2712 - Deformable Image Registration Validation for Re-Irradiation: A Landmark-Based Assessment across Diverse Disease Sites

Purpose/Objective(s): Re-irradiation (reRT) often involves substantial anatomical changes between treatment courses complicating accurate cumulative dose estimation. Deformable image registration (DIR) uncertainty in this setting is incompletely investigated and may lead to suboptimal clinical decision-making. This study aims to quantify DIR uncertainty using an anatomical landmarks (LM)-based approach across diverse reRT disease sites.Re-irradiation (reRT) often involves substantial anatomical changes between treatment courses complicating accurate cumulative dose estimation. Deformable image registration (DIR) uncertainty in this setting is incompletely investigated and may lead to suboptimal clinical decision-making. This study aims to quantify DIR uncertainty using an anatomical landmarks (LM)-based approach across diverse reRT disease sites.

Materials/Methods: We included patients undergoing > 1 course of reRT with clinically meaningful dose overlap in/near treatment regions. Anatomical LMs were manually identified on simulation CT scans from different courses based on distinctive surrounding image features. A clinically widely used B-spline-based DIR algorithm was evaluated using two parameter settings (A/B) and compared with rigid registration. Target registration error (TRE) was calculated at each LM and summarized overall and by anatomical category (airway, CNS, lung, bone and other). Fourteen patients were analyzed (including local recurrences and proximate multi-lesion scenarios across head and neck, thorax, breast, sarcoma, spine, and abdomen). In total, 190 LMs were identified (10-17 per patient).

Results: Overall mean TRE was 6.48 ± 6.73 mm with rigid registration, improving to 3.66 ± 4.09 mm with DIR setting A and 4.01 ± 4.06 mm with DIR setting B (DIR vs rigid, p<0.001). TRE results by anatomical category are summarized in the table. Despite overall improvement, large residual errors (TRE >10 mm) persisted, most notably in lung (12/43 LMs) and other soft-tissue landmarks (15/82), with fewer outliers in bone (4/45) and none in CNS (0/14).

Conclusion: B-spline DIR significantly improved landmark-based geometric accuracy over rigid registration for reRT planning, with the largest gains in the lung. However, substantial outliers reman, particularly in lung and other soft-tissue regions, underscoring the need for caution when using DIR for cumulative dose estimation. Larger-cohort studies are warranted to better characterize DIR performance across reRT scenarios and to quantify clinical and dosimetric consequences.