Liu Zhou, Yang Long, Liang JiuPing, Wen Binbin, He Zikun, Xie Yongsheng, Luo Honghong, Yang Qian, Liu Lijian, Luo Dehong, Li Li, Zhang Na
Department of Radiology, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Shenzhen, China.
Lauterbur Research Center for Biomedical Imaging, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China.
Eur Radiol. 2025 Jun;35(6):2968-2978. doi: 10.1007/s00330-024-11221-5. Epub 2024 Nov 27.
To investigate the incremental benefit of adding radiomic features to conventional semantic radiological feature-based differential diagnosis between benign and malignant lung nodules.
From May 2017 to March 2021, 393 patients with 465 pathologically confirmed lung nodules were enrolled with 54 patients with 54 lung nodules as external testing. Based on manually segmented lung nodules, 1409 radiomics features were extracted. Sixteen radiological features were obtained. The least absolute shrinkage and selection operator (LASSO) was used to select the most informative features from the two features set separately. Support vector machine (SVM) and logistic regression (LR) were used to build the models (radiomics model, radiological model, and combined model) with performance compared using the DeLong test.
After feature selection, six radiological features, including shape, vascular convergence sign (type III), margin, density, pleural traction sign, and spiculation, and nine radiomics features were selected. In the independent testing and external testing, combined models had significantly higher AUCs than the corresponding radiomic models for both the SVM classifier (AUC: 0.871 vs. 0.773, p = 0.029; 0.810 vs. 0.706, p = 0.037) and LR classifier (AUC: 0.871 vs. 0.742, p = 0.008; 0.828 vs. 0.712, p = 0.044), and the corresponding radiological model for both the SVM classifier (AUC: 0.871 vs. 0.803, p = 0.015; 0.810 vs. 0.730, p = 0.045) and LR classifier (AUC: 0.871 vs. 0.818, p = 0.034; 0.828 vs. 0.756, p = 0.040).
Radiomics features could add incremental benefits to the conventional radiological feature-based differential diagnosis.
Question Conventional semantic radiological feature-based differential diagnosis between benign and malignant lung nodules needs further improvement. Findings The model combining radiological features and radiomic features significantly outperforms a radiomic model and a radiological model. Clinical relevance Radiomic features could complement conventional radiological features to improve the differential diagnosis of lung nodules in the clinical setting.
探讨在基于传统语义放射学特征的良恶性肺结节鉴别诊断中加入放射组学特征的增量效益。
2017年5月至2021年3月，纳入393例有465个经病理证实的肺结节患者，其中54例有54个肺结节患者作为外部测试。基于手动分割的肺结节，提取1409个放射组学特征。获得16个放射学特征。使用最小绝对收缩和选择算子（LASSO）分别从两个特征集中选择最具信息量的特征。使用支持向量机（SVM）和逻辑回归（LR）建立模型（放射组学模型、放射学模型和联合模型），并使用德龙检验比较模型性能。
特征选择后，选择了6个放射学特征，包括形态、血管汇聚征（III型）、边缘、密度、胸膜牵拉征和毛刺征，以及9个放射组学特征。在独立测试和外部测试中，对于SVM分类器（AUC：0.871对0.773，p = 0.029；0.810对0.706，p = 0.037）和LR分类器（AUC：0.871对0.742，p = 0.008；0.828对0.712，p = 0.044），联合模型的AUC显著高于相应的放射组学模型；对于SVM分类器（AUC：0.871对0.803，p = 0.015；0.810对0.730，p = 0.045）和LR分类器（AUC：0.871对0.818，p = 0.034；0.828对0.756，p = 0.040），联合模型的AUC也显著高于相应的放射学模型。
放射组学特征可为基于传统放射学特征的鉴别诊断带来增量效益。
问题基于传统语义放射学特征的良恶性肺结节鉴别诊断需要进一步改进。发现放射学特征与放射组学特征相结合的模型显著优于放射组学模型和放射学模型。临床意义放射组学特征可补充传统放射学特征，以改善临床环境中肺结节的鉴别诊断。
