Diffusion Kurtosis Imaging of Neonatal Spinal Cord in Clinical Routine

Abstract

Diffusion kurtosis imaging (DKI) has undisputed advantages over the more classical diffusion magnetic resonance imaging (dMRI) as witnessed by the fast-increasing number of clinical applications and software packages widely adopted in brain imaging. However, in the neonatal setting, DKI is still largely underutilized, in particular in spinal cord (SC) imaging, because of its inherently demanding technological requirements. Due to its extreme sensitivity to non-Gaussian diffusion, DKI proves particularly suitable for detecting complex, subtle, fast microstructural changes occurring in this area at this early and critical stage of development, which are not identifiable with only DTI. Given the multiplicity of congenital anomalies of the spinal canal, their crucial effect on later developmental outcome, and the close interconnection between the SC region and the brain above, managing to apply such a method to the neonatal cohort becomes of utmost importance. This study will (i) mention current methodological challenges associated with the application of advanced dMRI methods, like DKI, in early infancy, (ii) illustrate the first semi-automated pipeline built on Spinal Cord Toolbox for handling the DKI data of neonatal SC, from acquisition setting to estimation of diffusion measures, through accurate adjustment of processing algorithms customized for adult SC, and (iii) present results of its application in a pilot clinical case study. With the proposed pipeline, we preliminarily show that DKI is more sensitive than DTI-related measures to alterations caused by brain white matter injuries in the underlying cervical SC.

Keywords: diffusion kurtosis imaging (DKI); diffusion tensor imaging (DTI); image processing pipeline; neonatal imaging; punctate white matter lesions; spinal cord.

Figure 1

Overall processing pipeline: the designed pipeline allows complete handling of diffusion kurtosis imaging scan of neonatal spinal cord from acquisition setup to preprocessing, processing, and postprocessing.

Figure 2

3D view of final diffusion acquisition scheme: directions of diffusion-sensitizing gradients relative to each b-value are displayed in three different colors as reported in the legend. Units are in s/mm². Markers indicate polarity: dots are the polarities in the set; asterisks are their opposite.

Figure 3

Visual inspection of denoising: The denoising of Patch2Self is compared against the original noisy image along with their corresponding residuals for each (A) b = 0, (B) b = 700, and (C) b = 2,100 s/mm² shells, respectively. Notice that Patch2Self does not show any anatomical structure in the corresponding residual plots and likely neither introducing structural artifacts.

Figure 4

Preprocessing: diffusion kurtosis imaging scan through preprocessing steps for one example subject: (A) field-of-view reduction, (B) motion correction, (C) segmentation: deep learning segmentation algorithm generally achieves satisfactory results in spinal cord detection, (D) example of artifactual slice due to poor fat saturation, causing the fat to alias on the spinal cord area, and (E) requiring manual correction of segmentation.

Figure 5

Diffusion kurtosis imaging scan overlaid on structural 3dT1w image: while both images are clearly not registered along the antero-posterior direction due to the very strong susceptibility artifact, the z-location is similar: see how the bottom tip of the cerebellum is consistent for the two scans.

Figure 6

Registration with PAM50 atlas and region-of-interest detection through atlas-based approach: (A) PAM50 atlas' cord segmentation binary mask; (B) white matter; and (C) gray matter probabilistic masks warped to the subject's diffusion kurtosis imaging motion-corrected mean image.

Figure 7

Effects of Patch2Self denoising on noise at different diffusion weightings. (A) Average signal-to-noise ratio computed on b = 0 images, (B) b = 700, and (C) b = 2,100 s/mm² increases in all the cohort, across C1–C4 vertebral levels under analysis, when including denoising with Patch2Self algorithm in the processing pipeline.

Figure 8

Box plots quantifying the increase in R² metric after fitting downstream the diffusion tensor imaging and MSDKI models for the whole spinal cord volume across all subjects. The R² improvements in each case are plotted by subtracting the scores of model fitting on undenoised data (raw) from R² of fitting each denoised output. Note that the consistency of microstructure model fitting on Patch2Self (P2S) denoised data is higher than that obtained from Marchenko–Pastur, especially as regards MSDKI model. Ns, 5.00e−02 < p ≤ 1.00e + 00; *, 1.00e−02 < p ≤ 5.00e−02; **, 1.00e−03 < p ≤ 1.00e−02; ***, 1.00e−04 < p ≤ 1.00e−03; ****, p ≤ 1.00e−04 in two-sided t-test with Bonferroni correction.

Figure 9

Mean signal kurtosis (MSK) decreases in neonatal periventricular white matter injuries: (A) white matter, (B) gray matter, and (C) cortico-spinal tract regions of interest (ROIs) overlaid on diffusion kurtosis imaging motion-corrected image. (D) Scatter plots of fractional anisotropy (FA) and MSK in group subjects across the aforementioned ROIs: colored spots indicate single subject's value for each metric; as reported in the legend, controls' measures are in blue, whereas the periventricular white matter injury group is in red. The units for MSK are in mm²/s, while FA is dimensionless. Error bars displaying mean (diamond) and standard deviation (bars) are overlaid on scatter plots.