UHF 7T MR scanners offers the possibility to acquire very high resolution in-vivo images, providing a new insight into human brain structural characterization. Nevertheless, in order to obtain highly contrasted and highly spatially resolved atlas, and to compensate for the drop in SNR related to reduction of the voxel size, averaging data among several subjects is needed. Construction of such an average template requires first to acquire high quality 3D T1-weighted images corrected for B1 heterogeneity before performing a robust spatial coregistration and normalization without overfitting the data that could lead to an artificial morphing. We present here an automatic pipeline that generates a high-resolution ((0.5mm)³ isotropic voxel) whole brain T1-weighted template built from MP2RAGE acquisitions obtained in 30 healthy controls at 7T.

Participants:

30 young healthy subjects without any neurological or psychiatric disorders gave their written consent to participate to the study, which was approved by the local ethics committee.

Data acquisition MRI:

MR explorations were performed on a 7T Magnetom step 2 system (Siemens, Erlangen, Germany), using a volume transmit/32-channel receive head coil (NOVA). In order to minimize the effects of the large spatial inhomogeneity of the transmit B1 field at 7T over the whole human brain, the MP2RAGE¹ sequence was chosen. After third order B0 shimming and B1 calibration, a sagittal T1-weighted 3D-MP2RAGE image was acquired (TR/TE=5000/3.13 ms, TI1/2=900ms/2750 ms, FOV=240 mm, voxel size=0.6*0.6*0.6mm3, flip angle1/2 = 6°/5°, slices = 256, GRAPPA = 3; TA=10.12min). Corrected T1-weighted images were computed from the combination of the images acquired with inversion times of TI1 and TI2 following the procedures of Marques et al.¹.

Automatic Pipeline of average template construction:

A fully automatic pipeline was built to generate a high quality MRI template (Fig 1). The method combines the multi-scale average model construction² and the Symmetric Group-Wise Normalization (SyGN^3,4) , that retains both the appearance and shape variation. In summary, each T1w volume is successively resampled from isotropic voxel size of 0.5, 1, 2, 4 to 8mm and blurred using gaussian kernel sizes from 1, 2, 4, 8 to 16 (fwhm). First iterations of non-linear registration (SyGN^3,4) are estimated from the images at the lowest resolution (8mm). All 8mm-downsampled source images are co-registered, subsequently producing, after 4 iterations the first 8mm average-template. This 8mm template is then over-sampled through trilinear interpolation to create a new 4mm-oversampled template, used as the next target for the non-linear registration of all 4mm-downsampled source images. This process is iterative until the construction of the template at the highest level of resolution (0.5mm). In this manner, large smooth deformations are recovered first, and finer, more local, deformations are recovered last. The result is a 3D asymmetric average template of MRI images from the whole database (N=30) that reduces dependence on individual anatomy and produces a template with high feature sharpness. All registrations were computed on a grid computing cluster running on Linux workstations (Ubuntu Linux 14.04) and using the SunGridEngine to distribute computations among the machines.

The final average template results (Fig. 2a) in a 3D T1w MR volume achieving i) high SNR, ii) good homogeneity over the whole brain, and iii) high resolution with voxel size of (0.5 mm)³.

Compared to a single subject volume (Fig. 2b top), the average template (Fig. 2b bottom), allows a clear depiction of the cerebellum branches (Fig 2b bottom_left), as well as the claustrum and different thalamic nuclei (Fig. 2b bottom_central) and the subparts of hippocampus (Fig 2b bottom_right). This direct comparison demonstrates the importance of averaging several data sets after an optimized spatial normalization using multi-scale and nonlinear registration to better depict substructures.

The 7TAMIbrain_{T1w_30} obtained from an optimized multiscale, nonlinear registration pipeline applied to B1-corrected whole-brain 3D MP2RAGE images of 30 healthy subjects, allows the accurate depiction of small brain structures such as the claustrum, thalamic nuclei, subparts of hippocampus and small cerebellum branches that may be used to create a high resolution atlas⁵. This smooth procedure to the construction of the high resolution template limits uses limiting amount of morphing preserving native geometrical attributes of the images.Such a template can be used for spatial registration of high resolution 7T images of individuals. It can be also used as a basis to create a precise atlas at 7T.