Building bridges between optical microscopy and magnetic resonance imaging is one major goal of structural imaging. Higher field strength allows higher signal to noise ratio (SNR) and in return enables increased spatial resolution. Increased spatial resolution of structural data may allow more accurate diagnostics and may lead to superior results [1], e.g. due to decreased partial volume effects and therefore potentially more accurate segmentations. However, high spatial resolution imaging with sufficient SNR prolongs time of acquisition. Even experienced subjects tend to move a few hundred micrometers to millimeters during an hour of scan time (Fig. 1), limiting the truly achievable resolution for in vivo imaging. Utilizing highly accurate prospective motion correction techniques even subtle motion can be corrected during image acquisition rendering ultra-high resolution imaging possible [2]. Here we demonstrate our work on the acquisition and reconstruction of the currently highest resolution MPRAGE at 7T.

We have acquired six ultra-high resolution MPRAGEs with an isotropic resolution of 250 µm of one Caucasian male subject (32 years) in four different sessions with full brain coverage. For comparison, one MPRAGE with an isotropic resolution of 500 µm has been acquired of the same subject. The experiment was performed with the approval of the ethics committee of the Otto-von-Guericke University, Magdeburg, Germany. Written informed consent was obtained from the subject prior to the scans.

Scanning has been conducted at a whole body 7T MRI (Siemens Healthcare, Erlangen, Germany) using a 32-channel head coil (Nova Medical, Wilmington, MA, USA). Scan parameters of the 250 µm MPRAGEs were: TR: 3580 ms, TE: 2.41 ms, TI: 1210 ms, FA: 5 °, BW: 440 Hz/px, slice partial fourier 6/8, matrix size: 880x880x640, ToA per average: ~53 min. The scan parameters of the 500 µm MPRAGE were: TR: 2740 ms, TE: 3.24 ms, TI: 1050 ms, FA: 5 °, BW: 130 Hz/px, matrix size: 416x416x352, ToA: ~19 min. All images have been acquired utilizing a prospective motion tracking system (MT384i, Metria Innovation Inc., Milwaukee, WI, USA). The tracking marker was attached to an individually created dental retainer, as described in detail in [2].

Image reconstruction of the ultra-high resolution data has been carried out offline as the raw data of each volume amounts to approximately 140GB making online reconstruction impossible. Combination of the different images of the 32-channels was performed by sum of squares (SoS) or adaptive reconstruction [3]. After reconstruction the images were co-registered using SPM and averaged. Noise outside the brain was removed by masking subsequently. Bias field correction of the high and ultra-high resolution images has been conducted with SPM. Additionally, noise was reduced in the ultra-high resolution images using a 2D adaptive Wiener filter with a 3x3 neighborhood in MATLAB (Fig. 2).

The figures 3 to 5 display the acquired images with an isotropic resolution of 500 µm and 250 µm. The 250 µm images have been averaged six times and reconstructed using adaptive combination and sum of squares (Fig. 3). After bias field correction and filtering it can be seen (Fig. 4) that reconstruction with adaptive combination compared to SoS results in a more homogenous intensity distribution. The ultra-high resolution images show much finer and sharper structures compared to 500 µm. In figure 5 for example the dura mater is outlined superiorly at ultra-high resolution. Furthermore, the amygdalo-hippocampal border can be differentiated better at higher resolutions, as indicated in [4], despite the SNR drop in the temporal lobe.The acquired volume probably is the highest resolution in vivo MPRAGE with full brain coverage until today. As expected, the reconstruction with adaptive combination outperforms sum of squares reconstruction for images with low SNR [3]. As image acquisition is exceptionally long, such high resolution is viable rather for the creation of more accurate atlases than for clinical routine. However, the higher spatial resolution also may lead to more precise results for the estimation of cortical thickness or for morphometry and volumetric techniques. Especially inferior parts of the brain (e.g. temporal lobes, cerebellum) suffer from low SNR due to inhomogeneous excitation. This can potentially be overcome by acquiring more averages, improvement of the sequence itself (e.g. B1-independent inversion pulse) or improvement in RF coil technology. More sophisticated noise filters should be applied to avoid inherent blurring effects of linear filters. Using a Wiener wavelet filter followed by speckle reducing anisotropic diffusion filter may additionally reduce the need for several averages [5].