Premium
Quantification of pulmonary microcirculation by dynamic contrast‐enhanced magnetic resonance imaging: Comparison of four regularization methods
Author(s) -
Salehi Ravesh M.,
Brix G.,
Laun F. B.,
Kuder T. A.,
Puderbach M.,
LeyZaporozhan J.,
Ley S.,
Fieselmann A.,
Herrmann M. F.,
Schranz W.,
Semmler W.,
Risse F.
Publication year - 2013
Publication title -
magnetic resonance in medicine
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.696
H-Index - 225
eISSN - 1522-2594
pISSN - 0740-3194
DOI - 10.1002/mrm.24220
Subject(s) - deconvolution , tikhonov regularization , magnetic resonance imaging , regularization (linguistics) , mathematics , blind deconvolution , nuclear magnetic resonance , algorithm , mathematical analysis , physics , inverse problem , computer science , medicine , artificial intelligence , radiology
Tissue microcirculation can be quantified by a deconvolution analysis of concentration–time curves measured by dynamic contrast‐enhanced magnetic resonance imaging. However, deconvolution is an ill‐posed problem, which requires regularization of the solutions. In this work, four algebraic deconvolution/regularization methods were evaluated: truncated singular value decomposition and generalized Tikhonov regularization (GTR) in combination with the L‐curve criterion, a modified LCC (GTR‐MLCC), and a response function model that takes a‐priori knowledge into account. To this end, dynamic contrast‐enhanced magnetic resonance imaging data sets were simulated by an established physiologically reference model for different signal‐to‐noise ratios and measured on a 1.5‐T system in the lung of 10 healthy volunteers and 20 patients. Analysis of both the simulated and measured dynamic contrast‐enhanced magnetic resonance imaging datasets revealed that GTR in combination with the L‐curve criterion does not yield reliable and clinically useful results. The three other deconvolution/regularization algorithms resulted in almost identical microcirculatory parameter estimates for signal‐to‐noise ratios > 10. At low signal‐to‐noise ratios levels (<10) typically occurring in pathological lung regions, GTR in combination with a modified L‐curve criterion approximates the true response function much more accurately than truncated singular value decomposition and GTR in combination with response function model with a difference in accuracy of up to 76%. In conclusion, GTR in combination with a modified L‐curve criterion is recommended for the deconvolution of dynamic contrast‐enhanced magnetic resonance imaging curves measured in the lung parenchyma of patients with highly heterogeneous signal‐to‐noise ratios. Magn Reson Med, 2013. © 2012 Wiley Periodicals, Inc.