Stacked CT radiomics, deep learning and clinical feature models for differentiating benign and malignant solitary pulmonary nodules

Solitary pulmonary nodules (SPNs) are common incidental findings on high-resolution CT, and differentiating benign from malignant lesions still depends mainly on invasive biopsy, highlighting the need for accurate non-invasive tools. In this retrospective multicenter study we analyzed 201 SPN patients (110 malignant, 91 benign) from three hospitals (2020–2024). We developed three single-modality classifiers—a clinical model incorporating demographic data, medical history, imaging characteristics, and seven serum tumor-associated autoantibodies, a radiomics model based on 17 LASSO-selected CT features extracted from manually segmented volumes of interest, and a deep learning visual classifier based on a modified Swin-Transformer architecture—and then fused their per-patient malignancy probabilities with a stacked ensemble using five meta-learners. The integrated model achieved an AUC of 0.936 (95% CI: 0.895–0.971) in the training set and 0.823 (95% CI: 0.699–0.934) in the independent test set, with 80.3% accuracy, 81.6% precision, a recall of 79.4%, and an F1-score of 0.797. Decision curve analysis showed that the fusion model had high clinical net benefit. Multimodal stacking of clinical, radiomic, deep-learning and serological information significantly enhances SPN risk stratification beyond single-modality approaches and may reduce unnecessary invasive procedures in routine practice.