Image Processing

Data processing was undertaken using harmonized protocols developed by the ENIGMA‐Ataxia consortium (http://enigma.ini.usc.edu/ongoing/enigma-ataxia/), based on publicly available and well‐validated software toolboxes. ²⁰

To measure the cross‐sectional area (CSA) and eccentricity, we employed the Spinal Cord Toolbox (SCT) version 4.2.2, an open‐source software package specifically designed to process spinal cord multimodal MRI data. ²⁰ In brief, automatic segmentation of the cervical spinal cord was conducted using a deep learning algorithm, ²¹ and if deemed necessary after visual inspection, the segmentations were manually corrected. Next, the C2 and C3 vertebral levels were manually marked at the posterior tip of the vertebral discs, which enabled the registration of subject images to a standardized template of the spinal cord and brainstem (the PAM50 template). ^{22 , 23 , 24} Lastly, the mean CSA and eccentricity were computed at each of the C1 to C4 vertebrae after correcting for the curvature of the spine. The CSA is quantified by the number of pixels in the set of axial slices defining each vertebral level of the segmented spinal cord, reported in millimeters squared. Eccentricity is computed by fitting an ellipse to each axial spinal slice and determining the deviation (ie, flattening) of the ellipse relative to a perfect circle. Mathematically, such a measure characterizes the shape of the spinal cord cross‐section defined as the square root of 1 − (d/D), ² where d and D are respectively defined as the smallest and largest diameter of the ellipse. Values closer to 1 indicate an anteroposterior flattening of the spinal cord. We assessed only the upper cervical spinal cord because we used MRIs centered on the brain with limited spinal cord coverage (Fig. 1). Because the spinal cord coverage was slightly different across individuals due to head size variability or field‐of‐view placement during data acquisition, different sample sizes were available for each vertebral level we examined (control subjects: C1 = 223, C2 = 223, C3 = 215, and C4 = 170; patients: C1 = 252, C2 = 252, C3 = 237, and C4 = 170).

Study design and imaging processing pipeline. For healthy control subject numbers, see Supporting Information Table S S1. N is the number of patients with Friedreich's ataxia (FRDA) enrolled by each site, and C1, C2, C3, and C4 are the sample sizes available for each vertebral level assessed. SCT, Spinal Cord Toolbox. [Color figure can be viewed at wileyonlinelibrary.com]