Optimization of cervical cord atrophy measurement using a real-world multi-center dataset in Multiple Sclerosis

Carsten Lukas^1,2*Barbara Bellenberg³Ferran Prados^4,5,6Paola Valsasina⁷Katrin Parmar⁸Iman Brouwer⁹Deborah Pareto¹⁰Alex Rovira¹⁰Jaume Sastre-Garriga¹¹Claudia Angela Gandini Wheeler-Kingshott^12,13,5Michael Amann^14,8Maria Assunta Rocca^15,16,7Massimo Filippi^{15,16,17,18,7}Marios C. Yiannakas⁵Eva M.M. Strijbis¹⁹Frederik Barkhof^20,5,9Hugo Vrenken⁹

• ¹Department of Neurology, St. Josef Hospital, Ruhr-University Bochum, Bochum, Germany
• ²Institute of Neuroradiology, St. Josef Hospital, Ruhr-University Bochum, Bochum, Germany
• ³Institute of Neuroradiology, St Josef Hospital, Bochum, Germany
• ⁴Centre for Medical Image Computing (CMIC), Department of Medical Physics and Biomedical Engineering, UCL University College London, London, United Kingdom
• ⁵Queen Square MS Centre, Department of Neuroinflammation, UCL Queen Square Institute of Neurology; Faculty of Brain Sciences, University College London, London, United Kingdom
• ⁶e-Health Centre, Universitat Oberta de Catalunya, Barcelona, Spain
• ⁷Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute, Milano, Italy
• ⁸Neurological Clinic and Policlinic, Department of Medicine, University Hospital Basel, Basel, Switzerland
• ⁹Department of Radiology and Nuclear Medicine, Amsterdam UMC, Amsterdam, Netherlands
• ¹⁰Section of Neuroradiology, Department of Radiology, Hospital Universitari Vall d'Hebron, Barcelona, Spain
• ¹¹Department of Neurology, Multiple Sclerosis Center of Catalonia (Cemcat), Hospital Universitari Vall d'Hebron, Barcelona, Spain
• ¹²Department of Brain & Behavioral Sciences, University of Pavia, Pavia, Italy
• ¹³Brain Connectivity Center, IRCCS Mondino Foundation, Pavia, Italy
• ¹⁴Medical Image Analysis Center (MIAC), Basel, Switzerland
• ¹⁵Neurology Unit, IRCCS San Raffaele Scientific Institute, Milano, Italy
• ¹⁶Neurorehabilitation Unit, IRCCS San Raffaele Scientific Institute, Milano, Italy
• ¹⁷Neurophysiology Service, IRCCS San Raffaele Scientific Institute, Milano, Italy
• ¹⁸Vita-Salute San Raffaele University, Milano, Italy
• ¹⁹Department of Neurology, Amsterdam UMC, Amsterdam, Netherlands
• ²⁰Centre for Medical Image Computing (CMIC), Department of Medical Physics and Biomedical Engineering, UCL, London, United Kingdom

Background: Cervical cord atrophy is linked to disability in multiple sclerosis (MS). Cervical cord cross-sectional area (CSA) measurement for atrophy quantification using magnetic resonance imaging (MRI) has been technically validated, but information about effects of methodological choices on associations of CSA with clinical variables is lacking. Aim: Assessing how image acquisition, cord level selection, CSA normalization and segmentation software affect measurement variance, separation of clinical groups, correlations with clinical scores, and to formulate recommendations for future study designs. Methods: Head and neck 3D-T1-weighted MRI of people with MS (pwMS, N=85) and healthy controls (HC, N=19) from five European centers. CSA measurements encompassed four methods (Active surface method ASM, NeuroQLab, SCT-Propseg and SCT-Deepseg), at two different levels of the cervical cord: C1-2 and C1-7 and normalization using four methods, based on cervical dimensions. Coefficient of variation (CV) of CSA was assessed in HC. In MS, Spearman correlations of CSA with EDSS were assessed. Separation between relapsing (rMS) and progressive MS (pMS) was quantified by area-under-the-curve (AUC) from receiver-operator-characteristic analysis. Results: For all combinations of imaging, cord level, and segmentation software, unnormalized CSA differed between HC and pMS. CV in HC varied between 10.5% and 13.5% for unnormalized CSA and was lower for CSA normalized by C1-C2 (range: 9.4-12.0%) and C1-C3 vertebral height (8.6-12.6%). Unnormalized and normalized CSA correlated with EDSS scores for all measurement combinations (Spearman’s rho between -0.646 and -0.372, all corrected p<0.001); correlations were stronger for CSA measured at vertebral level C1-7 than C1-2, and stronger for normalized than unnormalized CSA. Mean AUC for separating rMS from pMS ranged between 0.685 and 0.877, with higher AUC for CSA measured at the C1-7 than at the C1-2 vertebral level, and for normalized compared to unnormalized CSA. Conclusion: Clinical performance of CSA quantification regarding discrimination between rMS and pMS and correlations with EDSS was better for whole cervical cord (C1-7) than for C1-2 measurements, and for normalization by C1-C2 or C1-C3 vertebral height. Based on the quantitative results of this exploratory multi-center study and on previous literature, we formulated recommendations to support future study design decisions.