Assessment on clinical data of nonlinear stochastic deconvolution versus SVD and block-circulant SVD methods for quantitative DSC-MRI

Introduction. Dynamic Susceptibility Contrast – Magnetic Resonance Imaging (DSC-MRI) allows to quantify Cerebral Blood Flow (CBF), Volume (CBV) and Mean Transit Time (MTT) by measuring Arterial Input Function AIF(t) and tissue concentration C(t) and estimating by deconvolution the tissue Residue function R(t): ( ) [ ( ) ( )] C t CBF AIF t R t = ⋅ ⊗ (eq.1). Singular Value Decomposition (SVD) [1] is currently the gold standard deconvolution method, but bears some limitations: nonphysiological oscillations and negative values in R(t), estimated CBF dependence on selected threshold, delay/dispersion in AIF [2]. To overcome SVD limitations, a block-Circulant matrix SVD (cSVD) [3] and Tikhonov regularization (TIKH) [4] have been proposed, although the latter has never been extensively applied to clinical data. In [5] we proposed a Nonlinear Stochastic Regularization (NSR) method able to account for smoothness and non-negativity of R(t) and dispersion of AIF. NSR [6] is a nonparametric Bayesian method which considers ( ) ( ) ( ) l R t CBF R t d t CBF e ⋅ = ⊗ ⋅ (eq.2), where ) / t exp( / 1 ) t ( d 1 1 θ θ − = (eq.3) stands for the dispersion and