Distinguishing multiple primary lung cancers from intrapulmonary metastasis using CT-based radiomics

• Published in: European journal of radiology 2023-03, Vol.160, p.110671-110671, Article 110671
• Main Authors: Huang, Mei, Xu, Qinmei, Zhou, Mu, Li, Xinyu, Lv, Wenhui, Zhou, Changsheng, Wu, Ren, Zhou, Zhen, Chen, Xingzhi, Huang, Chencui, Lu, Guangming
• Format: Article
• Language: English
• Subjects: Humans; Intrapulmonary metastasis; Lung Neoplasms - diagnostic imaging; Lung Neoplasms - pathology; Multiple primary lung cancers; Neoplasms, Multiple Primary; Radiomics; Refined-radiomics; Retrospective Studies; ROC Curve; Tomography, X-Ray Computed
• ISSN: 0720-048X; 1872-7727
• DOI: 10.1016/j.ejrad.2022.110671

•Development of CT-based radiomic models that can efficiently distinguish between multiple primary lung cancers (MPLC) and intrapulmonary metastasis (IPM).•Development of radiomic algorithms aggregated tumor-level features into patient-level features representation.•Combining radiomic-clinical models for diagnosing multiple lung cancers diagnosis in real clinic scenario. To develop CT-based radiomics models that can efficiently distinguish between multiple primary lung cancers (MPLCs) and intrapulmonary metastasis (IPMs). This retrospective study included 127 patients with 254 lung tumors pathologically proved as MPLCs or IPMs between May 2009 and January 2020. Radiomics features of lung tumors were extracted from baseline CT scans. Particularly, we incorporated tumor-focused, refined radiomics by calculating relative radiomics differences from paired tumors of individual patients. We applied the L1-norm regularization and analysis of variance to select informative radiomics features for constructing radiomics model (RM) and refined radiomics model (RRM). The performance was assessed by the area under the receiver operating characteristic curve (AUC-ROC). The two radiomics models were compared with the clinical-CT model (CCM, including clinical and CT semantic features). We incorporated both radiomics features to construct fusion model1 (FM1). We also, build fusion model2 (FM2) by combing both radiomics, clinical and CT semantic features. The performance of the FM1 and FM2 were further compared with that of the RRM. On the validation set, the RM achieved an AUC of 0.857. The RRM demonstrated improved performance (validation set AUC, 0.870) than the RM, and showed significant differences compared with the CCM (validation set AUC, 0.782). Fusion models further led prediction performance (validation set AUC, FM1:0.885; FM2:0.889). There were no significant differences among the performance of the FM1, the FM2 and the RRM. The CT-based radiomics models presented good performance on the discrimination between MPLCs and IPMs, demonstrating the potential for early diagnosis and treatment guidance for MPLCs and IPMs.