QSM reconstruction challenge 2.0: A realistic in silico head phantom for MRI data simulation and evaluation of susceptibility mapping procedures

Abstract

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

QSM reconstruction challenge 2.0: Design and report of results

“…RC2 used a phantom with known ground truth susceptibility, which was derived from in vivo data through MR physics simulations 22 . While addressing shortcomings of RC1, attention was paid to ensure that using phantom data did not cause new weaknesses.…”

“…While addressing shortcomings of RC1, attention was paid to ensure that using phantom data did not cause new weaknesses. To avoid promoting piecewise smooth/contrast solutions, the susceptibility ground truth included physiological texture and realistic variations within structures, which was derived from R 1 and measurements 22 …”

“…A realistic numerical phantom made it possible to synthesize gradient echo acquisitions, which provided local field map estimates without background field remnants. Inconsistencies between the simulated data and known ground truth were restricted to complex Gaussian noise and intra‐voxel dephasing effects, simulated by down‐sampling from high‐resolution 0.65 mm to lower‐resolution 1 mm isotropic voxels 22 via k‐space cropping 7,23 . Due to the modular design of the simulations, other effects such as realistic background fields and shimming may be incorporated in a future challenge design.…”

“…Both in the first and second stages, two datasets were provided related to two different susceptibility models (namely SIM1 and SIM2) 22 . The main difference between both datasets was the presence of an intra‐hemispheric calcification (in SIM2), and different levels of susceptibility contrast between tissues, as seen in Figure 1.…”

“…For Stage 2, two datasets were generated, with peak SNRs of 100 and 1000, namely SNR1 and SNR2, respectively, for each susceptibility model (SIM1 and SIM2). The same k‐space cropping approach 22 was used to down‐sample the high‐resolution (0.65 mm to 1 mm) complex signal, ground truth susceptibility maps and segmentation labels. In the case of the susceptibility maps, the sharp edges between structures as well as the orders of magnitude larger susceptibility differences between air/bone and tissue resulted in severe Gibbs ringing artifacts, which were removed using sub‐voxel shifts 22,24 .…”