顾俊峰,张振岐,高晓坤,谢子明,陶广煜,倪其鸣,朱琳,于红,孙炎冰.混合特征筛选策略CT影像组学模型用于肺腺癌风险分层[J].中国介入影像与治疗学,2026,23(6):338-344
混合特征筛选策略CT影像组学模型用于肺腺癌风险分层
CT radiomics model based on hybrid feature selection strategy for risk stratification of lung adenocarcinoma
投稿时间:2026-03-08  修订日期:2026-03-13
DOI:10.13929/j.issn.1672-8475.2026.06.004
中文关键词:  肺肿瘤  腺癌  体层摄影术,X线计算机  影像组学
英文关键词:lung neoplasms  adenocarcinoma  tomography,X-ray computed  radiomics
基金项目:国家自然科学基金(82572212、82302188)、上海市卫生健康委员会医学新技术研究与转化种子计划(2025ZZ2071)、上海市科技计划(22Y11911100、24SF1904000)。
作者单位E-mail
顾俊峰 上海理工大学健康科学与工程学院, 上海 200090
上海市胸科医院放射科, 上海 200030 
 
张振岐 上海理工大学健康科学与工程学院, 上海 200090
上海市胸科医院放射科, 上海 200030 
 
高晓坤 上海理工大学健康科学与工程学院, 上海 200090
上海市胸科医院放射科, 上海 200030 
 
谢子明 上海理工大学健康科学与工程学院, 上海 200090
上海市胸科医院放射科, 上海 200030 
 
陶广煜 上海市胸科医院放射科, 上海 200030  
倪其鸣 上海市胸科医院放射科, 上海 200030  
朱琳 上海市胸科医院放射科, 上海 200030  
于红 上海市胸科医院放射科, 上海 200030  
孙炎冰 上海市胸科医院放射科, 上海 200030 13512183544@163.com 
摘要点击次数: 0
全文下载次数: 0
中文摘要:
      目的 观察基于混合特征筛选策略CT影像组学模型用于肺腺癌(LAC)风险分层的价值。方法 回顾性纳入280例LAC,按7∶3比例随机分为训练集(n=196)与测试集(n=84),并根据病理所示风险状态划分低风险亚组(n=180)与高风险亚组(n=100)。基于肺窗CT提取LAC影像组学特征;分别以最大相关最小冗余(mRMR)、递归特征消除(RFE)及绝对收缩和选择算子(LASSO),以及混合策略mRMR+RFE、mRMR+LASSO、RFE+LASSO及mRMR+RFE+LASSO筛选最优特征并构建模型,比较各模型用于LAC风险分层的效能。结果 分别基于mRMR、RFE、LASSO、mRMR+RFE、mRMR+LASSO、RFE+LASSO及mRMR+RFE+LASSO获得95、50、30、7、13、15及5个最优特征。以混合策略所获最优特征中,以mRMR+RFE及mRMR+LASSO所获权重分布较为平衡。相比其余模型,mRMR+LASSO模型用于训练集LAC风险分层的曲线下面积(AUC)最高(P均<0.05)。各混合策略模型在测试集的AUC均高于各单一策略模型(P均<0.05),而各混合策略模型AUC差异均无统计学意义(P均>0.05)。结论mRMR+LASSO CT影像组学模型用于LAC风险分层的效能上佳。
英文摘要:
      Objective To observe the value of CT radiomics model based on hybrid feature selection strategy for risk stratification of lung adenocarcinoma(LAC). Methods Totally 280 patients with LAC were retrospectively enrolled and randomly assigned to training set (n=196) and test set (n=84) in 7∶3 ratio, also divided into low-risk subgroup (n=180) and high-risk subgroup (n=100) according to pathological risk status. Then radiomics features of LAC lesions were extracted from lung window CT images. The optimal features were selected with single strategy (maximum relevance minimum redundancy [mRMR], recursive feature elimination [RFE], least absolute shrinkage and selection operator [LASSO]) and hybrid strategy (mRMR+RFE, mRMR+LASSO, RFE+LASSO, mRMR+RFE+LASSO), respectively, and the relative models were established, their performance for risk stratification of LAC were analyzed. Results Totally 95, 50, 30, 7, 13, 15 and 5 optimal features were selected with mRMR, RFE, LASSO, mRMR+RFE, mRMR+LASSO, RFE+LASSO and mRMR+RFE+LASSO, respectively. Among optimal features obtained with hybrid strategies, those obtained with mRMR+RFE and mRMR+LASSO had relatively balanced weight distribution. The area under the curve (AUC) of mRMR+LASSO model for risk stratification of LAC was the highest in training set, significant higher than that of all other models (all P<0.05). In test set, the AUC of each hybrid strategy model was higher than that of all sinlge strategy model(all P<0.05), while no significant difference was found between each two hybrid strategy models (all P>0.05). Conclusion CT radiomics model based on mRMR+LASSO had the best performance for risk stratification of LAC.
查看全文  查看/发表评论  下载PDF阅读器
关闭