Wren Alcott

I am a PACS application analyst for a mid-sized health system. Not a radiologist, and not the infrastructure team either — I am the layer in between. I build the hanging protocols, the worklist rules, the study routing and the report integration. When a radiologist says "it used to put the priors on the right," that is me.

Eleven years of it. Two PACS migrations, a RIS replacement, more DICOM tag morphing than anyone should admit to. For the last four years, an accelerating parade of AI pilots, all of which have to land somewhere in a workflow I own.

That is what I write about: not whether the model is clever, but where its output actually goes. Does the result arrive as a DICOM SR the viewer can use, or a secondary capture screenshot nobody can act on? Does it land on the right accession when the tech scanned the patient under the wrong order? Does the flag appear before the radiologist opens the study, or four clicks into a worklist they have already prioritised by hand?

Hal's line is that generation got cheap and verification did not. One layer down, mine is that buying the model got cheap and fitting it into somebody's day did not. Anyone can sign a pilot. Somebody then has to decide which worklist it writes to, what the radiologist sees when it is uncertain, and what happens to the study that was already reported before the result came back.

Some priors, so you know where I stand:

• The model is not the hard part. Getting its answer in front of the right person, at the point in their workflow where they can still act on it, is the hard part — and it is the part no demo shows.
• A DICOM conformance statement is optimistic fiction until you have tested it against your own study descriptions.
• Every vendor demos on a clean viewer with one study open. Ask to see it inside a hanging protocol, in a list of ninety, at the end of a shift.
• The output format is a clinical decision, not a technical one. A result that arrives as an image nobody can query is a result that quietly does not exist.
• A tool that is configured badly does not get escalated. It gets ignored, and six months later somebody reports the pilot as a failure of the model.