BACKGROUND AND PURPOSE: To optimize couch shifts based on multiple region-of-interest (ROI) registrations and derive criteria for adaptive replanning for management of deformations in head-and-neck (H&N) cancer patients. MATERIALS AND METHODS: Eight ROIs containing bony structures were defined on the planning-CT and individually registered to daily cone-beam CTs for 19 H&N cancer patients. Online couch shifts were retrospectively optimized to correct the mean setup error over all ROIs (mean correction) or to minimize the maximum error (MiniMax correction). Residual error distributions were analyzed for both methods. The number of measurements before adaptive-intervention and corresponding action-level were optimized. RESULTS: Overall residual setup errors were smallest for the mean corrections, while MiniMax corrections reduced the largest errors. The percentage of fractions with residual errors >5 mm was 38% versus 19%. Reduction of deformations by single plan adaptation was most effective after eight fractions: systematic deformations reduced from 1.7 to 0.9 mm. Fifty percent of this reduction can already be achieved by replanning 1/3 of the patients. CONCLUSION: Two correction methods based on multiple ROI registration were introduced to manage setup errors from deformations that either minimize overall geometrical uncertainties or maximum errors. Moreover, the registrations could be used to select patient with large deformations for replanning.