Data-driven separation of MRI signal components for tissue characterization

Abstract

Purpose: MRI can be utilized for quantitative characterization of tissue. To assess e.g. water fractions or diffusion coefficients for compartments in the brain, a decomposition of the signal is necessary. Imposing standard models carries the risk of estimating biased parameters if model assumptions are violated. This work introduces a data-driven multicomponent analysis, the monotonous slope non-negative matrix factorization (msNMF), tailored to extract data features expected in MR signals.

Methods: The msNMF was implemented by extending the standard NMF with monotonicity constraints on the signal profiles and their first derivatives. The method was validated using simulated data, and subsequently applied to both ex vivo DWI data and in vivo relaxometry data. Reproducibility of the method was tested using the latter.

Results: The msNMF recovered the multi-exponential signals in the simulated data and showed superiority to standard NMF (based on the explained variance, area under the ROC curve, and coefficient of variation). Diffusion components extracted from the DWI data reflected the cell density of the underlying tissue. The relaxometry analysis resulted in estimates of edema water fractions (EWF) highly correlated with published results, and demonstrated acceptable reproducibility.

Conclusion: The msNMF can robustly separate MR signals into components with relation to the underlying tissue composition, and may potentially be useful for e.g. tumor tissue characterization.

Keywords: Data-driven decomposition; Diffusion; Magnetic resonance imaging; Monotonous slope; Relaxometry; Tissue characterization; non-negative matrix factorization.