Text-prompted segmentation for the structure nobody contoured

A newer class of model takes a free-text description — a word or a clinical sentence — and returns a 3D mask, with no per-structure training. VoxTell is the current open example: trained on 62,000+ CT, MRI and PET volumes across more than 1,000 anatomical and pathological classes, with zero-shot performance on unseen datasets and tolerance for the way clinicians actually phrase things.

Where this is worth the trouble here. The structure set that comes with a case is fixed, and it never contains everything a plan should look at. The swallowing structures are the standard example — constrictors and larynx are usually in nobody's objective list, and sparing them costs nothing measurable. Prompting for a structure that was never contoured lets you look at a dose you would otherwise never see.

Two rules, and they are not optional.

1. Never score against a prompted contour. The leaderboard is computed on the published structure set. A structure you generated exists for your own optimization and your own understanding, and quoting a dose to it as though it were comparable is comparing your contour with someone else's.
2. Say which structures were generated in the write-up, with the prompt used. A number nobody can reproduce is an anecdote, and the prompt is the method section.

What it is bad at. The same weakness as every automated contour, sharper: unusual anatomy, post-surgical change, anything where the name and the picture disagree. It will confidently return a plausible mask for a structure that is not there.