Multivariate segmentation of brain tissues by fusion of MRI and DTI data

This paper proposes a method to improve brain-tissue segmentation, especially in subcortical region, by fusing the information in structural magnetic resonance (MR) images and diffusion tensor (DT) images in a sound statistical framework. The proposed method incorporates the information in DT images by parameterizing the space of diffusion tensors, in a principled and efficient manner, based on a set of independent orthogonal invariants. The proposed method couples the Markov tissue statistics of the structural-MR intensities with the tissue statistics of the DT invariants to define multivariate/joint probability density functions (PDFs) that differentiate brain tissues. The paper shows that while the information in DT images can allow improved differentiation between tissues in the subcortical region, which comprises anatomical structures having smooth (blob-like) shapes, it can produce unreliable results in the cortical regions that depict convoluted sulci/gyri. The proposed method exploits these characteristics of the images by introducing an appropriate anisotropic distance metric in the multivariate feature space.