Background: Low dosage computer tomography (LDCT) has been widely adopted as a sensitive method to detect early-stage lung cancer; however, debate regarding its accuracy and overdiagnosis is still ongoing. An accurate non-invasive test is needed to identify malignant nodules and reduce unnecessary invasive procedures. Studies show that plasma proteins and LDCT imaging features may be used to discriminate malignant pulmonary nodules from benign ones. We aimed to develop a combinatorial approach integrating serum protein markers and imaging features to improve the classification of pulmonary nodules. Methods: This study established a prospective research cohort, which enrolled 608 patients of pulmonary nodules. Plasma samples were collected to measure protein levels using Proximity Extension Assay (PEA) technology. The imaging features of the pulmonary nodules were extracted using the python 'radiomics' package. Following feature extraction, a deep learning networks model was built using training cases. The model was then tested in the testing set to evaluate its accuracy and robustness. Results: From the study cohort, 184 benign (BN) and 184 malignant (MT) samples matched for sex and age were chosen to have a representative training set. The rest 240 samples (81 BN and 159 MT) were used as a testing set. Image features were extracted from 448 patients (119 BN and 153 MT in training; 57 BN and 119 MT in testing) with raw LDCT image available. In the testing set, the model trained using only protein levels had an AUC of 0.83 [0.782-0.877] (sensitivity = 71.1% [95% CI 63.6-77.6]; specificity = 82.7% [73.1-89.4]) when classifying plasma samples of lung cancers from those of benign nodules. In comparison, the imaging features-only model had an AUC of 0.874 [0.821-0.916] (sensitivity = 81.5 [73.6-87.5]; specificity = 73.7 [61.0-83.4]). The combinatorial model integrating both protein and imaging features had an AUC of 0.878 [0.830-0.921] (sensitivity = 83.2% [75.5-88.8]; specificity = 77.2% [64.8-86.2]). Notably, the combinatorial model was highly sensitive for early-stage lung cancer, achieving a sensitivity of 93.1% [78.0-98.1] when classifying stage-0 lung cancer cases(n = 29) and 94.8% [85.9-98.2] for stage-I cases (N = 58). Its sensitivity for stage II-IV (n = 7) and clinically diagnosed lung cancers cases (n = 25) were 100% [64.6-100] and 40.0% [23.4-59.3], respectively. Conclusions: This study aimed to construct a model to distinguish malignant pulmonary nodules from benign lung diseases using protein levels and LDCT imaging features. The resulted combinatorial model showed by utilizing both types of features, it was able to accurately differentiate benign and malignant pulmonary nodules, suggesting it may provide guidance for the clinical management of these nodules.