AI Integration for Imaging Anomaly Review AI

The anomaly review workflow, often called the 'second read,' is designed to augment—not replace—the radiologist. AI fits into the PACS hanging protocol or worklist viewer as a discrete panel or overlay. When a study is opened, a pre-fetched AI result—delivered as a DICOM Structured Report (SR) or via a REST API—presents a list of detected anomalies (e.g., pulmonary nodules, intracranial hemorrhages, fractures) with bounding boxes, confidence scores, and relevant series/instance UIDs. This allows the radiologist to review the AI's findings as a separate, auditable source of information while conducting their primary interpretation of the native images.

Implementation requires a secure, low-latency pipeline. After a study is archived to the Vendor Neutral Archive (VNA) or PACS, a DICOM listener or HL7 ORU trigger sends it to an inference service. The AI model processes the images, and results are written back to the PACS as a DICOM SR object or stored in a secondary database. The PACS workstation or a zero-footprint viewer extension then queries for these results upon study load, rendering them as clickable annotations. This architecture keeps the primary image data pristine, maintains a clear audit trail of AI involvement, and allows for easy toggling of the AI overlay on/off.

Rollout and governance are critical. Start with a silent mode, where AI results are generated but not displayed, to establish baseline performance and build radiologist familiarity. Then, move to an assistive mode where findings are visible but require explicit radiologist verification and incorporation into the final report. This phased approach, coupled with RBAC controls to enable/disable the feature by user or workstation, manages change and gathers feedback. The workflow creates a formal feedback loop where radiologist confirmations or corrections are logged to continuously refine the AI models, turning a one-way integration into a collaborative system.

The core integration pattern connects an AI inference service to the PACS worklist and viewer via DICOMweb and HL7 FHIR APIs. When a study is flagged for AI review (e.g., all chest CTs), a DICOM Secondary Capture or Structured Report (SR) object containing the AI findings—annotated bounding boxes, confidence scores, and textual descriptions—is generated and stored in the Vendor Neutral Archive (VNA). This AI SR is linked to the original study via its Study Instance UID, creating a persistent, auditable record separate from the radiologist's final report.

Within the radiologist's workstation, the integration surfaces the AI findings as a toggle-able overlay layer and a side-panel finding list. Key implementation details include:

• Hanging Protocol Triggers: Automatically display the AI overlay when the relevant series is opened, without disrupting existing protocols.
• Interactive Review: The radiologist can accept, reject, or modify each AI finding. Each action is logged with user ID and timestamp.
• Confidence Thresholds: Configurable settings determine which findings are presented prominently (e.g., high confidence pulmonary nodules) versus those placed in a secondary review tab.
• Audit Trail: All interactions—AI result generation, radiologist verification, and final report inclusion—are captured in an immutable log for compliance (FDA 510(k), CE Mark) and model performance monitoring.

Rollout is typically phased, starting with a silent mode where AI runs in the background and results are logged but not displayed, establishing baseline performance. This is followed by a concurrent read pilot with a subset of radiologists, where the AI acts as a second reader. Governance is critical: a clear acceptance protocol must define how discrepancies between AI and the radiologist are escalated (e.g., to a subspecialist), and regular ground truth reconciliation sessions are used to refine AI thresholds and retrain models, closing the feedback loop for continuous improvement.

The anomaly review workflow is architected as a parallel, non-destructive layer. AI detection results—such as bounding boxes, confidence scores, and differential diagnoses—are delivered as a separate DICOM Structured Report (SR) or via a dedicated findings API. These results are presented as an overlay or sidebar in the PACS viewer (e.g., Sectra, Intelerad, Philips IntelliSpace), clearly distinguished from the primary study. This design ensures the original images and the radiologist's primary interpretation remain the single source of truth, with AI acting strictly as an assistive 'second read' that must be explicitly verified, modified, or dismissed.

Governance is enforced through role-based access controls (RBAC) within the PACS and integrated audit trails. For example, a department head can configure which AI models are active for which modalities or body parts, and mandate that junior radiologists view AI suggestions while allowing seniors to toggle them off. Every AI interaction—viewing, accepting, modifying, or rejecting a finding—is logged with user ID, timestamp, and action, creating a complete chain of custody for compliance and model performance monitoring. This audit trail is critical for regulatory submissions and internal quality assurance programs.

A phased rollout is essential for adoption and risk management. Start with a silent pilot, where AI runs in the background on a subset of studies (e.g., chest X-rays for pneumothorax) and its results are logged but not displayed to radiologists, establishing a baseline performance benchmark. Move to a guided pilot where AI findings are visible to a small, trusted group of radiologists in a non-critical setting, collecting feedback on UI/UX and workflow fit. Finally, proceed to controlled production, enabling the AI for specific high-value use cases across the department, accompanied by ongoing training and a clear protocol for handling AI discrepancies, documented in your PACS's integrated clinical decision support framework.