Skip to content
DEXPRADEXPRA home

Article · Applied AI

Imaging AI: six handoffs between the scanner and the chart

An imaging AI deployment is not one integration. It is six, each governed by a different standard, and the two nobody scopes are the last two: what the radiologist did with it, and who pays.

DEXPRA · · 9 min read

The stack exists and none of it is finished

The first wave of imaging AI arrived as products rather than as participants. Each one came with a viewer, a routing arrangement and its own idea of how a finding should reach a radiologist, which is how a reading room ends up with four screens and one patient. The standards that fix this now exist, and they cover the whole path: asking for the work, encoding the result, displaying it, recording what the reader did with it, putting the number in the chart, and watching the model afterwards.

What does not exist is a finished stack. Not one IHE profile on that path has reached final text: five are trial implementation, and the sixth is still out for public comment. profiles.ihe.net says so with dates against each one.

Every profile on the path, and the state IHE publishes it in.
ProfileWhat it coversStatus
AIW-IRequesting and managing inference over the DICOM worklistTrial implementation, August 2020
AIREncoding, storing, retrieving and displaying the resultTrial implementation, revised August 2025
AIRARecording the assessment of an individual resultTrial implementation, June 2025
IRAReporting applications inside the reading environmentTrial implementation, October 2023
IDRPackaging verified measurements into discrete report fieldsPublic comment, March 2026
MADOManifest-based access to the studies themselvesTrial implementation, revised September 2026

Trial implementation is not a quality warning. It means the text is stable enough to build against and IHE is still collecting what implementers hit, which is exactly the state a profile should be in while products are being written to it. It does mean two conformant systems can still differ in the corners, and it means an answer of "supports AIR" in a procurement response is the beginning of the conversation rather than the end of it. Ask which actor, which transactions, which optional elements, and when it was last taken to a Connectathon. Public comment is a stage below that, so a conformance claim against IDR today is a claim about a draft.

What follows walks the path in the order a study actually moves through it, and then the two layers that decide whether any of it survives contact with an operating budget.

Six handoffs

An imaging AI deployment is not one integration. It is six, and they fail differently: the early ones fail loudly, as a study that never arrives or a task nobody claims, and the late ones fail quietly, as a finding no clinician can act on or a model nobody is measuring.

SCANNER TO CHART · SIX HANDOFFS

Open a handoff for the standards that govern it and the payload that crosses it. Every tag, template and resource name is the real one; the payloads are shortened to the elements under discussion and none of them is a complete instance.

HANDOFF 01

Reaching the priors without moving the pixels

A comparison model is worth little without the prior study, and fetching every prior in full is how an imaging AI rollout takes the archive network down with it. MADO publishes a manifest instead: a FHIR document listing what exists, series by series, with the retrieve endpoint attached. The router reads the inventory, decides what it wants, and asks for that.

GOVERNED BY
IHE MADO, PIX/PDQ, ATNA
ON THE WIRE
WADO-RS Retrieve [RAD-107]

Identity comes first. PIX or PDQ reconciles the patient identifier across enterprise boundaries before anything is retrieved. A model that quietly compares two patients is worse than no model.

ATNA is not paperwork. An automated consumer reading imaging is an audit event and a node that has to authenticate before it gets one. The audit record is also the only evidence available later that the training set was assembled from studies the organisation was allowed to use.

APPLICATION/FHIR+JSON
{
  "resourceType": "Bundle",
  "type": "document",
  "entry": [
    { "resource": { "resourceType": "Composition",
        "type": { "coding": [{ "system": "http://loinc.org",
                               "code": "18748-4" }] } } },
    { "resource": {
        "resourceType": "ImagingStudy",
        "status": "available",
        "identifier": [{ "system": "urn:dicom:uid",
                         "value": "urn:oid:1.2.840.10008.5.1.4.1.1.2" }],
        "numberOfSeries": 4,
        "series": [{ "uid": "1.3.6.1.4.1.5962.99.1.2280943358",
                     "modality": { "code": "CT" },
                     "numberOfInstances": 318 }],
        "endpoint": [{ "reference": "Endpoint/wado" }] } }
  ]
}
Only handoff 03 carries the same object in three encodings at once, which is where the integration cost concentrates.

One finding, three encodings, and a version to check

The measurement a model produces gets written down three times. It is authored as a DICOM Structured Report on template TID 1500, because that is what the archive stores and the viewer understands. Supplement 219 is the same report as JSON, with business names bound to the coded tuples in a separate definition file so the payload reads as field names while the codes stay underneath it. The FHIR Observation is the same number again, for the chart.

Three encodings of one fact is a synchronisation problem, and the supplement is careful about it: the JSON supports a full-fidelity round trip, so a conformant binary SR can be regenerated from it. Treat the JSON as a second record that is merely similar and the two drift, which is the kind of defect nobody notices until an audit reads both.

The version trap is at the chart end. A measurement in the record is worth more when a clinician can get from the number back to the frames it came from, and in FHIR that link is ImagingSelection: the resource an Observation points at through derivedFrom to say which study, series, instances or image region produced the value. It is new in R5. It does not exist in R4, US Core still profiles R4, and R4 is what a hospital integration is most likely to hand you. On that endpoint the reference degrades to the study, the frame-level location ends up in a text note nothing can query, and the tether that was promised in the design review is not there.

This is worth finding out in week one rather than in acceptance testing. It is also the cheapest question on the list: ask for the capability statement and read the version off it.

The only ground truth a department produces for free

Every reading session generates labelled examples. A radiologist accepts a finding, corrects a volume, or rejects a detection outright, and each of those is a judgement about a specific model output on a specific study. In most deployments it is a click that goes nowhere.

AIRA is the profile that makes it data. It defines the recording and transfer of the outcome of assessing an individual AI result, by a human reader or by another process, as an object in its own right, defined across the range of DICOM objects a result might be. AIR is what gets the result in front of the reader in the first place, and sets the baseline an image display has to meet so the same finding looks the same on the next workstation.

Two things follow from capturing assessments, and both are worth the integration on their own. An unverified machine output can be kept out of the signed report, because the verification state is recorded rather than assumed. And the department acquires a local evaluation set that matches its own scanners, its own protocols and its own patients, which is the only evidence available when a vendor claims a new version is better.

Keep the original. A correction that overwrites the model output destroys the comparison. Store both values and which reader made the change, and the pair becomes the input to everything in the next section.

What a registry can and cannot see

The ACR runs Assess-AI inside the National Radiology Data Registry. A participating site submits de-identified algorithm output, report text and DICOM study metadata through ACR Connect. The registry derives a surrogate label from the report text using a prompting pipeline over commercial models held under a managed service, compares it with what the algorithm said, computes concordance centrally, and returns it on a dashboard. It covers a named set of use cases, intracranial haemorrhage and pulmonary embolism among them, rather than whatever a site happens to be running.

Two limits come with that design, and neither is a criticism of the registry. The reference is what the report said rather than what was there, so extraction error and model error arrive inside the same number. And a registry tells a site that something moved, not what moved it: which scanner, which software version, which population is a question for the local lineage record.

A third limit is arithmetic, and it is the one that catches change control plans. A concordance rate computed on a month of studies is an estimate with an interval around it, and a fall smaller than that interval is not a signal. At three hundred monitored studies a month and a baseline of 94%, the 95% interval on one month is about 2.7 percentage points wide either side. A trigger written at two points fires on the calendar.

The detection horizon

ALPHA 0.05 · POWER 0.80
94%
80% · NOISY REFERENCE99% · NEAR CEILING
3.0 pp
0.5 pp · STRICT10 pp · OBVIOUS
300
25 · ONE MODEL, ONE SITE1,500 · ENTERPRISE
CASES NEEDED
555
at 80% power
TIME TO CALL IT
1.9 months
at this volume
ONE MONTH, 95% CI
± 2.7 pp
Wilson, on one month
FLOOR IN A YEAR
1.2 pp
smallest detectable fall
Cases needed against the size of the fall, from a baseline of 94 percent. Catching a fall of 3.0 pp takes 555 monitored cases, which is 1.9 months at 300 studies a month. A year of that volume is 3,600 cases, so the smallest fall detectable within a year is 1.2 pp.CASES NEEDED10^110^210^310^410^50.52.04.06.08.010.0FALL IN CONCORDANCE (PERCENTAGE POINTS)ONE YEAR AT THIS VOLUME

A fall smaller than the interval is not a trigger. A dashboard that moves from 94.1% to 92.8% on a month of three hundred studies has moved less than its own uncertainty. Set the bound in a change control plan below the floor above and the plan commits to noticing something the monitoring cannot see.

Three things make the real horizon worse than this one. The reference label is extracted from a report rather than established, and error in a reference makes a true difference harder to detect. Checking every month is repeated testing, so some of the alarms are the calendar rather than the model. And a fall concentrated in one scanner or one subgroup is diluted in a pooled rate, which is the argument for stratifying before the volume gets spent.

Sample size for a one-sample test of a proportion, and the Wilson interval, both closed-form. Neither is a model of drift: they bound what the monitoring can see.

Read the panel the other way round and it becomes a design constraint. Decide the fall that would matter clinically, read off the cases needed, divide by the volume the site can actually monitor, and that is the earliest the plan can honestly promise to notice. A predetermined change control plan needs an observable trigger, and a trigger set below the detection floor commits a manufacturer to seeing something the monitoring cannot show.

Stratifying makes it worse before it makes it better. A fall concentrated in one scanner model or one subgroup is diluted in a pooled rate, so the pooled rate is the wrong place to look, and every stratum divides the volume that was already the binding constraint. That is an argument for choosing the strata that matter in advance, not for pooling.

Who pays, and what is on the table this autumn

None of the above gets built twice, so the second question in every business case is where the money comes from. There are four routes and they are not alternatives so much as different settings.

Four routes, and what each one turns on.
RouteSettingStandingWhat it turns on
CPT Category IIIOutpatient professionalTemporary, tracking onlyPriced by the local contractor, so payment varies by region and may be nothing at all
CPT Category IOutpatient professionalPermanent, nationally pricedPublished evidence and established use. CT-derived coronary FFR reached it as 75580 in the 2024 code set
CMS NTAPInpatientTwo to three yearsThe lesser of 65% of the technology cost or 65% of the amount by which the case exceeds the DRG payment
OPPS SaMSHospital outpatient facilityProposed for CY 2027Status indicator O1 on 36 designated codes, with 21 of them moved into new technology APCs

The fourth row is the live one. In the CY 2027 OPPS proposed rule, CMS renames the software codes Software as a Medical Service, designates 36 HCPCS codes as SaMS, creates status indicator O1 to mark them separately payable, and reassigns 21 of them from clinical APCs into new technology APCs at rates approximating what they are paid now. CMS-1850-P was published in the Federal Register on 7 July 2026 and comments closed on 31 August 2026, so the final rule is due in the next few weeks. A facility building a case for an algorithmic service this quarter is writing it against a payment category that does not exist yet.

No rate appears anywhere in this piece, and the omission is the point. A Category III code has no national rate to quote, because pricing is left to the contractor. SaMS is a proposal. A platform licence is whatever was negotiated. A calculator that multiplies an invented per-study figure by your volume returns a number that cannot survive a capital committee, and the number it returns is the one thing in the business case nobody will forget.

The catalog

Twenty-one entries: the profiles, the DICOM supplements, the FHIR resources, the vocabularies, and the governance and payment rules that sit over them. Search it or filter it by family. The five IHE profiles carry the state and date IHE publishes for them; nothing else carries a status, because there is no maturity score for a DICOM supplement that would not have been invented here.

TWENTY-ONE ENTRIES · PROFILES, SUPPLEMENTS, RESOURCES, RULES

21 OF 21 SHOWN

Every entry is in the document whether or not it matches the filter, so in-page search and a screen reader both reach all twenty-one.

What to build first

The order matters more than the choice of vendor, because two of these cannot be added afterwards.

  • Identity and audit, before the first retrieval. PIX or PDQ and ATNA. An automated consumer reading across boundaries needs the same identity discipline as a person, and the audit trail is the only later evidence of which studies a dataset was drawn from.
  • The worklist, before the second algorithm. One model can be integrated point to point. The second one is what turns a routing table into a maintenance problem, and AIW-I over UPS-RS is the answer that scales past it.
  • The structured report as the record, not the export. Author on TID 1500 and derive the JSON and the Observation from it. Deriving the SR from a proprietary result later is the expensive direction.
  • Assessment capture from day one. This is the retrofit that costs the most. A year of reading with no assessment capture is a year of ground truth that no longer exists, and no amount of later integration recovers it.
  • The chart mapping, once the FHIR version is known. Design for what the endpoint actually is rather than for what the specification allows.
  • Registry submission, and the monitoring plan it feeds. Decide the fall that matters and the volume available to see it before writing a trigger into anything a regulator will read.

The standards on this path are good. They were written by people who had watched the alternative, and taken together they describe a department where a finding travels from a scanner to a chart without anybody retyping it. None of that is the same thing as a stack a hospital can buy, and the six handoffs are where the difference lives.

If you want the whole of it, Imaging AI Standards: Everything Is Optional is the paper behind this piece. Twenty-six pages on the six layers and the document that governs each, the discovery supplement that was cancelled and what its absence costs, the fourteen exclusions in AIR and the two AIRA contradicts, the provenance chain DICOM has no attribute for, the three regulators and the one rule that actually binds, and twelve questions for a tender. It is free at /resources/imaging-ai-standards.

Related insights

Related resources

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