Introduction Proton resonance frequency (PRF)-shift MR thermometry is a promising tool for guiding thermal therapies in the treatment of liver tumors and heart arrhythmias, but is complicated by organ motion and respiration. To address motion, multi-baseline subtraction techniques have been proposed [1,2] that use a library of pre-treatment baseline images covering the cardiac and respiratory cycle. However, main field shifts due to lung and diaphragm motion can cause large inaccuracies in multi-baseline subtraction. In contrast, referenceless thermometry methods [3-5] based on polynomial regression of background phase are immune to motion and main field shifts. While referenceless methods can be accurate in most regions of the heart and liver, the background phase in some parts of these organs can require large polynomial orders to fit, leading to increased risk that the hot spot itself will be fit by the polynomial. We present a hybrid method for thermometry of moving organs that combines referenceless and multibaseline thermometry, and demonstrate that it estimates temperature with much lower error in volunteer heart and liver data than either method alone. Theory The algorithm is an extension of regularized referenceless thermometry [5]. We assume that three sources contribute to image phase during thermal treatment: 1) background anatomical phase (i.e., baseline image phase), 2) spatially smooth phase deviations from baseline caused by main field shifts (i.e., polynomial phase), and 3) focal, heat-induced phase shifts. This leads to the following treatment image model at voxel j: