Validation of deformable registration in head and neck cancer using analysis of variance.

Weekly cone beam CTs (CBCTs) of ten head and neck cancer patients undergoing five weeks of radiotherapy were used. An expert identified 23 anatomical features (landmarks) on the planning CT. The landmarks were automatically propagated to the CBCT using multiregion-of-interest (mROI) registration. Additionally, two human observers independently localized these landmarks on the CBCTs. Subsequently, ANOVA was used to compute the variance of each observer on the pairwise distance (PWD).

The PWD-ANOVA method accounts for human observer variation allowing a better estimation of the of DIR errors.

ANOVA based analysis demonstrated that a classical DTA approach underestimated the precision for the mROI due to human observer variation by about 25%. The systematic error (accuracy) of mROI ranged from 0.13 to 0.17 mm; the variability (1 SD) (precision) ranged from 1.3 to 1.5 mm demonstrating that its performance is dominated by the precision.

Deformable image registration (DIR) is often validated based on a distance-to-agreement (DTA) criterion of automatically propagated anatomical landmarks that were manually identified. Due to human observer variability, however, the performance of the registration method is diluted. The purpose of this study was to evaluate an analysis of variance (ANOVA) based validation to account for such observer variation.