Comparative evaluation of feature-based and intensity-based algorithms for dosimetric registration in radiotherapy quality assurance using electronic portal imaging devices.

Abstract

To evaluate image registration as a quantitative alternative to the gamma index for VMAT patient-specific quality assurance and to benchmark feature-based against intensity-based registration algorithms. We retrospectively analyzed 93 EPID-verified VMAT arcs using five registration algorithms: three feature-based (SIFT, Thin-Plate Splines, CONCORD) and two intensity-based (B-spline, Demons). We assessed the ability of each algorithm to align absolute dose distributions on a common coordinate grid. Dosimetric agreement was quantified using a dual-parameter metric combining linear regression slope (tolerance 95–105%) and the coefficient of determination [Formula: see text] Standard gamma analysis (3%/3 mm, 90% passing rate) served as the clinical reference. Agreement with gamma-based classification was evaluated using Cohen’s kappa [Formula: see text]. Feature-based methods significantly outperformed intensity-based models. The dose-specific CONCORD algorithm achieved the highest initial agreement with gamma analysis [Formula: see text]. Application of the dual-parameter metric further enhanced diagnostic specificity; specifically, CONCORD’s agreement improved to [Formula: see text] with 95.7% accuracy. In contrast, intensity-based deformable models exhibited poor performance and high false-negative rates. This behavior introduces non-physical dose warping that obscures clinically relevant delivery errors. Registration-based dose comparison offers a robust, quantitative alternative to the gamma index. By utilizing a global affine transformation and dose-specific topological features, the CONCORD algorithm effectively detects localized and systematic discrepancies without the error-masking risks inherent in generic deformable registration. This framework aligns with AAPM TG-307 objectives, offering a more granular and physically intuitive assessment of VMAT delivery fidelity.