Digital Health

AI in Radiology: Faster Reads, New Questions

Published 5 min read Acuitas Health Analytics

AI in Radiology: Faster Reads, New Questions

Radiology was one of the first parts of medicine to go fully digital. Every X-ray, CT and MRI is already a large set of numbers. That makes it a natural fit for AI, which learns patterns from images. It is now the field with the most regulator-cleared AI tools.

How many tools are there?

A 2026 study of the FDA's AI-enabled device list found 1,094 radiology records out of 1,430 in total, from 1995 to 2025. That is 76.5% [2]. The FDA publishes and updates the list so the public can see what has been authorised [1].

FDA AI-enabled device authorisation records, 1995–20251,430All AI device records1,094Radiology records
Source: Golshani and Joseph, analysis of the FDA list [2].

How AI reads scans

Most radiology AI is trained on thousands or millions of labelled images. It learns to link patterns in the pixels to findings, such as a bleed, a fracture or a lung nodule. Once trained, it can look at a new scan in seconds and mark areas of concern. Common uses include:

  • Triage: moving scans with a likely stroke, bleed or collapsed lung to the top of the reading list.
  • Detection: marking possible cancers on mammograms or nodules on chest CT.
  • Measurement: measuring organ size, heart function or tumour growth the same way every time.
  • Image quality: making clearer images from faster or lower-dose scans.
  • Reporting: drafting parts of the report for the radiologist to check.

Benefits for patients and doctors

Imaging use keeps growing, and many countries have too few radiologists. AI can help by sorting urgent cases faster, cutting time spent on routine measurements, and giving a second look. Faster reads matter most for time-critical conditions like stroke, where every minute of delay can mean more brain damage.

Regulation and safety

Each tool is cleared for a specific use. Using it outside that use is a risk. Tools can perform worse on images from different scanners, or for patient groups that were rare in their training data. WHO's AI ethics guidance calls for transparency, accountability, inclusiveness and safety [3]. In practice, hospitals need to test tools locally and keep checking them over time.

Adoption barriers

  • Payment: in many countries there is no separate payment for AI reads, so hospitals must justify the cost through time saved.
  • Integration: tools need to work inside existing image systems without extra clicks.
  • Trust: radiologists need to understand when a tool is likely to be wrong.
  • Too many alerts: false alarms waste time and can cause "alert fatigue".
  • Responsibility: who is liable when AI and a radiologist disagree is still being worked out.

Questions to ask before buying

  • What exactly is the tool cleared to do?
  • How did it perform on scanners and patients like ours?
  • How many false alarms should we expect?
  • Who monitors performance after go-live?

Key numbers at a glance

  • 1,094 of 1,430 FDA AI device records are in radiology.
  • That is 76.5% of all records.
  • Radiology, heart and brain together: 90.6%.

Go deeper

Want the full picture? Our report covers market size, growth, segments, countries and the key companies. Read the report or get a free sample.

Radiology's share over time

Radiology has led AI device clearances for years. Of 1,430 FDA authorisation records up to the end of 2025, three specialty panels, radiology, heart and brain, made up 90.6% [2]. The FDA's public list lets anyone search tools by name, company, specialty and date [1].

Screening and mammography

Breast screening is one of the busiest areas for AI. Screening programmes read millions of mammograms, and most are normal. AI can help sort normal from suspicious images, or act as a second reader. Screening services are testing whether AI can safely replace one of two human readers, which would ease staff shortages.

Stroke and emergency scans

Stroke is a leading cause of death and disability. WHO says a brain scan should be done as soon as possible when stroke is suspected, because clot-busting treatment works best within hours [4]. AI tools that flag likely strokes on CT scans and alert the stroke team aim to cut this delay.

What radiologists say they want

  • Tools that save time, not add clicks.
  • Clear confidence scores and highlighted areas.
  • Few false alarms.
  • Easy ways to give feedback when the tool is wrong.

Where to read more

All figures in this article come from public sources listed below. For more detail, start with the US FDA list of AI-enabled medical devices and WHO's guidance on AI ethics for health. Each is free to read online. If anything here raises a question about your own health, take this article to your doctor or pharmacist and talk it through together.

Training the next radiologists

As AI takes on routine measurements and sorting, radiologists in training still need to learn to read scans themselves. They must spot when a tool is wrong and handle unusual cases it was never trained on. Teaching programmes are starting to include how AI works, its limits and how to check its output. WHO's principle of responsibility and accountability means a human stays answerable for the final report [3].

What this means

AI is becoming a normal part of the radiology toolkit, mostly as a helper rather than a replacement. The biggest gains will come where radiologists are scarce and where speed saves lives. Proof of real-world benefit, simple integration and clear responsibility will decide which tools last.

Explore on Acuitas

Topics: Artificial intelligence · Imaging

Sources

  1. US FDA: Artificial intelligence-enabled medical devices (public list)
  2. Golshani and Joseph (Cureus, 2026), via BioSpace: three in four FDA AI device authorizations were in radiology
  3. WHO (2021): Ethics and governance of artificial intelligence for health
  4. WHO fact sheet: Stroke

Banner image: CT Scanner (NIH BioArt 373 - 632561) by Courtesy of NIAID Ryan Kissinger, Public domain, via Wikimedia Commons.

Topics: Artificial intelligence · Imaging

Back to Blog