Zusammenfassung
Introduction:
Brain morphometry studies typically utilize anatomical data with 1 mm3 isotropic resolution. However, the widespread availability of high-field MRI scanners and receive coil arrays have led to increased interest in submillimeter resolution data. While higher resolution may have an advantage (Bazin et al., 2014; Lüsebrink et al., 2013; Zaretskaya et al., 2017), lifting this constraint on voxel size opens the door to a variety of acquisition protocols with a wide range of image contrast and SNR. Here we aim to systematically investigate the effects of image SNR on surface reconstruction quality based on sub-millimeter MPRAGE data acquired at 3T.
Methods:
To systematically vary image SNR without changing voxel size, we acquired multiple repetitions of high-resolution (0.6 mm isotropic) MPRAGE data from 9 participants, and for each subject generated multiple image volumes with progressively higher SNR by averaging together increasing number of repetitions (i.e. averaging 1 to 6 (n=5) or 1 to 8 (n=4) repetitions, depending on the total number of repetitions acquired within the session).
Images were acquired on a 3T Siemens MAGNETOM Trio Tim system using a 32 channel coil and a 0.6 mm isotropic multiecho MPRAGE (van der Kouwe et al., 2008) protocol (TR/TE1/TE2/TI/FA/BW/ESP/matrix=2510 ms/2.88 ms/ 5.6 ms/1200 ms/7°/420 Hz/px/8.4 ms/400×400). To minimize blurring during the inversion recovery, we used acceleration in the partition direction (R=2) without partial Fourier in any direction and a slab-selective axial acquisition (slices per slab=224) to minimize the number of partition encoding steps.
Each average volume was used to generate cortical surfaces using FreeSurfer's native "hires" sub-millimeter reconstruction stream (Zaretskaya et al., 2017). To assess surface quality, for each set of surfaces (the gray-white and gray-CSF interface surfaces of each hemisphere) we computed the following parameters: (1) image SNR (the mean divided by the standard deviation of voxel intensities within the FreeSurfer white matter mask), (2) surface smoothness (defined as the median value of per-vertex local mean curvature), (3) number of topological defects in the initial surface, identified by FreeSurfer (Segonne et al., 2007), (4) median defect size. We also report gray-white matter contrast, derived from the reconstruction with 6 repetitions, and defined as the median of vertex-wise contrast values expressed as (white matter intensity – gray matter intensity)/(white matter intensity + gray matter intensity)⋅100.
Results:
Automatic reconstruction completed successfully in all but 2 cases. Both failed cases were comprised of a single repetition and could not complete automatically due to low SNR. The gray-white matter contrast of our data, based on averaging 6 repetitions for each subject, was 26.8 ± 0.4 (mean across subjects ± S.E.M). We observed a consistent increase in image SNR - accompanied by an increase in surface smoothness and by a decrease in the number of topological defects - with increasing number of repetitions included in the average. There was no change in the median defect size. While surface quality clearly improved with increasing SNR, gray-white surface smoothness and the number of defects did not reach a clear plateau even after 7 repetitions, suggesting that more repetitions may be needed to determine the SNR beyond which there is no noticeable improvement in surface quality. However, for this protocol and scanner, and across this small group of subjects, we find that qualitatively the performance of the surface reconstruction is adequate after averaging 5–6 repetitions.
Conclusions:
Overall we show that averaging multiple repetitions can compensate for the corresponding loss of image SNR due to increase in image resolution. High-quality reconstructions can hence be generated from voxels far smaller than conventional acquisitions, even at standard field strength, provided sufficient data quality.
