AI Integration for GE CardioPACS

AI integration for GE CardioPACS focuses on three primary surfaces: the Echo Report Module, the Cath Lab Hemodynamics Module, and the Advanced Visualization Workspace. The goal is to inject AI-derived measurements and findings—like automated LVEF, chamber volumes, valve gradients, or plaque characterization from cardiac CT—directly into the structured reporting templates and hanging protocols cardiologists already use. This is achieved by connecting inference services to the PACS via DICOM Secondary Capture and Structured Report (SR) objects, or through GE's Edison AI Platform APIs for validated, containerized algorithms. The AI acts as a silent first-pass analyst, populating quantitative fields and flagging studies that deviate from normal ranges for prioritized review.

A production implementation typically involves a secure, on-premises or hybrid cloud pipeline. DICOM studies are ingested from the CardioPACS Universal Data Manager or a listening service on the Centricity PACS broker. AI models (e.g., for echocardiogram quantification or coronary CTA analysis) run inference, generating results as DICOM SR. These SR objects are then routed back into the patient's study list within CardioPACS. For the user, this manifests as pre-populated measurements in the report, an AI findings panel in the viewer, or a color-coded priority flag on the worklist. Governance is critical; all AI-generated data must be stored as non-destructive overlays with clear provenance in the audit trail, and a human-in-the-loop verification step should be enforced before final report sign-off.

Rollout should be phased, starting with a single, high-value workflow like automated LVEF and volume calculation for echocardiograms. Begin with a pilot group of cardiologists, using the AI as an 'assistant' whose results they can accept, modify, or reject. This feedback is essential for tuning confidence thresholds and UI integration. Subsequent phases can expand to stress echo analysis, TAVR planning measurements from cardiac CT, or ischemia detection from nuclear studies. The key to adoption is minimizing disruption—the AI should feel like a natural extension of the existing CardioPACS tools, not a separate application. For broader context on integrating AI across multi-specialty imaging, see our guide on Enterprise Imaging AI, and for details on the underlying cloud or on-prem infrastructure, review AI Integration for GE Edison AI Platform.

Integration begins at the worklist and study ingestion layer. As DICOM studies (Echo, Cardiac CT, MR, Angio) arrive in CardioPACS, a lightweight service—deployed as a container or Windows service—monitors the DICOM node or listens for HL7 ADT/ORM messages. This service extracts key metadata (Accession Number, MRN, Study Description) and triggers an AI inference pipeline. For real-time analysis, the service can push a de-identified DICOM stream or a secure link to a GPU-accelerated inference service hosted on-premises or in a compliant cloud (e.g., AWS HealthLake Imaging, Azure Health Data Services). Common initial AI applications include automated left ventricular ejection fraction (LVEF) quantification, aortic valve calcium scoring, and plaque characterization from CCTA.

Results are returned as DICOM Structured Reports (SR) or JSON payloads containing measurements, contours, and confidence scores. These are injected back into CardioPACS via DICOM Store or a dedicated API (e.g., using GE's Centricity PACS APIs or the Edison AI Platform gateway). The AI-generated data is linked to the original study, making it available within the cardiologist's native review environment. Key integration surfaces include:

• The Advanced Visualization Workspace: AI contours and measurements overlay directly on 2D/3D images for interactive review.
• Structured Reporting Module: Quantitative results (e.g., chamber volumes, wall motion scores) auto-populate report templates, reducing manual data entry.
• Worklist Prioritization: Studies with AI-flagged critical findings (e.g., low EF, high-risk plaque) can be elevated in the reading queue via a rules engine.

For production rollout, a phased human-in-the-loop approach is recommended. Start with AI as a concurrent read or second reader, where results are visible but not auto-finalized, allowing cardiologists to verify, adjust, and provide feedback. This feedback loop is crucial for validation and model refinement. Governance requires configuring audit logs for all AI interactions, defining RBAC so only credentialed users can see AI outputs, and establishing a quality assurance (QA) workflow to periodically audit AI performance against ground truth. The architecture should support model versioning and A/B testing to safely deploy updates without disrupting clinical workflow. For health systems using multiple PACS, consider a central AI orchestration layer that can route studies to the appropriate AI models and deliver results to both CardioPACS and the enterprise EHR, ensuring a unified patient record.

Deployment begins with a secure, air-gapped testing environment mirroring your production CardioPACS instance. Here, AI models for chamber quantification or plaque analysis are validated against a curated dataset of historical studies to establish baseline accuracy and identify edge cases. Integration is scoped to specific, high-value workflows—initially targeting structured reporting for echocardiograms or automated LVEF calculation—using GE's APIs (like those for the Advanced Visualization module or Reporting tools) to pass anonymized DICOM data to the inference service and return structured results (DICOM SR) for review.

A human-in-the-loop governance model is critical. In the initial pilot phase, AI-generated measurements and findings are presented to the cardiologist as non-binding suggestions within the CardioPACS viewer or reporting interface, requiring explicit approval before being committed to the patient record. All AI interactions are logged to a separate audit trail, capturing the original image data, the AI inference payload, the clinician's action (accept, modify, reject), and a user identifier. This creates a feedback loop for continuous model refinement and meets compliance requirements for traceability.

Rollout follows a phased, risk-based approach: 1) Silent Mode: AI runs in the background on all eligible studies, generating results that are logged but not displayed, to monitor performance and drift. 2) Assistive Pilot: AI suggestions are enabled for a single reading room or a subset of credentialed cardiologists, with structured feedback collection. 3) Controlled Expansion: Based on validated performance and user comfort, the integration is expanded to additional workflows (e.g., aortic valve planimetry) and user groups. Each phase includes specific key performance indicators (KPIs), such as time-to-report, inter-observer variability reduction, and user acceptance rates, measured against the pre-AI baseline.

Security is architected around the principle of zero patient data persistence in the AI service. The inference pipeline is designed as a stateless service; DICOM images are streamed via secure, encrypted channels (often within the hospital network), processed, and immediately purged from the AI system's memory after the result is returned. Access to the AI service is controlled via the same Active Directory or LDAP groups used for CardioPACS, ensuring only authorized personnel can trigger or view AI analyses. This layered approach—combining phased validation, auditable human oversight, and strict data governance—enables the safe, scalable adoption of AI within the critical cardiology diagnostic pathway.