Tumor-conditioned inter-patient registration using planning computed tomography for voxel-based analysis to predict radiation pneumonitis in lung cancer patients.

Background and purpose: Deformable image registration (DIR) for voxel-based analysis (VBA) can be challenging in patients with non-small cell lung cancer (NSCLC) due to large variations in tumor size and location. This study aimed to assess whether a tumor-preserving inter-patient DIR approach improves VBA-based prediction of radiation pneumonitis (RP). Methods and materials: Three DIR methods were evaluated: deep learning-based Tumor-Aware Recurrent Registration (TRACER) and Patient-Specific Context and Shape (PACS), trained on a public dataset of 268 locally-advanced (LA) NSCLC patients, and iterative Symmetric Normalization (SyN). All methods were tested on 240 patients with LA-NSCLC. Geometric, dosimetric, and tumor preservation metrics were compared using the Wilcoxon signed-rank test. VBA was conducted with each DIR method to identify cohort-relevant regions (CRRs). Machine learning models incorporating clinical, dosimetric, and CRR dose features were used to predict grade 2 or higher RP. Results: TRACER best preserved tumor volume (1.39 %) and organ doses (mean 0.08 Gy) compared with PACS and SyN (p < 0.001). PACS showed higher geometric but worse dose preservation accuracy than TRACER. All DIR-based VBA methods identified the right lung as the CRR associated with RP. TRACER-derived CRR had slightly higher RP predictive performance (AUC 0.78 vs PACS 0.73 vs SyN 0.71), and outperformed the MLD-based ML model (AUC = 0.78 vs 0.69, p = 0.04; specificity = 0.62 vs 0.48). Conclusions: TRACER improved registration accuracy, with better tumor volume preservation and reduced OAR dose impact. Incorporating VBA-derived dose enhanced RP prediction accuracy compared with using MLD. CRRs identified through VBA were robust to the choice of DIR.