AI Integration for Radiology Study Triage and Prioritization

The primary integration point is the PACS worklist server, typically accessed via HL7 ADT/ORM messages for patient/order context and DICOM Query/Retrieve for study access. When a new study arrives in the PACS, an event-driven trigger (often via a DICOM C-STORE SCP or a monitoring service) pushes the study's metadata to an AI orchestration layer. This layer dispatches the relevant DICOM series to pre-configured AI models—for critical detection of intracranial hemorrhage, pulmonary embolism, or pneumothorax—running in a secure, GPU-accelerated inference environment. The results, returned as DICOM Structured Reports (SR) or a lightweight JSON payload, are then ingested back into the PACS to flag and re-order the worklist.

Implementation centers on configuring the worklist filter rules within the PACS (e.g., Sectra Workflow Manager, Intelerad PowerReader, Philips IntelliSpace Radiology) to surface AI-prioritized cases. A high-impact pattern is creating a dedicated "AI Critical" worklist bucket that overrides standard FIFO sorting. For each prioritized case, the AI findings can be presented as an overlay on the viewer or a sidebar panel, providing the radiologist with context before they open the study. This integration must be non-blocking; the study remains readable even if the AI service is delayed, ensuring clinical workflow continuity.

Rollout requires a phased, protocol-specific approach. Start with a single, high-volume, high-acuity protocol like Non-Contrast Head CTs for stroke in the ER. This limits variables, allows for clear outcome measurement (e.g., door-to-read time for positive cases), and builds clinical trust. Governance is critical: establish a multi-disciplinary AI committee (radiology, IT, compliance) to validate model performance on local data, define alert thresholds, and create a clear human-in-the-loop protocol where AI suggestions are always verified. Audit logs must track the AI inference, the worklist change, and the final radiologist interpretation to monitor for drift and measure impact.

The business impact is operational, not just clinical. By routing the 5-10% of studies with critical findings to the top of the queue, health systems can reduce critical result notification time from hours to minutes. This directly supports stroke and trauma center certifications, improves ER throughput, and reduces radiologist cognitive load during on-call shifts by providing a structured, prioritized safety net. For a deeper technical dive on connecting to specific platforms, see our guides for Sectra PACS and Philips IntelliSpace.

The integration architecture hinges on three core data flows: 1) Study Ingestion via DICOM C-STORE SCP listeners or HL7 ORM/O01 messages from your RIS, 2) AI Inference triggered by a rules engine that filters studies (e.g., all non-screener chest X-rays) and sends anonymized DICOM slices to containerized AI models via a secure API gateway, and 3) Result Integration where AI outputs—structured as DICOM SR (Structured Report) objects or simple JSON payloads—are posted back to the PACS. For platforms like Sectra or Philips IntelliSpace, this often uses their native AI Orchestrator or Universal Data Manager APIs to attach findings and set a priority score (e.g., Critical, Routine) on the study record itself.

Key implementation details include setting up a dedicated service queue (like RabbitMQ or AWS SQS) to handle peak modality output, ensuring studies are not lost during AI service latency. The AI service layer must be stateless, pulling studies from the queue, executing models (e.g., for pneumothorax, ICH, or mass detection), and writing results to a shared metadata store (like Redis) for low-latency worklist queries. For the radiologist's UI, the worklist is filtered or sorted via a custom column or flag injected through the PACS' web client API—Sectra's Workflow Manager or Intelerad's PowerReader APIs allow this—so a "Suspected Critical" tag appears next to the study, enabling one-click prioritization without leaving the native environment.

Rollout requires a phased silent mode where AI runs in the background for weeks, logging confidence scores against final reports to establish baseline accuracy and calibrate thresholds before any clinical alerts are shown. Governance is enforced through an audit log that tracks every AI inference, the user who overrode or accepted the suggestion, and the original image data for model retraining. This traceability is critical for compliance and continuous validation, especially when integrating multiple AI vendors from the GE Edison AI Platform or cloud-based algorithms. The final architecture should be redundant, with fail-open logic so that if the AI service is down, the worklist defaults to standard FIFO sorting, ensuring no disruption to clinical operations.

Production AI triage must operate within the existing clinical governance framework of your PACS (Sectra, Philips, Intelerad, GE). This means integrating at the worklist or routing engine level to flag studies for prioritization without altering original images or reports. AI-generated findings are typically delivered as DICOM Structured Reports (SR) or via a sidecar API, creating a clear audit trail. Critical design decisions include whether the AI runs on-premises, in a HIPAA-compliant cloud, or via the PACS vendor's own AI orchestrator, each with implications for data sovereignty, latency, and scalability.

A phased rollout is essential for managing risk and building trust. A common pattern is:

• Phase 1: Silent Mode & Validation. AI processes all incoming studies in the background, logging its triage scores and detection flags without impacting the radiologist's worklist. This generates a performance baseline and validates the AI against your local case mix.
• Phase 2: Assisted Triage. AI prioritization is introduced to a non-critical worklist (e.g., outpatient musculoskeletal MRI). The radiologist sees an AI-generated priority flag (e.g., "Potential Critical Finding") but the standard FIFO list remains available. This phase gathers feedback on UI integration and workflow impact.
• Phase 3: Integrated Prioritization. After validation and tuning, AI-driven re-ordering is activated for targeted high-acuity workflows like the ED or stroke protocol CTs. The system can be configured with rules-based overrides (e.g., manual re-prioritization, case complexity scoring) and integrated alerting via the PACS or middleware to critical care teams.

Governance is maintained through continuous monitoring. Key operational controls include:

• Human-in-the-loop verification: All AI-prioritized studies and flagged findings require final radiologist interpretation and sign-off.
• Performance dashboards: Track metrics like time-to-read for prioritized cases, AI flag false-positive/negative rates, and radiologist acceptance rates to monitor drift and clinical utility.
• Change management: Clear protocols for updating AI models, including re-validation against a hold-out dataset and communicating performance changes to the reading staff. This structured approach, integrated with your existing PACS audit logs and quality assurance programs, ensures AI augments the radiologist safely and effectively, turning data into faster, more consistent patient care.