Data acquisition
Ten healthy participants (Age: mean ± STD = 22.9 ± 1.2 years old; 3 females) were scanned using a Siemens 3T Prisma-fit with a 64 head-neck coil at three neck positions (extension, neutral, and flexion), as illustrated in Fig. 1. Informed consent was obtained from each participant and appropriate padding was used to maintain the head position and ensure comfort. The study was approved by the Comité d’éthique de la recherche du Regroupement Neuroimagerie Québec. All experiments were performed in accordance with the Declaration of Helsinki.
The field-of-view covered the top of the head down to at least the T1 vertebrae. A 3D sagittal T2w scan was acquired for each neck position with the following parameters: SPACE sequence, TR: 1.5 s, TE: 0.12 s, matrix: 72 × 384, in-plane resolution: 0.6 × 0.6 mm2, number of slices: 384, slice thickness: 0.6 mm, pixel bandwidth: 620 Hz/pixel).
Data processing
The SC was automatically segmented using the Spinal Cord Toolbox (SCT) v5.414 sct_deepseg_sc15 and the PMJ and vertebral levels were identified. The processing pipeline is available on GitHub (https://github.com/sct-pipeline/pmj-based-csa/releases/tag/r20230313) with its documentation (https://github.com/sct-pipeline/pmj-based-csa/tree/r20230313#readme).
Spinal segment labeling
Spinal segments were manually labeled on T2w images for each neck position by an expert rater (MB) using the spinal nerve rootlets to identify the mid spinal segment position (Fig. 2). The standardized procedure is available on GitHub (https://github.com/sct-pipeline/pmj-based-csa/tree/r20230313#manual-labeling-of-spinal-cord-rootlets-with-fsleyes).
Spinal segment labeling pipeline. (1) SC centerline is detected, (2) SC is straightened, (3) denoised, and manually labeled in the coronal view. (4) The inverse transformation is applied on the labels. The distances between the PMJ and each spinal segment (green), and between each spinal segment and corresponding disc (pink) were computed.
Briefly, to label the spinal segments, we propose straightening the SC to enhance visualization of the spinal nerve rootlets, as the natural curvature of the SC can impede their clear identification in the coronal view. Non-local means adaptive denoising is applied to denoise the image16. In summary, the non-local mean filter leverages the similarity between distinct image patches to decrease noise while preserving details. Furthermore, it integrates the spatial and frequency information present in the image to adapt the denoising process. Spinal segment labels are then placed at the center (superior-inferior) of where the rootlet emerges from the SC on the corresponding axial slice in the center of the SC. The labeling convention is indicated in Fig. 2, and all spinal segments with visible nerve rootlets (except C1) are labeled for all neck positions and participants. The labels are finally brought back to the curved space with the inverse transform.
Distance with PMJ, spinal segment and discs
To estimate the position of the spinal segments using the PMJ and the discs as references, we computed two types of distances along the SC centerline for each participant and neck position: between the PMJ and each spinal segment (Fig. 2.4 green), and between each intervertebral disc and its corresponding spinal segment (Fig. 2.4 pink).
To assess the reliability of the spinal segment location estimation across the different neck positions, we computed the standard deviation (STD) across neck positions for each participant. The STD was computed for each method (PMJ, discs) and for each level (2 to 7). The STD was chosen to assess variability, as opposed to COV, because the absolute value of distances PMJ-spinal segment becomes larger for lower spinal segments, which would have influenced the COV when compared to disc-spinal segment distances. To assess if there was a significant difference between the STD of the estimated spinal segment positions measured with the PMJ and the discs, we conducted a paired two-way ANOVA test on the STD across neck positions (dependent variable) of the 10 participants. The independent variables were the methods (PMJ, discs) and levels (from 2 to 7). A significance level of p-value = 0.05 was set. We used the ANOVA test to detect a global effect of the independent variables rather than specific group differences. No post-hoc analysis was conducted.
CSA computation
SC CSA was computed using sct_process_segmentation and averaged across 3 slices using three references: CSA PMJ, CSA spinal, and CSA discs (Fig. 3). CSA PMJ was computed at the distance between the PMJ and each spinal segment, averaged across the three neck positions. That distance was subject-specific. The reason we chose to average the distance across neck positions is the following. The main idea of this paper is to validate a PMJ-based method13 as a reference for computing CSA instead of relying on intervertebral discs due to the known variability in the spatial correspondence between the spine and SC. As a reminder, the PMJ-based method relies on a set distance from the PMJ along the SC centerline. In the Bédard et al.13 paper, that distance was set to 64 mm, which corresponded to the distance to the C2-C3 disc, averaged across 804 adult participants. In the present study, instead of computing a distance between the PMJ and the C2-C3 disc, we estimate a distance between the PMJ and each spinal segment (manually identified with the nerve rootlets). That distance is calculated for each participant and for each neck position. Our hypothesis is that, for each participant, regardless of the neck extension, the PMJ-based CSA will give similar values (as opposed to the disc-based CSA, because the SC moves along the superior-inferior axis relative to the discs, depending on the neck position). To verify this hypothesis, we set a subject-specific distance between the PMJ and each spinal segment to be the average across the three neck positions. Then, we compute the CSA using that distance, and we evaluate the stability of that CSA measure. CSA spinal is computed at the spinal segment location for spinal levels C2 to C8, and CSA discs at the intervertebral discs C2-C3 to C7-T1.
SC CSA computed with 3 references: (A) mean distance of each spinal segment from participant’s PMJ, (B) each spinal segment, (C) each intervertebral disc.
To estimate the variability of CSA across neck positions, we computed the COV, instead of the STD in order to aggregate the results across participants and levels, since CSA varies across participants and across spinal levels17.
The COV across neck positions was computed for each participant, method and level as following:
$$COV[i]= STD(neck\, position\, n=3)/MEAN(n=3)$$
A paired 3-way ANOVA test was performed to evaluate if CSA computed using the 3 different references (PMJ, Spinal, and Disc) yielded equivalent measures per participant. The independent variables examined in our analysis were the methods (PMJ, Spinal, and Disc), levels, and neck positions. Note that the ANOVA test was specifically conducted on the CSA rather than on the COV, in order to be able to account for the neck position as an independent variable. Furthermore, due to the paired nature of the ANOVA test, COV was not relevant.
Neck angle
The angle of the SC was computed for each neck position in the right-left axis. The angle was computed between the C1-C2 and C4-C5 intervertebral discs (Fig. S1)18,19. The C4-C5 disc was chosen instead of the C6-C7 disc as done in the previous studies18,19 since one participant presented a smaller field-of-view in the superior inferior direction resulting in the absence of the C5-C6 and lower discs.
Quality control
After running analysis, the quality of all SC segmentations, vertebral labeling and PMJ labeling were visually assessed. Problematic segmentations (under-segmentation/leaking) and labels were identified in SCT’s quality control report and manually corrected using ITK-SNAP. Disc labels were placed at the posterior tip of the disc for C1-C2 to C7-T1 and the PMJ was identified in the mid-sagittal plane. Corrections were added to the source dataset and the analysis was run again using the corrections instead of automatic labeling.