In the past, MR simulations have been used for answering particular methodological problems, optimizing imaging protocols and pulse sequences but also for training purposes. In order to allow for reasonable execution time in simulation procedure, MR simulators were usually confined to incorporation of simplified anatomical computational models of limited realism and/or simulations of basic MR physics concepts only.

Recent MR simulation studies have incorporated the extended Cardiac-Torso (XCAT) anthropomorphic computational phantom framework ¹ that allows for the generation of realistic human anatomy and physiology. However, realistic aspects of the experiment (such as excitation slice profile, B0 inhomogeneity, etc.) have not been investigated ^{2, 3} and compromises have been made (simulations not based on the Bloch equations).³ Moreover, those prior studies were focused to specialized MR applications only (such as in the area of Cardiovascular Magnetic Resonance - CMR).

The purpose of this study was the incorporation of the extended Cardiac-Torso (XCAT) anatomical phantom in MRISIMUL, a high performance multi-GPU (Graphics Processing Units) environment of realistic MR simulations.^{4, 5} Such a simulation platform would allow for general-purpose realistic MR simulations that make no assumptions with respect to the underlying parameter space.

MRISIMUL, a recently developed multi-GPU-based MR physics simulator, was utilized.^{4, 5} The XCAT model was used ¹ and a 3D computer model of the heart (myocardium and blood pool) at end-diastole was generated. The XCAT model consisted of 300 short-axis slices of 1536 x 1536 pixels with an isotropic voxel size of 0.5mm³. The resulting heart model consisted of 3854984 tissue voxels. A MATLAB framework was developed to define the spatial distribution of the cardiac tissues and assign MR properties to them (myocardium: T1 = 900 msec and T2 = 50 msec, blood: T1 = 1600 msec and T2 = 250 msec).

Two different CMR applications were studied. In the first case, a single-shot multi-slice bSSFP acquisition was performed for heart morphology evaluation. The sequence parameters for the simulation were the following: TR/TE = 2.06/1.03 msec, 490μsec sinc shaped RF pulse with 6 mm slice thickness and 70^o excitation flip angle, field of view = 360 x 270 mm, scan matrix size = 164 x 123, slice separation = 0 mm and receiver BW = 200kHz. The total number of time-steps was 1170400.

In the second case, a MOLLI pulse sequence was performed for T1 quantification of the myocardium and the blood pool on a mid-ventricular short-axis slice. A 5(3p)3 acquisition scheme was selected with initial TIs of 200 msec and 300 msec. The sequence parameters for the T1 mapping simulation were the following: TR/TE = 2.06/1.03 msec, the bSSFP readout used a 490μsec sinc shaped RF pulse with 6 mm slice thickness and 35^o excitation flip angle, slice thickness = 6mm, field of view = 360 x 270 mm, scan matrix size = 164 x 123 and receiver BW = 200kHz. The IR pulse was a hyperbolic secant adiabatic pulse with 4.74 msec duration. The total number of time-steps was 3300000.

In both aforementioned cases, a linear k-space trajectory was selected and a linear ramp up preparation of 10 pulses was used to reach steady state prior to the bSSFP readout. All the simulations were performed on a single-node, server style system consisting of 2 hexa-core (Intel E5-2630, 2.30 GHz) processors, 32 GB RAM and 4 Tesla C2075 GPU cards.

In Figure 1, twenty short-axis simulated images covering the entire heart from the apex towards the base are presented. Figure 2 depicts the raw magnitude images (cropped) sorted by inversion time and the resulting colored T1 map. The average myocardial T1 value was measured equal to 848±19 msec whereas the average blood T1 value was measured equal to 1574±12 msec. The underestimation of MOLLI on the true T1 values of blood and myocardium are in agreement with previous studies found in the literature.⁶Results show that the incorporation of realistic computational models of the human anatomy and physiology in a high performance multi-GPU MR simulation platform may benefit the design and optimization of MR clinical protocols and pulse sequences in the future. Although the examples presented in this study were limited within the field of CMR, realistic MR simulations could be performed for other MR areas as well. Last but not least, one should note that the wide availability of appropriate GPU hardware today allows for performing computationally demanding MR simulations on realistic anatomical models, which would not be the case only a few years earlier. Such realistic MR simulations could be further investigated for multi-modal imaging acquisitions (e.g. PET/MR).