High-resolution imaging at high fields (0.4-0.7mm isotropic) is typically acquired in over 10 minutes, after which even a very cooperative subject is expected to have moved on the order of 1-2 mm. Various motion tracking approaches have been proposed, such as image and k-space navigators¹, or motion tracking with external devices². MR-based methods avoid the need for additional hardware and synchronization, but tend to affect sequence timings and/or disturb the image contrast. Here, we demonstrate the ability to track Discrete Off-resonance MR markers with Three Spokes (“trackDOTS”)³ for retrospective motion correction in high-resolution anatomical imaging. This approach allows fast motion tracking (under 50ms per measurement) with negligible impact on the water signal.

A healthy subject was scanned on a 7T MRI scanner (Siemens) with a 32ch RF coil (Nova Medical), while wearing an adapted EEG cap fitted with 12 trackDOTS (Fig.1a). A high-resolution MP2RAGE sequence⁴ was acquired, interleaved with trackDOTS-selective 1D projections for motion tracking. Post-acquisition, motion parameter timecourses were estimated based on the projections, and used to realign the MP2RAGE k-space planes.

TrackDOTS: the 12 markers were 8mm-diameter hollow PEEK spheres (Fig.1b) filled with acetic acid doped with MnCl. One of the two acetic acid proton resonances peaks at 11.6ppm (the farthest from water and subcutaneous fat), and was chosen for the trackDOTS-selective excitations.

Image acquisition: initially, two quick 3D GRE images were acquired: a water image (TR/TE=3.5/1.5ms, α=5°, iPAT=4, res=3.5×3.5×1.75mm, acq.time=8.3s), and a trackDOTS-selective image (excitation freq. offset=2000Hz, TR/TE=10.0/3.1ms, α=10°, iPAT=8, res=1.75×1.75×1.75mm, acq.time=20s). The MP2RAGE was then applied (TR/TE/TI₁/TI₂=6000/4.94/800/2700ms, α₁/α₂=7°/5°, iPAT=3, res=0.6×0.6×0.6mm, acq.time=10min). This sequence was modified with two additional blocks before each inversion pulse: a FatNav block to track the head position based on fat-selective excitation⁵ (to be used as ground truth), and a “DotNav” block to measure the trackDOTS (Fig.1c). The DotNav comprised 3 orthogonal projections along the x, y and z-axes (freq. offset=2000Hz, TR/TE=15.0/3.1ms, α=30°, res=1.75mm, acq.time=45ms); “pre-phaser” gradients were also applied to give a modulation of π radians across one marker diameter, mitigating off-resonance contributions from larger structures.

Motion estimation and correction: the trackDOTS-selective 3D-GRE image was used for an initial identification of marker positions. Based on these positions, the water GRE image was used to optimize 12 coil combinations yielding the highest sensitivity to each marker (and the region within a 26mm radius) while suppressing the other regions. These combinations were then applied to the DotNav projections from the MP2RAGE, creating 12 marker-specific x, y and z-projections per time point. For each projection, the signal maximum was localized near the initial position (from the GRE), and the signal was integrated over an 8mm window to find its center of mass. Poor-quality projections (low peak amplitude/high signal contamination) were excluded from the set. To estimate head motion at each point t, 3 rotation and 3 translation parameters were defined in a 3×4 registration matrix M^t, mapping positions from a reference set of markers P⁰ to the positions at point t, P^t. P⁰ was created with the initial positions in the GRE, for which all coordinates were reliably known. With this choice, the minimization of P^t - M^t P⁰ allowed the use of partial marker coordinates, maximizing the available information. The estimated translations and rotations were then applied to each k-space plane of the MP2RAGE, and corrected images were reconstructed via NUFFT⁵.

The optimized coil combinations demonstrated visibly higher specificity than when using the nearest coil (Fig.2a). The pre-phasers also proved beneficial to reduce signal contaminations in the projections, as observed in a separate experiment (Fig.2b). As expected, phase profiles with periods below 16mm (the value used for the DotNavs) would have lead to a reduction in the marker signals as well, despite the extended artifact suppression. The DotNav projections yielded 18–24 reliable projection peaks per time point. The DotNav motion timecourses (45ms per measurement) followed very similar trends to the ground-truth given by the FatNavs (1.2s per measurement) (Fig.3), with an average absolute deviation between the two of 0.09±0.08mm (translations) and 0.20°±0.19° (rotations). After reconstruction, the motion-corrected images were clearly superior in quality to the uncorrected image, with visibly sharper features along the cortical surface and ventricles (Fig.4).The use of discrete off-resonance markers is suitable for head motion tracking during structural MRI acquisitions, and can significantly improve the quality of motion-corrupted high-resolution anatomical images. Additional trackDOTS measurements could still be incorporated in the MP2RAGE for tracking closer to the actual data acquisition (ex: between the two GRE readouts), for further improvements. Future work will also explore trackDOTS-based dynamic shimming in long-TE GRE acquisitions.