Co-registered hyperpolarized gas and proton MRI are required to calculate functional lung biomarkers using semi-automated pipelines. Automated registration between different spectral acquisitions is difficult due to differences in contrast and imaging features between different spectral images. Convolutional neural networks create abstract representations of images that may overcome these feature differences. We retrospectively pooled data sets previously registered using a semi-automated pipeline and applied random two-dimensional affine transformations and noise. Neural networks generated inverse transformation matrices from these data to correct the applied transformations. The trained network successfully corrected mis-alignment with an average error of less than one pixel.