Predicting pulmonary nodule growth from a single time point: a fusion model of radiomics and deep learning to optimize follow-up strategies.

Background: The growth of pulmonary nodules is an important predictor of malignancy. Their varied size makes them difficult to detect and monitor, and improper management may result in a high risk of malignancy. This study explores a non-invasive method for analyzing growth progression in pulmonary nodules between baseline and follow-up computed tomography (CT) scans. The approach integrates radiomics and deep features extracted from baseline imaging data. These features led to the development of a nodule growth prediction model to accurately differentiate growth status, thus providing an objective, quantitative basis for optimizing clinical follow-up strategies. Methods: We conducted a retrospective cohort study of patients with pulmonary nodules who underwent ≥2 chest CT examinations between 2021 and 2024, comprising 1,387 nodules (1,109 for internal training and 278 for internal testing) from baseline and follow-up CT scans. Nodules were classified into growth (volume increase ≥25%) and non-growth groups based on 1-year follow-up volumetric changes. From CT images, 2,264 radiomics features were extracted. Feature selection was conducted via least absolute shrinkage and selection operator (LASSO) regression, followed by principal component analysis (PCA) for dimensionality reduction. For deep feature extraction, we utilized a Swin Transformer architecture for nodule classification, extracting features from its penultimate layer which similarly underwent PCA-based dimensionality reduction. The extracted radiomics and deep features were integrated to develop logistic regression (LR) and random forest predictive models. Model performance was evaluated using the area under the curve (AUC), sensitivity, specificity, and additional metrics. Decision curve analysis (DCA) and calibration curves further evaluated clinical utility. Results: The combined model demonstrated strong predictive performance in predicting nodule growth in the Internal training set (n=1,109 nodules). It achieved an AUC of 0.902 [95% confidence interval (CI): 0.876-0.929], with a sensitivity of 0.815 and a specificity of 0.837, indicating a robust balance in diagnostic performance. The model significantly surpassed both radiomics-based (AUC =0.829; P<0.001) and deep learning-based (AUC =0.875; P<0.01) benchmarks. Furthermore, decision curve and calibration analyses consistently validated its clinical superiority and prediction accuracy. Conclusions: The integrated radiomics and deep learning model effectively stratifies pulmonary nodule growth risk. This approach provides clinicians with a quantitative tool for personalizing follow-up and intervention strategies, demonstrating significant potential for clinical translation.