Week 4: Notice 04: Diagnostic Radiology
The first expert/diagnostic notice — where the pitch is true and the audit still matters
15 March 2027
The pitch this week is, again, mostly true — sensitivity and specificity numbers on the published benchmarks really are strong. That’s exactly why the breaking-changes section has to do real work rather than just contradicting the pitch: the case against this migration isn’t “it doesn’t work,” it’s “what a training-pipeline argument costs a profession over a decade that a single-scan accuracy number can’t see.”
Outline#
- why “the pitch is true” and “the migration is fine” are different claims, and where this notice keeps them separate
- the specific mechanism (training-pipeline narrowing) linking this week’s economic and functional breaking changes together
- what a decade-scale rollback plan implies about how early a profession needs to notice this kind of migration to keep its options open
The pitch
A radiologist's read is pattern-matching against a reference set of prior scans with known outcomes. A model trained on millions of labelled scans across every modality now matches or exceeds specialist-level sensitivity on the findings it was trained to detect, returns a read in seconds instead of a worklist queue, and never has an off day at 2am on call.
Migration path
Scans route through the model first; findings above a confidence threshold generate a structured report directly into the record, and only borderline or flagged studies are queued for a radiologist's read.
Breaking changes
The unsatirized part. What the migration actually costs, plainly.
- Economic
- Radiology built a specific, well-compensated career path for one of medicine's few remote-friendly specialties, with training pipelines sized to that demand. Collapsing routine-read volume doesn't just cut cost per scan, it removes the case-volume junior radiologists need to train on, narrowing who can enter the specialty at all rather than just how many are employed once trained.
- Functional
- A model detects what it was trained to detect, at the prevalence and image quality of its training data; it degrades quietly on rare presentations, unusual patient anatomy, and equipment it wasn't calibrated against, without the confidence score reliably tracking that degradation. A radiologist under the same conditions doesn't stop reasoning about the specific patient in front of them just because the case is unusual.
Rollback plan
Read volume can be handed back to radiologists directly, and the profession itself doesn't disappear. But a training pipeline narrowed for a decade produces fewer radiologists able to absorb that volume back, and rebuilding that pipeline is a decade-scale project, not a policy reversal.
Teaching team
- Idris Fenn — tutor