Dose distribution comparison using image registration principles

Abstract

Radiotherapy relies on accurate dose distribution comparison methods, but current approaches have limitations. This study introduces a novel algorithm based on image registration principles to address these limitations. The algorithm uses a transformation matrix derived from image registration to align an evaluated dose distribution with a reference distribution. This transformation employs multiple steps: detecting keypoints, constructing descriptors, matching keypoints, and estimating an affine transformation matrix. The transformed distribution is then directly comparable to the reference through linear least squares regression. Validation on 174 dose distribution pairs demonstrated robust performance, with bias and precision within clinically acceptable limits. Linearity assessments confirmed consistent behavior across a wide range of dose intensities. Comparisons with gamma analysis showed substantial agreement (Cohen's Kappa: 0.77), while additional metrics highlighted its clinical suitability: precision (0.98), recall (0.95), accuracy (0.94), specificity (0.86), and F1-score (0.96). These results establish the algorithm as a promising complement to gamma analysis, with strong potential for clinical integration.