Dermatology Image Annotation FHIR Implementation Guide
0.1.0 - ci-build
Dermatology Image Annotation FHIR Implementation Guide
0.1.0 - ci-build
Dermatology Image Annotation FHIR Implementation Guide - Local Development build (v0.1.0) built by the FHIR (HL7® FHIR® Standard) Build Tools. See the Directory of published versions
This page provides a technical overview of the three Observation profiles defined by this IG. For formal definitions, see the Artifacts page.
The DermatologyImageAnnotation profile represents a single annotation result from clinical dermatology photography. It follows the US Core Observation pattern with component slicing for structured data elements.
Category pattern. Each instance requires two category entries: imaging from the standard Observation category system, and dermatology-annotation from this IG's AnnotationType CodeSystem. Category slicing uses a #pattern discriminator on $this with rules = #open.
derivedFrom requirement. Every annotation must reference its source image through derivedFrom, which accepts references to ImagingStudy, DocumentReference, or Media resources. This creates a traceable link from clinical finding to source evidence.
Method binding. The method element uses a required binding to AnnotationMethods, which captures how the annotation was produced (expert manual, clinician point-of-care, AI-assisted confirmed, AI-assisted corrected, or AI inference).
Component slicing. Components use a discriminator of type #pattern on path code with rules = #open, following US Core conventions. Required slices include skinFinding (1..*) with an extensible binding to DermatologyFindings. Optional slices cover confidence score, segmentation mask, bounding boxes, QA review status, and inter-annotator agreement.
The AIPrediction profile represents an AI model's classification output before clinician review. It is distinct from DermatologyImageAnnotation in several ways:
AIAST security tagging. Every AIPrediction instance must carry meta.security with the AIAST code from v3-ObservationValue, per the HL7 AI Transparency on FHIR IG. This ensures that downstream consumers can identify AI-generated content through a standard mechanism.
Device requirement. The device element is required and must reference the Device resource that identifies the AI model, including its name, manufacturer, and version. The performer element SHOULD NOT be populated for pure AI inference.
Status constraint. The status is bound to a restricted ValueSet that excludes final. AI predictions remain preliminary until a clinician acts on them. A clinician's acceptance produces a separate DermatologyImageAnnotation with method ai-assisted-confirmed or ai-assisted-corrected.
AI-specific components. Required slices include predictedFinding (1..*) and predictionConfidence (1..1). Optional slices cover heatmap (gradient visualization URL) and inferenceTimeMs (model latency).
The AnnotationSeries profile groups annotations for the same lesion over time using the hasMember pattern.
hasMember pattern. The hasMember element requires at least two references to DermatologyImageAnnotation resources. The series does not impose temporal ordering on the references themselves. Consumers MUST sort by the effectiveDateTime of referenced observations to reconstruct the correct temporal sequence.
Change detection components. Required slices include changeDetected (boolean indicating whether the classification changed between time points), earliestDate, and latestDate. An optional changeType slice uses a required binding to the ChangeType CodeSystem for structured change classification.
Status and performer. The series status is fixed to final because it represents a computed summary. The performer element SHOULD NOT be populated because the series is system-generated.