Skip to content
DEXPRADEXPRA home

Applied AI

Imaging AI Standards: Everything Is Optional

The standards for artificial intelligence in medical imaging are close to complete on paper. Almost none of them binds. Every IHE Radiology profile for AI is trial implementation or earlier, the supplement that would have let a system discover what AI is on the network was cancelled outright, and the one genuinely binding transparency rule in the United States attaches to certified EHRs rather than to the PACS and orchestrators where imaging AI actually runs. This paper is the map: who specified what, what each document explicitly declines to cover, and which of those gaps shows up in a procurement. It is the paper behind the article on the six handoffs between the scanner and the chart, and it goes considerably further — the technical stack from the worklist entry that triggers inference to the coded term that lands in the EHR, then governance, regulation across three jurisdictions, and who pays. Written from the primary specifications rather than from the coverage, with status and revision dates as published in September 2026.

What’s inside

  • The six layers a study passes through, the document that governs each, and its real maturity: no final text anywhere in the stack
  • AIW-I opened up: seven actors, nine transactions, and the four things it declares out of scope — two of which became profiles five years later and one of which was cancelled
  • The hole where discovery should be: why DICOM Supplement 224 mattered, and what every orchestrator does instead
  • Where AIR and AIRA disagree about what an AI result is, drawn object type by object type, and why trial implementation let it happen
  • AIR's fourteen exclusions verbatim, and the three at the end that a buyer assumes are included
  • The AIRA assessment vocabulary, and why Added and Unassessed are the two values home-grown feedback schemes leave out
  • The provenance chain from a finding to the clearance it was produced under, and the exact point where DICOM runs out of attributes
  • Leaving the department: the informative one-way mapping guide, ImagingSelection as R5-only, and why US Core skipping R5 is the part that costs money
  • The ACR–SIIM practice parameter section by section — what it asks for, and what kind of requirement a practice parameter is
  • PCCP adoption measured: 34 of 1,080 cleared radiology AI devices carry one, and what the public summaries of those 34 do and do not commit to
  • Regulation in the United States, Canada and the European Union, including the Digital Omnibus dates that moved
  • CPT Appendix S, the four payment routes, and the SaMS proposal that would make clinical software a payable category
  • Twelve questions for a tender that cannot be answered with a logo slide, and a glossary of the twenty-odd acronyms
PDF · 26 pages

Request this resource

Working on one of these problems?

Tell us what you’re working on in interoperability, data platforms, or clinical AI, and we’ll tell you what we think.

Get in touch