MajwareMAJWARE
Free templateEditable8 systems7 interfaces

Radiology AI Integration Architecture Diagram

Hospitals rarely run one AI algorithm — they run several, from different vendors, some on-premises and some in the cloud. Connecting each one directly to PACS does not scale. This diagram shows the orchestrator pattern: one AI platform receives studies, decides which algorithms should see them, de-identifies data that leaves the network, and returns results to PACS and the reading worklist in standard formats.

Open & edit in Diagram Builder →Download SVG

Free, no sign-up. The builder runs in your browser; nothing is uploaded.

Radiology AI Integration Architecture Diagram: RIS / Worklist, Modality, PACS, AI Orchestrator, AI Engine, Diagnostic Viewer, De-identification, Cloud AI
Radiology AI Integration Architecture Diagram — 8 systems and 7 labelled interfaces. Free to use in your own documentation.

What this diagram shows

It follows the direction of the IHE AI Workflow for Imaging (AIW-I) and AI Results (AIR) profiles.

Components

Modality and PACS

Acquire and store as usual; PACS forwards eligible studies to the orchestrator.

AI orchestrator

Filters studies by modality, body part and protocol, sends them to the right algorithms, collects results and routes them back.

On-prem AI engine

Runs inference inside the hospital network, so no data leaves the site.

De-identification

Removes or replaces patient identifiers before images go to an external AI service, and keeps the mapping needed to re-identify results.

Cloud AI

Vendor-hosted algorithm receiving de-identified studies over DICOMweb (STOW-RS).

RIS / worklist

Receives triage flags so suspected critical findings move up the reading list.

Diagnostic viewer

Displays AI overlays, measurements and findings from DICOM SR, SEG or presentation states.

How the data flows

  1. Acquire and store

    The modality sends images to PACS with C-STORE.

  2. Route

    PACS auto-forwards eligible studies to the orchestrator, which matches them against each algorithm’s criteria.

  3. Infer on-prem

    Studies go to the on-prem AI engine and results come back to the orchestrator.

  4. Infer in the cloud

    For cloud algorithms the orchestrator sends studies through de-identification, then to the vendor with STOW-RS. Results are re-identified on return.

  5. Return results

    Findings go to PACS as DICOM Structured Reports (SR), Segmentations (SEG) or secondary captures, and urgent findings raise a triage flag on the worklist.

  6. Review

    The radiologist sees AI results alongside the images in the diagnostic viewer and accepts or rejects them.

Interfaces in this diagram

Every connection in the diagram, ready to copy into an interface register or test plan.

#FromToProtocolDetail
1Modality (CT · MR · DX)→PACS (Image Manager)DICOMC-STORE
2PACS (Image Manager)↔AI Orchestrator (Routing rules)DICOMImages ⇄ SR
3AI Orchestrator (Routing rules)↔AI Engine (On-prem inference)DICOMInference
4AI Orchestrator (Routing rules)→RIS / Worklist (Reading priority)HL7Triage flag
5AI Orchestrator (Routing rules)→De-identification (PHI removal)DICOMC-STORE
6De-identification (PHI removal)→Cloud AI (Vendor SaaS)DICOMwebSTOW-RS
7PACS (Image Manager)→Diagnostic Viewer (AI overlays)WADO-RS—

Design considerations

  • Prefer standard result objects (DICOM SR TID 1500, SEG, presentation states) over burned-in secondary captures, so results stay measurable and searchable.
  • Keep a de-identification profile per AI vendor based on DICOM PS3.15 Annex E, and test it on real headers — burned-in text and private tags are the usual leaks.
  • Track every AI result against the final report to monitor performance over time.
  • Make sure AI results never overwrite or replace the original images.
  • Size the orchestrator for peak volume; a full day of CT can arrive within a few hours.

Majware software

DICOM Anonymizer Pro

For batch de-identification before data leaves your network: a Windows desktop tool that remaps UIDs consistently across studies, shifts dates and writes a tag-by-tag audit log.

See DICOM Anonymizer Pro →

Related reference

Frequently asked questions

What is an AI orchestrator in radiology?

An AI orchestrator (or AI platform) is middleware between PACS and multiple AI algorithms. It decides which studies each algorithm should process, handles transfers, collects results and returns them to PACS, the worklist and reporting in standard formats.

How are AI results returned to PACS?

Usually as DICOM objects: Structured Reports for findings and measurements, Segmentation objects for outlines, presentation states for annotations, or Secondary Capture images. Triage results may also go to the worklist as HL7 or FHIR messages.

Do images need to be de-identified for cloud AI?

That depends on your contract and local regulations. Many organisations de-identify data before it leaves their network using the DICOM PS3.15 Basic Application Level Confidentiality Profile, keeping a mapping so results can be matched back to the patient.

Make it match your site

Rename systems, add your AE titles and message types, then export PNG, SVG or PDF — or send a colleague a share link.

Open in Diagram Builder →

More architecture templates

All 12 templates →
PACS Architecture Diagram — preview
Imaging

PACS Architecture

Modalities, RIS worklist, PACS image manager, diagnostic and web viewers, and long-term VNA archive.

7 systems · 7 interfaces

Radiology Workflow Diagram (RIS–PACS) — preview
Imaging

Radiology Workflow

End-to-end order-to-report workflow: CPOE order, worklist, acquisition, PACS, reading and reporting.

6 systems · 7 interfaces

DICOM Modality Worklist (MWL) Workflow Diagram — preview
Imaging

Modality Worklist

How an HL7 order becomes a DICOM worklist entry, plus MPPS status and storage to PACS.

5 systems · 5 interfaces