Skip to content
Imaging Evidence Path

Chapter 5 — Imaging informatics

This chapter · ~2.5 h total across 8 lessons · How images and results move through healthcare systems · Energy: low · Setup: none

A quantitatively perfect model that cannot be retrieved in PACS, lands on the wrong patient, or loses its provenance is of little clinical value — and sometimes dangerous. A large share of real imaging-AI failures are informatics failures, not model failures. This chapter teaches the architecture and the failure-mode lens through workflows, not acronym lists.

The thread this chapter pulls

A CT does not travel from the scanner to a radiologist, or from an AI service to an EHR, by magic. It travels through a system of actors (modality, PACS, RIS, EHR, AI service) speaking specific protocols (DIMSE, DICOMweb, HL7/FHIR). Understanding that journey — and the identity, provenance and retrieval failures that happen at every arrow — is what separates an AI result that changes care from one that is clinically invisible. This chapter follows an image through that journey and asks, at each step, what can go wrong.

Lessons

  1. What happens after the scanner finishes a study? — modality, PACS, RIS, EHR, archive, viewer.
  2. How DICOM objects actually move — association, SCU/SCP, C-STORE, C-FIND, retrieve.
  3. Why C-MOVE is strange when you come from HTTP — the third-party destination.
  4. DICOMweb: the same imaging world through HTTP — QIDO/WADO/STOW.
  5. How AI enters an imaging workflow — retrieve, inference, return SEG/SR, provenance.
  6. Why a correct AI result can still be clinically invisible — viewer support, wrong study, workflow failures.
  7. Structured result vs screenshot — SEG/SR, machine usability.
  8. IHE: profiles as workflow recipes — AIW-I, AIR, AIRA.

Dense reference — the DIMSE/DICOMweb verb tables, the IHE profile versions, the HL7/ FHIR bridge, and the full failure-mode catalogue — lives in the informatics reference.

How this chapter connects

  • It assumes Chapter 1’s DICOM objects and identifiers.
  • Chapter 9 returns to the AI-result lifecycle (drift, monitoring, AIRA) once a model is deployed.

If you only have an hour, lessons 1, 5 and 6 are the core: the journey of an image, where AI plugs in, and why a correct result can still vanish.