Artificial intelligence in radiology: what it can do today and where its limits lie

The FDA has approved its thousandth AI tool for radiology, and the Swedish MASAI study demonstrated a 29% increase in cancer detection in mammographic screening with AI while halving radiologists' workload. The article distinguishes two approaches — specialized detectors trained on a single pathology, and general multimodal models capable of working with clinical context — and shows the limits and prospects of both in Czech clinical practice.

Late in 2025, the U.S. Food and Drug Administration (FDA) approved the thousandth tool using artificial intelligence for radiology. The Swedish randomized MASAI trial demonstrated that artificial intelligence in mammographic screening increases cancer detection by 29% while at the same time cutting the radiologists' reading workload nearly in half. In Lithuania, an autonomous system from the company Oxipit generates final reports for normal chest X-rays entirely without a physician's involvement. In the Czech Republic, the startup Carebot is testing its system in hospitals of the AGEL network.

At the same time, many radiologists describe frustration with systems that generate too many false alarms and that "see" pathologies where none exist. This frustration is often justified — but the problem usually does not lie in artificial intelligence as such, but in the fact that a particular workplace is using an outdated system corresponding to the technological level of 2018.

This article explains how artificial intelligence in radiology works, what the differences between the various approaches are, what systems can actually do today and where their limits remain — without the technical jargon of computer science, in the language of clinical practice.

In radiological practice today there are two fundamentally different approaches. Understanding them is key to enabling a radiologist to judge what they can expect from a particular system.

Most of the tools clinically deployed in radiology today belong to this category. These are systems designed and trained for one specific task: to find a fracture on an X-ray, to detect a pulmonary embolism on CT angiography, to measure the volume of a tumor on magnetic resonance imaging.

The principle of operation can be imagined through an analogy with a radiological atlas. Imagine you showed a resident two hundred thousand chest X-rays and for each one said: "here is a pneumothorax" or "there is no pneumothorax here." Over time, the resident would begin to recognize the pattern — an area without lung markings separated by a visceral pleural line. A neural network does the same thing, except that instead of visual perception it uses mathematical operations on image pixels.

The network processes the image gradually, layer by layer. The lowest layers detect simple patterns — edges, brightness transitions, textures. The middle layers assemble these patterns into more complex structures — the outline of a rib, the contour of the cardiac silhouette, the texture of the lung parenchyma. The highest layers recognize entire pathological patterns and their relationships. The output is a number — the probability that a given pathology is present on the image (for example: "pneumothorax: 94%"), possibly a segmentation mask (the precise outline of the lesion) or a localization box.

A key feature of these systems: they have no clinical context whatsoever. They do not know the patient's age, medical history, or laboratory results. They do not know why the examination was indicated. They cannot formulate a differential diagnosis or a recommendation. They are extremely precise pattern recognizers — nothing more, nothing less.

This approach is represented by systems such as Med-Gemini (Google), RadFM or CheXagent (Stanford). Unlike specialized detectors, these are models that process image and text simultaneously and generate free-text outputs — entire radiology reports, answers to clinical questions, differential diagnoses.

The principle of operation is fundamentally different. These models learned from hundreds of millions of radiology reports, textbooks, scientific articles and image–caption pairs. They were not trained on a single task, but on a general understanding of the relationships between medical images and clinical text.

When such a model is given a chest X-ray and the question "Describe the findings and propose further management," it does not give just a number. It produces coherent clinical text:

"On the PA chest radiograph there is a homogeneous opacity in the left costophrenic angle with a concave upper border, corresponding to a moderate-sized left-sided pleural effusion. The cardiac silhouette is within normal limits. Recommendation: correlation with the clinical condition; if exudate is suspected, consider diagnostic thoracentesis under ultrasound guidance."

This is not merely the chaining of probable words. From training on an enormous amount of data, structures form within the model's internal representations that encode the relationships between pathologies, anatomy and clinical context — no one explicitly inserted them, but the model had to learn them in order to be able to generate correct outputs. The professional community refers to this phenomenon as emergence — from a seemingly simple training process, abilities arise that the process did not directly prescribe.

A study published in Nature Medicine showed that reports generated by one of these models (Flamingo-CXR) were rated by board-certified radiologists, for 77.7% of outpatient chest X-rays, as comparable to or better than reports from human colleagues.

A specialized detector (a convolutional neural network) detects a specific pathology. The output is a number (probability) and a localization. It handles one to a few tasks. It does not integrate clinical context. There are hundreds of approved products on the market (both FDA and CE). The typical error is a false alarm, which is easily measurable. The processing speed is on the order of milliseconds.

A large multimodal model interprets the image in context. The output is free text — a report, a recommendation, a differential diagnosis. It handles dozens to hundreds of tasks. It integrates medical history and laboratory findings. As of February 2026 it has almost no regulatory approvals. The typical error is a hallucination — factually incorrect but convincingly sounding text, difficult to recognize. The processing speed is on the order of seconds.

In the clinical practice of 2026, only specialized detectors are realistically deployable — they have regulatory approval and proven safety. Multimodal models are the future, but for now they are predominantly in the research and pilot-project phase.

Understanding the training process helps explain why some systems work well and others do not.

The process resembles teaching a resident under strict supervision, but where the mentor never explains why, only shows what.

Data preparation. A radiologist goes through thousands to hundreds of thousands of images and labels the diagnosis for each one. For detection tasks they additionally enclose the pathology in a box, and for segmentation tasks they trace the precise outline of the lesion. This is the most expensive and most time-consuming phase — annotating a single prostate CT examination takes about thirty minutes.

Learning. The network is given an image, makes a prediction, and compares it with the correct answer from the radiologist. Based on the difference between the prediction and reality, the network's internal parameters are adjusted so that it will err less next time. This process is repeated a million times across the entire dataset.

Validation. The trained model is tested on images it has never seen before, ideally from other hospitals and devices. Only here does it become clear whether the model has learned to recognize genuine pathology, or merely the specific artifacts of one workplace.

The key lesson: the quality of a system depends on the quality and breadth of the training data. A model trained on ten thousand images from one hospital will be markedly worse than a model trained on a million images from a hundred centers.

The process is fundamentally different. These models do not learn from manually annotated data, but from an enormous amount of existing image–report pairs — millions of radiology reports already written by human radiologists, textbooks and scientific articles.

The model learns in two ways simultaneously. First, from the text itself: it is given the beginning of a radiology report and learns to predict what will follow. Second, from the linking of image and text: it learns that a particular visual pattern on an image corresponds to a particular description in the report. After training on billions of such examples, the model internalizes the relationships between visual patterns, anatomy, pathology and clinical context.

A fundamental difference from specialized detectors: the model does not need manual annotations from a radiologist (these are expensive and slow), but uses existing reports written in the course of routine clinical practice. The price for this is an enormous computational demand — training takes weeks to months on hundreds of specialized processors and costs millions of dollars.

The chest X-ray is the most thoroughly researched area. In Europe there are more than 170 commercial products with CE certification for this modality.

Clinically validated capabilities include the detection of pneumothorax, pleural effusion, consolidations, lung nodules, cardiomegaly, rib fractures and other pathologies. The most versatile system, Annalise.ai, detects 124 different findings on a single chest X-ray. A study on twenty thousand cases demonstrated that AI assistance increased radiologists' sensitivity by 26 percentage points for pneumothorax, by 14 for consolidations and by 12 for lung nodules. The average reading time fell by a third at the same time.

For the detection of fractures on skeletal X-rays, the Gleamer BoneView system achieves an accuracy of 97% with FDA approval for both adult and pediatric patients older than two years.

Pulmonary embolism. Aidoc achieves a sensitivity of 93% and specificity of 96% on CT angiography. Avicenna.AI reports an area under the curve (AUC) of 0.92 in multicenter validation.

Stroke. Viz.ai (the first product approved by the FDA specifically for the detection of large vessel occlusion of the brain) shows a sensitivity of 96% and specificity of 94% according to the manufacturer's data. In real-world operation the sensitivity is around 91% and the specificity 88%. RapidAI, deployed in more than 1,800 hospitals, achieves a sensitivity of 90–96% for occlusions of the M1 segment and the internal carotid artery.

Lung nodules. The sensitivity of artificial intelligence is 86–98% (higher than that of radiologists: 68–76%), and the specificity is 77–87% (lower than that of radiologists: 87–92%). The greatest added value is therefore the detection of nodules that a radiologist would miss — at the cost of a higher number of false-positive findings.

Prostate cancer. The PI-CAI study published in Lancet Oncology (10,207 examinations, 62 radiologists from 45 centers in 20 countries) demonstrated that the artificial intelligence system outperforms radiologists in the detection of clinically significant cancer: AUC 0.91 versus 0.86. The system detected 6.8% more clinically significant tumors at the same specificity while at the same time reducing the number of false-positive findings by 50%.

Brain tumors. The best segmentation systems at the BraTS 2025 competition achieve a Dice coefficient of 93% for the whole tumor — the overlap between the predicted and the actual segmentation is thus almost perfect.

Cardiac magnetic resonance imaging. Artificial intelligence automates the measurement of ejection fraction, the analysis of myocardial strain and the segmentation of cardiac structures.

The MASAI study represents the most robust evidence for the benefit of artificial intelligence in radiology. It is a randomized controlled trial involving more than 105,000 women in the Swedish national screening program, whose results were published successively in Lancet Oncology (2023), Lancet Digital Health (2025) and The Lancet (2026).

Key results: AI assistance led to a 29% increase in cancer detection without a significant increase in false-positive results. At the same time, the number of required readings fell by 44%. The full results of the study, published in January 2026, additionally demonstrated that the number of interval cancers (tumors diagnosed between two screening rounds) fell by 12%, while aggressive interval tumors decreased by 27%.

Simply put: artificial intelligence in mammography makes it possible to find more tumors with less work and without unnecessarily alarming healthy women.

No system replaces the radiologist's clinical judgment. Artificial intelligence does not see the patient, does not know their story, cannot ask follow-up questions. It cannot assess whether a finding is clinically relevant in the context of the patient's overall condition. It cannot take responsibility for the diagnosis.

Data distribution shift — the most common cause of failure. A model trained on data from one hospital or one type of device may fail markedly at another workplace. Studies have demonstrated that 81% of the 86 algorithms tested show reduced accuracy on external datasets, with roughly a quarter experiencing a drop of 10 or more percentage points.

Population bias. Models systematically underdiagnose Black patients on chest X-rays and generally show lower accuracy for populations that were underrepresented in the training data. Algorithms developed for adults do not work reliably in children.

Rare and atypical presentations. Artificial intelligence excels at typical findings but fails on rare presentations that it saw little or not at all in the training data. An experienced radiologist still clearly wins in this field.

One of the most disturbing research findings: incorrect AI suggestions worsen the performance of all radiologists, including highly experienced ones.

A study published in Radiology (2023) tested 27 radiologists on 50 mammograms. When the system suggested an incorrect BI-RADS assessment, the performance of all radiologists worsened without exception — even those with twenty years of experience. An analysis of nearly three thousand decisions by 92 radiologists confirmed that physicians tend to follow AI suggestions regardless of their correctness.

This is the paradox of artificial intelligence in radiology: a system that is supposed to help the radiologist can, under certain circumstances, worsen their performance if the radiologist loses critical distance. This leads to a practical recommendation: first form your own finding, and only then look at the system's output.

General models such as GPT-4V show, in radiological diagnosis, an accuracy of merely 8–37% — markedly below the level of an average radiologist. Specialized models (Med-Gemini, CheXagent) are substantially better, but still do not reach the level needed for autonomous diagnosis.

A fundamental problem is hallucination — the model can produce a clinically plausible but factually incorrect description. Unlike the false alarm of a specialized detector (which a radiologist easily identifies, because it is a box around a normal structure), a hallucination is wrapped in convincingly sounding expert text. This is precisely why it is crucial that every output undergoes critical assessment by a radiologist.

As of February 2026, no multimodal language model has regulatory approval for autonomous radiological diagnosis.

Not all systems are equal. Technological progress in this field is so rapid that a system five years old corresponds to an entirely different technological level than the current cutting edge. For a radiologist who judges the benefit of artificial intelligence on the basis of their own experience, it is essential to know which generation they are using.

These systems arose as the first clinical applications of deep learning. They were typically a simple classifier trained on a single dataset from a single hospital (usually fewer than 50,000 images) for the detection of one to three pathologies. The model was not updated after its deployment.

Typical accuracy: sensitivity 70–85%, specificity 60–80%. When the device or workplace changes, performance drops dramatically. There are so many false alarms that they lead to "alarm fatigue" — the radiologist begins to ignore them, by which the system loses its point. Explainability is zero — the system cannot show why it decided the way it decided.

If your experience corresponds to this description, you are with high probability using a first-generation system.

The current standard of clinically deployed systems. Representatives: Aidoc, Viz.ai, Lunit, Annalise.ai, Gleamer, Qure.ai. Trained on hundreds of thousands to millions of images from many centers. Multi-pathology detection. FDA or CE MDR certification on the basis of clinical studies. Visual explanation (heatmaps, bounding boxes — the radiologist sees where the model is "looking"). Continuous model updates in cloud solutions.

Typical accuracy: sensitivity 88–97%, specificity 85–96% depending on the type of pathology. Robust performance across various devices and populations.

Foundation models and multimodal systems. A single model handles dozens of tasks. It integrates image, text and clinical context. It generates reports and answers questions. In January 2026, the Aidoc CARE system became the first foundation model to receive FDA approval for 14 abdominal indications on CT, with an average sensitivity of 97% and specificity of 98%.

Ask yourself (or the information technology department) the following questions.

How many pathologies does the system detect? If one to three, it is probably first generation. If ten or more, it is second. If it generates a textual description of findings, it is third.

Does the system show why it decided? If not (only a yes/no output without visualization), it is first generation. Heatmaps and bounding boxes correspond to the second. Textual explanation, the third.

Is the model updated? If it was deployed years ago and has not been updated since, its performance probably does not correspond to current capabilities.

Does it have valid certification? FDA 510(k) clearance or CE MDR certification corresponds to the second and third generation. Older CE certification under the original MDD directive indicates an older product.

How does the system respond to images from a different device? If, after replacing the X-ray or CT machine, you notice a marked deterioration in performance, this is a symptom of an insufficiently robust model.

Aidoc (Israel) is currently the most comprehensive platform. Over 30 FDA clearances, deployment in more than 1,600 hospitals. It covers head CT (intracranial hemorrhage, stroke), chest CT (pulmonary embolism), abdominal CT (14 acute abdominal indications) and X-ray (cervical spine). In January 2026 it became the first manufacturer to receive FDA approval for a clinical foundation model (CARE).

Viz.ai (USA) specializes in emergent vascular pathologies. Over 13 FDA clearances, deployment in more than 1,400 hospitals. Its main strength: detection of large vessel occlusion of the brain on CT angiography with automatic notification of the interventional neuroradiologist — it demonstrably shortens the time from arrival to treatment in strokes.

Lunit (South Korea) dominates in the analysis of chest X-ray and mammography. INSIGHT CXR achieved, in a head-to-head comparison, the highest AUC of 0.93 for the detection of lung nodules, outperforming all other systems as well as human readers. More than 3,000 hospitals in 40 countries. Documentation: over 100 peer-reviewed scientific publications.

Annalise.ai (Australia) excels in breadth of coverage: Enterprise CXR detects 124 findings on a chest X-ray within a single CE certification. Over 1,000 facilities worldwide.

Gleamer (France) is a specialist in fracture detection. BoneView achieves an accuracy of 97% with FDA approval for both adult and pediatric patients. Over 2,000 institutions in 40 countries. ChestView covers the chest X-ray.

RapidAI (USA) is the standard for emergent neuroimaging. CT perfusion and CTA analysis for strokes. Deployment in more than 1,800 hospitals. Widely used as a triage tool in stroke centers.

Oxipit (Lithuania) operates the world's only fully autonomous system for radiology — ChestLink generates final reports for normal chest X-rays without any intervention by a radiologist. Sensitivity 99.9%, capable of automating up to 40% of cases (up to 80% in primary care). CE certification class IIb. Pilot deployment in Norway, Brazil and other countries.

Nuance AI Marketplace (now part of Microsoft) integrates artificial intelligence applications directly into the PowerScribe reporting system, used in approximately 80% of U.S. hospitals.

GE HealthCare Edison has 115 FDA clearances (the most of any manufacturer). It includes AIR Recon DL for magnetic resonance image reconstruction, Critical Care Suite for pneumothorax detection and Caption AI for cardiac ultrasound.

Siemens AI-Rad Companion is a cloud assistant integrated into the syngo.via platform. It covers lung nodules, coronary calcium, vertebral fractures, volumetric brain analysis and prostate segmentation. 86 FDA clearances. It analyzes images from any CT manufacturer.

Carebot is a Czech startup based in Prague and the first Czech company to obtain CE MDR class IIa certification for a radiology tool using artificial intelligence. Its main product, Carebot AI CXR, analyzes chest X-rays. Its portfolio also offers AI Bones (fractures) and, as a pilot, AI MMG (mammography). The system is being tested in hospitals of the AGEL network (Šumperk, Havířov, Nový Jičín) and in other centers, including Příbram.

As of September 2025, the FDA had approved 1,356 medical devices using artificial intelligence, of which 1,039 were for radiology. In 2025 alone a record 295 new approvals were added. Nevertheless, a large proportion of the approved products were not clinically tested with a human operator — a systematic review published in JAMA Network Open (2025) showed that clinical testing remains uncommon for AI-based radiology devices.

The Medical Device Regulation (MDR) significantly tightened the requirements for systems using artificial intelligence. Most radiology products fall into class IIa or IIb. As a consequence of the regulatory burden, the number of newly certified products on the European market fell — whereas in 2020, 50 products were newly certified, in 2024 it was a mere 3.

From August 2027, medical devices using artificial intelligence will additionally have to comply with the EU Artificial Intelligence Act (AI Act), which classifies radiology systems as high-risk, with requirements for transparency, human oversight and continuous monitoring.

The State Institute for Drug Control (SÚKL) carries out market surveillance over medical devices, but does not directly approve products using artificial intelligence — this is ensured by notified bodies within the CE certification process. Systems using artificial intelligence are deployed in 23 Czech stroke centers. A dedicated reimbursement pathway for artificial intelligence in radiology does not yet exist in the Czech Republic — the costs are borne by hospital budgets or innovation grants.

The greatest change of the near future will be the transition from dozens of single-purpose algorithms to a single universal model that can handle a range of tasks across modalities. Microsoft BiomedParse can segment 64 types of objects in 9 imaging modalities on the basis of a text prompt — the radiologist writes "lung nodule" and the model returns a segmentation mask. NVIDIA MONAI VISTA-3D segments over 126 anatomical classes from CT.

The expected horizon for the first regulatory approvals of foundation models for broader clinical deployment is 2026–2028.

The hospital network Northwestern Medicine in the USA carried out the world's first clinical deployment of generative artificial intelligence for radiology reports. Over 12 hospitals, approximately 24,000 reports over five months (study published in JAMA Network Open, 2025): an average efficiency improvement of 15.5% (over 40% for some radiologists) without a change in clinical accuracy. The system additionally flagged critical findings such as pneumothorax in real time, even before the radiologist's review.

Experimental systems can autonomously break down a clinical task into sub-steps. For example, given the instruction "evaluate acute stroke," such a system can approach the non-contrast CT, compute the ASPECTS score, open the CT angiography to identify the occlusion, run a perfusion analysis and synthesize the findings into a preliminary report. For now these are prototypes, but the direction is clear.

The most technologically mature application of artificial intelligence in radiology is image reconstruction using deep learning. Commercial solutions from GE, Canon and Siemens achieve a 40% reduction in noise on CT compared with traditional reconstruction. For magnetic resonance imaging, GE AIR Recon DL enables up to an 86% reduction in scan time. This directly affects daily practice — shorter examinations, lower doses, higher throughput.

Find out what you are using. If the system frustrates you with a high error rate, the first step is to find out the exact product name, version and date of the last update. A system from 2018 corresponds to a technologically entirely different reality than a current product.

Consider modernization. The difference between the first and second generation is vast. If your system detects one to three pathologies, has no visual explanation and has not been updated for years, an investment in a modern solution typically pays off in time savings and improved detection.

Maintain critical distance. Even the best system makes mistakes. First form your own finding, and only then look at the artificial intelligence output. Research has unambiguously demonstrated that incorrect suggestions worsen the performance even of experienced radiologists.

Make use of complementarity. Artificial intelligence makes mistakes differently than a radiologist — and it is precisely this difference that is valuable. The system catches a lung nodule that the radiologist overlooks when fatigued on a night shift. The radiologist recognizes an atypical presentation that the algorithm has never seen. Together they are demonstrably better than each one separately.

Calibrate the sensitivity thresholds. If the system generates too many false alarms, the problem may lie in the setting of the sensitivity threshold, not in the model as such. Most modern systems allow this calibration — a higher threshold means fewer alarms at the cost of a higher risk of a miss.

Follow regulatory developments. The EU Artificial Intelligence Act (fully in force from August 2027) will bring new requirements for transparency and human oversight. The European Health Data Space (EHDS) will, from 2031, open up cross-border sharing of medical images for research.

Artificial intelligence in radiology has reached the point where ignoring it is as irrational as accepting it uncritically. The evidence is strong: mammography (29% more cancers detected in the MASAI study), prostate (artificial intelligence outperforms radiologists in AUC in the PI-CAI study), chest (radiologists' sensitivity rises by 12–26 percentage points). At the same time, real limits remain: the hallucinations of multimodal models, automation bias, population bias and dramatic differences in quality between system generations.

A radiologist who refuses artificial intelligence loses a demonstrably effective tool. A radiologist who trusts it uncritically risks worsening their own performance. The optimal position is in the middle: the informed, critical use of modern, regulatorily approved systems as an extension of one's own judgment.

As Curtis Langlotz, president of RSNA 2024, summarized: "Radiologists who use artificial intelligence will replace those who do not use it."

This article is based on publicly available sources: peer-reviewed studies (The Lancet, Lancet Oncology, Lancet Digital Health, Nature Medicine, Radiology, JAMA Network Open), the FDA database, manufacturers' data and review articles published in the years 2024–2026. The specific numerical data on the accuracy of commercial systems are based on published validation studies; the real accuracy at a specific workplace may differ depending on the population, the device and the settings.

Transparency of creation:

The conception, structure and editorial line of the article are the work of the author, who prepared the content outline, established the key theses and directed the entire creation process. Generative AI (Claude, Anthropic) was used as a technical tool for research, fact-checking and the elaboration of the author's draft.

The author edited the outputs throughout, verified the key findings and approved the final wording. No part of the text was published without human review. All factual data were verified against the publicly available sources cited in the text.

The procedure is in compliance with the requirements of Art. 50 of EU Regulation 2024/1689 (AI Act) on the transparency of AI-generated content. #poweredByAI