AI Agentic Workflow for Radiology Image Pre-screening and Triage

This workflow automates the initial, repetitive review of incoming imaging studies (CT, MRI, X-ray) by orchestrating specialized AI inference engines. It directly addresses the operational bottleneck of unprioritized worklists, where critical findings risk delay. The business value comes from increased radiologist throughput, reduced time-to-intervention for urgent cases, and better utilization of specialist labor. Implementation requires deep integration with PACS and RIS, using an orchestrator like LangGraph to manage data flow, model execution, and exception handling.

The architecture is built for production, requiring controls for data quality, model confidence thresholds, and mandatory human review of all AI findings. The orchestrator manages API calls to cloud or on-prem inference services, ingests DICOM metadata, and routes low-confidence studies for immediate human read. A triage dashboard integrated into the existing RIS presents prioritized cases with AI annotations and reasoning, preserving the radiologist's final diagnostic authority. Rollout must be sequenced by modality and body part, with continuous monitoring of triage accuracy and system latency.

This workflow automates the initial processing of all incoming DICOM studies, a repetitive bottleneck that delays critical diagnoses. By deploying specialized AI agents for concurrent abnormality detection—targeting bleeds, fractures, and masses—the system generates preliminary findings and a urgency score before a radiologist opens the case. The operational upside comes from labor leverage: radiologists start their review with AI-generated annotations and a worklist sorted by clinical priority, reclaiming 15-30% of time otherwise spent on manual triage and basic measurements.

Implementation integrates directly with the hospital's PACS and RIS (e.g., Epic, Cerner) via DICOMweb and FHIR APIs. The orchestrator, built on LangGraph, manages agent execution, handles failures, and enforces audit trails. Critical controls include a human-in-the-loop review queue for low-confidence findings and ambiguous cases, with all AI annotations and priority scores visible in the radiologist's reading workstation. Deployment requires a phased rollout, starting with non-critical body parts, and continuous monitoring of model drift and triage accuracy against ground-truth outcomes.

This workflow automates the initial processing of incoming DICOM studies, running specialized AI models (e.g., for intracranial hemorrhage, fractures, masses) to detect and flag abnormalities. It directly addresses the radiologist bottleneck by pre-populating findings and applying urgency-based triage logic to the worklist. The operational upside comes from reducing time-to-diagnosis for critical cases by up to 70%, increasing department throughput, and allowing radiologists to focus cognitive effort on complex interpretations rather than initial screening.

Implementation follows a phased, safety-first approach. Phase 1 deploys a single AI model in a silent, parallel mode to validate performance against historical ground truth without affecting clinical workflow. Phase 2 introduces the triage dashboard into the PACS viewer, routing only high-confidence, non-critical findings for radiologist validation. Full orchestration and automated critical result escalation are activated in Phase 3, only after rigorous clinical validation and governance approval. Each phase includes immutable audit logs, mandatory human-in-the-loop review for low-confidence cases, and continuous monitoring for model drift against the PACS archive.

This workflow automates the initial processing of all incoming DICOM studies, running them through a configurable pipeline of specialized AI inference engines (e.g., for intracranial hemorrhage, pulmonary nodules, fractures). It directly addresses the radiologist's operational bottleneck of an unsorted, monolithic worklist. The business value comes from labor leverage—reducing the manual pre-read sorting effort—and from improved clinical outcomes by ensuring the most urgent and complex cases are surfaced first, compressing the diagnostic loop for critical findings like strokes or masses.

Implementation requires deep integration with the PACS (e.g., via DICOMweb) and RIS/EHR (via HL7/FHIR) to ingest studies and push prioritized worklists. The orchestrator, built with frameworks like LangGraph, manages state, routes studies to the appropriate AI models—which may be cloud-hosted or on-premise—and applies business logic for scoring urgency and complexity. Critical controls include confidence thresholding to route low-confidence predictions for human review, audit logging for all AI inferences, and configurable approval gates before any triage decision modifies the production radiologist schedule. Rollout is typically phased by modality (CT head first) and model, with parallel running to validate performance before full automation.

Deploying an AI triage agent into a production PACS/RIS environment requires a governance layer that enforces clinical safety while preserving operational gains. The architecture must embed explicit approval gates before any automated finding alters a radiologist's worklist, with all AI inferences logged to an immutable audit trail for quality assurance and regulatory review. Controls are implemented via a central orchestrator that manages data flow, applies confidence thresholds, and routes exceptions—such as low-confidence detections or discordant multi-model results—to a human-in-the-loop review queue staffed by senior radiologists or QA specialists.

A phased rollout mitigates risk and builds clinician trust. Phase 1 is a silent pilot where the AI runs in parallel, generating triage recommendations visible only on a separate dashboard for validation against historical ground truth. Phase 2 introduces the workflow for non-critical, high-volume studies like routine chest X-rays, with mandatory human sign-off on all priority changes. Phase 3 expands to more complex modalities and enables autonomous prioritization for high-confidence cases, maintaining the review queue for exceptions. Each phase is governed by KPIs on accuracy, time-to-diagnosis for urgent cases, and radiologist feedback, ensuring the system demonstrably improves throughput without compromising diagnostic integrity.