Motion corrected silent ZTE neuroimaging

Purpose To develop self‐navigated motion correction for 3D silent zero echo time (ZTE) based neuroimaging and characterize its performance for different types of head motion. Methods The proposed method termed MERLIN (Motion Estimation & Retrospective correction Leveraging Interleaved Navigators) achieves self‐navigation by using interleaved 3D phyllotaxis k‐space sampling. Low resolution navigator images are reconstructed continuously throughout the ZTE acquisition using a sliding window and co‐registered in image space relative to a fixed reference position. Rigid body motion corrections are then applied retrospectively to the k‐space trajectory and raw data and reconstructed into a final, high‐resolution ZTE image. Results MERLIN demonstrated successful and consistent motion correction for magnetization prepared ZTE images for a range of different instructed motion paradigms. The acoustic noise response of the self‐navigated phyllotaxis trajectory was found to be only slightly above ambient noise levels (<4 dBA). Conclusion Silent ZTE imaging combined with MERLIN addresses two major challenges intrinsic to MRI (i.e., subject motion and acoustic noise) in a synergistic and integrated manner without increase in scan time and thereby forms a versatile and powerful framework for clinical and research MR neuroimaging applications.


SI.1 | Phyllotaxis Trajectory
When designing the 3D phyllotaxis trajectory, the complete spoke distribution is produced by a fixed increment in the azimuthal angle by the golden angle ! ≈ 137.5 ∘ . Subsampling with a factor produce small angular steps if k is a Fibonacci number ( # ), which is demonstrated in Figure S1. The analytical expression valid for ∈ # is presented in Equation [1] of the main manuscript. Figure S2 shows Voronoi diagrams of the spoke distribution from the 3D spiral phyllotaxis as formulated by Piccini et al., 1 and compared to that used in MERLIN, demonstrating a non-isotropic FOV in the Piccini formulation.
Figure S1Azimuthal angle increment as a function of spoke subsampling factor. When is a Fibonacci number, the angular increment ( ! ) is small, which is required to maintain silent acquisition.

SI.2 | Navigator Resolution
An analysis of the optimal reconstructed resolution for the navigator images is shown in Figure S3. Motion corrupted data from one of the subjects in part 3 was processed with navigator resolution from 3-8 mm, followed by TGV reconstruction. The Average Edge Strength (AES) was then calculated for each image and the change relative to the 3 mm experiment was calculated. We observed reduced motion correction quality at lower navigator resolution ( Figure   S3A), which was supported by the quantitative AES analysis in Figure S3B.

SI.3 | Navigator Brain mask
A brain mask was generated using HD-BET 2 to mask out the brain for the registration. The brain mask was twice dilated using FSL maths 3 to include the skull. Example of a brain mask used in one of the experiments is shown in Figure S4. White arrows point to visible signal from the headrest and pads which we want to mask out.

Figure S4
Example of brain mask used for registration. The mask, here shown by outline only, is twice dilated to also cover the skull but to still exclude the signal from the headrest (white arrows) which will confound the registration since it does not move with the head, resulting in a non-rigid motion.

Average Edge Strength -AES
The average edge strength (AES) 4 is a global measure of the sharpness in the tissue interfaces. A binary edge mask was calculated using a Canny edge filter (with = 2) 5,6 for each axial slice on the first static image for each subject and the edge strength magnitude was calculated using a Sobel filter along all three axes 7 . The AES was then calculated as the mean edge strength within the brain mask for all voxels identified by the Canny edge filter. The AES is sensitive to global intensity scaling, so images were normalised by the 99% percentile image intensity before calculation. The AES is calculated individually for each image, in contrast to the SSIM which is a comparison between two images. We therefore used the percentage difference in AES between the reference and comparison images as our comparison metric

Structural Similarity Index Measurement -SSIM
The structural similarity index measure (SSIM) 8 is a compound measure which compares two images based on intensity, contrast and structure. It is evaluated for each voxel in an image based on a sliding window, from which a mean SSIM (mSSIM) can be calculated. The SSIM was evaluated within the brain mask using the parameters suggested by Wang et al. 8 (11 × 11 × 11 Gaussian kernel with = 1.5, $ = 0.01, % = 0.03), to produce an SSIM image, and the mSSIM was then calculated. The advantage of mSSIM is that it is a normalised metric between 0 and 1, and thus directly comparable to other studies. An mSSIM=1 is perfect agreement.

Figure S5
shows an expanded version of Figure 6 from the main manuscript, including the detrended motion traces from each acquisition. Figure S6-8 shows axial and sagittal slices from all three subjects, with and without motion correction, for all instructed motion paradigms. Images have been bias field corrected and windowed for optimal viewing quality.

Figure S6
Representative axial and sagittal slices from subject 1, from all motion paradigms, together with the estimated motion traces. The grey region in the time series indicates the short time window at the beginning of the acquisition which is not motion corrected including dummy segments and the low resolution WASPI acquisition.

Figure S7
Representative axial and sagittal slices from subject 2, from all motion paradigms, together with the estimated motion traces. The grey region in the time series indicates the short time window at the beginning of the acquisition which is not motion corrected including dummy segments and the low resolution WASPI acquisition.

Figure S8
Representative axial and sagittal slices from subject 3, from all motion paradigms, together with the estimated motion traces. The gray region in the time series indicates the short time window at the beginning of the acquisition which is not motion corrected including dummy segments and the low resolution WASPI acquisition.