Chapter 10 — Imaging Informatics: Objects, Workflows, and Provenance

This chapter opens Part IV by following an imaging examination beyond image formation. A model can produce a technically convincing result and still fail to affect care if the result cannot be retrieved, is associated with the wrong study, or cannot be explained later. Those are not properties of the model alone. They are properties of a distributed path through objects, identifiers, transports, workflow state, viewers, records, and people.

The central problem is easy to miss when an imaging system is represented as a file transfer. A scanner does not simply produce a file that a radiologist opens. An order is scheduled, an examination is acquired, objects are identified and stored, a report is created, and results cross boundaries between imaging and hospital information systems. An AI service may enter that path as another consumer and producer of imaging objects. At each hand-off, the bytes can be valid while the relationship between the bytes and the intended clinical context is wrong or unusable.

An examination is a distributed workflow

It is useful to read the system by function rather than by a supposedly universal hospital topology. One installation may separate a PACS, a vendor-neutral archive, a radiology information system, a viewer, and an electronic health record. Another may combine some of those roles or put an AI service inside an existing product. The physical arrangement is an implementation choice. The functions still need to be accounted for: an imaging modality produces objects; an archive stores and serves them; order and reporting systems coordinate work; a viewer presents the study; the wider clinical record receives findings; and an AI or quantitative service may retrieve source material and return derived results.

That functional map gives the chapter its spine. First, separate the identity of an object from the state of the examination that produced it. Then follow the transport mechanisms that move objects, beginning with classic DICOM DIMSE and continuing with DICOMweb. Next, follow an AI result from input selection through return and provenance, including the difference between a structured object and a rendered picture. Finally, ask what interoperability claims, custody promises, and confidentiality transformations actually mean.

The distinctions matter because they fail independently. A study can have valid pixel data but carry the wrong workflow identity. A result can reach an archive but not be retrievable or renderable by the viewer. A transfer can be acknowledged without establishing later safekeeping. A technically careful confidentiality profile can reduce identifiers without deciding whether a particular release is legally anonymous. These are related problems, but no single identifier, protocol, conformance statement, or de-identification flag solves all of them.

The rest of the chapter therefore treats an examination as a chain of explicit questions. Which object is this? Which study and series does it belong to? Which transport carried it? Which workflow state does the receiving system know? What source and processing history must accompany a derived result? Who accepted the object, who is responsible for keeping it, and who can use it? Only after those questions are separated does the route from an image to a clinical result become legible.

The first place to apply that separation is immediately after acquisition, where the examination becomes a distributed state rather than a completed file.