Logo PTB

Large-Scale data analysis for Magnetic Resonance Fingerprinting

Working group 8.42

 Content

Introduction

Magnetic resonance imaging is a standard medical imaging method where typically qualitative images are acquired for diagnostic purposes. Currently new imaging techniques are developed which aim at a quantitative characterization of the tissue parameters and which can be applied in clinical praxis (Working group 8.13). Quantitative imaging methods may potentially lead to improved diagnostics, e.g., when monitoring therapies. A promising quantitative imaging approach is provided by Magnetic Resonance Fingerprinting [Nature 495(7440): 187-192, 2013] which enables the simultaneous determination of several tissue parameters. The method employs a pseudo-random sequence of excitations of the magnetization together with an undersampling based imaging scheme. In this context, a challenging data analysis problem emerges where a huge number of tissue parameters (>100 000) needs to be determined from the recorded data. Until now, template matching methods have been applied for this task which utilize a library of test patterns generated by a physical model.

Within working group Data Analysis and Measurement Uncertainty a novel data analysis technique has been developed where the tissues parameters are determined by solving a large-scale regression problem. The new approach is specified in terms of a statistical model and provides improved estimates for the tissue parameters. In addition, the new method also yields uncertainties for the obtained tissue parameter estimates.

Figure 1: Magnetization dynamic of different brain tissue types for a random excitation sequence.

In Figure 1 the basic principle of the Fingerprinting approach is illustrated in terms of different magnetization dynamics as resulting for different tissue types. Figure 2 shows reconstruction results obtained in a simulation. As can be seen, the novel regression based approach yields smaller reconstruction errors when compared to the template method. Moreover, the remaining reconstruction errors are reasonably characterized by the uncertainties resulting from the novel data analysis technique.

Figure 2: Comparison of reconstruction results obtained within a simulation for the T1 relaxation time (in ms) by the template matching scheme (Template Method) and by the new approach (Regression Method). The new approach not only yields an improved quantitative imaging, but also provides corresponding uncertainties (Uncertainty). The simulation results demonstrate that these uncertainties reasonably characterize the accuracy (Absolute Error) of the novel quantitative imaging scheme.

Publication

  • Selma Metzner, Gerd Wübbeler, Clemens Elster, Approximate large-scale Bayesian spatial modeling with application to quantitative magnetic resonance imagingAStA Adv Stat Anal (2019), 103(3), 333-355, Opens external link in new windowhttps://doi.org/10.1007/s10182-018-00334-0
  • Gerd Wübbeler, Clemens Elster, A Large-Scale Optimization Method using a sparse Approximation of the Hessian for Magnetic Resonance FingerprintingSIAM Journal on Imaging Sciences, 10(3), 979-1004 Opens external link in new windowhttps://doi.org/10.1137/16M1095032