Multimodal deep learning framework integrating deep radiomics, clinical features, and EGFR Status to predict growth trends of pulmonary ground-glass nodules

Pulmonary ground-glass nodules (GGNs) are commonly identified in low-dose computed tomography (CT) screenings. However, their indolent growth patterns often require prolonged follow-up to differentiate benign from malignant, presenting significant challenges for clinical decision-making and patient well-being. This study aims to construct a multimodal deep learning framework that integrates radiomic features, clinical variables, and epidermal growth factor receptor (EGFR) mutation status to enable early prediction of pulmonary nodule growth trends, thereby facilitating precision diagnosis and treatment. We retrospectively analyzed data from 421 patients with pulmonary nodules between December 2021 and October 2024, comprising 147 cases with nodule growth and 274 stable cases. A three-dimensional convolutional neural network (ResNet3D) was utilized to extract spatial features from CT images, while clinical variables were modeled using a random forest algorithm. These data were integrated via a probabilistic averaging fusion strategy. Additionally, 107 radiomic features were extracted using PyRadiomics, with 15 representative features selected through recursive feature elimination. Features associated with nodule growth for exploratory analysis were identified using the Wilcoxon rank-sum test and logistic regression analysis. TotalEnergy, ClusterShade, and LeastAxisLength were identified as radiomic metrics independently correlated with nodule growth via exploratory analysis, and these features were strongly associated with nodule type, EGFR mutation status, and typical imaging manifestations. On the independent test set, the fusion model yielded an AUC of 0.9123. DeLong test demonstrated that this fusion model was significantly superior to the single ResNet3D model, while no statistically significant AUC difference was observed compared with the RF model incorporating EGFR status. Although EGFR mutations were more common in growing nodules, the predictive improvement brought by EGFR information was only marginal and failed to reach statistical significance. Subgroup analysis demonstrated an extremely low sensitivity for progressive pGGNs, with merely 4 progressive pGGN positive cases in the test set, rendering this subgroup analysis exploratory. The multimodal deep learning framework developed in this study effectively predicts the growth trends of pulmonary GGNs by integrating deep radiomics, clinical features, and molecular data. This modeling strategy provides tentative evidence supporting clinical risk evaluation and may assist clinicians in streamlining follow-up protocols appropriately.