The Radiology Queue: AI in Medical Imaging Workflows

Medical-imaging AI is often presented as a contest between an algorithm and a clinician. In practice, value appears—or disappears—inside a workflow: the correct study must reach the correct model, the output must reach the intended user at the right time, and the user must understand what the output can and cannot support. A high-performing model with broken routing can be clinically useless or unsafe.

Medical and regulatory notice: This article is about system design and is not medical advice. Imaging findings must be interpreted by qualified professionals in clinical context. Device status, permitted claims, user population, and post-market duties vary by intended use and jurisdiction. An organization must use the authorized version within its labeling and applicable local requirements.

Start with the clinical workflow, not the demo

An imaging pathway spans order, protocol, patient preparation, acquisition, quality control, transfer, worklist, interpretation, reporting, communication, and follow-up. AI can support several distinct functions:

• detect incomplete views, motion, positioning, or acquisition-quality problems;
• reconstruct images or reduce noise within a validated acquisition setting;
• segment anatomy or quantify a measurement;
• flag a suspected finding for worklist prioritization;
• present a detection, probability, measurement, or heat map during interpretation;
• compare current and prior studies;
• draft report language for clinician review;
• track whether a critical-result communication was completed.

These functions have different risk, evidence, latency, and regulatory needs. “AI for radiology” is not an adequate intended-use statement. Define modality, body region, protocol, patient population, user, output, clinical role, exclusions, and response to failure.

Treat DICOM and workflow context as part of the product

DICOM is the international standard for transmitting, storing, retrieving, processing, and displaying medical-imaging information. Its current edition covers information objects, services, encoding, network communication, media, security profiles, web services, and more. Conformance to DICOM does not guarantee that every implementation uses fields consistently.

A production integration must validate:

• patient, study, series, and instance identity;
• modality, anatomy, view, laterality, protocol, contrast, and acquisition parameters;
• complete-series arrival and late images;
• prior-study retrieval and patient-merge events;
• image orientation, pixel spacing, compression, and transformations;
• burned-in identifiers and secondary captures;
• model result objects, presentation states, measurements, and provenance;
• PACS, VNA, RIS, EHR, worklist, and reporting-system behavior.

Do not route solely from one free-text description. Use tested combinations of coded and technical attributes, with an exception queue. Show the source series and model version with each result. Our AI data-quality observability guide provides patterns for freshness, completeness, and schema monitoring.

Know what authorization does—and does not—mean

The FDA's AI-Enabled Medical Device List identifies devices FDA has recognized as AI-enabled and authorized for marketing in the United States. FDA explicitly says the list is not comprehensive, is updated periodically, and links to device-specific database entries. It is not an endorsement of AI in general and does not authorize off-label claims, a different model version, or a hospital's custom workflow.

Read the device's public decision material and labeling. Determine:

• submission number and exact product/version;
• intended use, indication, user, population, modality, and exclusions;
• whether the function is detection, triage, quantification, reconstruction, or another role;
• required acquisition environment and compatible systems;
• warnings, limitations, performance evidence, and cybersecurity obligations;
• whether the output is intended to be independently reviewed and how.

Outside the United States, use the applicable regulator and conformity pathway. Do not assume FDA authorization supplies authorization elsewhere—or vice versa.

Design triage so the queue remains safe

A triage function generally changes worklist priority; it does not provide a definitive diagnosis or permit a study to disappear. A negative model result should not remove a study from required interpretation unless an explicitly authorized workflow says otherwise.

Define what happens when:

• the model flags a study already being read;
• many studies are flagged simultaneously;
• the model is unavailable or returns no result;
• an unsupported protocol is received;
• duplicate studies or corrected images arrive;
• a flag conflicts with clinical priority already assigned;
• the priority message reaches one system but not another;
• a user dismisses or overrides the flag.

The interface must distinguish “AI priority,” “clinical stat,” and ordinary queue order. Record timestamps for acquisition completion, routing, inference, result delivery, queue change, first open, preliminary action, and final report. Those timestamps expose whether the AI actually changed care or merely produced fast inference.

Keep the human-AI team inside the evidence model

The 2025 final IMDRF Good Machine Learning Practice principles emphasize the total product life cycle and the performance of the human-AI team. IMDRF's SaMD clinical-evaluation framework distinguishes valid clinical association, analytical validation, and clinical validation.

That means evaluation cannot stop at sensitivity and specificity on a curated test set. Study whether the intended user can find the output, interpret it correctly, recognize limitations, and recover from failure under realistic workload. Measure automation bias: does a false positive pull attention away from more urgent work? Does a false negative reassure a reader improperly? Does a heat map narrow the search too early?

Use meaningful human approval: show source images, compatible priors, result type, uncertainty where validated, and known limits. A clinician must be able to ignore, correct, and report a problem without losing access to the original study.

A worked example: suspected intracranial hemorrhage triage

Consider a non-contrast head CT routed to an authorized triage algorithm.

1. The router confirms modality, protocol, anatomy, series completeness, and supported acquisition characteristics.
2. Images are sent through the approved integration, and the inference request receives a traceable identifier.
3. The model returns a priority notification linked to the exact study and version. It does not alter image pixels or populate a final diagnosis.
4. The worklist marks the study with a distinct AI-priority indicator. Existing clinical-stat status remains visible.
5. A qualified reader opens the original series, reviews the complete examination and clinical context, and makes an independent interpretation.
6. The normal critical-results process is followed when warranted; AI notification is not treated as completed clinician-to-clinician communication.
7. The system records delivery, display, user action, final report, correction, and any discrepancy report.

If the model times out, an unsupported protocol arrives, or a patient merge occurs, the study remains in the ordinary queue and the exception is visible. The fail-safe state is continued interpretation, not silent exclusion.

Validate locally before changing care

Vendor evidence is necessary but may not predict performance at a new site. Scanner manufacturer, field strength, reconstruction kernel, dose, protocol, population, prevalence, technologist practice, and reporting culture can all affect results.

Create a staged evaluation:

1. Technical verification: connectivity, routing, supported inputs, output rendering, timing, identity, audit, and failure behavior.
2. Retrospective local validation: representative historical studies with an appropriate reference standard, prespecified analysis, subgroup slices, and no leakage.
3. Shadow deployment: live data and timing without changing the clinical queue.
4. Limited prospective use: defined sites, shifts, users, and rollback criteria.
5. Scaled use: only after workflow, safety, and equity evidence meets approved gates.

The peer-reviewed DECIDE-AI guideline provides reporting recommendations for early-stage live clinical evaluation of AI decision-support systems. It is a reporting guideline, not a substitute for regulatory authorization or a complete study-design standard, but its focus on clinical workflow and human factors is valuable.

Measure clinical workflow and safety

Model metrics should include sensitivity, specificity, positive and negative predictive values at local prevalence, calibration when probabilities are shown, unusable-output rate, and confidence intervals. Subgroup evaluation should be clinically and statistically justified rather than a token checklist.

Workflow measures include:

• time from acquisition completion to validated model result;
• time from result to reader opening and final report;
• position change in the queue and displacement of unflagged studies;
• false reprioritization and missed-priority events;
• user acceptance, override, and discrepancy reasons;
• report corrections and critical-result communication;
• system uptime, unsupported-study rate, and stale-result rate;
• added clicks, interruption burden, and reader trust over time.

Patient-important outcomes may require larger prospective studies and careful causal design. Faster report time is not automatically better outcome, and before-after comparisons can be confounded by staffing, demand, protocol, or policy changes.

Manage drift and controlled change

Medical-imaging environments change continuously: scanners are upgraded, reconstruction settings move, new protocols appear, population mix shifts, and integrations are patched. Monitor input distribution and workflow performance, but do not treat every statistical shift as clinical drift or automatically retrain a regulated model.

FDA's final guidance on Predetermined Change Control Plans for AI-enabled device software functions was issued in August 2025. It gives recommendations for describing planned modifications, the methodology for developing and validating them, and impact assessment in a PCCP. It does not allow arbitrary unreviewed updates; the authorized plan and applicable requirements govern.

Hospitals need their own change control too. A model update, PACS migration, scanner software release, routing-rule edit, display change, or new site can trigger re-verification or revalidation. Freeze deployment during investigation when thresholds or safety events require it. Preserve the previous working configuration and rollback path.

Cybersecurity, privacy, and operational continuity

Imaging data is highly identifying, and medical-device integrations create privileged paths through clinical networks. Use network segmentation, mutual authentication where supported, least privilege, encryption, key management, logging, vulnerability handling, vendor access controls, and tested backup or recovery. Remove unnecessary data from model requests, but do not de-identify in a way that breaks patient matching or clinical use.

Threat modeling should include altered DICOM attributes, study substitution, result replay, malicious files, unauthorized queue changes, compromised vendor support, and denial of service. A downtime procedure must keep studies visible and interpretable without AI. Users should know when the service is degraded and which results may be delayed.

For a broader release review, use our AI operational-readiness checklist and maintain audit evidence for every critical control.

Release gates

Require documented evidence for:

• exact authorized product, version, intended use, and local regulatory status;
• DICOM conformance and tested local routing semantics;
• technical, local, shadow, and prospective evaluation proportional to risk;
• human-factors testing and clear result presentation;
• ordinary-queue continuity for negative, unavailable, or unsupported cases;
• subgroup, scanner, protocol, and site performance;
• cybersecurity, privacy, access, and vendor controls;
• post-deployment monitoring, complaint, incident, and vigilance processes;
• controlled changes, revalidation triggers, and rollback;
• named clinical, technical, safety, and executive owners.

No single metric should open the gate. The decision is a safety case connecting intended use, evidence, controls, residual risk, and accountable approval.

Source notes

Reviewed 2026-07-30. Primary sources are the current DICOM standard; FDA's periodically updated AI-Enabled Medical Device List; the 2025 final IMDRF GMLP principles; IMDRF's SaMD clinical-evaluation framework; FDA's August 2025 final PCCP guidance; and the peer-reviewed DECIDE-AI reporting guideline. The FDA list is not comprehensive and authorization is product-, version-, and intended-use-specific. DECIDE-AI guides reporting of early clinical evaluation; it is not regulatory authorization.

Questions imaging leaders should ask

Does FDA listing mean a product is approved for every use?

No. Inspect the device-specific authorization and labeling. The list itself is a transparency resource, not a blanket claim.

Can a triage model rule out disease?

Only if that exact role is supported by applicable authorization and evidence. Many triage tools prioritize review and do not replace interpretation or remove negative studies.

What is the most useful first test?

End-to-end technical verification with real local workflow exceptions. A correct model output delivered to the wrong patient, series, or queue is still a failed system.

Should a hospital retrain when drift appears?

Not automatically. Investigate the cause, clinical significance, authorized change pathway, and validation need. A routing or scanner problem may be the real cause.

Make the whole queue safer

Imaging AI earns trust when it improves the complete path from acquisition to action while keeping original images, clinician authority, and failure recovery intact. The goal is not to make the model look intelligent. It is to help the imaging service find important work, interpret evidence, communicate reliably, and learn from discrepancies without creating a new invisible source of risk.