Segmentation-based retrospective correction of intensity non-uniformity in multi-spectral MR images

Intensity non-uniformity in magnetic resonance (MR) images is an adverse phenomenon, which manifests itself as slow intensity variations of the same tissue over the image domain. It may have serious implications for MR image analysis. For example, intensity non-uniformity increases the overlap between intensity distributions of distinct tissues and therefore makes segmentation more difficult and less precise. Because correction of intensity nonuniformity and segmentation are inherently related problems, we propose a novel method, which interleaves them, so that they support each other and gradually improve, until final correction and segmentation is reached. We derive a parametric non-uniformity correction model in a form of a linear combination of non-linear basis functions. The non-uniformity correction is based on iterative minimization of class square-error, i.e. within-class scatter, of intensity distribution that is due to non-uniformity. For this purpose we employ a non-parametric segmentation method presented in MI 4684-41. We consider inter-spectral independent non-uniformity effects and provide corresponding non-uniformity correction models and algebra for computing the parameters. The proposed method is tested on simulated and real, single- and multi-spectral, MR brain images. The method does not induce additional intensity variations in simulated uniform images and efficiently removes non-uniformity of simulated and real MR images and thereby improves the results of segmentation.