Radiology AI result routing is the process of sending the correct study to the correct algorithm, returning structured outputs to the correct systems, and logging every acceptance, rejection, and override. ACR–SIIM guidance describes routing rules using DICOM, Health Level Seven, and Fast Healthcare Interoperability Resources, while the ACR–SIIM practice parameter urges practices to document those rules and conformance claims before deployment [1]. The technical choices are not vendor-neutral plug-ins; they depend on PACS behavior, metadata reliability, workflow drivers, and the intended review path [1][3][5]. This guide answers the professional research question by explaining the named workflow, identifying decision points where the literature is explicit, stating where evidence is limited, and walking through a hypothetical evaluation example. The goal is not a checklist for purchasing; it is a map of the routing decisions a practice must own, regardless of the AI vendor involved.
What Result Routing Encompasses
A simplified routing flow begins when the PACS sends studies to an AI orchestrator. The orchestrator applies predefined routing rules for each examination type and directs images to the appropriate AI solutions, which return results for distribution to other systems [4]. The ACR–SIIM practice parameter requires documentation of relevant routing rules, including DICOM, HL7, and FHIR specifications, along with inclusion and exclusion criteria [1]. Those criteria are not only about exam labels. Real-world orchestration may also consider sequence-level content, such as whether appropriate MRI sequences exist for a model's training or test set, and anatomy, since an abdomen or pelvis CT may include pulmonary parenchyma or pulmonary vessels despite not being labeled as a chest CT [4].
Routing also includes what happens after inference. Practices are advised to log acceptance, rejection, and overrides of AI outputs in a standardized form, a step that connects routing decisions to peer learning and quality improvement [1]. Monitoring should cover routing rule adherence, accuracy, calibration, drift, and subgroup performance [1]. That means a routing workflow is not complete when a study reaches an algorithm; it is complete only when the routed output is assessed, recorded, and reviewable. A multi-society statement notes that data orchestration involves ensuring the right slices from the correct series of the relevant study for the right patient in the right setting are sent in an optimized time [5].
One recurring limitation is metadata. The ACR, CAR, ESR, RANZCR, and RSNA statement warns that relying on DICOM metadata alone is often insufficient because of high variability and labile content [5]. A routing engine that filters only on modality or study description may silently send incomplete series or miss findings in unlabeled anatomy. The sources do not provide a universal metadata correction standard; they instead point to the need for content understanding and structured routing criteria [4][5].
Where Routing Interfaces with Worklist Prioritization
AI triage models can inform worklist prioritization by flagging cases with unexpected or serious findings. A chest radiograph flagged as positive for pneumothorax may prompt a worklist manager to move that study toward the top, reducing the time before a radiologist confirms the finding [3]. But study prioritization depends on many factors beyond AI output: study availability, completeness, ordering department, chronicity, and service-level agreements such as required turnaround time [3]. Standalone AI screening tools cannot claim the top position on a worklist because they are unaware of those competing factors and the relative importance of other pending studies [3]. Routing logic therefore has two distinct jobs: selecting the right algorithm for the study and delivering the result into a broader prioritization context.
The Prioritization of Worklists for Reporting profile, released for trial implementation, defines transactions for AI models, the radiology information system, and other relevant systems to feed a worklist driver application such as PACS or EHR [3]. The source describes the profile's scope as balancing factors across many studies rather than dictating a single prioritization formula. This separation matters in routing design. A result routed only to the PACS may be visible to referring physicians who have not been educated about the model's details and accuracy, which is not always optimal [5]. The same statement notes that PACS currently offers limited modes for AI results integration, and in most instances the radiologist cannot modify AI results in PACS [5]. Routing destinations should therefore be evaluated as carefully as routing triggers.
Documenting Local Routing Controls
A practical routing record should capture the DICOM or HL7 fields used as inclusion and exclusion criteria, the version and intended population of each connected model, and the conformance claims for relevant IHE profiles such as AI Workflow for Imaging or AI Results [1]. The record should state whether an output will be recorded in the medical record and how a qualified end-user reviews it [1]. It should also name the escalation path for routing failures and the person or role authorized to take a model offline [1]. This documentation supports the periodic monitoring that the ACR–SIIM parameter requires and provides the audit trail needed when a study is misrouted.
Technical Decision Points and Documentation Requirements
The ACR–SIIM practice parameter separates deployment records from ongoing monitoring. Before use, practices should document version or build, deployment date, intended use, how outputs will be reviewed by qualified end-users, whether and how outputs will be recorded in the medical record, intended populations, any FDA-authorized predetermined change control plan, acceptance testing results, and routing rules [1]. After deployment, practices should establish continuous monitoring with periodic reports on routing rule adherence, accuracy, calibration, drift, and subgroup performance [1]. A further control is assigning authority to take an AI solution offline for remediation [1]. The sources do not specify one governance structure for all practices, but they consistently place routing decisions inside a review group that includes qualified end-users.
Regulatory documentation adds another layer for triage and notification software. Under 21 CFR 892.2080, design verification and validation must include a detailed description of notification and triage algorithms, algorithm inputs and outputs, and how the algorithm affects clinical practice or patient care [2]. Performers must also provide a summary of triage effectiveness, such as improved time to review of prioritized images for pre-specified clinicians, with statistical uncertainty and subgroup analyses [2]. This requirement is relevant to routing because triage effectiveness is not the same as diagnostic accuracy; a routing workflow can reduce review time for a specific clinician group without improving sensitivity and may do so only under defined conditions [2].
The technical method can be tag-based or content-aware. Mayo Clinic's enterprise FAST team defines DICOM tag-based routing rules in its DEWEY orchestrator, which listens for new studies, classifies them by tags, launches tag-specific workflows, and wraps AI outputs into DICOM objects such as structured reports [6]. As algorithms reach more sites and users, those DICOM tag-based rules are refined to ensure proper series routing [6]. This is one organizational example and should not be read as a market-wide requirement. It does, however, show that routing rules are treated as a living configuration, not a one-time setup.
Monitoring Routing Failures and Performance
Routing failure usually appears first as a change in run counts or rejection patterns. The Mayo FAST team reports monitoring weekly algorithm run counts, including statistical process control charts, and beginning root cause analysis by investigating image series routing from the scanner [6]. The team has rolled out 17 internally developed medical image-based algorithms, averaging over 10,000 algorithm runs weekly, with algorithms classified as processing and perception tasks [6]. That denominator is specific to one enterprise and does not establish industry-wide failure rates. It does illustrate that routing problems often surface as operational signals before they appear as clinical complaints.
The ACR–SIIM parameter adds that complications or adverse events potentially related to AI use should be reviewed in peer-learning programs, with findings or conclusions made actionable [1]. Potential safety events should be triaged for medical device reporting, and a documented escalation pathway should be maintained [1]. Acceptance, rejection, and overrides of AI outputs should be logged in a standardized form, and practices should benchmark against qualified registries such as Assess-AI [1]. These monitoring obligations apply to routing because a misrouted study can trigger a false-negative absence of AI output, an event that is harder to detect than an incorrect positive result. The sources do not provide a validated threshold for what routing adherence rate is acceptable; they specify the records and reviews that should exist.
A Concrete Evaluation Example
Consider a hypothetical radiology group evaluating triage routing for intracranial hemorrhage on noncontrast head CT. The proposed flow assumes that the PACS sends each new head CT series to an AI orchestrator, which checks DICOM tags and then confirms that the correct axial noncontrast series exists. If the series is present, the orchestrator routes the study to the hemorrhage triage model. If the model returns a positive flag, the result is delivered to the worklist driver with a timestamp, confidence or uncertainty indicator where allowed by regulation, and the study identifiers used for inference [1][4]. This is a suggestion, not a tested installation, and the group would need to validate every step locally before clinical use.
The evaluation would begin with acceptance testing of the routing rules, including whether a head CT labeled with a noncontrast protocol but containing the wrong series is rejected and logged. The group would then run a small sample of positive and negative cases through the full routing path, recording whether each study reached the correct model, whether the output returned to the intended destination, and whether the radiologist override rate was stable. The ACR–SIIM parameter supports logging overrides in a standardized form, and the Mayo example supports reviewing routing first when run counts change [1][6]. The group would also need to define how flagged cases interact with worklist priority, because the triage model alone cannot account for other pending studies, service-level agreements, or ordering department factors [3]. Finally, the group would document the model version, intended population, rejection reason codes, and the person authorized to disable routing if harm is suspected [1].
Sources & references
- acr–siim practice parameter for imaging artificial intelligence (ai) · gravitas.acr.org
- 21 CFR § 892.2080 Radiological computer aided triage and notification software. · www.ecfr.gov
- Integrating and Adopting AI in the Radiology Workflow: A Primer for Standards and Integrating the Healthcare Enterprise (IHE) Profiles | Radiology · pubs.rsna.org
- AI for Radiology: A Primer Part II. Interacting with AI Results | Radiology · pubs.rsna.org
- Developing, Purchasing, Implementing and Monitoring AI Tools in Radiology: Practical Considerations. A Multi-Society Statement From the ACR, CAR, ESR, RANZCR & RSNA - Journal of the American College of Radiology · www.jacr.org
- State of the AI: Post-Deployment Monitoring of Radiology-Focused Internally Developed AI - PMC · pmc.ncbi.nlm.nih.gov
- Intelligent radiology workflow optimization with AI agents | Artificial Intelligence · aws.amazon.com