AI胸片算法标记肺癌紧急怀疑:优化特征选择与性能
本研究基于苏格兰西部2887张前瞻性及983张回顾性胸片,比较三种特征选择方法对商业胸片AI算法标记肺癌紧急怀疑(USC)性能的影响。结果显示基于AIC和LASSO回归的特征选择(四个特征)优于临床专家选择的五个特征,且回归加权联合评分优于个体特征阈值。结论提示商业胸片AI需本地校准和任务特异性优化。
要点
- 回归特征选择模型优于专家选择特征,特征更少但敏感度更高。
- 联合评分单阈值优于个体特征阈值,便于控制阳性标记率。
- 商业AI算法需本地校准以适配人群和任务。
核心结论
- 基于AIC和LASSO回归的特征选择模型优于临床专家选择的特征,敏感度最高。
- 回归加权联合评分性能优于个体特征阈值。
- 商业胸片AI算法需本地校准和任务特异性优化以检测USC。
研究设计与样本
- 研究类型为算法性能评估,使用2887张前瞻性收集的胸片(全科医生和门诊常规护理)及983张回顾性收集的胸片作为额外验证集。
- 索引测试为算法概率评分;参考标准为USC,由盲法专家放射科医生小组对胸片复审得出。
- 比较三种特征选择方法:放射科医生选择特征+商业个体特征阈值、放射科医生选择特征+回归加权联合评分、AIC和LASSO回归特征选择模型。
- 性能评估在Youden指数和15%阳性标记率两个阈值下进行。
诊断标准与关键数值
- AIC/LASSO回归模型选择四个特征:实变、结节、结核和空洞;临床专家选择五个特征:空洞、结节、肿块、肺门增大和纵隔增宽。
- AIC/LASSO模型敏感度最高,但受15%标记率阈值显著影响。
- 回归加权联合评分性能优于个体特征阈值。
临床应用与阅片要点
- 日常阅片和报告中,可关注AI算法标记的实变、结节、结核和空洞等特征,但需注意算法需本地校准。
- 使用联合评分单阈值有助于控制阳性标记率,优化工作流程。
局限性
- 研究基于苏格兰西部人群,结果可能受人群和本地实践影响。
- 回顾性验证集样本量较小(983张)。
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 investigate the impact of feature selection and thresholding on the performance of a commercial chest X-ray (CXR) algorithm for flagging urgent suspicion of lung cancer (USC) in a West of Scotland population.
MATERIALS AND METHODS: The datasets comprised 2887 prospectively collected CXRs from general practitioners and outpatient routine care in the West of Scotland. An additional validation dataset included 983 retrospectively collected CXRs. The index test was an algorithm probability score; the reference standard was USC, derived from a blinded expert radiologist panel review of CXR. Three feature selection approaches were compared: radiologist-selected features with commercial thresholding of individual features, radiologist-selected features with regression weighting for the combined score, and a model using Akaike Information Criterion (AIC) and LASSO regression approaches for feature selection. Performance was evaluated at two thresholds: the Youden index and 15% positive flag rate.
RESULTS: Regression-based feature selection using the AIC and LASSO approaches resulted in a four-feature model (consolidation, nodule, tuberculosis and cavity) compared with the five features (cavity, nodule, mass, hilar enlargement and mediastinal widening) selected by clinical experts. Sensitivity was highest in the AIC/LASSO model, but was significantly impacted by a 15% flag rate threshold. Regression weighting for the combined score demonstrated superior performance compared to individual feature thresholds.
CONCLUSION: In this study, regression-based feature refinement against an expert-derived reference standard outperformed clinically intuitive selection. Commercial CXR algorithms require local calibration and task-specific optimisation for USC detection. Workflow and post-market deployment considerations can be addressed by constraining positive flag rates through a unified model threshold.
KEY POINTS: Question How should users select and combine AI chest X-ray algorithm features to achieve optimal deployment performance for detecting urgent suspicion of cancer in routine practice? Findings Regression-based feature selection outperformed expert-selected features, and combining features under a single regression-weighted threshold outperformed individual feature thresholds for urgent cancer suspicion detection. Critical relevance statement Optimising chest X-ray AI for urgent cancer suspicion requires data-driven, regression-based feature selection rather than clinical intuition; health systems therefore need local capability to validate and calibrate algorithms for their population and task.
引用
- Duncan SF, Dudman K, Odigbo S, et al. Flagging urgent suspicion of lung cancer with AI CXR algorithms: optimising feature selection and performance. Insights Imaging. 2026. PMID: 42966828. doi:10.1186/s13244-026-02438-7
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。