Area A—Regulated Diagnostic AI
XR Long Bone Labeling
Deployed v1.0.0 · legacy migration
Six bone segmenters in one self-routing container, and the migration that retired a legacy TensorFlow 1 pipeline.
- validation Dice, six segmenters
- 0.79–0.97
- validation Dice, six segmenters
- V&V cases per release
- 52 + 8
- V&V cases per release
01/Problem
Radiologists must localize long-bone findings to specific anatomical regions — manual, slow, inconsistent work. Separately, a legacy TensorFlow 1 pipeline blocked multi-class expansion and cloud/on-prem deployment.
02/Approach
A BiT-ResUNet segmentation suite across humerus, femur, forearm and lower leg — six trained segmenters in a single container that self-selects the right model per DICOM instance, orchestrated on GKE via Cloud Functions and Pub/Sub. Every input is validated (view, modality, age, body part, pixel readability) with each failure mapped to a numbered error code. Output consolidates to one per-study DICOM SR.
03/Outcome
- Validation Dice 0.79–0.97 across six segmenters — all clearing the 0.75 clinical release bar.
- Deployed to five GCP environments across two regions (US + Canada).
- 52-case verification + 8-case validation matrices executed end to end against every environment on each release.
- Legacy pipeline retired; multi-class expansion and dual-region deployment unblocked.
04/My role
- Directed the migration as a platform priority, tied to the dual-region deployment requirement.
- Made automated V&V a hard release gate.
- Approved single-container self-routing packaging over six separately deployed services.
- Set the Dice ≥ 0.75 bar and held weaker models to it rather than shipping early.