S977
Clinical - Oligometastatic cancer
ESTRO 2026
aimed to predict overall survival (OS) in patients with oligometastatic NSCLC based on a longitudinal analysis of FDG-PET/CT imaging features. Material/Methods: This analysis included a retrospective cohort of 61 OMD NSCLC patients, all with ≥ 10 FDG-PET/CT scans acquired before or within three years after OMD diagnosis; later scans were excluded. The modeling endpoint was 4-year OS from the first OMD diagnosis. PET images were converted to body-weight– normalized standardized uptake values using vendor- specific approach [1]. PET-visible lesions were segmented with a GLOW-FDG model.A total of 25 longitudinal PET imaging features were extracted, derived from the dynamics of total tumor burden (TTB), total lesion glycolysis (TLG), and the number of lesions. Highly correlated features (Kendall’s τ >0.8) were resolved. Univariate analyses used the Mann– Whitney U test for continuous variables (significance level: p<0.05). For multivariate modeling, features were ranked by selection frequency in (Least Absolute Shrinkage and Selection Operator) LASSO logistic regression. The final model parameters were determined using the Akaike (AIC) and Bayesian (BIC) information criteria. Predictive performance was evaluated using nested repeated stratified cross- validation, with calibration assessed via the Brier score. The final logistic regression model was trained using repeated stratified cross-validation. Results: After correlation removal, 14 features were left (Table. 1). Univariate analysis (Table 1) showed that 11 features were significant (p<0.05), with AUC values between 0.67 and 0.83. Based on the LASSO selection (Table 1), the top two features were: fraction of cancer positive scans (97% selection frequency) and mean change in TLG between adjacent scans normalized by time between scans (mean TLG velocity) (95%). AIC and BIC were the lowest (Fig. 1A) for the two-parameter model, nested cross validation showed performance with mean AUC of 0.87 (95% CI: 0.83 – 0.91) (Fig. 1A) and mean Brier score of 0.16 (95% CI: 0.13 – 0.18) (Fig. 1B). The final model weights were -1.04, -1.01, and 0.33 for fraction of cancer positive scans, mean TLG velocity, and intercept respectively.
Conclusion: A two-parameter logistic regression model achieved robust cross-validated performance, suggesting that longitudinal FDG-PET image biomarkers may provide prognostic information for OMD in NSCLC. References: 1. M. Fritsak et al. Technical Note: Vendor-Specific
Made with FlippingBook - Share PDF online