By Guillaume Herpe, Ingrid Millet, Mathieu Lederlin, and Joseph Benzakoun
As France’s national radiology congress, JFR 2026, kicked off on Thursday in Paris, AI was prominent on the agenda, notably its use in A&E and how it can be practically integrated into daily practice.
The dedicated session on AI in emergency which took place on the afternoon of the first day of the meeting provided a practical overview of its applications in emergency imaging: chest, abdomen, neuroradiology, automated detection and decision support, its real benefits for staff and patients, and its limitations in practice.
These specialty presentations and roundtable aimed to equip participants to assess for themselves, within their own organizations, the suitability of a tool both before adopting it and after its implementation, noted the experts.
While there are varying levels of maturity, prospects and applications of AI in A&E, ranging from automated detection to decision support and the coordination of patient pathways, transition from proof of concept to everyday clinical practice remains the most challenging stage, they agreed.
At a time when more than 200 AI solutions dedicated to medical imaging are now authorized in Europe and nearly 1,000 in the United States, emergency imaging remains one of the most eagerly anticipated – and most demanding – areas for their deployment.
High patient volumes, the need for speed, night shifts, life-threatening conditions are precisely where the promise of AI ‘that sorts, detects and prioritizes’ should be felt most keenly, according to session presenters but do these tools really change daily practice, or do they remain merely a technological promise?
Field lessons
In his presentation on ‘proof of concept to practice’, Dr. Guillaume Herpe (Poitiers University Hospital), focused not on ‘what can AI detect?’, but ‘how can it be implemented within a department?’. He illustrated his talk with local data from Poitiers University Hospital, at times challenging the notion of impact.
Out of 3,700 patients, software for the automated detection of fractures reduced discrepancies in interpretation between A&E doctors and radiologists by 17% (5.1% compared with 6.6% without AI), without, however, altering the rate of clinically significant discrepancies. For post-traumatic intracranial hemorrhages detected on CT scans, combining a junior doctor with AI on night duty (682 scans) achieved a negative predictive value of 99.3% and an F1-score of 97%, standardizing diagnostic performance between day and night shifts.
A four-year old boy fell from a ladder. The initial report identified a fracture of the ulna. The fracture detection software (Gleamer BoneView) additionally identified a butter-slice fracture of the distal radius, which was not mentioned in the report — illustrating the role of AI as a safety net in pediatric trauma.Images courtesy of Guillaume Herpe and SFR.
Finally, he pointed to a study currently being drafted at Le Havre University Hospital which demonstrates that a change in the patient population attending A&E directly impacts the performance of a fracture detection algorithm – a deviation known as algorithmic drift.
Performance figures published in the literature therefore never guarantee high or stable performance once the tool has been deployed locally, hence the importance of continuous monitoring, including of any drift over time.
A performance result published in the literature therefore never guarantees high or stable performance once the tool has been deployed locally, hence the importance of continuous monitoring, including of any drift over time, noted the authors.
The question of adoption remains. A survey of staff at Poitiers University Hospital reveals a high level of trust in AI and a strong intention to use it, but a more moderate perception of its usefulness: the main obstacle is no longer acceptance, but demonstrating a real clinical impact.
He pointed to a recent randomized trial published in The Lancet, involving 700,000 patients evaluating an AI-assisted stethoscope: despite good algorithmic performance, 40 per cent of practitioners had stopped using the tool after twelve months, and it was the actual usage rate — rather than the algorithm’s accuracy — that determined its impact on the population. This is an issue of particular sensitivity for A&E radiologists, given that tools specifically designed for their discipline alone account for more than half of the market for AI solutions already available in medical imaging.
Prof. Mathieu Lederlin (Rennes University Hospital) opened the session with use of AI in chest imaging for detecting pneumothorax and areas suggestive of pneumonia, and for the prioritization of suspected pulmonary embolism on CT angiography. In turn, Prof. Ingrid Millet (Montpellier University Hospital) discussed acute abdominal conditions where AI is making greater strides as a decision-support tool rather than as a standalone diagnostic tool. Closing the subspecialty presentations, Dr. Joseph Benzakoun (GHU Paris – Sainte-Anne) addressed emergency neuroradiology, and the way in which these tools help to coordinate the patient’s pathway, from the initial alert to the revascularization procedure.
Editor's note: This is an edited version of a translation of an article published in French online by the SFR in its congress newsletter JFR Actu’. Translation by Frances Rylands-Monk. To read the original version, go to JFRActu’.
