Bayesian model selection and classification: application to brain tissues through T distribution
Iftah Gideoni · cIRcle (University of British Columbia) · 2009
A Bayesian procedure for model selection, parameter estimation and classification, using models of non-orthogonal basis functions, is applied to the problem of T2 decay rate distributions in brain tissues. The feasibility of generating reliable synthetic images of tissue-classified pixels is examined. The work determines, for the first time, the Bayesian probability of existence of short (5-15ms) T2 component in the brain tissues, and found it to be higher than 99% for all white matter tissues and higher than 80% for all gray matter tissues except Cortical Gray . The probability of having no more than three components of decaying exponents in the Ti distributions of the brain tissues, is found to be higher than 90% for all the tissues. We arrive to these findings through the use of models which are parameterized by highly coupled parameters, and the use of multi-dimensional search in the space of these models.