Fast Joint Estimation and Selection of Mixture Models via Model Averaging: Application to Diffusion Compartment Imaging of the White Matter Microstructure

• Activity Number: 243
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Aymeric Stamm*+ and Olivier Commowick and Patrick Pérez and Christian Barillot and Simon K. Warfield
• Companies: Harvard Medical School/Boston Children's Hospital and Visages INSERM/INRIA U746, IRISA - UMR CNRS 6074, Rennes, France and Technicolor and Visages INSERM/INRIA U746, IRISA - UMR CNRS 6074, Rennes, France and Harvard Medical School/Boston Children's Hospital
• Keywords: diffusion MRI ; white matter microstructure ; mixture model ; model selection ; model averaging ; diffusion compartment imaging

Abstract

The white matter contains bundles of unioriented axons or fascicles in which water diffusion is restricted along the main axis, which can be used to infer white matter microstructure. Diffusion MRI indirectly measures diffusion at low pixel resolution and generally conflates multiple contributions arising from non-parallel fascicles. Mixture models are often used to disentangle these processes but require prior knowledge of their number. We propose a joint model estimation and selection to this problem. We fit a set of nested models with increasing complexity to the data and compute the associated Akaike weights (probability of a model to be the best Kullbach-Leibler model). We then extend each model by component duplication and re-indexing to make parameters comparable across models. We compute weighted average of each parameter using Akaike weights and simplify the resulting mixture by modularity clustering to remove component duplicates. The analysis of two brains scanned under same protocol yielded averaged models within seconds from the set of original nested models with a resulting number of fascicles in agreement across subjects and with generalization error model selection.