AI Integration for Clinical Decision Support in Imaging

Effective AI-powered CDS integrates at three key points in the imaging workflow: study triage, active interpretation, and report finalization. During triage, AI can pre-analyze incoming studies against appropriateness criteria (e.g., ACR guidelines) and flag potentially mismatched orders or missing prior comparisons for the radiologist. Within the active reading session on platforms like Sectra Workstation or Philips IntelliSpace, a context-aware AI agent can surface relevant differential diagnoses, pertinent literature, or institutional protocols based on the hanging protocol, patient history from the EHR (via FHIR), and preliminary AI detection findings. This moves CDS from a separate lookup tool to an integrated, workflow-native copilot.

The technical integration hinges on a secure middleware layer that orchestrates data flow. When a study is opened, the PACS viewer sends a secure payload (study UID, modality, body region) via a REST API or DICOM Service Class User (SCU) to a governed AI inference service. This service retrieves context from the VNA and EHR, runs inference through specialized models (e.g., for guideline checking or literature retrieval), and returns structured results as DICOM Structured Reports (SR) or a JSON payload. These results are rendered as a non-obtrusive sidebar or overlay within the PACS, allowing the radiologist to accept, modify, or ignore suggestions without breaking their flow. Critical to adoption is designing these prompts to be specific—suggesting "consider apical pleural thickening vs. early fibrosis in this asbestos-exposed patient" rather than a generic list.

Rollout requires a phased, governance-first approach. Start with non-diagnostic CDS use cases like automated Fleischner Society guideline checks for pulmonary nodule follow-up or ACR incidental findings management recommendations. Implement a closed-loop feedback system where radiologist interactions with AI suggestions are logged to an audit trail for model refinement and compliance. This builds trust and provides clear ROI by reducing missed follow-ups and standardizing care. Ultimately, this architecture turns the PACS from a passive viewer into an intelligent clinical partner, embedding evidence-based guidance where decisions are made.

The integration is triggered when a study is COMPLETED in the PACS worklist or arrives in the designated CDS routing queue (e.g., a DICOM Modality Worklist or an HL7 ORM/ORU message). The system extracts key context: the study's Accession Number, Modality, Body Part, and the Clinical Indication from the order. This metadata is packaged with a secure, de-identified DICOMweb link to the images and sent to the AI inference service. The AI service, which may host multiple specialized models (e.g., for chest CT pulmonary embolism, brain MRI MS lesions, or mammography BI-RADS scoring), runs the appropriate algorithms based on the study context.

Results are returned as a structured DICOM Structured Report (SR) or a FHIR Observation bundle. This includes AI-generated findings with confidence scores, references to relevant clinical guidelines (e.g., ACR Appropriateness Criteria, Fleischner Society nodules), and suggested differential diagnoses. These results are injected back into the PACS in two key ways: 1) As an SR object linked to the original study, viewable in a side-panel or overlay within the radiologist's workstation (Sectra, IntelliSpace Portal, etc.). 2) As discrete data points pushed via HL7 to the Reporting Module to pre-populate draft report sections or macro suggestions, reducing manual entry and ensuring guideline adherence.

Governance is critical. All AI interactions are logged with a full audit trail—study ID, user, model version, inference timestamp, and result. A human-in-the-loop review step is enforced; AI suggestions are clearly flagged as assistive and require radiologist verification before finalization. The system supports feedback loops where radiologist overrides or corrections are sent (anonymized) back to the AI ops platform for model retraining. Rollout follows a phased pilot: start with a single high-volume, well-defined workflow (e.g., PE detection in CTPA) in a non-interruptive "silent mode" to validate accuracy and workflow fit before enabling active, in-worklist CDS alerts.

Governance begins with model validation and clinical integration testing. Before any CDS suggestion reaches a radiologist, AI outputs must be validated against your institution's historical cases and clinical guidelines. This involves creating a sandbox environment within your PACS (e.g., a test partition in Sectra, Philips IntelliSpace, or Intelerad) to run inference on prior studies, comparing AI-generated appropriateness criteria or differential diagnoses with the original radiology reports and pathology follow-ups. Key technical steps include establishing audit logs for all AI inferences, mapping AI outputs to standardized ontologies like RadLex, and defining confidence thresholds for when a suggestion is surfaced versus suppressed.

Safety is engineered through human-in-the-loop workflows and clear attribution. AI suggestions for differential diagnosis or guideline adherence should be presented as non-interruptive, context-aware annotations within the radiologist's native reading pane—never as autonomous actions. For example, an AI module integrated via the PACS API might display a discrete panel with "Considerations: Based on ACR Appropriateness Criteria® and lesion characteristics..." All suggestions must be traceable: the final report should log if an AI-derived insight was viewed, ignored, or incorporated, creating a defensible record. This requires tight integration with the reporting module to capture this metadata within the report object itself.

A phased rollout mitigates risk and builds trust. Start with a non-diagnostic pilot, such as using AI to prepopulate the 'Clinical History' field from the EHR or to suggest relevant prior comparisons, which has immediate workflow benefit without direct diagnostic liability. Phase two introduces CDS for low-acuity, high-volume cases (e.g., routine follow-up chest X-rays for COPD) where AI can suggest follow-up intervals per Fleischner Society guidelines. The final phase expands to complex decision support, like differential diagnosis for indeterminate lung nodules on CT, involving a concurrent read protocol where the AI output is reviewed by a specialist before the primary read. Each phase requires continuous monitoring of adoption metrics, suggestion acceptance rates, and time-to-report, with feedback loops to refine prompts and integration points.

Why Inference Systems for this integration? We architect these systems with clinical safety as the first constraint. Our implementations use the PACS vendor's official APIs and IHE integration profiles (like IHE AIW) to ensure stability and vendor supportability. We build the necessary orchestration layer to manage multiple AI models, handle fallback scenarios, and enforce access controls (RBAC) so CDS tools are only available to credentialed users. This results in an AI integration that radiologists trust because it respects their workflow, clinicians value because it provides actionable support, and compliance officers approve because it's built for auditability and control. Explore our foundational guide for AI Integration for Radiology PACS Systems to understand the core patterns that enable this governed approach.