Normalization of joint image-intensity statistics in MRI using the Kullback-Leibler divergence

We describe a novel algorithm for altering global statistics of magnetic resonance images (MRI) to fit a model distribution while preserving local feature contrast. Our algorithm estimates a multiplicative correction field that alters the intensity statistics of an image or set of images to best match those of a model. This is achieved by minimizing the Kullback-Leibler divergence between the observed and desired intensity distributions. This procedure is effective for the discovery and removal of undesirable intra-individual and inter-individual signal intensity changes caused by developmental processes, disease processes or MR scanner intensity artifacts. Ultimately our goal is to improve the quality of segmentations obtained by classification of tissues on the basis of signal intensities by removing undesirable signal differences both within a subject, where tissue of the same composition may image differently in different parts of the acquisition volume, and between subjects in cases where both inter-subject and inter-acquisition variability are confounds. Validation experiments with synthetic data indicate the algorithm can successfully remove typical signal intensity inhomogeneities, and illustrative results demonstrate successful intensity normalization applied to a segmentation problem.