Area B—Production ML Platform
Imaging Inference Platform
The serving backbone · 4 models, 5 environments
One shared framework so each new clinical model ships as a thin plug-in, not bespoke infrastructure.
- dissimilar AI services, one platform
- 4
- dissimilar AI services, one platform
- model repos, one CI/CD pipeline
- 7
- model repos, one CI/CD pipeline
01/Problem
Every clinical model shares the same error-prone plumbing — study events, DICOM retrieval, eligibility, decoding, inference, standards-compliant output — identically across dev, stage, pre-prod and two production regions. Re-implementing it per model is slow and risky.
02/Approach
A shared platform library built on a template-method base class that owns the whole lifecycle — eligibility, prioritization, download, preprocess, infer, postprocess, DICOM SR/GSPS output — so each model is a thin subclass. It hosts a CNN ensemble, a classical registration cascade, cascaded classifiers and a landmark detector feeding a measurement engine. Hardening includes a multi-decoder DICOM fallback chain, per-order thread isolation and an integration suite run against a live DICOM store.
03/Outcome
- Four architecturally dissimilar services kept live across five environments, including two production regions.
- Every new model ships as a thin subclass — no reimplemented DICOM I/O, serving, eligibility or testing.
- Survived a TensorFlow 2.11 → 2.20 upgrade without losing class-activation-map findings.
- The decode-fallback chain turned a class of runtime crashes into recoverable paths.
04/My role
- Directed consolidation of model serving onto this framework as an organizational standard, so the marginal cost of each launch fell.
- Prioritized maintenance and framework-upgrade work against features, protecting the fleet from dependency drift.
- Required integration tests against live DICOM stores rather than mocks.