Playtime is over: radiology AI in 2026

The radiology AI market has gone through significant development in recent years. Where once startups with promising demos and marketing grabbed the spotlight, we now see a more mature market emerging. Still, there are also new players, new dynamics, and familiar challenges. What have we seen in 2025 and what does it tell us about what’s to come in 2026?

A changing field of providers

The number of CE-marked AI solutions continues to grow slowly, but the number of active companies in the market is actually declining¹. This is mainly due to consolidation: acquisitions set the tone. For example, DeepHealth recently acquired both Kheiron and iCAD after snapping up Aidence and Quantib in 2022; Gleamer absorbed French company Pixyl, and South Korea’s Lunit bought Volpara. At the same time, players such as Behold.ai were forced to shut down operations.

The lesson: just having an approved product is not enough; scalability, a strong brand, and financial strength are indispensable to survive as an independent player.

At the same time, new providers are also entering the scene. For example, hardware manufacturer United Imaging (China) has been launching 27, yes you are reading it right, 27 CE-marked solutions. Turkish company hevi.ai recently released seven clinically deployable applications, of which four together form the hStroke suite.

Today, half of the available AI solutions is provided by only 15% of the companies with the largest portfolio’s.

Dark blue: number of new AI solutions brought to market in Europe (CE-marked) per year. Light blue: number of published papers per year related to Health AI Register’s product pages.

So, yes, the number of solutions is still growing, but they increasingly come from fewer, larger, and better-positioned players.

Growing interest in AI platforms

The era of “point solutions” seems to be over. Instead of stand-alone tools for a single task, hospitals are increasingly opting for an AI platform. By now, there are more than twenty providers of such AI platforms, coming from different kinds of companies: PACS vendors (e.g., Sectra and Agfa), independent platform companies (e.g., deepc, Incepto), AI vendors that also offer third-party AI-products (e.g., Aidoc, Gleamer), and OEMs (e.g., Siemens, Philips).

Until recently, several large industry players - Bayer, GE, and Philips - had chosen to rely on Blackford as their infrastructure partner rather than further developing their own platform technology, resulting in a relatively similar technical backbone across offerings. However, in September 2025, Bayer announced it would discontinue its Calantic Digital Solutions AI platform and wind down Blackford services, just five years after entering the radiology AI platform space. Seen in the context of a crowded and highly competitive platform landscape, this move may mark an early signal of market correction.

At the same time, the move toward platform-first strategies is reinforced by regional or even national initiatives and by large-scale tenders in countries like the Netherlands, Norway, Germany, France, and Poland.

Together, these developments show that AI in radiology is no longer about isolated applications but increasingly forms a structural part of the hospital’s digital setup. For healthcare organizations, the choice of a platform partner therefore becomes a fundamental decision: it largely determines which applications are available, how well they integrate, and how flexible the hospital can remain in the future - particularly in a market where even well-resourced players may reconsider their commitment. The strategic implications of choosing an AI platform partner are further explored in a recent interview by Kicky van Leeuwen with radiologist Inez Verpalen, reflecting on the experiences of two hospitals.

The ten platform providers with the largest number of AI applications offered.

Comprehensive AI and foundation models

The technology behind AI continues to evolve, along with ambition. The need is shifting from support for a single finding to complete image analysis. That calls for broader, more versatile systems. Some companies are actively addressing this: Annalise.ai explicitly positions itself as a provider of “comprehensive AI,” with more than a hundred findings in a single product. Newcomer a2z Radiology AI is also moving in this direction, although they don’t yet have CE-marked products.

One technological development supporting this trend is the rise of foundation models. These are base models trained on enormous amounts of data and can then be applied to multiple tasks. A concrete example: Aidoc received FDA clearance for a triage tool for 14 indications based on its own CARE foundation model. This approach makes AI solutions potentially more flexible and scalable. At the same time, it requires new ways of working, e.g. moving away from AI results in secondary captures to more interactive widgets. And how do you validate locally that the application performs well enough across all findings?

Moreover, foundation models clash with current regulations (especially in the U.S.), which are primarily designed to assess narrowly defined models with a single, clear task. Foundation models, with their broad applicability and adaptive nature, call for new evaluation frameworks to ensure their safe and responsible clinical adoption.

From image to report with AI

More and more AI companies realize that they need to deliver a workflow solution to make an impact and build a sustainable business case. An AI result as a secondary capture is often not enough. As a result, there is a clear trend toward optimizing reporting.

Some AI products can already draft a preliminary report for a given case. Companies like Jacobian (previously SmartReporting and Fluency for Imaging), Rad AI, and Radpair specialize in reporting and leverage LLM technology to accelerate the reporting process for all studies while directly incorporating third-party AI results. AI platforms themselves are increasingly offering such functionality as well.

This is a positive development, since improving reporting improves the entire workflow, not just the diagnosis of a single condition, but the way every radiologist works. However, it requires a significant change for radiologists: who dares to take the leap?

Tasks being addressed by the 291 radiology AI solutions. One solution can address multiple tasks.

Science still focuses on performance instead of impact

As can be seen in the first figure, 2025 was a record year with more than 170 published works related to the commercial AI solutions. However, recent research² from the Radboud University Medical Center shows that still, the majority of studies are focused on technical performance (such as sensitivity and specificity), not on clinical impact or cost-benefit analysis - yet the latter is precisely what’s needed to push adoption forward and justify investments.

Without solid clinical evidence, it often remains difficult to convince policymakers and executives to commit to large-scale implementation as a return of investment is not guaranteed for many diagnostic applications. That is why the focus increasingly seems to be shifting from clinical AI applications to workflow-supporting solutions that deliver direct efficiency gains.

A recent study from Prof. Curt Langlotz³ illustrates the potential impact on the general workload of a radiology department. Rather than evaluating diagnostic accuracy, the study attempts to model the potential effect of AI on radiologists’ day-to-day workload. The analysis concludes: “The model projects a 33% reduction in hours worked by radiologists in 5 years, with a range of 14% to 49%. The main effects are due to radiology report drafting for all modalities and study delegation for radiography and mammography.”

In other words, the potential impact of AI in radiology is evident, but robust, real-world evidence proving this impact at scale is still largely missing.

Levels of efficacy evaluated in peer-reviewed papers included in the 2020 and 2023 analyses. The values next to each bar represent the percentage of papers at each efficacy level, relative to the total number of papers analysed in that year (2020: n = 237; 2023: n = 421). A single paper could address multiple levels of efficacy; therefore, the percentages do not add up to 100% Source: Antonissen, N., [..], van Leeuwen, K. et al. Artificial intelligence in radiology: 173 commercially available products and their scientific evidence. Eur Radiol 36, 526–536 (2026). https://doi.org/10.1007/s00330-025-11830-8

And yet… what are we waiting for?

Radiology AI has become significantly more mature: platforms are scaling up, clinical applications are becoming more diverse, and ambitions are growing. But despite this progress, true transformation remains for many centres just out of reach. Why is that?

Orchestration remains complex. Most AI solutions still require customized integrations, workflow adjustments, and ongoing support. We are not yet at the point where implementation is truly plug-and-play. This limits scalability and presents a barrier for many hospitals.

The financial picture is also unclear. Business cases are hard to substantiate, and ROI remains uncertain. Unless AI can demonstrate that it not only works, but also works better, faster, and cheaper than current practice, it will replace nothing and mainly add costs and complexity. The big question remains: will we ever reach a point where AI is trusted and accepted enough for radiologists to truly let go of certain tasks?

The long-awaited leap to multimodal AI - where imaging is combined with clinical data and other sources into one intelligent system - remains largely theoretical for now. Yet several companies are actively working to bridge the gap between promise and practice.

With a continued focus on seamless integration and clearer value propositions, the puzzle pieces are gradually falling into place. The AI hype is over, playtime is over, now the real work begins!

By Kicky van Leeuwen

¹ https://healthairegister.com/radiology
² Antonissen, N., Tryfonos, O., Houben, I.B. et al. Artificial intelligence in radiology: 173 commercially available products and their scientific evidence. Eur Radiol (2025). https://doi.org/10.1007/s00330-025-11830-8

³ Langlotz, C.P., The Effect of AI on the Radiologist Workforce: A Task-Based Analysis. medRxiv (2026). https://doi.org/10.64898/2025.12.20.25342714

A variation of this article named 'Radiologie-AI anno 2025: Playtime is over' was previously published in Dutch in the Memorad.