Skip to content

Imaging informatics — reference

Reference · Dense lookup material — the verb tables, profile versions, HL7/FHIR bridge, and failure-mode catalogue. Teaching versions live in the lessons.

Actors (functions, not fixed wiring)

Actor Function
Modality scanner producing DICOM objects
PACS store, query, distribution, viewer/workstation
VNA standards-based (DICOM) archive role (longer-term/cross-vendor)
RIS scheduling, ordering, worklists, reporting, status
EHR hospital-wide clinical record
AI / quant service consumer of images, producer of derived objects

Real deployments blur PACS/VNA, fold RIS into EHR, and place the AI service inside or outside the archive.

DIMSE verbs (classic networking)

AEs (AE Title + host/port); SCU = client, SCP = server; association negotiates Presentation Contexts (SOP Classes + Transfer Syntaxes).

Operation Verb What it does
C-STORE send push an object to an SCP
C-FIND query search by keys
C-MOVE retrieve (via move) source SCP sends to a destination AE (third party)
C-GET retrieve (to self) source SCP sends back to me (often disabled)
Q/R C-FIND + C-MOVE/C-GET find-then-pull loop
MWL worklist C-FIND against the worklist

C-MOVE is three-party: asker tells source to push to a separate destination AE. Final status is precise (PS3.4 Table C.4-2): 0000 Success = all C-STOREs succeeded; B000 Warning = one or more sub-operations failed; A801 Failure = Move Destination unknown (no C-STOREs attempted); other failures include A701/A702 (out of resources) and Cxxx (unable to process). A destination/connectivity/storage failure is therefore never 0000 — it surfaces as B000 or a Failure code. “Nothing arrived despite Success” = delivered to a different named destination, not the asker (expectation mismatch, not a protocol lie).

DICOMweb (HTTP re-skin, parallel transport)

Service Maps to What it does
QIDO-RS C-FIND query studies/series/instances via HTTP
WADO-RS C-GET retrieve objects/pixels/metadata via HTTP
STOW-RS C-STORE upload objects via HTTP
UPS-RS workitem/workflow REST interface to the DICOM Unified Procedure Step

Auth (HTTP/OAuth vs DIMSE/TLS/ACL), paging, and batching differ by transport. Viewer capability problems (e.g. cannot render SEG) are transport-independent.

Structured results

  • DICOM SEG — segmentation referencing source series/SOPs + geometry (Ch. 2).
  • DICOM SR — tree of coded findings/measurements referencing source images.
  • Measurements / provenance — SR can encode measurements with source refs; capture provenance (model/version/inputs/time) or the result is unauditable.
  • Screenshots/PDFs are display artefacts: not queryable/computable/linkable.

IHE profiles (recipes; all Trial Implementation; verified profiles.ihe.net)

Profile Full name Addresses
AIW-I AI Workflow for Imaging (Rev 1.1, TI 2020) inference workflow on UPS-RS — dispatch/track tasks
AIR AI Results (Rev 1.3, TI 2025) storage/retrieval/display of structured AI results
AIRA AI Result Assessment for Imaging (PC→TI 2025) assessment/feedback/monitoring/drift (post-market)

HL7 / FHIR bridge

DICOM = images/imaging objects; HL7 = orders/results/demographics. Order: RIS→modality (MWL), RIS↔EHR (HL7). Result: reporting→RIS→EHR (HL7, increasingly FHIR). FHIR matters because discrete measurements/AI findings need to land as structured resources (e.g. Observation), not PDFs.

Failure-mode catalogue

Silent resampling · identity mismatch / wrong patient · lost provenance · worklist collision · AI result not retrievable · no audit trail · de-id leaks. Diagnose “deployed but ignored” via viewer capability, study linkage, timing, provenance — before blaming the model.

TRACE-CT connection

Single-machine/local today, so networking is future — but it practises informatics discipline: the QC-bundle schema (schemas/visual-qc-bundle-v1.schema.json) encodes provenance (SHA-256 of inputs/outputs, pipeline version, geometry entries), aggregate_qc_reviews.py produces privacy-safe aggregate evidence, and the SEG objects it decodes (decode_seg_file()) are exactly the structured-derived-object handling described here. The EuSoMII AI-assisted workflow pipeline (notes §3) is the deployment shape this chapter explains.

What to retain (chapter summary)

  1. Actors: modality → PACS/VNA → AI service → derived object → reporter → EHR; RIS owns order/worklist.
  2. DIMSE verbs (C-STORE/C-FIND/C-MOVE/C-GET, Q/R, MWL) vs DICOMweb (QIDO/WADO/STOW, UPS) — same concepts, two transports. C-MOVE is three-party.
  3. Structured results (SEG, SR, measurements + provenance) ≠ screenshots/PDFs.
  4. A large share of imaging-AI failures are identity/provenance/retrieval failures; IHE profiles (AIW-I, AIR, AIRA) are the integration recipes.
  5. Imaging is one subsystem in the hospital’s HL7/FHIR fabric; results must cross that bridge to affect care.