RCR publishes a second edition of its guidance on fitting image-analysis AI into reporting
On 21 October 2026 the Royal College of Radiologists (RCR) published the second edition of its guidance on integrating image-analysis AI with radiology reporting systems (RIS and PACS), replacing the 2021 first edition. It now includes a section on training.
What's new
The guidance is written for "reporting authors", a term it uses for radiologists, reporting radiographers, sonographers and other trained professionals who report images. It covers the stage after the images are taken, when AI software analyses them alongside or before the human report. It does not recommend particular products and does not cover the ethics of AI.
Its core standards say AI must be used only for its intended purpose, its diagnostic accuracy must be clearly declared, and the AI analysis should be finished and available in PACS before a human starts reading. Vendors must state which abnormalities each algorithm looked for and its sensitivity and specificity for each one, and the RCR warns against assuming that algorithm output is always correct. It also asks departments to monitor AI after deployment, because software faults, drift or a new version can change performance.
The human report stays central. Reporting authors should understand the limits of an algorithm, may challenge its output, and should still issue a report tailored to the patient. Reports drafted by large language models must be checked by the reporting author. If AI analysis arrives after a report and disagrees with it, a radiologist should review the images while staying alert to anchoring and automation bias.
The new training section is the part most relevant to learners. It says AI output should be shown in a way that lets readers form their own interpretation first. That favours AI markings that can be switched off for the first read and revealed afterwards for feedback. AI marks burned permanently into the image (so-called secondary capture) "cannot be unseen", which the RCR says risks anchoring bias and undermines perceptual learning for trainees.
What it means for learners
- Read the film before you look at the AI. The RCR's advice matches how skills are built: run your own search pattern first, such as the one in how to read a chest X-ray or how to read a skeletal X-ray, then compare your answer with the AI markings and learn from the differences.
- Know what the tool was built to find. An AI fracture tool that says "no fracture" has not looked for effusions, dislocations or a lung lesion on the edge of the film. Check the declared findings and accuracy for the tool your department uses.
- The report is still yours. Whether you are a student presenting a film or a reporter issuing one, AI output is decision support, not a second reader. Follow your local policy on AI-flagged studies and senior review. See also night-shift fatigue and AI in chest X-ray reading.
Source
- The Royal College of Radiologists. Integrating image analysis artificial intelligence with radiology reporting workflows (RIS and PACS), second edition. London: RCR, October 2026. Working group chaired by Prof Amaka C Offiah. rcr.ac.uk: guidance page and PDF
- The Royal College of Radiologists. News, 21 October 2026: second edition published. rcr.ac.uk: announcement
Also in What's new: Night-shift fatigue and AI in chest X-ray reading · UKHSA updates X-ray dose figures · NICE extends AI fracture detection evidence period.
For education only. Not a substitute for a formal report, senior review or your local policy. Never use this page to make decisions about a real patient.