Fully-integrated framework for registration of spinal cord white and gray matter
Sara Dupont1, Benjamin De Leener1, Manuel Taso2,3, Nikola Stikov1,4, Virginie Callot2,3, and Julien Cohen-Adad1,5

1Neuroimaging Research Laboratory (NeuroPoly), Institute of Biomedical Engineering, École Polytechnique de Montréal, Montréal, QC, Canada, 2Centre de Résonance Magnétique Biologique et Médicale (CRMBM), UMR 7339, CNRS, Aix-Marseille Université, Marseille, France, 3Centre d'Exploration Métabolique par Résonance Magnétique (CEMEREM), Hôpital de la Timone, AP-HM, Marseille, France, 4Montreal Heart Institute (MHI), Montréal, QC, Canada, 5Functional Neuroimaging Unit, CRIUGM, Université de Montréal, Montréal, QC, Canada


The spinal cord (SC) white and gray matter can be affected by a large number of pathologies. Being able to segment precisely the SC internal structure would be useful to better understand SC diseases, improve diagnosis and assess treatment efficiency. We introduce a complete framework for (i) multi-atlas automatic segmentation of the gray-matter, (ii) accurate registration to the MNI-Poly-AMU template and (iii) extraction of quantitative metric using partial volume information. Results showed improved accuracy of template registration when adding prior automatic gray-matter segmentation. The proposed method is freely available and provides an unbiased framework for quantitative analysis of SC MRI data.


The spinal cord (SC) is the main relay between the brain and the peripheral system and contains neurons notably responsible for complex functions such as locomotion. The SC white matter (WM) and gray matter (GM) can selectively be affected by a large number of pathologies such as multiple sclerosis, amyotrophic lateral sclerosis or trauma. Being able to precisely segment the WM and GM would be useful for (i) obtaining more specific knowledge about the origin of SC atrophy and (ii) improving the accuracy of registration of already-existing WM/GM templates and atlases of the SC4,5. However, due to the small size of the WM and GM and the relatively poor contrast in conventional MR sequences (e.g T1/T2-weighted), it is often difficult to properly segment the WM and GM.

The goals of the study were to: (i) improve the accuracy of an existing multi-atlas-based method for automatic segmentation1 by including prior shape information based on vertebral labeling, (ii) make segmentation independent from image contrast (e.g., T2*-weighted, magnetization transfer, etc.) and (iii) integrate WM/GM segmentation in a registration pipeline to the MNI-Poly-AMU template3.



For generating the dictionary required by the multi-atlas segmentation method, 38 healthy subjects were scanned on 3T MRI systems (TIM-Trio and Verio, Siemens Healthcare) using the standard head, neck and spine coils; multi-echo GRE T2*-weighted (T2*-w) images were acquired (mean age 25.9±4.2 y.o., 21M-17F). For validation, T2-w, T2*-w, magnetization transfer (MT) and diffusion-weighted images were acquired from ten other healthy subjects (mean age 37.6±21.0 y.o., 7M-3F) on the TIM-Trio. Parameters are listed in Table I.

Multi-atlas based segmentation method:

A multi-atlas slice-based model was built out of 38 subjects to represent the variability of WM and GM morphology. Building from an existing method1, the accuracy of segmentation was increased by integrating information about vertebral levels into the model. Our method was adapted to several contrasts using a linear normalization between median intensity values of WM/GM based on SC internal structure pre-registration (combining our multi-atlas based model and vertebral levels information). This method outputs probabilistic segmentations of WM/GM.

Registration framework:

The proposed registration framework includes (see Figure I): (i) registration of the MNI-Poly-AMU template to the T2-w, and T2*-w or MT images, (ii) automatic WM/GM multi-atlas based segmentation on T2*-w or MT images, (iii) registration of the template WM probabilistic map to the automatic WM segmentation; the warping field obtained here is used to correct the template registration from previous steps, (iv) registration of the MNI-Poly-AMU template (integrating the internal structure correction) on MR metric images (here DTI), (v) extraction of MRI metrics from the probabilistic WM tract atlas (thirty tracts).


The automatic probabilistic multi-atlas based segmentations were thresholded at 0.5 and compared with manual segmentations using 2D Dice coefficient (DC) on WM and GM, along with Hausdorff distance (HD) and maximum median distance (MD) on skeletonized GM segmentations. Fractional anisotropy (FA) was extracted in five WM tracts (gracilis, cuneatus, lateral corticospinal, spinocerebellar, and lemniscus tracts) using maximum-a-posteriori to account for partial volume effect4 in both the classical approach (registration of the MNI-Poly-AMU template to DTI without information about SC internal structure) and the proposed registration framework. Metric values were compared between registration approaches using paired Student’s t-tests (p=0.05).


Figure II shows examples of GM automatic segmentation results compared to manual segmentations for T2*-w and MT images. Associated DC, HD, and MD mean values and standard deviation are also displayed in Figure II.

Visual inspection suggests that registration of the white matter atlas was improved in six subjects over ten using the proposed framework compared to the classical approach. The diffusion metrics extracted from five WM tracts are consistent with the literature2 (Figure III). However, using the classical approach or the proposed framework did not show any statistically significant difference in the metrics values.


GM and WM were accurately segmented on ten healthy subjects using two contrasts (T2*-w and MT). Performance was better on the T2*-w images thanks to the higher in-plane resolution (0.5x0.5mm2 versus 0.91x0.91mm2). Using automatic WM/GM segmentations, we proposed a method for registering the MNI-Poly-AMU template to MRI metrics while taking into account the topology of the SC internal structure.

This proposed framework demonstrated better accuracy for extracting metrics within specific WM tracts. By allowing the extraction of relevant metrics into specific WM tracts, the proposed framework will help to improve specificity of SC atrophy measures and to quantify metrics of WM microstructure. The method is freely available in the Spinal Cord Toolbox: http://sourceforge.net/projects/spinalcordtoolbox/.


This study was supported by the Canadian Institute of Health Research (CIHR FDN-143263), the Sensorimotor Rehabilitation Research Team (SMRRT), the Fonds de Recherche du Québec - Santé (FRQS 28826), the Fonds de Recherche du Québec - Nature et Technologies (FRQNT 2015-PR-182754), Quebec Bio-Imaging Network (QBIN) and the Natural Sciences and Engineering Research Council of Canada (NSERC)


1. Asman AJ, Bryan FW, Smith SA, et al. Groupwise multi-atlas segmentation of the spinal cord’s internal structure. Medical Image Analysis 2014;18(3):460-71.

2. Cohen-Adad J, Samson RS, Smith AK, et al. Multisite DTI of the spinal cord with integrated template and white matter atlas processing pipeline. Proceedings of the 22th Annual Meeting of ISMRM, Milan, Italy, p.1727.

3. Fonov VS, Le Troter A, Taso M, et al. Framework for integrated MRI average of the spinal cord white and gray matter: the MNI-Poly-AMU template. NeuroImage 2014;102(2):817-27.

4. Lévy S, Benhamou M, Naaman C, et al. White matter atlas of the human spinal cord with estimation of partial volume effect. NeuroImage 2015;119:262-71.

5. Taso M, Le Troter A, Sdika M, et al. Construction of an in vivo human spinal cord atlas based on high-resolution MR images at cervical and thoracic levels: preliminary results. Magnetic Resonance Materials in Physics Biology and Medicine 2014;27(3):257-67.


Table I: Acquisition parameters for T2-weighted (T2-w), T2*-weighted (T2*-w), magnetization transfer (MT) and diffusion-weighted images (used for DTI reconstruction).

Figure I: Proposed framework for the registration of the MNI-Poly-AMU template to a metric image integrating additional information about the spinal cord internal structure from WM/GM automatic segmentation.

Figure II: Automatic GM segmentation results for T2*-w (top) and MT (bottom) images. The probabilistic automatic segmentations (right column) were thresholded at 0.5 and compared to manual segmentations (middle column). Numerical results are displayed for each slice as well as averaged across ten healthy subjects.

Figure III: FA map registered to the MNI-Poly-AMU template while accounting for WM/GM morphology and averaged per vertebral level (C2 to C5) across ten healthy subjects. Five white matter tracts are overlaid: gracilis (blue), cuneatus (yellow), corticospinal (green), spinocerebellar (pink), and lemniscus (red).

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)