Pathology-robustmr intensity normalizationwith global and local constraints

The intensity values in magnetic resonance (MR) images are not standardized. This prevents intensity comparison of different MR volumes that may be needed for visualization, intensity-based processing, and diagnosis. To this effect, this paper introduces a novel, pathology-robust MR intensity normalization algorithm that improves over the literature in three major aspects: 1) Pathology robustness: We achieve this by comparing the input MR volume with a reference volume, identifying the modes of their joint intensity distribution by the mean shift algorithm, and assigning each voxel a confidence value based on the distance of its intensity to the nearby mode. 2) Global and local analysis: We improve both the global similarity of intensities by matching the input and the reference histograms with histogram specification, and the local intensity similarity by minimizing the mean voxel intensity difference with dynamic programming. 3) Structure-preserving fusion of global and local approaches: The optimal fusion of global and local metrics is achieved by preserving the structures (defined as edges) in the normalized data compared with those in the input. We show the effectiveness of the proposed method with both visual and quantitative results.