Development and validation of a phase-stable computed tomography radiomics model for brain metastasis risk stratification in non-small cell lung cancer.
Authors
Affiliations (1)
Affiliations (1)
- Department of Radiology, the Key Laboratory of Advanced Interdisciplinary Studies Center, National Center for Respiratory Medicine, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou, China.
Abstract
Brain metastasis (BM) is a frequent complication of non-small cell lung cancer (NSCLC), but individualized risk stratification remains limited. We evaluated whether computed tomography (CT) radiomic features reproducible across non-enhanced and contrast-enhanced phases could improve BM risk prediction and whether integration with clinical biomarkers and semantic CT findings provided additional value. This retrospective study included 226 patients with pathologically confirmed NSCLC who underwent pretreatment multiphase chest CT and contrast-enhanced brain magnetic resonance imaging. BM was defined as metastasis detected at baseline or during follow-up. Patients were randomly divided into training (n=181) and held-out test (n=45) sets. Radiomic features reproducible between non-contrast-enhanced CT (NE-CT) and venous-phase contrast-enhanced CT (CE-CT) were defined as phase-stable using an intraclass correlation coefficient greater than 0.80. Radiomic features were reduced using univariable filtering, correlation analysis, and least absolute shrinkage and selection operator (LASSO) with 5-fold cross-validation. Clinical biomarkers and semantic CT features were integrated with phase-stable radiomics. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), classification metrics, calibration, and decision curve analysis. Among single-sequence models, NE-CT showed the best discrimination (AUC, 0.901; 95% confidence interval (CI): 0.858-0.945 in training and 0.765; 95% CI: 0.618-0.912 in testing). The phase-stable radiomics model achieved AUCs of 0.832 (95% CI: 0.773-0.891) and 0.793 (95% CI: 0.657-0.929), respectively. Adding clinical biomarkers improved test performance (AUC, 0.909), whereas further addition of semantic CT findings yielded similar discrimination (AUC, 0.908). The fully combined model had test accuracy of 0.867, sensitivity of 0.667, specificity of 0.939, positive predictive value of 0.800, and negative predictive value of 0.886. Within thin-slice CT protocols of 2 mm or less, differences in contrast phase and reconstruction kernel had limited influence on model discrimination. Phase-stable CT radiomics provided reproducible BM risk stratification in NSCLC. Integration with clinical biomarkers improved discrimination, whereas semantic CT findings provided limited additional gain. These findings support further multicenter validation of phase-stability-guided CT radiomics for BM risk assessment.