Datasets

The three datasets described below were used to evaluate the method.

The first is the 80 unrelated subjects subset of the Human Connectome Project (HCP80) dataset. All subjects are healthy adults. The following scans were acquired on a Siemens Skyra 3 T system with a 32-channel head coil (Van Essen et al., 2012, Van Essen et al., 2013):

All of these images were corrected for gradient nonlinearity induced distortions by the HCP pipeline (Glasser et al., 2013). The pipeline additionally corrected the intensities of the T₁- and T₂-weighted images for B₁ inhomogeneities. Geometric distortions due to B₀ inhomogeneity and eddy currents were corrected by the pipeline using EDDY and TOPUP, making use of the additional images with reversed phase encoding (Andersson et al., 2003, Andersson et al., 2012) Fractional anisotropy (FA) images were calculated using FSL's DTIFIT.

The HCP80 dataset contains 77 usable subjects, 10 of which were used for experimentation and 10 were manually labelled for the quantitative evaluation of our method. This left a training group of 57 subjects. All results shown in this paper were obtained by training the model on this training group and segmenting the subjects in the labelled group.

The second set of structural data were the young subjects of the publicly available dataset described by Forstmann et al. (2014). These data were acquired using a 7 T Siemens Magnetom MRI using a 24-channel head array Nova coil (NOVA Medical Inc., Wilmington MA):

The imaging slab for the limited field-of-view acquisition was 64 mm thick for the FLASH protocol and 76.8 mm for the MP2RAGE protocol and positioned to capture the subcortical structures. Both slab sequences consisted of axial slices tilted − 23° to the true axial plane in scanner coordinates. The QSM volume was calculated using the phase information of the FLASH MRI sequence and the method proposed by (Schweser et al., 2013).

The 7 T dataset consisted of 29 usable subjects; one subject was not used because registration of the whole-brain MP2RAGE volume to the FLASH slab failed. Fourteen of these subjects were male and 15 were female. Their mean age was 23.8 years with a standard deviation of 2.3 years. For quantitative evaluation, a ‘leave 5 out’ scheme was used, where per block of 5 or 4 subjects the method was trained on the other 24 or 25 subjects.

To investigate how well the method performs on data acquired using more typical protocols and in the presence of pathology, it was also applied to the dataset from a Huntington's disease (HD) study. HD is an informative test case as the disease is associated with severe atrophy of subcortical grey matter. These data and the manual segmentations are described in detail in the original papers (Douaud et al., 2006, Douaud et al., 2009) and acquired as part of the MIG-HD (Multicentric Intracerebral Grafting in HD) project.

The dataset consists of ten patients and six control subjects from the original study. The number of control subjects included was limited to six, as this was the only consistent subset of the control data that had both identical acquisition protocols for the structural images and for which diffusion data were available. For each subject the following scans were obtained on a General Electric 1.5 T Signa system with a birdcage head coil with 40 mT/m maximum gradient strength: