A Machine Learning Approach to Diffusion MRI Partial Volume Estimation

Tissue-type partial volume modelling is generally an ill-posed problem in single-shell diffusion MRI. On the other hand, T1w images are typically acquired along with the diffusion data, and allow for an accurate estimation of the tissue partial volume fractions (PVFs). We propose in this paper to compare different data driven approach to predict the T1w-derived PVFs from the diffusion data. The aim is to alleviate the within subject mis-registration between the two modalities. Our experiments show that the random forests is the most accurate and scalable method for predicting the tissue partial volume fractions. Additionally, such predictions can be used to inform the fitting of the two-compartment model to retrieve a diffusion tensor that is not biased by partial volume effects or constraints.