Magnetic Resonance (MR) temperature mapping can be used to monitor temperature changes during minimally invasive thermal therapies. However, MR-thermometry contains artefacts caused by phase errors induced by organ motion in inhomogeneous magnetic fields. This paper proposes a novel correction strategy based on a Principal Component Analysis (PCA) to estimate magnetic field perturbation assuming a linear magnetic field variation with organ displacement. The correction method described in this paper consists of two steps: a magnetic field perturbation model is computed in a learning step; subsequently, during the intervention, this model is used to reconstruct the magnetic field perturbation corresponding to the actual organ position which in turns allow computation of motion corrected thermal maps.