Six million CT images become a reusable starting point.

The human problem

Building and labeling a separate model for every finding on a CT scan is costly and can fragment clinical evidence.

The prior constraint

Medical-imaging systems often need task-specific labels and may transfer poorly beyond the data used to build them.

AI’s actual role

A foundation model learned shared representations from CT images and associated clinical information, then supported downstream tasks.

The documented result

The Nature paper reports training on about six million images from 15,331 CT scans and evaluates downstream capabilities; final copy must quote task-specific benchmarks directly from the paper.

Why it may matter

A reusable representation can lower the research cost of asking new questions of scans, while every clinical use still needs its own validation.

Limitations

Training scale does not guarantee clinical usefulness or fairness.

Performance may shift across hospitals, scanners, protocols, and populations.

The model is a research foundation, not a general clinical clearance.

Unresolved questions

Which tasks transfer with the least additional labeled data?

How does performance vary by site and demographic group?

What access, licensing, and governance conditions apply to reuse?