Radiology AI has a peculiar profile: FDA-cleared tools have existed for years, academic evidence supports clinical value, and yet deployment is patchy and integration is the main bottleneck. This post is the integration reality — PACS workflow, worklist prioritization, radiologist acceptance dynamics, and why the tools exist but the integration is the hard part.
The tools exist
FDA-cleared AI in radiology: over 700 devices by early 2026, spanning modalities (CT, MRI, X-ray, mammography, ultrasound) and use cases (nodule detection, triage, quantitative measurement).
Clinical evidence substantial for many use cases. Stroke triage (large vessel occlusion detection), pulmonary embolism detection, breast cancer screening — all have strong evidence.
Vendor ecosystem mature. Aidoc, Zebra Medical, Rad AI, Viz.ai, GE Healthcare, Philips, Siemens — dozens of companies with commercial products.
Integration is the bottleneck
PACS workflows built pre-AI. Adding AI means modifying established clinical workflows that radiologists have used for decades.
Multiple AI vendors per site common. A hospital might have 5-10 different AI tools from different vendors; each has its own integration, UI, workflow. No unified experience.
Reimbursement unclear. Some AI use cases have specific reimbursement codes; most don't. Unclear business case for broad deployment.
Radiologist acceptance variable. Some find AI helpful; some find it disruptive; few have strong evidence yet of meaningful productivity gains per radiologist.
Integration patterns
Worklist prioritization. AI flags critical findings; those cases jump the queue for radiologist review. Highest-value integration pattern for time-critical findings (stroke, PE).
Report draft generation. AI pre-fills structured report sections based on image findings. Radiologist edits rather than writing from scratch. Maturing pattern.
Overlay annotations. AI marks findings directly on images for radiologist to confirm or reject. Common for nodule detection, fracture detection.
Quantitative measurement. AI performs measurements (tumor size, cardiac function) that radiologists would do manually. Saves measurement time; integrates into reports.
AI orchestration layers
Multi-vendor orchestration platforms emerging: Blackford, Ferrum, various startups. Single UI for multiple AI vendors; consistent workflow; unified billing.
PACS-integrated AI: GE, Philips, Siemens, Agfa all embedding AI directly in their PACS. Convenient but vendor-locked.
Cloud vs on-prem debate. On-prem for latency and privacy; cloud for scalability and model updates. Hybrid deployments increasingly common.
Radiologist acceptance dynamics
Worklist prioritization: high acceptance. Radiologists see clear value when AI surfaces urgent cases.
Detection assistance: mixed acceptance. Good tools are welcomed; poor tools (many false positives) are disabled or ignored.
Second-read automation: most resistance. Radiologists view this as replacing expertise; cultural and professional concerns. Slow adoption.
Productivity claims scrutinized. AI vendors often claim 20-40% productivity gains; radiologists often see much smaller improvements in practice.
Regulatory framework
FDA 510(k) clearance for most devices. Substantial equivalence standard, not full efficacy. Clearance ≠ proof of clinical value.
De Novo pathway for novel AI use cases. More rigorous review; longer process.
Predetermined Change Control Plans (PCCP). FDA guidance allows AI vendors to update models within defined parameters without full re-clearance. Reduces update friction.
ROI patterns
Stroke triage: clear ROI through faster treatment windows, better outcomes, reduced long-term care costs. One of the clearest cases.
Lung screening: ROI through higher screening volumes with same staffing; early cancer detection saves downstream costs.
General productivity AI: ROI harder to demonstrate. Without clear patient impact, cost justification is marginal.
Outlook
Consolidation likely. Many vendors, few viable businesses. Expect M&A and closures through 2026-2027.
Orchestration layer matters increasingly. Single UI for multi-vendor AI is the real infrastructure play.
See multimodal AI post for how foundation models are starting to compete with specialized radiology AI.