%、准确率97.2%。
- 两名读者间无显著系统差异(Wilcoxon W=1;P=.66),半手动分析中位时间分别为3.47分钟(IQR 3.06–3.97)和5.21分钟(IQR 3.57–6.65)。
临床应用与阅片要点
- 深度学习工具可全自动定量fSAD,与半手动法一致性极好,适合在日常阅片中快速评估小气道病变。
- 对≥28%肺受累的fSAD,自动工具敏感度100%、特异度96.8%,可作为筛查或分诊依据。
- 自动工具可减少手动操作步骤,半手动法需3.47–5.21分钟,自动法有望提升工作效率。
局限性
- 本研究为回顾性设计,可能存在选择偏倚。
- 半手动分析时间仅在20例随机选取病例中测量,样本量较小。
- 研究未报告多中心外部验证结果。
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 To evaluate the accuracy and time efficiency of a deep learning (DL)-based tool for automated quantification of functional small airway disease (fSAD) at chest CT compared with a conventional semimanual method. Materials and Methods This retrospective study included paired inspiratory and expiratory chest CT examinations performed from January 2016 to July 2022, with fSAD defined according to established criteria. Semimanual fSAD assessment (fSADman) was performed using commercial postprocessing software with multiple manual steps, whereas automatic fSAD assessment (fSADauto) used a DL-based tool to fully automate the analysis. Time for fSADman was measured by two cardiothoracic radiologists in 20 randomly selected cases. The Wilcoxon signed rank test assessed systematic differences between reader measurements. Agreement between fSADman and fSADauto was evaluated using Spearman correlation and Bland-Altman analysis. Classification accuracy for clinically relevant fSAD (≥28% lung involvement) was also assessed. Results The study included 249 CT examinations from 196 patients (median age, 56 [IQR, 48-65] years; 120 [61.2%] male). Median lung volumes were 5196 mL (IQR, 4315-6192 mL) at inspiration and 2802 mL (IQR, 2401-3451 mL) at expiration. Agreement between fSADman and fSADauto was excellent (Spearman r = 0.93 [95% CI: 0.89, 0.96]; P < .01), with a small bias of -1.7% on Bland-Altman analysis. For detecting fSAD of 28% or greater, fSADauto achieved 100% sensitivity, 96.8% specificity, and 97.2% accuracy. No significant systematic difference was observed between readers (Wilcoxon W = 1; P = .66). Median semimanual analysis times were 3.47 (IQR, 3.06-3.97) minutes and 5.21 (IQR, 3.57-6.65) minutes, respectively. Conclusion DL-based automated quantification of fSAD demonstrated excellent agreement with semimanual assessment and high diagnostic accuracy and may serve as an efficient and reliable alternative to the traditional semimanual method. Keywords: Pulmonary, Lung, Machine Learning, Workflow Optimization, Segmentation Supplemental material is available for this article. © RSNA, 2026.
引用
- Gherca S, Yang S, Bremerich J, et al. Fully Automated Quantification of Functional Small Airway Disease at Inspiratory and Expiratory Chest CT Using Deep Learning. Radiol Cardiothorac Imaging. 2026. PMID: 42206980. doi:10.1148/ryct.250215
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。