Clinical validation of an AI system for MRI-based detection of lumbar degenerative pathology
Introduction
● Low back pain is the leading cause of years lived with disability worldwide, and MRI is the reference standard for patient assessment [1, 2]
● Lumbar spine MRI interpretation is challenging due to the multiplicity of classification systems and their reliance on subjective visual grading [3]
● Existing AI algorithms are limited in scope or too complex for routine clinical implementation [4, 5]
What is the performance of a novel AI solution for detecting lumbar degenerative pathology on MRI?
Methods
Data Collection:
● Retrospective collection of adult lumbar spine MRIs from French private practices
● MRI protocols included sagittal T1, T2 (DIXON/STIR) and axial T2-weighted sequences (2D or 3D)
Study Design:
● Standalone AI performance
● Reader study: 1 radiologist read all cases twice with and without AI assistance
Ground Truth:
● Two expert musculoskeletal radiologists reviewed all exams (T8-T9 to L5-S1)
● Eight pathologies assessed: disc herniation, disc bulging, foraminal stenosis, facet arthropathy, spondylolisthesis, disc degeneration, endplate degeneration, and c central canal stenosis
AI Software:
● LumbarMR (Gleamer, Paris, France) generates a structured report of degenerative pathologies across all intervertebral levels
Statistical Analysis:
● Observation-wise sensitivity and specificity (observation = patient × intervertebral level × pathology)
● Per-pathology sensitivity and specificity
● Interpretation time and diagnostic performance with and without AI
Results
Overall AI performance across 10,647 analyzable observations:
Overall sensitivity
● 78.4% (95% CI: 76.2–80.6)
Overall specificity
● 94.4% (95% CI: 93.9–94.9)
Per-pathology AI performance is shown in Figure 1
Subgroup analyses
● AI sensitivity was significantly higher in older patients (86.0% vs. 73.5%, p < 0.001)
Reader study
● Significant reduction in radiologist interpretation time (116.4 s to 97.2 s, p < 0.001)
● Preserved sensitivity (p = 0.87) and specificity
(p = 0.25)
Conclusions
● The AI solution demonstrated robust standalone performance for detecting lumbar degenerative pathology on MRI
● Performance was highest for central canal stenosis and endplate degeneration
● AI assistance reduced radiologist interpretation time by 16.2%
[1] Buchbinder et al. Lancet 2018. [2] de Souza et al. Clinics 2019. [3] Esposito et al. JOR Spine 2025. [4] Soydan et al. Glob. Spine J. 2025. [5] van der Graaf et al. Eur. Radiol. 2025.
