用虚拟人体生成模拟CT，提升肺结节检测与分类的AI性能。
Virtual Patients, Real Gains: Digital Twin-Based Simulated CT for Multitask Lung Nodule Analysis
- 基于解剖结构的数字孪生生成1044张带标注的虚拟CT。
- 结节检测灵敏度提升至0.56，分类准确率AUC达0.87。
- 适合缺乏真实标注数据的医学AI模型训练与验证。
基于AI的肺癌筛查受限于稀缺且标注完整的CT数据，尤其是罕见结节表现。本文探究基于物理的、解剖结构引导的模拟CT是否能提升AI在肺结节三项任务中的表现：检测、分割和恶性程度分类。利用虚拟肺部筛查试验框架，生成174个数字人体孪生体（XCAT3），嵌入512个程序化控制的结节（X-Lesions，大小4-30mm），在两种扫描仪配置下模拟生成CT（DukeSim），共获得1,044张带标注图像。结合临床数据，训练了检测（MONAI）、分割（VISTA3D、nnU-Net）和分类（Med3D）模型，并在外部测试集上评估。检测敏感度在1个假阳性/扫描时从0.37升至0.56（p<0.001）；分割效果小幅提升（Dice从0.61→0.64，0.66→0.69）；分类AUC从0.78增至0.87（p<0.001）。基于物理的虚拟成像试验可缓解医学AI中的数据匮乏问题。
原文摘要 · Abstract (English)
AI-based lung cancer screening is constrained by scarce, annotated CT data, particularly for rare nodule presentations. We investigate whether physics-based, anatomy-informed simulated CT can improve AI performance across three lung-nodule tasks: detection, segmentation, and malignancy classification. Using the Virtual Lung Screening Trial framework, we generated 174 digital human twins (XCAT3), embedded 512 procedurally controlled nodules (X-Lesions, 4-30 mm), and simulated CT (DukeSim) under two scanner configurations, yielding 1,044 annotated scans. Combined with clinical data, these trained models for detection (MONAI), segmentation (VISTA3D, nnU-Net), and classification (Med3D), evaluated on external test sets. Detection sensitivity at 1 FP/scan rose from 0.37 to 0.56 (p < 0.001); segmentation improved modestly (Dice 0.61 to 0.64, 0.66 to 0.69); classification AUC rose from 0.78 to 0.87 (p < 0.001). Physics-based virtual imaging trials can help address data scarcity in medical AI.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。