AI Integration for GE Edison AI Platform

The GE Edison AI Platform serves as a centralized hub for validating, deploying, and managing AI applications within the GE imaging ecosystem. AI fits into three primary integration surfaces: the Model Validation & Deployment layer, the DICOM Integration Gateway, and the Clinical Workflow Embedding APIs. For developers, this means packaging algorithms into Docker containers that meet Edison's specifications, which are then ingested, validated for clinical safety and performance, and registered within the platform's catalog. The platform's orchestration engine handles the scaling of inference across on-premises or cloud infrastructure, managing GPU resources and ensuring high availability for production workloads.

Once deployed, AI models connect to imaging data via the platform's DICOM Integration Gateway. This gateway listens for incoming studies from modalities, PACS (like Centricity), or VNAs, applying configurable rules to route specific study types (e.g., non-contrast head CTs) to the appropriate AI container for inference. Results are returned as structured DICOM-SR (Structured Reporting) objects or secondary capture images, which are then sent back to the PACS or VNA. This enables AI findings to be embedded directly into the radiologist's hanging protocol on a GE Advanced Visualization workstation or a zero-footprint viewer, appearing as an overlay, measurement, or prioritized finding list alongside the native images.

The final fit is within the clinical user's workflow. Through Edison's APIs, health systems can build custom applications that pull AI results into reporting tools, populate structured report templates, or trigger downstream actions—like alerting a stroke team via HL7 ADT messages. Governance is managed through the platform's dashboard, providing audit trails of every inference, model performance monitoring for drift, and RBAC controls to ensure only authorized clinicians can see AI outputs. A successful rollout typically starts with a single high-impact use case (e.g., ICH detection for stroke triage) in a pilot reading room, using the platform's tools to measure AI adoption and radiologist interaction before scaling to other modalities and sites.

The core integration pattern uses the Edison Developer Services (EDS) API to register, validate, and deploy AI applications as Docker containers within the Edison ecosystem. Each model is packaged with its dependencies and a standardized inference service endpoint. Incoming DICOM studies are routed from connected modalities or PACS (like Centricity) to the appropriate AI container via the Edison DICOM Adapter, which handles secure ingestion, de-identification if required, and study queuing. Results are returned as DICOM Structured Reports (SR) or secondary capture images, which are then injected back into the PACS workflow for radiologist review.

For clinical workflow embedding, the integration leverages Edison's Clinical Workflow Engine to define rules for AI triggering—such as auto-running a brain bleed detection model on all non-contrast head CTs from the ED. Results can be configured to generate HL7 alerts to the EHR or prioritize studies on the radiologist's worklist in GE's Advanced Visualization tools. A key governance layer is the Edison AI Validation Suite, where model performance (inference speed, accuracy drift) is continuously monitored against a ground-truth dataset, ensuring clinical safety before any algorithm is promoted from a sandbox to a production cluster.

Rollout follows a phased, use-case-driven approach. We typically start with a single AI application in a non-critical workflow (e.g., automated chest X-ray triage for follow-up studies) within one department. This allows validation of the data pipeline, user acceptance, and operational support procedures. Successive models are then onboarded using the same containerized pattern, managed through the Edison platform's centralized dashboard for version control and resource allocation. This architecture ensures each AI module is isolated, scalable, and compliant with healthcare data governance standards, providing a clear path from pilot to enterprise-scale AI operations across the imaging network.

Production deployment on Edison begins with a validation environment separate from your live clinical PACS. Here, you run inference on a curated, representative dataset—spanning modalities, protocols, and patient demographics—to establish baseline performance metrics (sensitivity, specificity, PPV) and confirm the algorithm's output aligns with clinical expectations. This phase should include integration testing with your Centricity PACS or HealthCloud DICOM nodes to verify study routing, result ingestion via DICOM Structured Reports (SR), and display within the radiologist's workflow using Edison AI Apps or custom viewer extensions.

Governance is managed through Edison's Model Management APIs, which handle version control, containerized deployment, and runtime monitoring. For each model, you define access controls (RBAC) to restrict which users or sites can trigger AI analysis and establish an audit trail logging every inference, user interaction, and model update. Critical to this is implementing a human-in-the-loop review pattern, where AI findings are presented as non-binding annotations or structured data within the report draft, requiring radiologist verification before finalization. This builds trust and creates a feedback loop for continuous model improvement.

A phased rollout mitigates risk. Start with a silent mode, where AI runs in the background on all incoming studies but results are only logged, not displayed. This validates real-world performance without impacting workflow. Next, enable assistive mode for a pilot group of radiologists on non-critical, routine studies (e.g., outpatient chest X-rays), using AI to prioritize the worklist or generate report drafts. Finally, after monitoring accuracy and user feedback, expand to full clinical integration across modalities and user groups, with ongoing performance monitoring for concept drift and regular re-validation against updated clinical guidelines.