Quantitative susceptibility mapping (QSM) is a recent in vivo magnetic resonance imaging (MRI) technique that provides quantitative information about the bulk magnetic susceptibility distribution in tissues, a promising measure for studying brain iron¹. Hitherto, susceptibility maps have been analyzed almost exclusively using so called region-of-interest (ROI)-based analysis techniques, either by laborious manual segmentation or employing segmentation algorithms using T₁-weighted (T₁w) images². The primary disadvantage of this approach is that it only allows the study of a few pre-defined anatomical regions.

A voxel-based analysis (VBA) of susceptibility maps would facilitate a better understanding of the intricate anatomical structure (e.g. sub-nuclear regions) of deep gray matter and its relation to diseases and normal aging. However, this requires a susceptibility brain template, which is currently not available.

In the present work, we developed and quantitatively assessed six strategies for creating a susceptibility brain template for VBA based on ANTs³, representing the first step toward an understanding of sub-nuclear susceptibility changes without the need for a priori information.

We compared six strategies for creating QSM templates: The direct approach (D) used susceptibility maps as an input to the template generation algorithm. The conventional approach (C) created a T₁w template and applied the involved warps to the individual susceptibility maps. The resampled conventional approach (rC) resampled the T₁w images to the same voxel size as the QSM before the processing. The rescaled QSM approach (rQ) used as an input for the algorithm susceptibility maps that were rescaled to a similar dynamic range of image intensities as present in conventional T₁w images. The hybrid approach (H) used a linear combination of T₁w images and susceptibility maps as input. The multi-modal approach (M) used the multi-modal ANTs template generation algorithm with T₁w images and rescaled susceptibility maps as input. For methods rQ, H, and M the optimal parameters were determined before comparing with other methods by using a wide range of parameters.

Data acquisition and reconstruction: We applied all strategies to a cohort of 10 healthy adults (5 male, 50±1 years). Subjects were scanned on a 3T GE Signa Excite HD 12.0 using a multi-channel head-neck coil. We used a 3D GRE sequence to acquire data for QSM (matrix 512x192x64, 256x192x128mm³, TE/TR=22ms/40ms, BW=13.9kHz, flip=12°) and a 3D magnetization-prepared FSPGR sequence to acquire T₁w images (TE/TI/TR=2.8ms/900ms/5.9ms, flip=10°, 1mm isotropic). Susceptibility maps were reconstructed from raw k-space data using scalar-phase-matching⁴, gradient unwarping⁵, best-path unwrapping⁶, V-SHARP^1,7, and HEIDI⁸.

Analysis: Since an objective analysis of templates is difficult, we decided to assess the results by visual rating. Three blinded raters with several years of experience in neuroimaging assessed the templates in four anatomical regions: basal ganglia, thalamus, venous vasculature, and motor cortex. All templates were compared pair-wise (side-by-side) in a win-lose fashion with respect to visual similarity to the single-subject susceptibility maps. Finally, the number of wins were summed for each strategy and region.

Figure 1 shows exemplary slices of the templates obtained with the different strategies on the level of the basal ganglia. Figure 2 summarizes the quantitative comparison. For strategy D the algorithm reproducibly delivered unusable results. Strategies rQ, M, and H resulted in templates with visually relatively similar quality. Method rQ (closely followed by method M) obtained the best ratings in the basal ganglia and the thalamus. In the vasculature and the cortex, methods M and H yielded the best ratings. The multi-modal approach (M) was the best overall compromise with relatively high ratings in all regions. Figure 3 shows exemplary slices in the respective regions. The conventional methods C and rC resulted in substantially lower ratings.

This is the first study that proposed and systematically compared different strategies for brain template generation based on quantitative susceptibility maps. We found that the optimal strategy depended on the region of interest in the brain. The best techniques for each region utilize images with the best contrast in the respective anatomical areas, i.e. T₁w in the cortex and QSM in the basal ganglia. Failure of strategy D can be attributed to the optimization of the ANTs algorithms for “conventional” images; compared to those images, susceptibility maps have relatively small intensities (sub-ppm) equally distributed around zero.

The proposed pre-processing schemes (methods rQ and H) allow the use of existing template generation methods to create magnetic susceptibility brain templates with high visual quality (Fig. 3). This sets the foundation not only for VBA of susceptibility in the human brain but also for the creation of high quality atlases of the brain’s susceptibility distribution.