Structured results and downstream use

The return step of an AI integration determines whether the output remains data or becomes only a display. A reader may need a rendered overlay or report for an immediate decision, but a later measurement, audit, query, or research analysis may need the structured object from which that display was produced. The distinction is not that a screenshot has no clinical use. It is that a screenshot or PDF is primarily a presentation of a result, whereas a structured DICOM object can carry the relationships and semantics needed for further computation through a structured DICOM object.

What structured objects preserve

DICOM SEG represents a segmentation as an object with structured pixel data and references to source images or series, together with the spatial information needed to interpret the mask. DICOM SR represents coded content such as findings or measurements in a structured report tree and can reference source images. The precise information available depends on the information-object definition and the implementation, so the useful claim is bounded: these object types provide a machine-readable place for information that a rendered picture alone does not carry.

The distinction is easiest to see in a downstream request. Suppose a collaborator asks for per-lesion volumes, the source series identifiers, or a re-analysis of a segmentation boundary. A PNG does not by itself contain voxel labels, physical geometry, or source references in a form that permits those operations. A PDF may communicate the result clearly to a human while still leaving those machine-level inputs outside the document. If the original SEG, SR, or equivalent structured representation was discarded, the missing information cannot generally be recovered from the rendered image.

Choose the return object for the next use

This is why result design belongs inside the workflow rather than at the end of a model demo. A structured segmentation can support an overlay, a volume calculation, or a later audit if the receiving systems preserve and interpret its references. A structured measurement can be queried or passed to another system if its coded meaning, subject, study relationship, and provenance survive the hand-off. A rendered output can still be the right view for a person, but it should not be mistaken for a complete substitute for the source result when downstream use requires structure.

The choice is difficult to reverse. Once a workflow retains only a picture, a later process cannot infer all discarded labels, geometry, source identity, and processing history with certainty. Keeping both a structured object and an appropriate human rendering may be necessary, but that is a workflow and custody decision rather than a universal requirement that every output use SEG or SR.

The result therefore has two audiences that must not be confused. The viewer needs a usable presentation; other services, auditors, and future analyses need an object whose identity, relationships, and semantics are explicit. The next layer of the chapter asks how standards-based workflow profiles coordinate those audiences and the hand-offs between them.