Benveniste, Pierre-Louis.; L茅tourneau-Guillon, Laurent.; Araujo, David.; Chougar, Lydia.; Fetco, Dumitru.; Hori, Masaaki.; Kamiya, Kouhei.; Messina, Steven.; Tsagkas, Charidimos.; Audoin, Bertrand.; Bakshi, Rohit.; Bannier, Elise.; Blezek, Daniel.; Brisset, Jean-Christophe.; Callot, Virginie.; Charlson, Erik.; Chen, Michelle.; Ciccarelli, Olga.; Demorti猫re, Sarah.; Edan, Gilles.; Filippi, Massimo.; Granberg, Tobias.; Granziera, Cristina.; Hemond, Christopher C.; Keegan, B. Mark.; Kerbrat, Anne.; Kirschke, Jan.; Kolind, Shannon.; Labauge, Pierre.; Lee, Lisa Eunyoung.; Liu, Yaou.; Mainero, Caterina.; McGinnis, Julian.; Laines Medina, Nilser.; M眉hlau, Mark.; Nair, Govind.; O鈥橤rady, Kristin P.; Oh, Jiwon.; Ouellette, Russell.; Prat, Alexandre.; Reich, Daniel S.; Rocca, Maria A.; Shepherd, Timothy M.; Smith, Seth A.; Stawiarz, Leszek.; Talbott, Jason.; Tam, Roger.; Tauhid, Shahamat.; Traboulsee, Anthony.; Treaba, Constantina Andrada.; Valsasina, Paola.; Vavasour, Zachary.; Yiannakas, Marios.; Lombaert, Herv茅.; Cohen-Adad, Julien. (2026).听.听Multiple Sclerosis Journal.听
Magnetic resonance imaging, or MRI, is an important tool for finding and tracking spinal cord lesions in people with multiple sclerosis (MS), which are areas of damage caused by the disease. But automatic computer methods for detecting and outlining these lesions often work well only for one MRI type or one hospital鈥檚 scanning setup, which makes them less useful in real clinics where scan methods vary a lot. To address this, the researchers developed a more robust segmentation system, meaning a model that can automatically identify lesion boundaries, across many MRI contrasts and imaging sites. They trained and tested it on a large dataset of 4,428 annotated images from 1,849 people with MS across 23 imaging centers, using six different MRI contrast types and scans taken at 1.5, 3, and 7 tesla, which refers to the strength of the MRI scanner. Compared with existing methods that are designed for only one contrast type, the new model generalized better across different scan settings, according to neuroradiologist ratings. It also remained strong when tested across different spinal cord levels, image resolutions, threshold settings, and external datasets. Overall, the study shows that this approach can detect spinal cord MS lesions accurately and reliably across diverse MRI data, which is an important step toward making automated lesion analysis more useful in everyday clinical care.
Figure 1. Sankey diagram of annotated MRI scans across clinical sites. Line thickness is associated with the number of scans.
MRI scan distribution is clustered per acquisition type (3D, 2D sagittal, or 2D axial) and per MRI contrast, for each site.