Artificial intelligence in medical imaging: beyond performance, the challenge of trust

The recent position statement by the Académie nationale de médecine (French National Academy of Medicine) on artificial intelligence in medical imaging marks an important turning point. It does not call into question the value of these technologies - it redefines the conditions for their use.

The question is no longer whether AI performs well. In many use cases, it already does.

The real issue lies elsewhere: can it be trusted in clinical practice?

1. A technological promise already widely demonstrated

Artificial intelligence has profoundly transformed analytical capabilities in medical imaging. Anomaly detection, automatic organ segmentation, lesion quantification, decision support: the performance achieved on certain tasks is now remarkable.

In concrete terms, AI is already able to identify fractures on an X-ray, detect a pulmonary embolism on a CT scan, flag an intracranial haemorrhage in an emergency setting, assist breast cancer screening and analyse prostate MRI, sometimes with performance comparable to that of expert radiologists. It can also automatically segment organs such as the liver, heart or lungs, measure tumour burden or compare successive examinations to monitor how a condition progresses.

These advances address a major challenge. Imaging volumes are growing exponentially while medical resources remain limited. Each examination produces more information, making its analysis ever more complex. In this context, AI is a valuable aid for reducing certain errors caused by inattention, harmonising practices and automating particularly time-consuming tasks.

But this technical performance, however impressive, is not enough to guarantee relevant use in real-world conditions.

2. The real challenge: integration into clinical reasoning

This is probably the main lesson of the Académie nationale de médecine's report: AI cannot be used as a standalone tool. It must fit within a demanding clinical framework, in which the physician remains at the centre of the decision.

The quality of an algorithm cannot be reduced to its accuracy rate. An AI that performs well in one facility may produce less reliable results in another if the equipment, acquisition protocols or patient populations differ from those used during its training.

Beyond the quality of the training data, an AI must also be able to integrate with the facility's information system, access prior examinations, take the clinical context into account and present results that the physician can understand. Without this integration, even an excellent algorithm risks losing much of its value.

The question is therefore no longer solely one of model accuracy, but of their ability to support medical reasoning, to be understood by their users and to fit naturally into everyday practice.

3. New risks to manage

Introducing AI into medical practice does not eliminate risks: it creates new ones.

An algorithm's performance remains directly tied to the quality of the data it was trained on. Used outside its area of competence or in a different context, it can produce errors, or even aberrant results. The most recent models are not immune to "hallucinations" or inappropriate interpretations either when they encounter unfamiliar situations.

The report also highlights a more insidious risk: automation bias. The better an algorithm appears to perform, the more the physician may be tempted to accept its output without fully exercising critical judgement. In the long term, this dependence could lead to a gradual loss of certain diagnostic skills, particularly in interpreting normal examinations or atypical situations.

On top of these issues come essential questions of cybersecurity, health data protection, integration with hospital information systems and maintaining performance over time. The Académie stresses in particular the need to put in place genuine algorithmovigilance, that is, continuous monitoring of algorithms after deployment in order to detect possible drift, analyse their errors and ensure their continuous improvement.

AI therefore cannot be regarded as a "turnkey" solution. Like any medical device, it requires a framework for use, permanent human supervision and a precise understanding of its limitations.

4. Building trustworthy medical AI

In the face of these challenges, one conviction stands out: the future of artificial intelligence in healthcare will not be decided solely by the performance of algorithms, but by the ability to build systems that are reliable, transparent and integrated into clinical practice.

The Académie nationale de médecine's report points out that several conditions are now essential: systematic validation of results by the physician, transparency about training data and model limitations, seamless integration into hospital information systems, rigorous protection of health data and ongoing training for the professionals who will use these tools.

The Académie also stresses the need to monitor the performance of algorithms after deployment. Unlike conventional software, an AI operates in a constantly changing environment: new imaging modalities, new acquisition protocols, changing populations, the emergence of new diseases or therapeutic innovations can gradually alter its performance. This continuous monitoring, known as algorithmovigilance, is becoming an essential condition for ensuring the reliability of tools over time.

Ultimately, it is no longer just a matter of developing a high-performing AI, but of building an AI capable of earning the lasting trust of healthcare professionals. It is this trust that will make it possible to move from technological demonstration to large-scale adoption.

5. Doctreen's choice: a hybrid AI rooted in medical practice

At Doctreen, this vision has guided our technological choices from the outset. We chose to develop a hybrid AI, which combines the power of generative artificial intelligence with symbolic intelligence based on decision trees built with expert physicians.

This approach reflects a simple conviction: in healthcare, an answer is not enough. It must fit within coherent medical reasoning.

In concrete terms, when a physician dictates their examination or when an image is analysed, the AI does not simply generate text automatically. It progressively qualifies the observations, answers structured medical questions, checks the consistency of the information collected and relies on reference frameworks built with domain experts to produce a report that is reliable, consistent and in line with clinical practice.

This hybrid architecture reconciles two complementary approaches: the ability of generative models to understand and structure medical language, and the robustness of explicit reasoning grounded in clinical expertise. It thus limits certain risks inherent in generative models used on their own, while always keeping the physician at the heart of the decision.

This vision aligns directly with the recommendations made by the Académie nationale de médecine. A truly useful AI cannot be a black box. It must be understood, supervised, integrated into the information system, able to evolve with medical practice and always validated by the healthcare professional.

Our ambition is therefore not to replace the physician, but to provide them with an intelligent assistant capable of reducing the documentation workload, harmonising practices and making medical output safer, so that they can devote more time to their clinical expertise and their patients.

6. Rethinking the value of AI in healthcare

Artificial intelligence opens up considerable prospects for medical imaging and, more broadly, for all specialties that produce medical documents.

The progress made in recent years already demonstrates its potential to improve diagnostic quality, automate repetitive tasks and support professionals in their everyday practice.

But the next step will not be a mere race for performance.

The real challenge now is to build AI that can integrate sustainably into healthcare organisations, support clinical reasoning without replacing it and retain the trust of the professionals who use it.

As the Académie nationale de médecine points out, the results produced by an AI will always need to be placed in their clinical context, validated by the physician and evaluated over time. This requirement is not a limit on innovation; on the contrary, it is the condition for its success.

In healthcare, then, the value of an AI is not measured solely by its algorithmic accuracy. It is built over time, through the quality of its integration, the transparency of how it works, the trust it inspires in healthcare professionals and its concrete impact on the quality of care.

The question is no longer how far artificial intelligence can go, but how to design an AI that physicians can trust, for the long term.

Académie nationale de médecine, Rapport 26-01 — Apport de l'intelligence artificielle en imagerie médicale (Contribution of artificial intelligence to medical imaging), Bulletin de l'Académie nationale de médecine, 2026.