The reconstruction of water and fat images based on
chemical shift-encoded MRI has been under intensive investigation
ever since 1980's, due to its significance in biomedical and
clinical researches. Various approaches have been proposed, but
robust reconstruction with respect to high and ultra-high MRI
remains challenging. Major challenges include: (1) Ill-posedness
arising from the non-linear MR signal model. Multiple solutions may
exist at pixels, whereas improper selection of solutions leads to
severe estimation errors. (2) Large field inhomogeneity variation in
MR field complicates reconstruction. Approaches validated on low
field inhomogeneities may disfunction in the presence of large field
inhomogeneity variation. (3) Critical demanding of biomedical and
clinical applications on accuracy and robustness. Estimates must be
at minimal error rate to be acceptable.
Since estimation of MR field map is crucial to water and fat image
reconstruction, this dissertation mainly focuses on robust recovery
of field map. The approach developed with this thesis research
consists of a Markov random field (MRF) based energy model for
casting the estimation of field map to extensively studied MRF
energy optimization, and a novel Iterated Conditional Modes (ICM)
algorithm providing high-performance MRF energy optimization.
Compared to MRF energy model seen elsewhere, the proposed NLR-MRF
model is characteristic of non-linear least residual data cost terms
and background masking accounted for improvement of accuracy and
efficiency. The major components of the novel ICM algorithm are the
stability tracking (ST) and median initialization algorithms. The
stability tracking algorithm dynamically keeps track of iterative
stability at each pixel to avoid redundant computations. It is
demonstrated that with an optimal configuration, the ST algorithm
can substantially speed up the ICM iterative computations with
accuracy compromise. Median based algorithms are developed to
address the high sensitivity of ICM to initialization. By assuming
the quasi-unimodality of MR field inhomogeneity, the median-based
initialization algorithm identifies all unambiguous pixels in the
whole MR field, estimates the median and uniformly or block-wisely
sets the median as the initial guess. Experimental validation on
synthetic and a large group of in vivo 7 Tesla mouse datasets
demonstrate the robustness of the proposed approach.