Machine learning models based on clinical-radiomics features to identify SCLC in ≤ 3 cm peripheral solid lung nodules: a multicenter study

Background Current guidelines recommend upfront surgery without biopsy for patients with strongly suspected early-stage lung cancer, but this approach may be inappropriate for small cell lung cancer (SCLC). A noninvasive method for preoperatively distinguishing peripheral SCLC (p-SCLC) from peripheral non-small cell lung cancer (p-NSCLC) is crucial. Methods This multicenter study included 510 patients with peripheral solid lung nodules ≤ 3 cm, allocated to training set (n = 327), internal test set (n = 141), and prospective external test set (n = 42). After tumor segmentation on pretreatment computed tomography (CT) images, 1,223 radiomic features were extracted. Features were selected using the least absolute shrinkage and selection operator (LASSO) and maximum relevance minimum redundancy (mRMR) algorithms. Based on the filtered features, eight machine learning classifiers were trained on the training set to construct radiomics models. The best-performing radiomics model was subsequently combined with clinical variables to build a hybrid clinical-radiomics model. Performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration analysis, and decision curve analysis (DCA). Results The Logistic Regression-based radiomics model showed the highest discrimination, with AUCs of 0.950 (95% confidence interval [CI] 0.916–0.977), 0.938 (0.888–0.978), and 0.919 (0.763–1.000) in the training, internal test, and external test sets, respectively. The hybrid model provided significantly better discrimination, with corresponding AUCs of 0.976 (0.953–0.992), 0.976 (0.949–0.995), and 0.983 (0.944–1.000). Calibration curves and DCA indicated better calibration and greater clinical net benefit for the hybrid model. The hybrid model also maintained good discrimination in subgroup analyses of patients with nodules ≤ 2 cm and patients with negative lymph nodes. Conclusion The hybrid clinical-radiomics model noninvasively distinguished p-SCLC from p-NSCLC with high accuracy and may support individualized preoperative evaluation and treatment planning.