CT radiomics combined with neural networks predict the malignant degree of pulmonary grinding glass nodules.

Background: This study investigates the use of CT radiomics combined with convolutional neural networks (CNN) to predict the malignancy of lung ground glass nodules (GGN), which are challenging to diagnose due to their ambiguous boundaries. The goal is to improve diagnostic accuracy and support personalized treatment planning. Methods: Retrospective data from 670 patients with pulmonary nodules (2019-2023) were analyzed. CT images were preprocessed using Gaussian filtering and manually segmented to define regions of interest (ROI). A CNN model was trained using MATLAB's Deep Learning Toolbox, and its performance was compared to the Mayo and Brock models. Results: Key predictors of malignancy included nodule diameter, volume, mean CT value, and consolidation-to-tumor ratio (CTR). The CNN-based model achieved an AUC of 0.887, with 82.4% sensitivity and 75.5% specificity, outperforming existing models (Mayo: AUC = 0.655; Brock: AUC = 0.574). Validation accuracy reached 85.07%. Conclusion: In this single-center retrospective study, integrating CT radiomics with CNN depicted promising potential for GGN malignancy prediction, though external validation remains necessary. These findings warrant verification in multicenter prospective cohorts.