Anatomical misregistration and background phase variation due to physiological motion are two major sources of inaccuracy in PRF thermometry of the liver. We propose a pipeline consisting of 3D segmented EPI acquisition, 3D deformable motion correction and PCA-based background phase correction for liver thermometry. 3D acquisition results in an increase in SNR and makes 3D registration possible. PCA-based background phase correction improves ΔT quality over a more commonly used dictionary-based method. The proposed pipeline was used to achieve 75th percentile ΔT bias and standard deviation values of 1.70°C and 1.65°C respectively over whole-liver in one healthy volunteer.