Medical imaging AI integration: intended use, PACS/RIS interfaces and evaluation evidence

Medical imaging AI integration starts with who receives an output, when it appears, how it affects work and who handles failure. Model performance, interface contracts, user review and intended-use evidence need separate assessment.

This is a technical preparation guide, not a verified hospital deployment report or diagnostic or treatment advice. Drafts, measurements, image sorting and finding suggestions have different effects. An assistance label does not establish that a function is safer or exempt from medical-device requirements.

Describe actual functions, not only product labels

TFDA's device-scope guidance considers purpose, functions, mechanism and risk. Claims must match actual behavior; a signature or draft label does not independently settle classification. The official computer-aided detection/diagnosis update concerns specific evaluation work.

This is a use-question list, not classification. Do not categorically call measurement or triage administrative software, or substitute revised marketing text for assessment. Seek authority clarification with concrete functional information when attributes are uncertain.

Inspectable output is not validated interpretation

For a hypothetical report-draft workflow, establish approved inputs and outputs. Retain sources, versions, generation time and pending-review status. Text, measurements and marks should lead to the relevant image or record for designated reviewers.

Pixels and citations supply inspection locations, not proof of correct measurement or interpretation. Leave unresolved material as unresolved. Assess false positives, material omissions, incorrect measurements and wrong-patient or study association for the intended task. This guide does not rank clinical harm across every situation.

Agree who may edit, use, sign or return outputs and how responsibility is assigned. Review is one control; it does not alone establish device clearance, unchanged liability or fitness for care.

Confirm the PACS/RIS interface contract

Agree hospital and vendor support before selecting a format. DICOM SR and secondary capture are not interchangeable annotation formats. DICOM's Secondary Capture definition describes image objects converted from non-DICOM formats. An image object does not automatically provide reusable structured measurements or reports.

For HL7/FHIR or existing notification paths, verify actual versions, messages, fields, permissions and acknowledgements. A standard's name does not prove interoperability. This guide adds no autonomous diagnosis or urgent-alert service promise.

Data availability is not permission

Existing retrospective data are not automatically available for testing or training. Establish purpose, medical-data conditions, recipients, access and safeguards under Taiwan's PDPA, with retention and deletion owners.

DICOM confidentiality profiles do not guarantee whole-object de-identification through attribute processing alone. Pixels, other attributes and context can remain identifying. Inspect these in the approved export process rather than treating removal of a name field as sufficient. On-premises deployment is not blanket data-use permission.

Prepare evaluation for the defined use

Specify population, equipment, imaging workflow and users, then establish data-use conditions and representative evaluation material. Separate development and held-out tests. Appropriate authorized reviewers create versioned references, recording disputes and exceptions.

Assess errors, material misses, review burden and integration failures according to function. Vendor reports are inputs whose data, version and use need comparison with this project; do not assume all vendor evidence is mismatched. Owners determine acceptance criteria and any further clinical evidence under applicable requirements.

Before a pilot, agree monitoring, issue ownership, stopping and recovery. Reassess affected work after model, interface, equipment or workflow changes. Successful connection, usage, physician signatures and clinical acceptance are different evidence and cannot substitute for one another.

The existing healthcare scope covers administrative and reviewed preparation. The chart-summary and candidate-code guide examines source, version and review in another integration context.