Skip to content
Imaging Evidence Path

How AI enters an imaging workflow

Tonight · ~20 min · read · energy: low · setup: none

An AI model is not a website someone visits. It is a service that must be wired into the imaging workflow: triggered by the right study, given the right inputs, and made to return a result that re-enters the system with provenance. This lesson is that wiring, and the points where it goes wrong.

The anchor: retrieve, infer, return, prove

A deployed AI service does four things, in order, and each is a failure surface:

  1. Retrieve the right study/series (via DICOMweb or DIMSE), triggered by a routing rule (e.g. “all head CTs on arrival”).
  2. Infer — run the model, producing a derived object (a SEG, a measurement, a finding).
  3. Return that object into PACS, linked to the source study (STOW-RS or C-STORE), so a viewer can retrieve and display it.
  4. Provenance — record what model/version/inputs produced the result, when, at which site. Without this, the result is unauditable.
flowchart LR
    P["PACS"] -->|"retrieve study<br/>(QIDO/WADO or C-FIND/MOVE)"| A["AI service"]
    A -->|"infer"| D["SEG / SR / measurement"]
    D -->|"return, linked to source<br/>(STOW/C-STORE)"| P
    A -.->|"provenance:<br/>model, version, inputs, time"| PR["audit store"]

Where this wiring fails

Each arrow is a place a correct model output becomes clinically useless or dangerous (lesson 6 expands this):

  • Wrong study / wrong series retrieved — the routing rule fired on the localizer instead of the axial volume (Ch. 1), so the model ran on the wrong input.
  • Result attached to the wrong patient/study — an identity mismatch lands the SEG on the wrong StudyInstanceUID.
  • Result not retrievable by the viewer — the SEG/SR exists in the archive but the reading workstation cannot query or render it.
  • No provenance — nobody can reconstruct which model version produced which result, so a later failure cannot be traced.

Why structured return objects matter

The output of “infer” determines whether the result is machine-usable. Returning a DICOM SEG (a segmentation that references its source series and carries geometry, Ch. 2) or a DICOM SR (a tree of coded findings/measurements referencing source images) makes the result queryable, computable, and linkable to the source voxels. Returning a PNG screenshot or PDF does none of that — it is a picture of a result, not a result. Lesson 7 draws this contrast sharply; here, the point is that the return step is where you decide which kind of object you produce.

TRACE-CT, in miniature

TRACE-CT is single-machine/local today, so this networking is future — but it already practises the discipline that makes a future deployment safe: its QC-bundle schema encodes provenance (SHA-256 of inputs/outputs, pipeline version, geometry entries), its SEG decode (decode_seg_file()) is exactly the structured-derived-object handling this lesson describes, and aggregate_qc_reviews.py produces privacy-safe aggregate evidence. The EuSoMII AI-assisted workflow pipeline (notes §3) is the deployment shape this chapter explains.

Stop and think — then reveal

An AI service runs perfectly on the right study, produces a correct tumour volume, and STOWs a PNG screenshot of the segmentation into PACS. The radiologist sees it and agrees. What has been lost that a DICOM SEG would have preserved?

Almost everything that makes the result usable downstream: the source-image references (which series/slices it was drawn on), the per-frame geometry (so the mask can be overlaid or measured in a tool), machine-readability (you cannot query, aggregate, or re-measure from a PNG), and provenance of the voxels (the screenshot does not say which voxels are tumour, only what the picture looks like). A future step — recomputing a feature, auditing the boundary, or aggregating volumes across patients — is impossible from the screenshot. The SEG is a first-class data object; the PNG is a display artefact (lesson 7).

What to retain

  1. A deployed AI service does retrieve → infer → return (linked to source) → provenance. Each step is a failure surface.
  2. Routing can fire on the wrong series; identity mismatches attach results to the wrong study; viewers may not retrieve/display the object; missing provenance makes results unauditable.
  3. The return object decides machine-usability: SEG/SR are first-class; PNG/PDF are display artefacts (lesson 7).
  4. TRACE-CT already practises the discipline (QC-bundle provenance, structured SEG handling) that a future deployment needs.

Next: the painful lesson — why a correct AI result can still be clinically invisible.