Data Processing in Multidimensional MRI For Biomarker Identification: Is It Necessary?

Abstract

Multidimensional MRI (MD-MRI) is an emerging technique that holds promise for identifying tissue characteristics that could be indicative of pathologies. Before these characteristics can be interpreted, MD-MRI measurements are converted into an spectrum. These spectra are then utilized to obtain some understanding of the underlying tissue microstructure, often through the use of statistical, machine learning, and mathematical modeling methods. The aim of this study was to compare outcomes of using unprocessed MDMRI signals for statistical regression in comparison to the corresponding spectra. Backed by a theoretical argument, we described an experimental procedure regressing both MDMRI signals and spectra to histological outcomes intrasubject. Through using multiple conventional ML methods, and a proposed method using convex sets, we aimed to see which yielded the highest accuracy. Both theory and experimental evidence suggest that, without a priori information, statistical regression was best performed on the MDMRI signal. We conclude, barring any a priori information regarding tissue changes, there is no significant advantage to performing regression analysis on reconstructed spectra in the process of biomarker identification.