MR Lumbar Spine Labeling & Abnormality Detection

Manual measurement of lumbar vertebrae on MRI is slow, varies between readers, and doesn't scale. The model had to support clinical measurement — not just classification — validated against multi-radiologist ground truth and packaged as a standards-compliant service within an FDA-submittable device suite.

A DSNT heatmap-regression network on a ResNet-34 encoder predicts 36 landmarks (6 points on each of L1–S1), trained with a Jensen–Shannon + coordinate-MAE loss for sub-pixel, end-to-end differentiable localization. Geometry is un-corrected back to the original image so error is true anatomical millimetres. A clinical engine derives vertebral and disc heights and grades them against age- and sex-normative models. Shipped as a MONAI Deploy pipeline emitting DICOM GSPS overlays and Structured Reports, hard-gated by an upstream T1/T2 sequence classifier.

• 2.07 mm mean error, 98.12% per-vertebra and 97.33% per-series accuracy on internal evaluation.
• 2.43 mm / 96.23% / 93.75% on a 256-series facility-disjoint bench set — sites entirely absent from development.
• Both inside measured radiologist inter-reader variability of 1.4–3.5 mm.
• Consistent across sex, age, manufacturer and site; 5,043 studies curated into stratified splits with documented leakage controls.

• Owned the product-line roadmap; Spine Labeling achieved Health Canada clearance and fed the FDA 510(k) submission.
• Directed detection and remediation of a data-independence defect — facilities reused between development and bench sets across a four-model device family — with re-alignment and re-annotation of hundreds of series before submission.
• Established three-radiologist consensus ground truth (Dice agreement ≥ 0.85) and required error be reported against inter-reader variability.
• Chose landmark regression over cheaper classification-only, accepting a longer timeline for a clinically defensible product.