Abstract / Summary
The incremental diagnostic value of delayed-phase ¹⁸F-FDG PET/CT for indeterminate pulmonary nodules remains uncertain. We compared conventional and extended models incorporating delayed metabolic retention. This single-center retrospective study included 612 patients with one index pulmonary nodule measuring 8–30 mm, divided chronologically into training (2018–2022; n = 312) and temporal-validation (2023–2025; n = 300) cohorts. Reference diagnoses were established by histopathology or ≥ 24-month imaging follow-up for benign solid nodules. Early and delayed scans were acquired approximately 60 and 120–180 min after injection, respectively. A conventional logistic model included seven clinical, CT, and early-phase PET predictors selected using least absolute shrinkage and selection operator regression. The nested extended model added the retention index of maximum standardized uptake value (RI-SUVmax). Internal validation repeated model development in 1,000 bootstrap resamples; locked models were assessed for discrimination, calibration, and diagnostic performance in temporal validation. Malignancy was confirmed in 369 patients (60.3%). Temporal-validation areas under the receiver operating characteristic curve were 0.891 (95% CI, 0.854–0.928) for the conventional model and 0.903 (95% CI, 0.868–0.938) for the extended model; the paired difference was 0.0120 (95% CI, − 0.0003–0.0243; P = 0.056). Corresponding Brier scores were 0.134 and 0.127, and calibration slopes were 0.557 and 0.482, indicating substantial miscalibration. At training-derived cutoffs, sensitivity increased from 76.7% to 84.4%, while specificity decreased from 86.7% to 81.7%. Exploratory decision curves showed threshold-dependent differences in net benefit. Adding RI-SUVmax yielded a small, statistically uncertain improvement in discrimination, with persistent miscalibration and a sensitivity–specificity tradeoff. Recalibration and multicenter external validation are required before clinical implementation. Not applicable.