AI Integration for GE Women's Health Imaging

AI integration for GE Women's Health Imaging connects to specific functional surfaces within the clinical workflow. The primary integration targets are GE Senographe Pristina/Senobright mammography systems, LOGIQ and Voluson ultrasound platforms, and SIGNA Architect/MR breast coils, feeding data into the Centricity PACS or HealthCloud Imaging ecosystem. AI models typically interface via the Edison AI Platform or direct DICOM Secondary Capture/Structured Report (SR) feeds to deliver findings into the reading worklist. Key data objects include DICOM images, patient demographics, prior exam comparisons, and BI-RADS assessments, which AI uses to generate supplemental findings for lesion detection, density assessment, and risk scoring.

Implementation follows a staged, human-in-the-loop pattern. For screening mammography, AI runs as a pre-read service, analyzing incoming DICOM studies and attaching an SR with findings (e.g., suspicious calcifications, asymmetries, masses) and a confidence score. This SR is ingested by Centricity PACS, where it populates a dedicated AI findings panel within the Universal Viewer Zero Footprint Client. The radiologist's workflow remains central; AI highlights regions of interest on the mammogram hanging protocol, but all decisions and final reporting stay with the clinician. For breast MRI and ultrasound, AI can be triggered post-acquisition to provide automated lesion segmentation and kinetic curve analysis, with results overlaid on the advanced visualization tools for quicker quantification.

Rollout requires careful governance, starting with a pilot in a single reading room. Integration points must be validated for HL7 message flow (ORM/ORU) to ensure AI results are correctly associated with the right patient and study in the RIS. A QA dashboard for monitoring AI performance (e.g., false-positive rates, case review metrics) should be integrated into the clinical admin console. Given the regulated nature of breast imaging, all AI outputs must be audit-logged and non-destructive, preserving the original images and report. Successful adoption hinges on configuring the PACS to present AI insights without disrupting the radiologist's efficiency, treating AI as a consistent, silent second reader that prioritizes workflow rather than attempting to replace clinical judgment.

The integration architecture connects to three primary data surfaces within the GE ecosystem: the Centricity PACS/Universal Viewer for study access and display, the Edison AI Platform for model orchestration and validation, and the Women's Health Advanced Visualization tools (e.g., for breast MRI and tomosynthesis). The core data flow is triggered by a new DICOM study arrival in the PACS. Using DICOMweb or a modality worklist hook, studies matching specific modalities (FFDM, DBT, Breast MRI, Ultrasound) are routed to a secure, HIPAA-compliant inference queue. AI models—whether third-party algorithms for lesion detection or custom models for breast density classification—are containerized and deployed either on-premises adjacent to the PACS or within a private cloud, interfacing with the Edison AI Platform for lifecycle management and result validation.

Processed AI results are returned as DICOM Structured Reports (SR) or as discrete data objects via API, containing findings, confidence scores, and, for biopsy planning, 3D coordinates. These are injected back into the PACS study as a secondary capture or linked series. For the radiologist, integration occurs at the workstation level: AI findings can be displayed as hanging protocol overlays on mammograms or as a side-panel findings list in the Universal Viewer, with clickable links to navigate to suspicious slices. For biopsy planning, key measurements and coordinates can be pushed directly into the GE biopsy device interface or the Invenia Automated Breast Ultrasound (ABUS) review station, streamlining the path from detection to intervention. This closed-loop flow ensures AI is a contextual aid, not a disruptive silo.

Governance and rollout require a phased approach. Start with a non-interruptive pilot, where AI results are stored but not displayed, to establish baseline performance and radiologist trust. Use the Edison AI Platform's tools for model drift monitoring and ground truth reconciliation. For production, implement role-based access control (RBAC) to determine which radiologists see AI prompts and configure escalation rules in the workflow manager for high-probability findings. A successful implementation reduces manual search patterns, standardizes BI-RADS assessments, and shortens the time from screening to diagnostic procedure, directly impacting patient throughput and clinical consistency.

Integrating AI into GE's mammography, breast ultrasound, and breast MRI workflows requires a security-first architecture that respects the clinical environment. This typically involves deploying AI inference containers within the hospital's secure network, connecting to the GE Centricity PACS or HealthCloud Imaging via DICOMweb and HL7 interfaces. Patient data never leaves the protected health information (PHI) boundary; AI models are brought to the data. All AI-generated findings, such as lesion probability scores or BI-RADS density assessments, are written back as DICOM Structured Reports (SR) or HL7 observations, creating a permanent, auditable trail linked to the original study within the GE archive.

A phased rollout is critical for clinical adoption and risk management. Phase 1 often begins in a silent mode, where AI runs in the background on all incoming studies but results are only visible on a separate dashboard for radiologist education and algorithm validation. Phase 2 introduces concurrent read support, where AI findings (e.g., lesion markers, risk scores) are presented as a non-intrusive overlay in the GE Advantage Workstation or Universal Viewer, requiring explicit user action to accept or dismiss. The final phase enables worklist prioritization, where the PACS worklist is dynamically sorted based on AI-identified critical findings (e.g., high-probability masses, architectural distortions), ensuring the most urgent cases are read first.

Governance is established through a multi-disciplinary team including breast radiologists, IT, compliance, and clinical engineering. This team defines the acceptable use policy for AI outputs, mandates regular algorithmic drift monitoring against ground-truth pathology reports, and establishes a clear human-in-the-loop protocol where AI serves as an assistive tool, not an autonomous decision-maker. All AI interactions are logged, enabling retrospective review of model performance and user engagement, which is essential for quality assurance and meeting FDA SaMD (Software as a Medical Device) or CE mark compliance requirements for clinical AI tools.