基于分割的深度学习肺气肿CT定量优于LAA-950
该研究开发并验证了一种基于深度学习的胸部CT肺气肿分割算法,使用虚拟数据和多中心临床数据进行预训练与微调。结果显示该模型在分割准确性、跨成像条件一致性与影像/肺功能相关性上均优于传统LAA-950阈值法。
要点
- 深度学习分割法对肺气肿定量更准且更稳定,报告可优先参考其百分比结果
- 模型与放射科医生Fleischner视觉评分及DLCO、FEV1pp、FEV1/FVC相关性更强
- 跨扫描仪、剂量和重建核条件下,深度学习法一致性优于LAA-950
核心结论
- 该深度学习分割模型在肺气肿定量准确性和稳健性上优于LAA-950,更贴合放射科医生评估和肺功能指标。
研究设计与样本
- 回顾性研究,利用虚拟数据集(20例人体模型,平均年龄43岁±11,10名男性)进行预训练。
- 临床数据集包括多中心队列C1(101例,57岁±8,54名男性)、C2(23例,57岁±7,14名男性)和C3(1159例,65岁±9,586名男性)用于微调与验证。
- 以专家标注的肺气肿掩膜为参考标准,将模型肺气肿百分比与Fleischner视觉评分和肺功能指标进行相关性分析。
关键影像表现/征象
- 基于深度学习的肺气肿分割在CT上显示与专家标注更接近的病灶范围,Dice系数高于LAA-950。
- 在不同扫描仪、剂量和重建核条件下,深度学习模型所得的肺气肿百分比一致性优于LAA-950。
- 深度学习模型的肺气肿百分比与Fleischner视觉评分及DLCO、FEV1pp、FEV1/FVC的相关性均高于LAA-950。
诊断标准与关键数值
- 虚拟数据Dice系数:深度学习模型76.6%±8.8,LAA-950为51.5%±23.5。
- C1队列Dice系数:深度学习模型48.4%±24.2,LAA-950为22.5%±20.2;C2队列Dice系数:64.8%±9.3对32.7%±16.8。
- C1队列偏倚与一致性限度:深度学习模型1.0%±4.3,LAA-950为2.1%±15.1。
- C1队列与视觉评分相关性:深度学习模型0.77,LAA-950为0.47;C3队列DLCO多变量R²:深度学习模型0.31-0.32,LAA-950为0.21-0.25。
临床应用与阅片要点
- 日常阅片中对CT肺气肿定量可采用该深度学习分割结果,尤其需跨设备或跨重建参数比较时。
- 该模型结果与视觉分级及肺功能更一致,可用于辅助肺气肿严重程度分级和随访评估。
- 注意LAA-950在不同成像条件下的偏差较大,传统阈值法结果解读需谨慎。
局限性
- 研究为回顾性设计,虚拟数据人体模型平均年龄43岁,与临床队列年龄存在差异。
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
OBJECTIVES: To develop a deep learning segmentation algorithm to enable accurate, reliable emphysema quantification while improving agreement with radiologist assessments and pulmonary function tests.
MATERIALS AND METHODS: The model was developed using retrospective virtual and clinical datasets. Virtual data enabled pre-training using ground truth across controlled parameters, including scanners, doses, and reconstruction kernels, while clinical data enabled fine-tuning with expert-annotated emphysema masks. Segmentation accuracy was quantified using the Dice coefficient, and robustness was quantified using emphysema percentage consistency across imaging conditions. Model-based emphysema percentage was correlated with Fleischner visual scores (ordinal; 0-5) and pulmonary function tests (DLCO, FEV1pp, and FEV1/FVC) and compared to LAA-950. Statistical analysis included univariate/multivariate correlations. Quantitative assessment included Dice, bias, limits of agreement, and reproducibility coefficient.
RESULTS: Virtual data included 20 human models (mean age: 43 years ± 11 [SD], 10 men), and clinical data included multi-center cohorts of 101 patients (C1; 57 years ± 8, 54 men), 23 patients (C2; 57 years ± 7, 14 men), and 1159 patients (C3; 65 years ± 9, 586 men). The model outperformed LAA-950 in segmentation accuracy, achieving higher Dice scores across virtual (76.6% ± 8.8 vs 51.5% ± 23.5), C1 (48.4% ± 24.2 vs 22.5% ± 20.2), and C2 (64.8% ± 9.3 vs 32.7% ± 16.8) cohorts. Furthermore, analysis showed improved bias and limits of agreement (C1; 1.0% ± 4.3 vs 2.1% ± 15.1) and stronger correlations with visual scoring (C1; 0.77 vs 0.47) and pulmonary function tests (C3; DLCO multivariate R²: 0.31-0.32 vs 0.21-0.25).
CONCLUSION: The proposed model provided a promising, clinically relevant alternative to LAA-950, improving accuracy and consistency in emphysema quantification and further aligning with radiologist assessments and lung function metrics.
KEY POINTS: Question Can deep learning-based emphysema segmentation on CT improve the accuracy and robustness of emphysema quantification relative to the traditional LAA-950 biomarker? Findings The deep learning model outperformed LAA-950 in terms of accuracy and reproducibility and demonstrated stronger correlations with visual emphysema scores and pulmonary function tests. Clinical relevance The improved performance of the deep learning approach enables more reliable emphysema quantification across imaging conditions, with closer agreement to radiologist assessments and pulmonary function tests.
引用
- Sotoudeh-Paima, Ghojogh Nejad, O'Sullivan-Murphy, et al. Segmentation-based deep learning emphysema quantification using chest CT: improved accuracy and robustness vs LAA-950. Eur Radiol. 2026. PMID: 42717029. doi:10.1007/s00330-026-12800-4
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。