The aim of this study was to develop and demonstrate a supervised learning algorithm to reconstruct MR images acquired in highly in-homogeneous magnetic fields. Brain images were used to train a deep neural network. This was performed for image sizes of 32 x 32 and 64 x 64. Results obtained demonstrate REMODEL’s ability to reconstruct the images obtained in in-homogeneous magnetic fields of up to ±50 kHz with high fidelity. The root-mean-square-error for these reconstructions compared to the uncorrupted ground truth was lesser than 0.15 and significantly lesser than the corrupted images.