KEY POINTS
- Three publicly available deep-learning auto-segmentation models were applied retrospectively to 460 patients with stage III NSCLC from NRG Oncology/RTOG 0617, a multi-institutional dataset originally comparing 60 versus 74 Gy definitive thoracic radiotherapy. Median follow-up in the source trial was 5.1 years.
- The analysis tested whether AI-derived left anterior descending coronary artery V15 Gy could reproduce previously reported survival separation obtained from manually contoured LAD structures. Manual LAD contours themselves were unavailable to the investigators, preventing direct patient-level geometric comparison.
- With a common LAD V15 Gy cutoff of 10%, all three models significantly separated overall survival: HRs were 1.40 (95% CI 1.02–1.93; p=0.038), 1.32 (1.03–1.68; p=0.026), and 1.37 (1.03–1.84; p=0.033) for Models 1–3, respectively.
- Model performance differed substantially despite all three producing significant survival separation. Mean LAD volumes were 0.95, 2.22, and 2.51 cm³ for Models 1–3, and Model 1 showed marked superior LAD under-contouring, demonstrating why geometric plausibility remains important even when an outcome association is preserved.
- Model 3 aligned most closely with previously published manual-contour survival curves. Using its optimized percentage cutoff of V15 Gy >8.3%, two-year overall survival was 48.7% versus 63.4% and the univariable HR was 1.45 (95% CI 1.07–1.96; p=0.017); mean absolute error versus manual survival curves was 5.8%.
- In multivariable analysis, Model 3-derived LAD V15 Gy >8.3% remained associated with worse survival (HR 1.37, 95% CI 1.01–1.87; p=0.046), alongside 74 versus 60 Gy prescription dose (HR 1.30; p=0.030) and grade ≥3 esophagitis (HR 1.51; p=0.012). The model-specific threshold was optimized within this cohort and should not be interpreted as a validated universal LAD constraint.
CLINICAL TAKEAWAY
AI-generated LAD contours can retain clinically meaningful dose-outcome information at the cohort level, potentially enabling large-scale cardiac substructure analyses that would be impractical manually. However, model-to-model contour differences were substantial, manual contours were unavailable for paired validation, and the optimized V15 threshold is not ready for treatment-planning use.