Deep learning extracts upper spinal cord measurements from brain MRI
Published in Artificial Intelligence, Brain/Neurology, Journals, Scanners and Imaging, Tools.
Published November 1, 2026.
A study in npj Digital Medicine describes an automated method for identifying upper cervical spinal cord levels C1–C3 and extracting measurements from T1-weighted brain MRI.
Jarola and colleagues developed deep-learning models using data from 998 controls and 879 people with hereditary ataxias across 32 sites. External validation involved 2,737 MRI examinations from four public datasets. A separate longitudinal cohort included 51 people with multiple sclerosis.
The best labeling model, nnU-Net, achieved Dice overlap scores of 0.85 for C1 and 0.89 for C2 and C3. A quality-control classifier achieved 97.9% accuracy and 98.7% sensitivity for identifying labeling errors. These metrics evaluate image processing, not disease diagnosis. The pipeline also detected longitudinal spinal cord atrophy in the multiple sclerosis cohort.
The work could support larger multicentre imaging studies and biomarker research; it does not demonstrate improved treatment outcomes. The paper discloses consortium industry support and several authors’ commercial relationships. This peer-reviewed accepted version remains subject to editorial changes.
Source: Jarola et al., npj Digital Medicine, October 23, 2026.