DICOM SR (Structured Reporting)

Unlike free-text reports, DICOM SR organizes findings into a strict hierarchical tree structure of Content Items. Each node represents a single observation, such as an image reference, a spatial coordinate, or a coded measurement. This tree uses parent-child relationships to link a detected lesion to its properties, like its bounding box and SNOMED-CT coded morphology, enabling deterministic parsing by downstream systems.

DICOM SR encodes the precise location of a finding using SCOORD and SCOORD3D Content Items. A 2D bounding box is represented as a polyline graphic, typically specifying the top-left and bottom-right corners in pixel or millimeter units relative to a referenced image. For volumetric data, SCOORD3D defines a point in the patient-based coordinate system, allowing exact 3D localization of a lesion.

To ensure semantic interoperability, DICOM SR mandates the use of controlled medical terminologies. Findings are not stored as raw text but as coded triples (Code Value, Coding Scheme Designator, Code Meaning). For example, a 'malignant tumor' is encoded using a SNOMED-CT code, while measurement types like 'Long Axis' use DICOM Controlled Terminology, eliminating ambiguity in multi-vendor environments.

A key feature for CADe integration is the ability to store graphical annotations as presentation state information directly within the SR document. The SCOORD module defines the exact pixel coordinates of a bounding box or ellipse that was displayed to the radiologist. This links the quantitative measurement to the visual markup, providing a complete audit trail of what the CADe algorithm highlighted.

DICOM SR uses Template Identifiers (TIDs) to constrain the structure of a report for a specific clinical use case. A mammography CADe report conforms to TID 4000 (Mammography CAD) , which strictly defines the allowed Content Items, such as 'Finding Site' and 'Probability of Malignance'. This templating ensures that every CADe system outputs a predictable, validated data structure that a PACS can reliably ingest.

The semantic meaning between nodes in the content tree is defined by Relationship Types. Common types include CONTAINS (parent-child grouping), HAS PROPERTIES (linking a lesion to its size), and INFERRED FROM (linking a conclusion to the evidence image). This explicit graph of relationships allows an AI system to programmatically navigate the logic of a diagnostic report.