A deep learning framework for automated cervical spinal cord vertebral labeling and morphometric extraction from brain MRI with large-scale multicenter validation

This study aimed to develop and validate a fully automated pipeline for vertebral level identification (C1–C3) of the upper cervical spinal cord (SC) in T1-weighted brain MRI with a quality control module to assess the generated masks. MRI data from 998 controls and 879 patients with hereditary ataxias (32 sites; age 5–85 years; 46% male) were used to train deep learning models (nnU-Net, U-Net with deep supervision, SwinUNETR) for automatic vertebral labeling of C1–C3 levels. A secondary model based on VGG, ResNet, and EfficientNet architectures was trained to identify vertebral labeling errors.

The pipeline also generates image quality metrics including signal-to-noise ratio, contrast-to-noise ratio, coefficient of joint variation, and entropy focus criterion. External validation was performed using 2737 MRIs from four public datasets, and clinical utility was demonstrated in a longitudinal cohort of 51 patients with multiple sclerosis. The nnU-Net achieved the highest performance (Dice: C1 = 0.85, C2 = 0.89, C3 = 0.89), labeling 95.8% of mislabeled cases from the Spinal Cord Toolbox.

The VGG-based classifier achieved 97.9% accuracy and 98.7% sensitivity. External validation showed labeling accuracies exceeding 92% in four of five datasets and sensitivity to longitudinal SC atrophy in multiple sclerosis. This pipeline enables scalable, reliable vertebral labeling and morphometric extraction of the upper cervical SC for multicenter studies and supports biomarker development in neurodegenerative diseases.

The authors gratefully thank all of the individuals for their participation in the study and the Friedreich’s Ataxia Research Alliance (FARA, USA) for their help with FRDA participant recruitment. The authors thank and acknowledge FARA, Takeda Pharmaceuticals Company, Novartis Gene Therapies, IXICO plc, PTC Therapeutics, and Larimar Therapeutics for support in funding and for their participation in the TRACK-FA Neuroimaging Consortium, including providing advice and feedback on study design, implementation, and analysis of outcomes. We thank Coordenação de Aperfeiçoamento de Pessoal de Nível Superior for the Financial Support.

Study data were collected and managed using REDCap electronic data capture tools hosted and managed by Helix (Monash University, Australia)9,28. Study data were collected, stored, and managed using XNAT, a software framework for managing neuroimaging laboratory data hosted by Monash (MXNAT) and Monash Biomedical Imaging (MBIXNAT) (Monash University, Australia). The authors thank the research coordinators across all TRACK-FA and READISCA sites for their assistance in participant recruitment and testing.

The following collaborators opted to have their names acknowledged: Pedro Marcelino Mendes Novaes Melo, Department of Neurology, University of Campinas, Campinas, Brazil; Jennifer Farmer, Friedreich’s Ataxia Research Alliance (FARA), Downingtown, Pennsylvania, United States of America; Tobias Linding, Department of Diagnostic and Interventional Neuroradiology, University Hospital Tübingen, Tübingen, Germany; Thomas H. Mareci, Department of Biochemistry and Molecular Biology, University of Florida, Gainesville, Florida, USA; Henry Paulsen, Department of Neurology, University of Michigan, Ann Arbor, MI, USA. Schulz, Department of Neurology, RWTH Aachen University, Aachen, Germany.