AI in Radiology: From Hype to Practice

Radiology has repeatedly adapted to technological change, and AI follows the same path. Early tools showed promise but often failed in practice because they were too narrow, poorly integrated and difficult to justify operationally. Adoption is now accelerating as integrated platforms address concrete clinical and workflow needs. In this context, AI becomes part of radiology’s response to growing imaging demand, rising case complexity and workforce constraints.

Key Points

• Belgian adoption is strongest where AI solves urgent workflow problems.
• Triage tools are prioritising time-critical findings across multiple centres.
• Neurology AI supports volumetry and lesion tracking beyond visual review alone.
• LLM reporting tools remain limited by PACS and RIS integration gaps.
• The structural imbalance between growing imaging demand and a constrained radiologist workforce makes AI not optional, but necessary.

Not a 'New' Story

When people talk about AI in radiology, it is often considered as a recent development. In reality, the field has been evolving for decades — from early computer-aided detection systems in the 1990s to the deep learning wave of the past decade.

Radiology has always been at the forefront of computational and digital disruption. We have seen this before. Multiple times.

Consider what the speciality has absorbed in the past thirty years alone.

Film development gave way to full digital imaging. The entire workflow — acquisition, storage, distribution, reporting — was transformed. Radiologists went from reading films on lightboxes to working on digital workstations with picture archiving and communication systems (PACS). No more physical film and no more darkrooms. This resulted in faster imaging processing, an increase in imaging volumes, and easier communication of results, as well as heightened expectation of faster results. That transition alone reshaped every aspect of how a radiology department operates.

At the same time, reporting changed. Voice recognition for radiology dictation was first described as early as 1981. Early systems required hours of voice training and had suboptimal performance. By the early 2000s, automated speech recognition was standard in radiology departments across the Western world. It replaced the medical transcriptionist — which was largely the secretary’s job — who typed out dictated tapes, speeding up result communication even further. The prediction at the time was that this would result in needing fewer secretaries. That did not happen. Similar to the digitalisation of imaging itself, voice recognition software enabled faster result communication and further increases in imaging volume, requiring as many, if not more, administrative staff than before.

There is a Belgian chapter in this story worth mentioning. Lernout & Hauspie, founded in Ypres (Ieper) in 1987, became a global leader in speech technology. They also acquired PowerScribe — one of the first dedicated voice recognition platforms for radiology reporting. Unfortunately, the company collapsed in 2001 in one of the largest corporate fraud scandals in Belgian history. But the technology survived. The Dragon product line and core speech engines were acquired by ScanSoft, later rebranded as Nuance Communications — whose technology ended up in radiology departments worldwide. Nuance was itself acquired by Microsoft in 2021. A Belgian startup from a provincial city left a lasting mark on how radiologists work globally.

Then came computer-aided detection (CAD). The first CAD systems for mammography appeared as early as the late 1980s and early 1990s. In June 1998, the FDA approved the first commercial CAD system — the R2 ImageChecker M1000. These systems were widely adopted, particularly in the United States, driven by their positioning as a safety tool (“extra pair of eyes”) and supported by reimbursement. However, their clinical impact proved limited, with high false-positive rates and poor workflow integration reducing their practical value.

And teleradiology changed where radiologists could work — enabling remote reporting, cross-border consultations and 24/7 coverage models that simply did not exist before.

Each of these waves was, at the time, described as disruptive. In practice, they reshaped how radiology operates. The speciality has not stood still. It has continuously evolved and remained at the forefront of technological innovation.

Bold Predictions, Different Reality

In November 2016, Geoffrey Hinton — one of the founding figures of modern deep learning — declared that “people should stop training radiologists now,” predicting deep learning would outperform radiologists within five years. The statement made headlines.

This was not the first time such predictions were made about our speciality — but it was arguably the loudest. He has since acknowledged he was wrong on timing, clarifying he was referring only to image analysis. His revised prediction: most medical image interpretation will eventually be performed by a combination of AI and radiologist, making both more efficient and more accurate (Lohr 2025).

The numbers tell a different story from the original prediction. A 2025 systematic review in JAMA Network Open found that of all FDA-authorised AI and machine learning medical devices, 76% are radiology applications (Sivakumar et al. 2025). No other medical speciality comes close. Radiology has not been a passive target of AI development — it has been its primary proving ground.

That concentration reflects a structural reality. Imaging volumes continue to grow faster than the radiologist workforce. In the United States, imaging demand is estimated to increase by approximately 3–5% per year, with CT and MRI volumes growing even faster due to expanding indications and technological advances. In contrast, the radiologist workforce grows at only 1–2% annually, creating a structural imbalance between demand and capacity (Christensen et al. 2025; Rozenshtein et al. 2025).

A similar dynamic is observed across Europe. While precise data are less consistently reported, workforce expansion is constrained by training capacity and policy decisions, and in some countries — including Belgium — the number of training positions has been deliberately limited in recent years. At the same time, imaging studies themselves are becoming more complex. Multiparametric MRI, advanced CT techniques such as perfusion and spectral imaging, and the increasing importance of longitudinal follow-up all contribute to a higher interpretation workload per case.

This is therefore not simply a question of increasing volume, but of increasing complexity per examination. The result is a widening gap between demand and available expertise — a structural trend, rather than a temporary fluctuation.

AI is not arriving to replace radiologists; rather, it is an essential technological advancement that assists in addressing the growing imbalance between the demand for imaging and the available resources to interpret them.

Why the First Wave Disappointed

From around 2015 onwards, deep learning-based AI tools arrived in radiology in force. Early applications focused on well-defined detection tasks — lung nodule detection, chest X-ray abnormality classification, intracranial haemorrhage detection and fracture identification — areas where large, annotated datasets were available and visual patterns were relatively consistent. The technology was genuinely promising.

In our practice, we evaluated a chest CT AI tool for detecting lung parenchymal abnormalities such as consolidation and interstitial disease patterns. While the diagnostic performance was acceptable, the incremental diagnostic value was limited. This came at the cost of poor workflow integration: every flagged finding required additional clicks, review step and time. In a busy radiology department (which today is effectively all of them), even one additional minute per study is not a minor inconvenience: it is operationally unworkable. The tool was not worth its financial or time cost. We were not alone in reaching that conclusion.

The same story played out across many early implementations. Tools were either not accurate enough, too narrow in scope or — most commonly — poorly integrated into the existing digital infrastructure of a radiology department: the radiology information system (RIS) that manages scheduling and reporting, and the PACS that stores and distributes images.

In addition, performance reported in published studies often did not translate into real-world practice, where variability in data quality, workflows and patient populations is significantly higher than in controlled study environments.

The technology worked in isolation. It did not work in practice.

From Single Tools to Platforms

There was a second structural problem. Most early AI vendors sold one solution for one problem: a nodule detector, a bone age tool, a stroke triage algorithm. It was comparable to buying a separate app for every single task on your phone — each requiring its own installation, subscription and maintenance. Buying, validating and maintaining multiple single-purpose tools from different vendors was complex and expensive, and the return on investment for any individual tool was difficult to justify.

This has changed in recent years and has contributed to the acceleration in adoption. Most established AI vendors now offer multi-application platforms — suites of tools covering multiple modalities and indications within a single integrated environment. A platform that handles triage across chest CT, musculoskeletal imaging and neuroradiology simultaneously is a fundamentally different proposition. It finally begins to make the cost–benefit equation work.

Real-World Adoption: The Belgian Experience

Belgium has a strong track record in adopting new radiology technologies. Over the past two years, there has been a marked acceleration in AI adoption, with widespread testing and increasing clinical use.

The tools gaining the most traction follow a clear logic: they solve an urgent operational problem without disrupting the workflow.

Triage and critical finding detection has been the most widely adopted category. CADt (computer-aided detection for triage) applications automatically detect and flag time-critical findings — intracranial haemorrhage, pulmonary embolism, major fractures — and prioritise them in the reporting queue. The clinical rationale is simple: a patient with a brain bleed should not wait behind routine studies. These tools are now in active use across multiple Belgian centres.

Stroke care has seen particularly integrated AI deployment. Tools for automated clot detection and brain perfusion CT analysis are now embedded in stroke workflows, alongside communication platforms that enable rapid sharing of findings between radiologists, interventional radiologists and neurologists — all within a single workflow. In a condition where minutes determine outcomes, this kind of end-to-end integration has genuine clinical value.

Plain radiograph AI has also gained strong uptake. Fracture detection across the appendicular and axial skeleton is now a clinical reality. Automated orthopaedic measurement tools — scoliosis angles, lower limb alignment, foot posture assessment — address a specific pain point: these measurements are time-consuming when done manually, prone to variability, and now done automatically and consistently.

CT lung nodule detection and follow-up tools are increasingly adopted, particularly as lung cancer screening expands across Europe. Automated detection, characterisation and longitudinal tracking address both a clinical need and a capacity problem simultaneously.

Neurological follow-up represents one of the more compelling use cases. Automated brain volumetry for dementia follow-up, as well as automated MS lesion load quantification on MRI, are areas where AI genuinely does something a radiologist cannot do reliably by hand. Counting and comparing scattered MS lesions across multiple sequences and timepoints is a difficult task visually. Subtle volume changes in brain atrophy are below the threshold of what the human eye can detect on routine review. AI makes both processes measurable, reproducible and clinically actionable — directly informing therapeutic decisions.

Vendor-integrated platforms from major players — Siemens, GE, Canon — are another important channel. These companies now offer AI suites via their own platforms, combining proprietary and third-party tools under a single licence model. Lung nodule analysis and prostate MRI assessment using the PI-RADS scoring system (a standardised method for classifying the likelihood of clinically significant prostate cancer) are among the more widely used applications delivered through this route.

LLM-based report structuring and correction tools — where large language models (LLMs) automatically structure and refine radiology reports within the digital workflow — exist and are technically mature. Their clinical uptake in Belgium remains limited for now, primarily because most PACS and RIS providers have not yet built the necessary integrations. This represents a key bottleneck that will likely resolve over time. Once integrated, these tools are expected to deliver significant productivity gains and facilitate the broader adoption of structured reporting. When that happens, they may have a major impact on reporting workflows, similar to the impact of voice recognition.

What the currently adopted tools share is a practical cost-benefit profile that finally makes sense. They save time, reduce the risk of missed critical findings, or enable measurements that were previously impractical. They integrate into existing workflows without requiring radiologists to fundamentally change how they work.

Reimbursement remains an open question in Belgium. Unlike some neighbouring countries, there is currently no structural form of government reimbursement for AI tools in radiology, which likely contributes to slower or more limited adoption. Instead, adoption is driven by departmental or institutional decisions, weighing vendor pricing against demonstrated benefit. This creates variation: larger academic centres and private group practices tend to be further ahead, while smaller departments face higher barriers to entry. This will need to evolve if Belgium is to capture the full potential of these technologies at scale.

What Is Coming Next: Foundation Models and Beyond

The tools described above are, by and large, narrow AI — each designed for a specific task on a specific modality. They work well precisely because they are focused. But the next generation of AI in radiology is moving in a different direction.

Foundation models — large-scale AI systems trained on vast and diverse datasets — are beginning to enter the medical imaging space. Unlike narrow tools, these models can generalise across tasks, modalities and clinical contexts. If current narrow AI tools are like specialised instruments, each designed for one task, foundation models are closer to a generalist consultant — capable of integrating multiple sources of information and reasoning across them. A single foundation model may be capable of analysing a chest radiograph, generating a structured report, flagging incidental findings and integrating clinical context from the patient record. All within one system.

Early examples are emerging. A recent model developed at the University of Michigan illustrates this shift clearly: acting as a “co-pilot,” it integrates multiple MRI sequences with clinical context to generate a targeted diagnosis or differential diagnosis across 52 neurological conditions. It also performs automated urgency triage and generates referral recommendations — for example, directing cases to a stroke neurologist versus a neurosurgeon. Trained on over 220,000 MRI studies and validated prospectively on nearly 30,000 studies, it demonstrates how AI is moving beyond single-task detection towards integrated clinical reasoning (Lyu et al. 2026).

This also reflects a broader pattern in the current AI landscape: systems are developed and validated in research settings but increasingly move rapidly towards clinical deployment. Several vendors are already translating these concepts into early clinical products, with first-generation “co-pilot” functionalities and integrated AI workflows being introduced into routine practice following regulatory approval. While still evolving, this rapid transition from research to real-world use is a defining characteristic of the current wave of AI in radiology.

For radiology, the implications are significant. Automated report generation — not just structuring or correcting existing reports but generating a first draft from image analysis. This is a near-term reality for routine studies. Multimodal AI that integrates imaging findings with laboratory results, clinical history and prior reports will move the speciality closer to a decision-support model rather than a pure image-reading model.

None of this is without challenge. Validation requirements are substantial. Regulatory pathways for adaptive and generalised AI systems are still evolving. Liability frameworks have not kept pace. And the risk of over-reliance — radiologists rubber-stamping AI output without genuine oversight. This risk is real and needs to be managed proactively.

But the direction is clear. The question is no longer whether AI will be a structural part of radiology practice. It already is. The question is how fast the transition to more generalised, integrated systems will occur.

An Evolving Profession

Radiology has never stood still. From film to digital, from typed reports to voice recognition, from isolated reading rooms to integrated multidisciplinary teams. The profession has continuously redefined what a radiologist does and how they do it.

AI is the next step in that evolution.

The context matters here. As stated earlier, imaging volumes continue to grow significantly faster than the radiologist workforce — a structural imbalance that is both global and local. Projections suggest the US alone could face a shortage of tens of thousands of radiologists by 2033 (AAMC 2021). In Belgium, the situation is compounded by a deliberate policy decision to reduce the number of radiology training positions in recent years, resulting in a stagnating and in some regions declining radiologist workforce — at precisely the moment when demand is accelerating.

A recent Stanford University analysis projected that AI could reduce the hours spent on routine radiological tasks by 14 to 49% over a five-year horizon, primarily through automated report drafting and study delegation for standard examinations (Langlotz et al. 2025). Furthermore, detection tasks that algorithms handle reliably will increasingly be automated. That frees capacity for what requires genuine clinical expertise: complex and less common diagnoses, integration of imaging findings with clinical context, multidisciplinary consultation and interventional work.

The workflow will evolve accordingly. AI will propose findings, triage cases and draft initial reports. Radiologists will validate, correct and contextualise. This enables higher throughput while preserving, and in fact expanding, the time for complex, data-dense examinations that genuinely require expert judgement.

The advisory and connecting role of the radiologist also becomes more important in this model. Bridging imaging findings with clinical specialities and primary care, guiding appropriate use of imaging and taking ownership of AI governance and quality oversight. These are not peripheral responsibilities. They are central to what radiology will look like going forward.

Radiology has absorbed every previous wave of technological change and emerged with a more capable, more integrated practice.

Conclusion

Radiology has always evolved alongside technological innovation. Early AI demonstrated clear technical potential but failed to deliver meaningful value in practice, largely due to limited integration, narrow scope and unfavourable cost–benefit profiles.

What has changed in recent years is not the underlying technology alone, but its implementation. AI is moving from isolated, single-purpose tools to integrated platforms that fit within existing workflows and solve concrete operational problems. This shift explains the current acceleration in adoption.

At the same time, imaging volumes continue to grow, studies are becoming more complex, and workforce expansion remains limited. In this context, AI is not a standalone innovation, but part of an ongoing evolution in how radiology manages increasing demand and complexity.

As with previous technological transitions, the impact will ultimately be defined not by the tools themselves, but by how effectively they are integrated into clinical practice.

Conflict of Interest

None.