← Archive

Foundation · Healthcare AI programme · NDA

Spinal cord segmentation under scarce labels

Limited labels, high clinical stakes. A confident wrong segmentation is worse than a slow human read — accuracy without trust criteria is theatre.

Lead outcome

90%+

Dice ~0.9 accuracy at handoff

Focus

Medical ImagingHealthcare AINDA
StackNVIDIA MONAIDeep Learning

What I owned

I kept clinical trust criteria visible in every gate while transfer learning closed the accuracy gap — validation discipline equal to the model work.

Narrative

USA medical imaging client under NDA. The engineering problem was segmentation with thin labels. The programme problem was a buyer who could not afford confident wrongness.

My seat: keeping validation, failure modes, and clinical sign-off on the critical path — not as a week-twelve surprise after a demo glow.

Same operator pattern as every other pillar: absorption and accountability before polish.

Validation equals model work

If the trust criteria are weaker than the network, the programme is not done.

Scarce labels, scarce shortcuts

Transfer learning is a tool. Clinical gates decide whether it ships.

90%+

Dice ~0.9 accuracy at handoff

Transfer

learning under scarce labels

Clinical

trust criteria I held as gates

Next step

Want the confidential layer — names, constraints, steering artefacts? That conversation stays over email.

Email me about this case