KEY POINTS
- The retrospective cohort included 142 patients with stage III lung cancer receiving definitive thoracic radiotherapy, generally 50–66 Gy and most commonly 60 Gy in 30 fractions.
- At approximately three months, 102 patients were classified as responders and 40 as non-responders. Median progression-free survival was 19.6 months in responders versus 9.9 months in non-responders.
- The dataset contained 540 analysable CBCT scans, with a median of four per patient. The primary model used 20 planning-CT features, 10 dose features, and 15 features from the first CBCT obtained during the first one to two treatment weeks.
- Combining planning CT, dose, and the first CBCT produced an AUC of 0.858 ± 0.078, accuracy of 0.760 ± 0.074, sensitivity of 0.785 ± 0.070, and specificity of 0.700 ± 0.187.
- The same model achieved a progression-free-survival concordance index of 0.672 ± 0.058, compared with 0.612 ± 0.075 when only planning CT and dose were used.
- Adding further CBCTs reduced response-classification AUC to 0.831 with two scans, 0.798 with three, and 0.762 with four or more. Survival concordance changed only marginally, peaking at 0.685 with three scans.
- The study used five-fold internal cross-validation but no independent cohort. CBCT tumour contours were propagated from planning CT rather than redelineated, and important variables including delivered dose, PET, smoking history, PD-L1, and molecular biomarkers were unavailable.
CLINICAL TAKEAWAY
The first on-treatment CBCT may contain useful early response information without requiring additional imaging. The decline in performance after adding scans is more likely a warning about dimensionality and overfitting than evidence that later imaging lacks biological value.