Hybrid CT features predicted symptomatic pneumonitis after locally advanced NSCLC radiotherapy

KEY POINTS

• The retrospective two-centre study included 90 patients in the development cohort and 57 independent patients for external validation. Symptomatic grade 2 or higher pneumonitis occurred in 22/90 (24.4%) and 12/57 (21.1%), respectively.
• Patients had unresectable locally advanced NSCLC treated with 6 MV IMRT to 54–66 Gy, with median dose 60 Gy in 30 fractions. Patients receiving immunotherapy or targeted therapy were excluded.
• Three planning-CT regions were studied: lung minus PTV, lung minus GTV, and PTV minus GTV. Each generated 1,046 handcrafted radiomic features and 512 deep-learning features extracted using a pretrained 3D ResNet-34.
• Clinical characteristics and conventional lung dose-volume metrics were not significantly associated with pneumonitis in the development cohort. V5 showed only a borderline trend (p = 0.055).
• Internal validation showed hybrid models outperforming either feature family alone. AUCs were 0.838 for lung-minus-PTV, 0.901 for lung-minus-GTV, and 0.919 for PTV-minus-GTV.
• The PTV-minus-GTV hybrid model performed best and achieved an external-validation AUC of 0.828 (95% CI 0.696–0.960). Calibration testing was acceptable, and decision-curve analysis suggested positive net benefit across relevant thresholds.
• The finding that a relatively narrow peri-tumoural region outperformed whole normal-lung regions suggests that pretreatment tissue phenotype close to the target may contain information not captured by standard DVH parameters.
• The cohorts were retrospective and relatively small, CT slice thickness was 5 mm, and only one model underwent external validation. Whether prediction can guide lung avoidance or treatment modification without worsening tumour coverage remains untested.

CLINICAL TAKEAWAY

SOURCE

Radiation Oncology