Artificial intelligence is no longer a promise for the future: it is already the present. In medical imaging departments, it has taken root quietly but durably, to the point that the question is no longer whether AI will transform radiology, but how that transformation is actually unfolding in practice. Does it occupy the place and the importance that was imagined? Has the initial enthusiasm been confirmed in the face of clinical reality?
Far from science-fiction scenarios, the reality of AI in imaging is more nuanced. Promising faster and more efficient workflows, AI can be applied to tasks ranging from administrative work to technical tasks, from medical secretaries to radiologists. But the conditions of its real impact are not what one might expect.
As the field of medical imaging sees ever more innovative technologies emerge, AI is opening new avenues for optimizing and refining the diagnostic process. But understanding what it truly is, and what it does not yet do, is the prerequisite for any informed use.
AI in Imaging: A Generic Term That Conceals Great Diversity
Artificial intelligence is a broad term encompassing several families of techniques. The first is machine learning: mathematical models draw relevant information from a training dataset to learn how to perform a task. Once trained, the model can be deployed to process new data.
Its subcategory, deep learning, is based on artificial neural network systems with multiple hidden layers of neurons. This type of learning involves a very large number of parameters and consequently requires a much greater volume of training data. It is the technology behind most of the medical image analysis tools currently available.
Finally, natural language processing (NLP) is a multidisciplinary field combining linguistics, computer science, and AI. Its objective is to create tools capable of interpreting and synthesizing text: automatic structuring of radiology reports, coding assistance, and extraction of relevant clinical information from patient records.
In medical imaging, these techniques translate into growing automation of previously manual tasks: exam preparation, acquisition protocol selection, image analysis and post-processing, and equipment maintenance. The question is no longer “is it possible?” but “which tasks, under what conditions, with what safeguards?” Informed use of these tools, grounded in sufficient knowledge of how they work and what their limitations are, is a prerequisite for realizing the true benefit of automation.
From Dream to Reality: AI Is Not Here to Replace the Radiologist
Real-world evidence reframes the most sensationalist claims. If AI is not being presented as a substitute for the radiologist, it is precisely because performance data measured under real conditions confirm this: the radiologist combined with AI systematically outperforms either one taken alone. This model of “augmentation” rather than replacement is observed across studies in mammography, fracture detection in emergency settings, stroke, oncology, and cardiac MRI.
Clinical reality also places adoption, ergonomics, and integration into existing systems ahead of algorithmic precision. A technically high-performing tool that is poorly integrated into the workflow will be less useful than a slightly less precise tool perfectly adapted to daily practice. This is one of the major lessons from real-world deployments.
That said, very real risks accompany this rise in power. The first is the risk of skill erosion: excessive reliance on AI, particularly in the training of future radiologists, could undermine fundamental competencies in image analysis and interpretation. AI must always be regarded as a clinical decision support system: it is the healthcare professional who remains responsible for the diagnosis.
The second risk relates to algorithmic opacity. Most state-of-the-art systems rely on deep learning, whose internal mechanisms are not accessible to users. These “black-box algorithms” make decisions without explaining their reasoning, raising legitimate questions of oversight, accountability, and trust. Radiologists, clinicians, and other imaging professionals must familiarize themselves with these technologies and actively contribute to their ethical development and responsible implementation.
Chest Radiography: A Model Case for AI-Assisted Double Reading
Chest radiographs represent a privileged testing ground for evaluating the contribution of AI in clinical practice. They are among the most frequently performed examinations in the world and account for a significant portion of the daily workload of many radiologists. Yet errors are possible: pulmonary nodules, pneumonias, and pneumothoraces can be missed, particularly in high-volume settings or when fatigue is a factor.
Double reading, performed by two radiologists, is an established method for reducing these errors. But it remains costly in terms of time and human resources. This is precisely where an AI-assisted double reading system finds its value.
The optimal application of this technology lies after the radiologist has validated the report, in asynchronous post-reading mode. The AI assistant then functions as an additional “safety layer,” without disrupting the workflow. When a discrepancy is identified, a rapid review by a second reader can be triggered, enabling the radiologist to be notified in a timely manner. This AI-assisted double reading model does not seek to correct the diagnosis but to reduce the risk of omissions on key findings.
Radiologists supported by these tools have demonstrated improved detection performance across a wide range of lesions.
Artificial intelligence in medical imaging is not a mirage: it delivers measurable time savings, strengthens diagnostic confidence, and provides a presence “after the report” without adding to the daily workload. Its added value is greatest when it operates in a clearly defined support role: detection assistance, alerting for omissions, and help with prioritizing urgent cases.
But this value is only fully realized with trained professionals who are capable of exercising critical judgment over algorithmic outputs and of understanding the situations where AI may fall short. Far from weakening radiology, a reasoned use of AI can strengthen its practice: by freeing professionals from repetitive, low-value tasks, it allows them to focus on what the algorithm does not yet know how to do, namely integrated clinical reasoning, patient communication, and expertise on complex cases.
What Comes Next: The Era of Generative AI
The next step in this transformation promises to run even deeper. Generative artificial intelligence, the kind that creates content (text, images, summaries) rather than merely classifying, is beginning to enter clinical environments. Large language models applied to radiology reports, systems capable of automatically generating a structured synthesis from imaging data, and tools to assist non-specialist clinicians with preliminary analysis in emergency settings are among the many potential applications.
But this new wave raises specific challenges. By virtue of their probabilistic nature, these systems can generate inaccurate results that nonetheless appear plausible, a phenomenon known as algorithmic “hallucinations.” In a clinical context, this risk is particularly critical. Practitioners are logically calling for greater transparency regarding the mechanisms of these tools: the gradual move away from black-box algorithms toward explainable and traceable systems is one of the conditions for clinical acceptability.
For radiologists, software vendors, and regulatory bodies, the challenge of the coming years will be to define the validation, supervision, and accountability frameworks appropriate for these new generative tools. The reflection underway on AI governance in healthcare is still in its early stages, but it will lay the foundations for an integration that is both ambitious and responsible. The radiologist of tomorrow will be neither replaced by AI nor independent of it: they will be the one who knows how to get the best from it, while keeping the upper hand.
Sources
Potočnik J, Foley S, Thomas E. Current and potential applications of artificial intelligence in medical imaging practice: A narrative review. J Med Imaging Radiat Sci. 2023 Jun;54(2):376-385. doi: 10.1016/j.jmir.2023.03.033. PMID: 37062603. https://pubmed.ncbi.nlm.nih.gov/37062603/
Topff L, Steltenpool S, Ranschaert ER, Ramanauskas N, Menezes R, Visser JJ, Beets-Tan RGH, Hartkamp NS. Artificial intelligence-assisted double reading of chest radiographs to detect clinically relevant missed findings: a two-centre evaluation. Eur Radiol. 2024 Sep;34(9):5876-5885. doi: 10.1007/s00330-024-10676-w. PMID: 38466390. https://pubmed.ncbi.nlm.nih.gov/38466390/
ECR 2026. https://connect.myesr.org/ecr-2026/programme/?tabref=0
