人工智能辅助肺超声检测系统性硬化症与炎性肌病相关间质性肺病
本研究将深度学习应用于系统性硬化症(SSc)或特发性炎性肌病(IIM)患者的肺超声(LUS),以胸部CT及LUS-ILD-24判读为标签训练卷积神经网络。140例患者共3920张LUS图像分为开发集(74例)与独立测试集(66例),VGG-16在患者层面表现最佳(AUC 0.972,敏感度97.4%,特异度92.6%),并与肺功能及CT严重度相关。
要点
- VGG-16患者层面AUC 0.972、敏感度97.4%、特异度92.6%,表现与LUS-ILD-24判读相当或更优。
- Grad-CAM显示胸膜特征为主要预测区域,即使B线不频繁亦可识别ILD。
- 模型输出与肺功能检查及CT的CALIPER指标具有强相关性。
核心结论
- 深度学习用于肺超声可在结缔组织病中准确检测间质性肺病,并可辅助专家判读。
- 可解释人工智能提示胸膜特征足以可靠识别ILD,即使B线不频繁。
研究设计与样本
- 纳入SSc或IIM伴或不伴ILD、且有配对LUS与胸部CT的患者。
- LUS图像以CT结果及LUS-ILD-24人工判读作为标签。
- 对InceptionV3、ResNet-50、VGG-16三种预训练CNN进行迁移学习微调,并开发轻量级新架构LUS-Net。
- 共140例患者、3920张LUS图像,分为开发集74例与独立测试集66例。
- 以AUC、敏感度、特异度及与专家判读一致性在图像层面和患者层面评估模型。
主要发现
- VGG-16取得最佳患者层面表现:AUC 0.972,敏感度97.4%,特异度92.6%。
- CNN输出与肺功能检查及基于CT的CALIPER指标具有强相关性。
- Grad-CAM显示胸膜特征为影响模型预测的主要区域。
- CNN表现与LUS-ILD-24判读相当或更优。
关键数值与推荐
- VGG-16患者层面AUC 0.972,敏感度97.4%,特异度92.6%。
- 数据集共140例患者、3920张LUS图像,开发集74例、独立测试集66例。
临床怎么用
- 在SSc或IIM患者中,可考虑用深度学习辅助LUS判读以识别ILD并与专家判读互为补充。
- 模型预测关注的胸膜特征提示阅片时应重视胸膜异常,即使B线不频繁。
- 模型输出与肺功能及CT严重度指标相关,可作为评估ILD严重度的辅助信息。
局限性
- 摘要未报告模型在前瞻性外部人群中的验证结果。
- 摘要未报告LUS-Net等其他架构的具体性能数值。
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)。
原文
OBJECTIVE: Lung ultrasound (LUS) is a sensitive, low-cost, and radiation-free modality for interstitial lung disease (ILD) detection. We previously developed and validated LUS interpretation criteria in systemic sclerosis (SSc) and idiopathic inflammatory myopathy (IIM) showing excellent diagnostic performance and correlations with ILD severity. In this study, we applied deep learning to evaluate whether convolutional neural networks (CNNs) can accurately detect ILD and its severity on LUS. METHODS: Patients with SSc or IIM ± ILD and paired LUS and chest computed tomography (CT) were included. LUS images were labeled using CT results and human LUS-ILD 2024 (LUS-ILD-24) interpretation. Three pretrained CNN architectures (InceptionV3, ResNet-50, VGG-16) were fine-tuned via transfer learning, and a de novo lightweight architecture (LUS-Net) was developed. Model performance for ILD detection was assessed at image and patient levels using area under the curve (AUC), sensitivity, specificity, and agreement with expert interpretation. CNN outputs were correlated with pulmonary function tests (PFTs) and CT-based Computer-Aided Lung Informatics for Pathology Evaluation and Rating indices. Gradient-weighted Class Activation Mapping (Grad-CAM) visualized regions driving predictions. RESULTS: A total of 140 patients representing 3,920 LUS images were included and split into development (74) and independent test sets (66). VGG-16 achieved the best patient-level performance (AUC 0.972, sensitivity 97.4%, specificity 92.6%) showing strong correlations with PFTs and CT severity. Grad-CAM highlighted pleural features as the primary regions influencing model predictions. CNN performance matched or exceeded LUS-ILD-24 interpretation. CONCLUSION: Deep learning applied to LUS enables accurate ILD detection in connective tissue disease and can enhance expert interpretation. Explainable artificial intelligence suggests pleural features, even when B-lines are infrequent, are sufficient for reliable ILD recognition.
引用
- Fairchild RM, Deluna MD, Fazli M, et al. Artificial Intelligence-Aided Lung Ultrasound Detection of Interstitial Lung Disease in Systemic Sclerosis and Inflammatory Myopathy. Arthritis Care Res (Hoboken). 2026. PMID: 41940478. doi:10.1002/acr.80054
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。