Multi-center Validation of Pulmonary Nodule Classification Model for Lung Cancer Screening

Background/Objectives: Early screening of at-risk patients by detecting pulmonary nodules before they deteriorate is critical for timely intervention and treatment. Furthermore, accurately classifying nodule types on computed tomography (CT) images helps in planning appropriate follow-up and treatment strategies. Methods: We collected CT data from 238 patients through a medical data collection agency in the United States, where two radiologists annotated 1,904 pulmonary nodules. After excluding part-solid and segmentation-failed nodules, a total of 1,888 nodules were included. Of these, 1,416 nodules were used for training and 472 for internal validation. External validation was conducted on two independent public datasets: LUNA (n=1,122) and ISBI (n=220). Using segmentation masks automatically generated by a deep learning-based segmentation model, radiomic features were extracted and used to train an AI-based model for classifying solid and ground-glass nodules. Results: The proposed model achieved strong performance across both internal and external validation datasets. The area under the receiver operating characteristic curve (AUC) was 0.976 (95% confidence interval [CI]: 0.962–0.990) on the internal validation set. On external datasets, the model achieved an AUC of 0.962 (95% CI: 0.945–0.979) on LUNA and 0.951 (95% CI: 0.925–0.976) on ISBI, demonstrating high generalizability across different cohorts. Feature-level analysis showed that higher-percentile voxel intensities, especially from axial slices, contributed most to classification. Conclusions: This study presents a classification model that distinguishes between solid and ground-glass nodules with high performance on both internal and external datasets, supporting its potential utility in early lung cancer screening and personalized follow-up strategies.