Generative AI in Radiology: Report Drafting, Triage and What Is Cleared

Generative AI in radiology has moved from research demos to the first real deployments. Hospitals are now using models that read an image and draft the report, and the FDA has cleared triage tools built on large foundation models instead of one narrow algorithm per finding. At the same time, no generative tool that writes a full diagnostic report on its own has been cleared in the US as of late 2026.

That gap matters if you run a radiology department, a diagnostic centre or a teleradiology service. The useful question is not "will AI replace radiologists" but "which parts of the reporting workflow can AI take on today, with what evidence, and under what rules".

This guide covers what has changed in 2025-2026, what the evidence shows, the regulatory status, and a practical way to start.

What changed in radiology AI in 2025-2026

For most of the last decade, radiology AI meant single-purpose detectors: one algorithm for intracranial haemorrhage, another for pulmonary embolism, another for lung nodules. Each needed its own validation, integration and contract. That is still most of the market. Independent trackers that count the FDA's list put radiology at roughly three quarters of all AI-enabled devices the agency has authorised.

Three things are different now.

• Foundation models for triage. In January 2026 Aidoc received FDA 510(k) clearance for an abdominal CT triage tool built on its CARE foundation model, combining 11 newly cleared indications with three earlier ones in one workflow. One model, many findings.
• Report drafting in production. Northwestern Medicine built a generative model in-house that drafts radiograph reports for radiologists to edit, and published results in JAMA Network Open in June 2025.
• Breakthrough designations for draft reports. In 2026 the FDA gave Breakthrough Device designation to chest X-ray report-drafting systems from Mosaic Clinical Technologies (part of Radiology Partners) in March and from Aidoc in June. Breakthrough status speeds up review; it is not clearance.

The FDA has also said, on its AI-enabled medical device list, that it will explore ways to tag devices that use foundation models and LLM-based functions. That is a signal the agency expects many more of them.

How generative report drafting works

A report-drafting system is a vision-language model. It takes the image (and ideally prior studies and the clinical indication) and produces structured text: findings, then an impression. The radiologist opens the study, sees the draft, edits it and signs.

The key design choice is that the draft is a starting point, not an answer. The radiologist still owns the read. That is also how the FDA designations are framed: drafts for radiologist review.

What the evidence shows so far

The strongest published deployment evidence is the Northwestern study. Across its 11-hospital network, the team analysed roughly 24,000 radiograph reports over five months in 2024. Report completion efficiency improved by 15.5% on average, with some radiologists seeing gains of up to 40%, and peer review found no loss of clinical accuracy. The model was built on Northwestern's own data rather than adapted from a general chatbot.

Read the numbers carefully:

• It was radiographs (X-rays), the highest-volume and simplest modality. CT and MRI reports are longer and harder.
• 15.5% is a productivity gain on report completion, not a diagnostic accuracy gain.
• The model was trained on the same institution's reports and style. A model trained elsewhere may need adaptation to your templates and terminology.

For triage tools, vendors report fewer false alerts with foundation models than with older point solutions. Those are company claims; ask for the validation data behind them.

Regulatory status: US and India

Different parts of the radiology AI stack sit in different regulatory places.

In India, CDSCO's guidance on medical device software classifies standalone software by the condition it addresses and how much its output drives clinical decisions, and it asks AI developers to describe training data and bias across Indian sub-populations. We cover both regimes in our guide to AI medical device regulation under FDA and CDSCO.

Where generative radiology AI fits in an Indian hospital

Indian imaging departments face the same pressure as everywhere else: rising CT and MRI volumes, too few radiologists outside big cities, and heavy reliance on teleradiology at night. That makes three use cases attractive.

• Night and weekend triage. Flagging suspected bleeds, free air or pneumothorax so the on-call radiologist reads those first.
• X-ray report drafting for high-volume OPD and health-check work. Normal and near-normal chest X-rays dominate volume; drafts help most here.
• Structured reporting. Even without full drafting, an LLM can turn dictated free text into a structured template and check that required fields are present.

All three depend on plumbing more than on the model. The AI needs DICOM images from your PACS, the order and clinical indication from your HIS or RIS, prior reports for comparison, and a way to send results back. If orders still arrive on paper slips, fix that first. Our HL7 FHIR integration guide explains the standards involved, and our diagnostic centre software guide covers the RIS side.

Costs and effort

Pricing for radiology AI varies widely: per study, per scanner, per site per year, or bundled into a platform fee. What drives the real cost is usually not the licence.

• Integration: connecting PACS, RIS and HIS, routing studies, returning results to the worklist. Often weeks of work per site.
• Validation: running the tool silently on your own studies before go-live, with radiologists grading the output.
• Change management: radiologists need time to calibrate trust in drafts and triage flags.
• Monitoring: tracking performance after go-live, because scanner upgrades and protocol changes can shift accuracy.
• Data infrastructure: storage and compute if you host anything on-premise.

For a mid-sized Indian hospital, the integration and validation effort is often larger than year-one licence cost. Budget for it explicitly.

How to evaluate a radiology AI vendor

Use this checklist in vendor conversations.

1. Which indications are cleared or approved, in which countries, and for which modalities and scanner makes?
2. What was the validation population? Does it include Indian patients and the scanner types you run?
3. Can you run it in shadow mode on your own studies for four to eight weeks before go-live?
4. How does it integrate: DICOM routing, HL7 or FHIR results, worklist re-prioritisation, report templates?
5. What happens when the model is updated? Is there a predetermined change control plan, and will you be told?
6. What does it log? Can you audit which drafts were accepted, edited or rejected?
7. Where is data processed and stored, and how does that fit the DPDP Act and your contracts?

A practical way to start

1. Pick one modality and one problem. Chest X-ray turnaround in OPD, or CT triage at night. Not "AI for radiology".
2. Measure the baseline. Turnaround time, reports per radiologist hour, critical-finding notification time, discrepancy rate.
3. Fix the data path. Orders, indications and priors must reach the AI digitally.
4. Run in shadow mode. The tool reads in parallel; radiologists grade its output without it touching patient care.
5. Go live with guardrails. Drafts clearly marked as AI-generated, radiologist sign-off required, easy feedback buttons.
6. Monitor monthly. Accuracy samples, edit distance, turnaround, and any missed findings reviewed in your quality meeting.

If you need custom pieces, such as structured reporting on top of your existing PACS or a small model to route studies, that is ordinary healthcare software development and machine learning work. At Algon we build those integration layers for hospitals; the cleared clinical models themselves should come from vendors who carry the regulatory responsibility.

What to measure after go-live

Review these monthly with your radiology lead. If turnaround improves but discrepancies rise, slow down and re-check the workflow.

Common mistakes

• Buying a tool for a finding you rarely see, while the real bottleneck is routine X-ray volume.
• Skipping local validation because the vendor has an FDA clearance from another population.
• Letting AI flags live in a separate viewer nobody opens during a busy shift.
• Treating a breakthrough designation or CE mark as proof the tool works on your scanners.
• No plan for what happens when the vendor updates the model.

Conclusion

Generative AI in radiology is real in 2026, but in specific forms: foundation-model triage that is cleared and in use, and report drafting that has published productivity results and is moving through regulatory review. The gains come from faster, better-prioritised reporting with the radiologist firmly in charge.

Start with one modality, validate on your own studies, wire it properly into PACS and HIS, and measure. That approach works whether you buy a cleared product or build supporting tools around one. For a wider view of where AI helps a hospital beyond imaging, see AI in hospital management.