Here we propose a hybrid scheme of Parallel Tempering
(PT) and Levenberg-Marquardt (LM) approaches, named
PT/LM, for optimization problems of non-linear models.
The aim is to design an efficient and stable pipeline
for non-Gaussian diffusion metrics estimation from noisy
diffusion-weighted MRI data. Diffusional kurtosis (K)
and stretched exponential (
)
models were investigated. Our numerical and experimental
results, performed on ex-vivo healthy mouse brain at
11.7T, demonstrate that the proposed novel hybrid scheme
improves the efficiency and stability of conventional LM
based pipelines, providing less grainy non-Gaussian K-
and
-maps,
with higher contrast-to-noise ratio.