Longitudinal spinal cord atrophy in multiple sclerosis using the generalized boundary shift integral.

Objective: Spinal cord atrophy is a clinically relevant feature of multiple sclerosis (MS), but longitudinal assessments on magnetic resonance imaging using segmentation-based methods suffer from measurement variability, especially in multicenter studies. We compared the generalized boundary shift integral (GBSI), a registration-based method, with a standard segmentation-based method.
Methods: Baseline and 1-year spinal cord 3-dimensional T1-weighted images (1mm isotropic) were obtained from 282 patients (52 clinically isolated syndrome [CIS], 196 relapsing-remitting MS [RRMS], 34 progressive MS [PMS]), and 82 controls from 8 MAGNIMS (Magnetic Resonance Imaging in Multiple Sclerosis) sites on multimanufacturer and multi-field-strength scans. Spinal Cord Toolbox was used for C2-5 segmentation and cross-sectional area (CSA) calculation. After cord straightening and registration, GBSI measured atrophy based on the probabilistic boundary-shift region of interest. CSA and GBSI percentage annual volume change was calculated.
Results: GBSI provided similar rates of atrophy, but reduced measurement variability compared to CSA in all MS subtypes (CIS: -0.95 ± 2.11% vs -1.19 ± 3.67%; RRMS: -1.74 ± 2.57% vs -1.74 ± 4.02%; PMS: -2.29 ± 2.40% vs -1.29 ± 3.20%) and healthy controls (0.02 ± 2.39% vs -0.56 ± 3.77%). GBSI performed better than CSA in differentiating healthy controls from CIS (area under the curve [AUC] = 0.66 vs 0.53; p = 0.03), RRMS (AUC = 0.73 vs 0.59; p < 0.001), PMS (AUC = 0.77 vs 0.53; p < 0.001), and patients with disability progression from patients without progression (AUC = 0.59 vs 0.50; p = 0.04). Sample size to detect 60% treatment effect on spinal cord atrophy over 1 year was lower for GBSI than CSA (CIS: 106 vs 830; RRMS: 95 vs 335; PMS: 44 vs 215; power = 80%; alpha = 5%).
Interpretation: The registration-based method (GBSI) allowed better separation between MS patients and healthy controls and improved statistical power, when compared with a conventional segmentation-based method (CSA), although it is still far from perfect. ANN NEUROL 2019 ANN NEUROL 2019;86:704-713.
(© 2019 American Neurological Association.)