An improved EM algorithm for solving MLE in constrained diffusion kurtosis imaging of human brain

Abstract

The displacement distribution of a water molecular is characterized mathematically as Gaussianity without considering potential diffusion barriers and compartments. However, this is not true in real scenario: most biological tissues are comprised of cell membranes, various intracellular and extracellular spaces, and of other compartments, where the water diffusion is referred to have a non-Gaussian distribution. Diffusion kurtosis imaging (DKI), recently considered to be one sensitive biomarker, is an extension of diffusion tensor imaging, which quantifies the degree of non-Gaussianity of the diffusion. This work proposes an efficient scheme of maximum likelihood estimation (MLE) in DKI: we start from the Rician noise model of the signal intensities. By augmenting a Von-Mises distributed latent phase variable, the Rician likelihood is transformed to a tractable joint density without loss of generality. A fast computational method, an expectation-maximization (EM) algorithm for MLE is proposed in DKI. To guarantee the physical relevance of the diffusion kurtosis we apply the ternary quartic (TQ) parametrization to utilize its positivity, which imposes the upper bound to the kurtosis. A Fisher-scoring method is used for achieving fast convergence of the individual diffusion compartments. In addition, we use the barrier method to constrain the lower bound to the kurtosis. The proposed estimation scheme is conducted on both synthetic and real data with an objective of healthy human brain. We compared the method with the other popular ones with promising performance shown in the results. Show more