AI Integration for GE Imaging Systems

AI integration for GE Imaging Systems connects at three primary layers: the data ingestion and orchestration layer (GE HealthCloud, Universal Data Manager), the clinical workflow layer (Centricity PACS, CardioPACS, AW Server), and the AI execution and validation layer (Edison AI Platform). The most common pattern is to use the Edison AI Platform as the central hub for validated AI applications, which then push results—as DICOM Structured Reports (SR) or secondary capture images—back into the PACS worklist or directly into the radiologist's hanging protocol on the Advanced Visualization (AW) workstation. For custom or third-party algorithms, integration typically occurs via DICOMweb services for study retrieval and result submission, and HL7 FHIR APIs for patient context and reporting workflows.

High-value use cases are tied to specific GE modules and workflows:

• Centricity PACS RIS: AI-driven study triage can re-prioritize the reading worklist based on critical finding detection (e.g., ICH, pneumothorax), routing urgent cases to the top.
• GE AW Server: AI-powered segmentation tools (e.g., for lung nodules, liver volumes) can be embedded as one-click plugins, generating 3D models and measurements within the radiologist's native post-processing environment.
• CardioPACS: AI algorithms for automated chamber quantification or plaque analysis can populate structured report templates, reducing manual measurement time for cardiologists.
• Edison AI Platform: Serves as the governed deployment environment, handling model validation, monitoring inference performance, and managing the secure data pipeline between PACS and the AI application, whether it's hosted on-premise or in GE HealthCloud.

A production rollout requires a phased approach, starting with a non-diagnostic "concurrent read" workflow where AI results are presented as a separate finding list for radiologist verification. This builds trust and creates audit trails. Governance is critical; the integration must support RBAC to control which users see AI prompts, maintain detailed audit logs of AI inferences linked to the original study, and enable seamless human-in-the-loop feedback to retrain models. The end goal is a seamless workflow where AI acts as an embedded assistant, prioritizing work, reducing repetitive tasks, and surfacing quantitative insights without disrupting the radiologist's primary diagnostic flow.

Integration with GE's ecosystem typically follows a hub-and-spoke model centered on the GE Edison AI Platform and Centricity PACS. The primary data flow begins when a completed DICOM study is sent from a GE modality (like a Revolution CT or SIGNA MRI) or a VNA to a designated receiving node. Using DICOM C-STORE SCP services or monitoring a DICOM directory, an integration service can capture the study and route it—along with relevant HL7 ADT data for patient context—to containerized AI inference engines. Results are returned as DICOM Structured Reports (SR) or Secondary Capture objects, which are then injected back into the PACS. For workflow orchestration, the Centricity PACS API or GE HealthCloud APIs are used to update the radiologist's worklist, applying priority flags or prefetching prior studies based on AI findings.

For advanced visualization and quantitative tools, integration extends to GE AW (Advantage Workstation) Server or the CardioPACS environment. Here, AI-generated segmentations or measurements (e.g., tumor volumetry, ejection fraction) can be pushed as DICOM Segmentation or Parametric Map objects, making them available within the radiologist's or cardiologist's native 3D review session. Governance is enforced through RBAC at the API layer and by maintaining a full audit trail of AI inferences linked to the original study UID. This architecture supports use cases like automated pneumothorax detection on chest X-rays routed to the ED worklist or brain bleed triage on head CTs that trigger an alert in the reading room.

Rollout requires a phased approach, starting with a single modality or body part in a QA environment. A critical step is validating the AI output's alignment with GE's hanging protocols and display formats to avoid disrupting radiologist efficiency. Successful production deployment hinges on integrating with existing dose monitoring and quality assurance workflows in systems like GE DoseWatch, ensuring AI operations are part of the department's continuous improvement cycle. For health systems using GE HealthCloud, the pattern shifts to leveraging cloud-native services for scalable, managed AI inference, with data residency and egress costs being key architectural considerations.

A production integration with GE Imaging Systems and the Edison AI Platform is built on a secure, auditable pipeline. DICOM studies are typically routed from Centricity PACS or HealthCloud to a dedicated inference queue via DICOMweb or a modality worklist trigger. AI models, whether third-party algorithms from the Edison Marketplace or custom models, run in isolated, HIPAA-compliant containers. Results are returned as DICOM Structured Reports (SR) or Secondary Capture objects, stamped with the model's unique device identifier (UDI) for full traceability. All data exchanges are encrypted in transit, and access is controlled via integration with the hospital's existing identity provider (e.g., Active Directory) for role-based access control (RBAC).

We advocate for a phased, use-case-led rollout to build clinical trust and operational proof. Phase 1 often starts with a non-diagnostic, operational AI use case like automated protocoling or image quality control (e.g., slice positioning check for brain MRI). This de-risks the integration and validates the data pipeline. Phase 2 introduces a diagnostic-assist tool, such as a pneumothorax detection algorithm for chest X-rays, deployed in a "silent mode" where it runs in the background and its findings are logged but not displayed, allowing for retrospective validation against radiologist reports. Phase 3 activates clinical integration, displaying AI findings as a prioritized worklist in Centricity PACS or as an overlay in Advanced Workstation (AW), but requiring explicit radiologist confirmation before inclusion in the final report.

Governance is maintained through a continuous feedback loop. All AI inferences, along with the radiologist's final interpretation, are logged to a secure audit database. This creates a ground-truth dataset for periodic model performance review, enabling the detection of concept drift or performance degradation on your specific patient population. A clear escalation and downtime procedure is documented, defining how to bypass AI routing if the service is unavailable. This structured, incremental approach ensures AI augments the radiologist's workflow safely, maintaining the standard of care while introducing efficiency gains where they are most defensible.