Spinal cord MRI has great potential for the diagnosis and prognosis of neurodegenerative and traumatic diseases. However, processing of spinal cord MRI data is challenging due to the lack of standard tools and guidelines¹. Particularly, a standard reference space (template) would facilitate multi-center and large group studies, and would avoid user bias when quantifying MR data into small regions of the spinal cord. Recently, we introduced the MNI-Poly-AMU template, based on T₂-weighted data from 16 subjects, covering C1 to T8 vertebral levels². In this manuscript, we propose the MNI-Poly-AMU50, an unbiased left-right symmetric template based on 50 subjects, available for T₁-, T₂- and T₂*-weighted MR contrast, that cover the entire spinal cord and brainstem, and that is merged with a probabilistic white and gray matter atlas³.50 healthy subjects (mean age: 27+-7 y.o., 31 men and 19 women) were scanned in Montreal (n=33) and Marseille (n=17) on 3T systems (TIM Trio and Verio, Siemens Healthcare) using the standard head, neck and spine coils. Each subject had a 3D T₁-weighted (T₁w) and T₂-weighted (T₂w) scan covering the full spinal cord and brainstem. This large coverage was achieved by acquiring two FOVs per contrast (1: head and cervical spine; 2: cervical, thoracic and lumbar cord), stitched together using off-line software tools provided by the manufacturer’s MRI console after correcting for image bias field. Acquisition parameters are presented in Figure 1. Total acquisition time was 22 min. Pre-processing for template generation is illustrated in Figure 2 and included (i) stitching of images, (ii) cropping to cover from brainstem to L2/L3 level, (iii) extraction of spinal cord centerline using PropSeg⁴, (iv) robust slice-based non-linear normalization of spinal cord intensity (average set to 1000), (v) straightening the image using multi-dimensional b-spline interpolation, (vi) unidimensional b-spline-based vertebral alignment using an average model of vertebral distance⁵ based on the manual labeling of the following landmarks: rostral pons (RP), pontomedullary junction (PMJ), and intervertebral disks from C2 to L1. Template creation: Following pre-processing of the 50 T₁w and 50 T₂w straight images of the spinal cord, unbiased left-right symmetric templates were independently constructed using hierarchical group-wise image-registration method described in ⁶. Following template creation, both T₁w and T₂w templates were co-registered and the final template space was computed so as to correspond to the middle deformation point between T₁w and T₂w template. Parameters were: two-step diffeomorphic transformation (greedy SyN), mutual information metric, 100x100 iterations. Then, the T₂*-weighted template and the white and gray matter probabilistic atlases were merged to the template as in ². We established a proof-of-concept by first registering the proposed template on a T₂w image from one healthy subject (female, 23-year-old) using the Spinal Cord Toolbox⁷, then registering the T₂w image with a diffusion MRI image, which has been resampled to 1x1x1 mm3, and finally extracting diffusion metric (here fractional anisotropy) into white and gray matter regions averaged over specific vertebral levels (C5-T2 and T5-T9). Dice coefficient between manual segmentation and registered template segmentation was calculated to assess the registration accuracy. This proof-of-concept can be fully reproduced using the script available at http://www.neuro.polymtl.ca/downloads.Figure 3 shows both generated templates with overlay vertebral labeling as well as white and gray matter atlases. Each template has 200x200x1100 voxels of 0.5x0.5x0.5 mm3 resolution and is oriented as Right-to-Left/Posterior-to-Anterior/Inferior-to-Superior (RPI). Figure 4 shows the results of registration of anatomical and diffusion MRI data with the template as well as FA measures extracted in gray and white matter for specific vertebral levels. Note that the poor resolution and SNR of the FA image introduce a high bias in white matter measurements due to CSF contamination. However, despite this poor resolution issue, SCT managed to correctly register the MNI-Poly-AMU50 template and atlases on the diffusion MRI data and extract useful MR metrics. Dice coefficient between registered template spinal cord segmentation and manual segmentation of subject’s T₂-weighted image was 0.79.We created an unbiased symmetric T₁- and T₂-weighted template of the brainstem and spinal cord, based on 50 subjects. Furthermore, we demonstrated the ability of the Spinal Cord Toolbox to register new data to this template and to extract useful MRI metrics into specific regions of the spinal cord (e.g., white and gray matter, selected vertebral levels). These templates provide a common framework for co-registering multi-parametric data and hence facilitate the conduction of multi-center and large group studies. All these developments are part of the Spinal Cord Toolbox⁷. Further developments will include atlases of white matter pathways⁸ and gray matter sub-regions in the template, as well as probabilistic atlases of spinal levels⁹.