AI Radiology in 2026: Vendors, Clearances, and Reading Workflow

A trauma scan lands in the PACS worklist at 02:14. Before the on-call radiologist opens it, Aidoc has already flagged a probable pulmonary embolism, pushed an alert to the ED clinician, and reordered the worklist so this study floats above the routine outpatient backlog. The radiologist confirms in ninety seconds. The patient is on heparin before sunrise. This is not a future-state slide deck. It is the production workflow at hundreds of hospitals across the US, Europe, and the Middle East in 2026.

Radiology is the medical specialty where AI has moved furthest from pilot to production. The reasons are structural: imaging is digital end-to-end, ground truth is recorded, and the bottleneck (radiologist time) is so expensive that even modest productivity gains pay for the software within a quarter. This post walks through the vendor landscape, the regulatory state, the integration realities, and where the actual gains are showing up.

The vendor landscape worth knowing#

The FDA’s running list of authorized AI/ML-enabled medical devices crossed roughly a thousand entries in 2026, with radiology accounting for the dominant share. A handful of vendors anchor the market.

Aidoc built the always-on triage layer — chest CT for PE and intracranial hemorrhage, then expanded to abdominal, spine, and cardiac findings. Their model runs against every study as it lands and reorders the worklist. Health systems like Cedars-Sinai, UMass Memorial, and Sheba have deployed it across modalities.

Viz.ai anchored its position around large-vessel occlusion stroke detection — the original Viz LVO clearance opened the door to a CMS New Technology Add-On Payment in the US, which is still a rare reimbursement story for radiology AI. They have since expanded into pulmonary embolism, aortic disease, and a multi-condition cardiology suite.

Rad AI focuses on report drafting and impression generation rather than image interpretation — they sit alongside the dictation workflow, ingesting findings and generating consistent impressions, follow-up recommendations, and patient-friendly summaries. Productivity studies report meaningful time-per-report reductions.

Annalise.ai ships a comprehensive chest X-ray and head CT triage suite with broad finding coverage — over a hundred annotations on a single chest study. Strong adoption in Australia, the UK NHS, and parts of Asia.

Lunit (Korea) built mammography and chest X-ray detection products with strong published reader-study results and meaningful European and Asian footprint.

GE Healthcare Edison and Siemens AI-Rad Companion are the OEM-bundled stacks — each vendor curates an in-house plus third-party model marketplace tightly coupled to their scanners and PACS. The OEM strategy matters because integration is the hardest part of the deployment, and bundling reduces that friction.

Other meaningful players: Nuance Precision Imaging Network (Microsoft) as a model marketplace, Bayer Calantic Digital Solutions, Subtle Medical for image enhancement, Heart Flow and Cleerly for cardiac CT (covered in the cardiology piece), and a long tail of single-finding vendors clearing 510(k) submissions every quarter.

How the technical architecture actually looks#

Most production radiology AI is a convolutional or transformer-based segmentation/classification model trained on tens of thousands of annotated DICOM studies. The training data problem is non-trivial — every site’s scanners, protocols, contrast timing, and patient population differ, so generalization gaps appear when a model trained primarily on Western European data is deployed in South Asia or Latin America. The serious vendors now publish performance breakdowns by site type and have multi-site validation in their 510(k) submissions.

The inference path is where engineering matters. A study lands in the PACS — usually Sectra, Philips IntelliSpace, Change Healthcare, Visage, or an OEM stack — and a DICOM router (sometimes vendor-supplied, often Compass or Laurel Bridge) forwards a copy to the AI orchestration layer. Many sites run Nuance Precision Imaging Network or Blackford Analysis as the orchestration layer that fans out to multiple vendor models and writes results back as DICOM Secondary Capture or Structured Report objects, with worklist alerts surfaced inside the reading workflow.

Latency targets are aggressive: triage findings need to be back in the worklist within roughly two minutes of acquisition, otherwise the radiologist has already opened the study and the alert is wasted.

What production deployment actually looks like#

The honest productivity picture in 2026 has matured beyond the hype. Triage tools deliver real wins for time-sensitive findings (LVO stroke, PE, intracranial hemorrhage, pneumothorax) where the workflow reordering shaves dozens of minutes off door-to-treatment time. The clinical evidence is published and the reimbursement story (CMS NTAP, private payer codes, NHS AI Award funding in the UK) is more credible than it was three years ago.

Report-assist tools like Rad AI deliver clearer per-radiologist productivity gains — measurable reduction in reporting time per study, lower transcription cost, more consistent impressions. Hospitals that staff around productivity are seeing the gains land in the staffing model rather than as ROI on the software line item.

Generalist detection AI — find everything on every chest X-ray — has the murkier story. False positive volume creates alert fatigue, and the radiologist still has to confirm every flagged finding. Where it works well: as a safety net for low-suspicion outpatient studies, or as a teaching adjunct for residents.

Integration with existing PACS and reporting#

The integration debt is the part nobody writes white papers about. Hospitals run on Sectra, Philips IntelliSpace, GE Centricity, Change Healthcare Workflow Intelligence, Visage 7, or a regional vendor — each with its own worklist semantics, hanging protocols, DICOM SR conformance quirks, and security posture. A new AI vendor typically requires:

• DICOM routing from the modality or PACS to the AI inference endpoint
• A method to return results — Secondary Capture image, DICOM SR, or worklist priority update
• A way to surface alerts in the reading client without requiring radiologists to leave their primary viewer
• Audit logging that satisfies the security review
• Single sign-on and identity mapping
• Performance monitoring so model drift is detectable

Health systems that deploy radiology AI well have either built an internal orchestration layer or standardized on Nuance PIN/Blackford as the integration substrate. Going vendor-by-vendor with point integrations is the failure pattern.

Regulatory state and the Predetermined Change Control Plan#

The FDA’s Predetermined Change Control Plan framework — finalized in 2024 and now a routine part of 510(k) submissions for AI/ML devices — lets vendors pre-specify the kinds of model updates they can ship without a new clearance. This unlocks continuous improvement workflows that the old static-device paradigm prevented. EU MDR and the EU AI Act add a parallel high-risk-device track for European deployment, with conformity assessment and post-market surveillance requirements that vendors must navigate separately from the US 510(k) path.

For deploying health systems, the regulatory posture matters because procurement and clinical governance now ask hard questions about model versioning, monitoring, and the change-control plan well before signing.

Where pdpspectra fits#

We help radiology departments and imaging service lines design the integration layer between PACS, AI vendors, and reading workflow — DICOM routing, result write-back, worklist priority, monitoring, and SSO. We are vendor-neutral on the AI itself; we own the orchestration and operational scaffolding that lets the clinical wins actually land. See our AI and LLM integration practice.

Related reading#

• AI in clinical trials in 2026
• Healthcare AI playbook: pilot to production
• Voice AI in clinical documentation

If you are scoping a radiology AI deployment and want a pragmatic read on vendor selection, integration cost, and the operational lift, reach out. We have done this work in academic centres, community hospitals, and tele-radiology operations across multiple geographies.