A spinal-cord AI worked across thousands of brain scans—its clinical role remains untested
The multicentre pipeline labelled upper cervical levels and extracted research measurements from routine brain MRI. External validation was large, but the multiple-sclerosis demonstration involved only 51 people and no care decision.
The short version
- What happened
- Training used brain MRI from 998 controls and 879 people with hereditary ataxias across 32 sites; external validation used 2,737 scans from four public datasets.
- Who it affects
- People with neurodegenerative disease may eventually benefit if existing brain MRI can supply reliable spinal-cord measurements without extra scanning, but no change in diagnosis, treatment or outcome was tested.
- What remains unproven
- The longitudinal clinical demonstration included only 51 people with multiple sclerosis
Keep in mind
The longitudinal clinical demonstration included only 51 people with multiple sclerosis
No diagnosis, prognosis or treatment decision was evaluated
Performance was not above 92% in one of five external datasets
Extensive consortium and industry relationships require transparent independent replication
Source analysis · 1 linked source. Primary analysed source date: 23 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The pipeline turns a brain scan into a repeatable spinal-cord measurement
A large international team trained three segmentation approaches to find and label the C1-C3 levels of the upper cervical spinal cord in ordinary T1-weighted brain MRI. The development data included 998 controls and 879 people with hereditary ataxias from 32 sites, aged 5 to 85. A separate image classifier was trained to flag labelling errors, while the pipeline also calculates basic image-quality measures.
The best model, nnU-Net, recorded Dice overlap scores of 0.85 at C1 and 0.89 at C2 and C3. The quality-control classifier reported 97.9% accuracy and 98.7% sensitivity. Across 2,737 external scans from four public sources, labelling accuracy exceeded 92% in four of five evaluated datasets. These results support a scalable research tool across heterogeneous scanners and populations.[1]
Detecting change in 51 people is a feasibility result
The team applied the measurement pipeline to longitudinal scans from 51 people with multiple sclerosis and detected upper-cervical atrophy over time. That is evidence the tool can recover a plausible research signal; it is not evidence that the measurement predicts progression, changes treatment or improves care.
A next-stage assessment should lock the pipeline before testing, report failures by scanner and patient group, compare automated measurements with expert review and evaluate a pre-specified clinical question. Public and charitable agencies funded the work, while companies supported the TRACK-FA consortium. Several investigators disclosed industry relationships unrelated to the manuscript; the model developers reported no competing interests.[1]
The distinction that matters
The evidence ladder
The study moves from multicentre development to external validation and a small longitudinal demonstration.
Development
1,877 people across 32 sites
Train C1-C3 labelling and quality control
External validation
2,737 public-dataset scans
>92% accuracy in four of five datasets
MS demonstration
51 people followed over time
Atrophy signal, not clinical utility