Robust organ mapped dose: using multiple image registrations to identify deformation uncertainty in radiation dose mapping

TL;DR

This paper introduces ROAD, a method to improve radiation dose estimation in reirradiation by accounting for uncertainties in image registration.

Contribution

The novel ROAD method uses multiple image registrations to create a dose resampling kernel that captures deformation uncertainty in organ dose mapping.

Findings

ROAD reduced underestimation of near-maximum doses in regions with large anatomical variation.

The method demonstrated greater robustness compared to fixed-kernel approaches in pelvic reRT cases.

Additional kernel expansions captured residual errors beyond inter-DIR discordance.

Abstract

Objective. To assist with reirradiation (reRT) treatment planning, we propose a robust organ-mapped dose (ROAD) method for cumulative dose estimation within critical organs-at-risks (OARs), incorporating deformable image registration (DIR) uncertainty via a dose resampling kernel derived from organ-specific independent DIRs. Approach. The discordance among three distinct DIRs, each of unknown accuracy, was used to estimate spatial uncertainty. For each voxel within an OAR, the discordance generated a per-voxel dose-resampling kernel. Two additional kernel expansions incorporated uncertainties not captured by inter-DIR discordance: the first ensured all returned dose originated within the OAR, while the second ensured all OAR dose voxels were sampled. The maximum dose within the kernel–OAR intersection was assigned to each voxel to yield a robust dose map. The approach was demonstrated for five pelvic, five head-and-neck, and five thoracic reRT cases using DIR-mapped background doses. Kernel generation was analysed by tracking kernel magnitude and its correlation with mean distance to agreement (MDA) and Hausdorff distance. Resulting dose distributions were compared with baseline mapped doses and a fixed-kernel robustness method. Main results. Analysis confirmed generally well-chosen DIRs but revealed residual errors beyond DIR discordance, detected by the additional kernel expansions. ROAD produced dose distributions comparable to fixed-kernel methods under low deformation uncertainty but demonstrated greater robustness in regions with large anatomical variation, particularly in the pelvis. ROAD reduced instances where mapped near-maximum doses underestimated original values, without increasing overall dose, by capturing uncertainty from organ filling and positional changes missed by fixed-kernel accumulation. Significance. Accurate cumulative dose estimation is critical for safe and effective reRT planning. The proposed ROAD framework explicitly incorporates voxel-level DIR uncertainty, providing a more reliable OAR dose estimate in regions with substantial anatomical change. This enhances confidence in reRT dose assessment and offers a practical, robust tool for clinical evaluation of cumulative organ doses.

Figures