DICOM3D AI Hub

This contest edition demonstrates DICOM3D’s native InterSystems AI Hub interoperability workflow using the previously qualified Babylon Lite imaging interface. The broader DICOM3D platform continues independently with its Three.js/R3F-based v0.7.0-alpha.1 research prerelease.

Contest installation and demonstration · Native agent source · Broader platform prerelease

This contest edition demonstrates a bounded decision-making task: an IRIS Interoperability production uses native %AI.Agent to explain a recorded public-research segmentation finding, retrieve restricted FHIR R5 evidence, validate its exact references and select the corresponding lesion in the imaging viewer. A synthetic source-identity mismatch requests review instead of inventing a correction.

The integration reuses DICOM3D’s existing Babylon Lite interface. The viewer alone is not the contest agent; the observable native evidence/decision/validation workflow is the contest-specific improvement. DICOM3D as a broader project predates the contest.

Start with the portable packet instructions. No private patient data, model weights, credentials or restricted IRIS image are included. The approved public H000 research example is explicit. Authorized Community AI Hub Build154 acquisition and separately supplied model weights are prerequisites. Local fixture authentication is not a production clinical authorization system. This is a research prototype, not diagnosis or clinical deployment.

Qualification status: The native explanation/validated lesion-selection and synthetic mismatch paths were demonstrated on the earlier frozen packet. This male-only source derivative has a different packaging identity. Intermittent CPU inference failures remain unresolved, and a five-run repeatability gate was not completed. Read the tested scope and known limitations before installing or demonstrating.

The male-only derivative has subsequently completed a fresh isolated Build 154 and FHIR R5 startup, production start/stop, nine-resource readback and authorization-denial checks using the historical pinned model. No new inference was run in that startup check. The runtime guide documents the exact model and image identities; this historical fixture is not a claim of general provider support.

The imaging agent uses its restricted native toolsets in-process. A separate historical Build 154 test qualified authenticated MCP with synthetic tools, but this repository does not yet provide a DICOM3D-specific authenticated MCP endpoint.

Local Safe Demo

The Safe Demo source builds a standalone, browser-only anatomy and imaging presentation. Its AI outcomes are simulated; it does not connect to the native agent, IRIS or a model. It is not currently deployed from this repository. The existing viewer requires WebGPU; a WebGL fallback is not qualified. The demo’s expanded male reference and the native frontend’s reference are separately packaged.

Software: Apache-2.0. Existing notice and asset-specific attribution remain separate.