Keywords radiology - education - clinical practice - artificial intelligence
Introduction
The 2024 Nobel Prize in Physics was awarded to Geoffrey Hinton for his groundbreaking
work in developing neural networks, which underpin many of today’s applications using
artificial intelligence (AI). A leading thinker in AI research, Hinton is also considered
a polarizing figure based on his statements about the potential impact of AI on medicine
and, in particular, radiology. “People should stop training radiologists now ,” he stated provocatively in 2016. “It’s just completely obvious that within five years deep learning is going to be better
than radiologists”
[1 ]. Later, he added that radiologists are “already over the edge of the cliff, but you haven't yet looked down. There’s no ground
underneath”
[2 ]. These statements sparked a broad discussion among experts that ranges from euphoria
about the technology’s potential to concern about the future of an entire field [3 ]
[4 ].
While perceived as dystopian by some, others view such statements as a catalyst: For
example, radiologist Curtis P. Kelly describes how Hinton's words personally motivated
him, the start of his specialist training, to closely examine the pros and cons of
AI in radiology – not out of fear, but from the desire to play a role in actively
shaping the transformation [5 ]. From this perspective, AI does not necessarily mean replacement; instead, it can
also be seen as an opportunity to re-imagine and develop the field. This is also reflected
in Curtis P. Langlotz’s oft-quoted statement: “Radiologists who use AI will replace radiologists who don’t”
[6 ]. From this perspective, AI is understood not as a replacement, but as a tool – that
is intended to expand and strengthen human diagnostic capabilities.
Against this background, the question arises not only of how AI and other digital
innovations will change radiology, but also of how the field itself is dealing with
this transformation. Due to its high level of digitalization, the standardization
of imaging procedures, and its vital role in clinical management, radiology seems
very well-suited as a field for implementing digital innovations at an early stage
and evaluating them critically. At the same time, it exemplifies many of the challenges
that modern medicine faces in the digital age: data protection, regulatory frameworks,
questions of medical responsibility and, last but not least, uncertainties among the
next generation of medical professionals.
The aim of this narrative review is therefore to present the current state of research
on the use of AI methods in radiology and to highlight opportunities and challenges
in a critical and evidence-based manner. This study is intended to motivate prospective
radiologists to actively take part in shaping the digital transformation and to reinforce
the importance of ongoing training.
Radiology leads the way in the digital transformation
Radiology leads the way in the digital transformation
Due to its high level of digitalization, radiology is playing a leading role in the
digital transformation in medicine. Through the introduction of digital imaging systems,
in particular, picture archiving and communication systems (PACS), radiology has been able to develop comprehensive expertise at an early stage
in handling and managing complex data structures [7 ]. This is partly due to the nature of radiological data: Imaging techniques primarily
generate digital, standardized data formats, such as DICOM (Digital Imaging and Communications
in Medicine), for storing and transmitting image data, including related metadata,
which form the basis for seamless digital processing. Important standards for imaging
interoperability include DICOM HL7 (Health Level Seven, i.e. reporting standards for
sharing clinical information) and FHIR (Fast Healthcare Interoperability Resources),
a modern, web-based standard for structured data exchange. These standards define
the format, the availability of relevant metadata, and the interface behavior, and
they are therefore essential for teleradiology, image networking, and AI workflows
[8 ].
Thanks to radiology’s high level of interdisciplinarity and the immediate digital
availability of standardized radiological image data, the field is ideal for an integrated
approach to diagnostics. This makes it possible to systematically combine radiological
information with clinical, pathological, laboratory chemical and molecular biological
data to provide a comprehensive picture of individual patients.
The availability of standardized digital data formats in radiology, such as DICOM,
fundamentally simplifies the technical storage and transmission of radiological information.
It should be noted, however, that it is not a trivial matter to use this data for
AI-powered applications. Although radiological image data comes in a digitized and
standardized format, it remains a significant challenge to analyze and interpret this
content using AI algorithms. Unlike structured, tabular data, such as laboratory values,
from which diagnoses like diabetes mellitus can be derived relatively easily and automatically,
radiological image data are high-dimensional, complex, and often poorly annotated.
For example, it is significantly more complex to train an AI to recognize arbitrary
pathologies in a thorax-abdomen CT scan than to identify disease patterns from fully
numerical laboratory data sets.
The much-cited AI-friendliness of radiology should therefore be viewed with a critical
eye: While the high level of digitalization reduces the technical barriers to entry,
it does not replace the complex content preparation, annotation, and validation of
image data needed to use AI successfully.
Overall, radiology, due to its systematic use of standardized digital formats, provides
ideal conditions for technological innovations, and it is positioning itself in the
long term as a strategic discipline at the intersection of technological advances,
clinical care, and research. Radiology takes on the role of managing diagnostic data,
so that interdisciplinary cooperation and ongoing training create new options for
applications, for example, through inclusion of radiological data in multicenter studies.
[Fig. 1 ] illustrates the vital role of radiology as an integrative platform in digital medicine
and shows its connection to important technological and clinical advances. Increasingly,
the focus is shifting to the technical infrastructure, which forms the basis for efficiently
using, storing, and analyzing large amounts of image data.
Fig. 1 Radiology provides a platform for the digital transformation in medicine. The graphic
shows the integrative role of radiology at the intersection of technological progress,
clinical applications, and interdisciplinary collaboration. DICOM: Digital Imaging
and Communications in Medicine; HL7 FHIR: Health Level Seven – Fast Healthcare Interoperability
Resources; PACS: Picture Archiving and Communication System; RIS: Radiology Information
System.
Digital infrastructure
Digital infrastructure provides the backbone of modern radiological practice, and
it enables practitioners to use and manage medical image data in an efficient, secure,
and reliable manner. In this context, a variety of software systems play an important
role. One key component is PACS, which allow for centralized, standardized digital
storage, processing, and distribution of medical image data. In addition, radiology
information systems (RIS) are used, which digitally map administrative and organizational
processes such as scheduling appointments or writing and distributing reports. Thanks
to ongoing technological developments, these systems are able to ensure seamless integration
in everyday clinical practice and enable cross-site access to data.
Another essential component of the digital infrastructure is the international DICOM
standard, which supports the exchange of manufacturer-independent medical image data
between different imaging systems and software solutions. This process is being augmented
increasingly by the HL7/FHIR standard, which is designed specifically for sharing
data between different healthcare facilities and to develop interoperable databases
[9 ]. These standards significantly improve collaboration between different medical facilities
and disciplines by ensuring, through standardized interfaces, that different clinical
systems can be connected efficiently in a network.
In response to the growing trend to store and manage the ever-increasing volumes of
data, a large number of facilities are implementing vendor-neutral archives (VNAs) [10 ]. VNAs are data format-independent archives that can store and manage not only radiological
images but also other medical data. Their flexibility regarding the manufacturer technology
used supports sustained, cross-institutional use, which in turn protects investments
and provides efficient archiving and access control in the long term.
Teleradiology demonstrates the great potential of digital solutions in a particularly
impressive way. Teleradiological networks allow medical image data to be shared promptly
and independent of location, which benefits regions lacking infrastructure and rural
areas, in particular [11 ]. However, when deployed internationally, for example, in parts of Africa, there
are often different data protection and regulatory requirements compared to those
in Germany and the European Union. This may limit the direct comparability of the
data collected there.
Use of cloud-based technologies is another increasingly relevant component of digital
infrastructure [12 ]. These solutions not only provide major benefits in terms of scalability and flexibility,
but also allow for on-demand use of computing power, which is essential, especially
for data-intensive AI applications [13 ]. In addition, cloud-based solutions make it easier to collaborate in clinical and
scientific contexts, as data and analytic tools can be accessed regardless of location.
It is useful to distinguish, in this context, between Infrastructure as a Service (IaaS), which provides basic IT resources such as storage and computing power, Platform as a Service (PaaS), which additionally offers a development environment, and Software as a Service (SaaS), where finished applications can be used directly [14 ]. These advances enable more flexible care, allowing radiology to explore new research
areas and medical applications through cloud-based solutions.
In addition to image-based AI applications, image-independent systems are also gaining
in importance – for example, for optimizing administrative processes in everyday clinical
practice. One example of this is AI-powered patient scheduling tools, which can help
to reduce wait times, improve resource utilization, and strengthen patient satisfaction
[15 ].
However, digitalization also faces numerous challenges, in particular, with regard
to data protection and regulatory requirements. Data protection regulations, such
as the European Union's General Data Protection Regulation (GDPR) and the complementary
German Federal Data Protection Act (BDSG), establish strict framework conditions for
processing and storing medical data, especially as part of cloud-based infrastructures
and on AI platforms [16 ]. This results in additional requirements for IT security, data protection concepts,
and transparent information and consent procedures for patients and medical users.
These regulatory requirements can complicate and delay implementation processes. As
a result, it can be complex and costly to meet legal requirements for using digital
technologies. Nevertheless, past experience shows that digital solutions can be integrated
in clinical processes in a compliant and secure manner, if regulatory and ethical
requirements are taken into account during the early development stages. One example
of successfully meeting such requirements is the implementation of a cloud-based teleradiology
network in African countries [11 ]. This network sustainably improved care while at the same time meeting the highest
standards of data security, taking into account data protection and regulatory requirements.
AI: Research, innovation, and clinical practice
AI: Research, innovation, and clinical practice
In medicine, AI encompasses a variety of algorithmic approaches to recognize patterns
in large data sets, analyze complex relationships, and support diagnostic and therapeutic
decisions. Methodologically, AI models can be divided into different categories [17 ]. Supervised learning uses annotated training data, in which, for example, tumors in MRI images were manually
marked, to specifically train the algorithm to detect them. In contrast, unsupervised learning works without such markings: The algorithm analyzes the image data independently and
searches for patterns or abnormalities – for example, to identify previously undetected
tissue structures or potential subgroups of tumors. Deep learning refers to methods using deep neural networks (e.g. convolutional neural networks (CNNs)), which gradually extract features from raw data and thereby learn complex,
hierarchical representations. Deep learning models are used predominantly in supervised
settings (with labeled data). More recently, however, self-monitored and unmonitored
approaches have also been developed that require less or no explicit labeling [18 ]. Recurrent neural networks (RNNs) are used additionally for processing sequential data. For complex language
processing, transformer-based networks are now primarily used; these include large language models (LLMs) such as ChatGPT or GPT-4 [19 ]. There are also generative adversarial networks (GANs) that can be used, for example, to create synthetic training data [20 ].
Several AI technologies have already become established in clinical routine and are
being deployed in a range of diagnostic scenarios. One of the first and now a widely
used application is LungCAD, an AI-powered system for detecting lung pathologies in
X-ray and CT images; it is used, in particular, to detect pneumothoraces, pulmonary
nodules, and interstitial lung diseases [21 ]
[22 ]. Automated fracture detection is also now being used in many hospitals.
AI-powered mammography, which contributes to increasing sensitivity and reducing false-negative
findings in screening programs, is another clinically established field [23 ]. [Fig. 2 ] illustrates, as an example, clinical applications of AI in radiological diagnostics,
especially in lung and breast imaging, as well as its methodological foundations and
further perspectives.
Fig. 2 Examples of clinical applications using AI in radiology. The graphic shows established
AI applications in lung and breast imaging and their methodological principles [42 ]
[43 ]
[44 ]
[45 ]. Icons adapted from Flaticon (www.flaticon.com ). CT: computed tomography; AI: artificial intelligence; RADS: reporting and data
system.
AI systems are also used to optimize the acquisition of image data [24 ]. In magnetic resonance imaging (MRI), reconstructive deep learning algorithms can
help to ensure high-quality images with significantly reduced scan times. Such algorithms
not only increase efficiency but also help to reduce motion artifacts, improve diagnostic
accuracy, and increase patient comfort by reducing the burden of immobility and potential
discomfort through shorter scan times [25 ].
Many AI-powered methods are still in the research and development stage. [Fig. 3 ] provides an overview of key AI applications in radiology, methodological foundations,
areas of application in clinical practice, as well as current challenges and solutions.
Fig. 3 Systematic overview of AI-powered innovations in radiology. Important AI methods and
clinical applications are shown here – from image analysis, acquisition, and workflow
optimization to radiomics, promoting young talents, and finding solutions for existing
challenges. AG IT: Information Technology Working Group; CAD: computer-aided detection;
DRG: German Radiological Society; AI: artificial intelligence; LLM: large language
model; MRI: magnetic resonance imaging: NKLM: German National Competency-based Learning
Objectives Catalog in Medicine; NLP: natural language processing.
Automated organ segmentation in CT and MRI data is particularly promising in this
regard. It is used, for example, in body composition analysis to quantify fat and muscle distribution; it is also used to assess metabolic or malignant
diseases [26 ]. Another important area of research is structured reporting based on natural language processing (NLP) and LLMs. Current studies show that LLMs are already capable of coherently structuring
and summarizing findings, which could reduce the amount of variability in findings
and improve the clinical workflow [27 ].
The applicability of AI-powered models is also being increasingly tested in diagnostics.
For example, radiomics – a method for quantitative analysis of radiological image
data that identifies patterns not visible to the naked eye by extracting and mathematically
evaluating image-based features – is being researched to characterize the molecular
properties of tumors [28 ]. Radiomic models are able to predict IDH mutations in gliomas with high accuracy
[29 ], which could reduce the need for invasive biopsies in the future. Similarly, AI-powered
image analysis was used to estimate the response to immune checkpoint inhibitors in
lung cancer [30 ], which can be of great importance for personalized patient management, especially
in targeted (immuno)therapies. One important factor limiting many radiomic applications
is that the method requires segmentation or annotation of the regions of interest . Manual segmentations can exhibit significant inter-observer variability and are
very time-consuming without (semi)automated methods. This can affect both reproducibility
and scalability. For this reason, standardized segmentation protocols, validation
by multiple annotators, the use of AI-powered segmentation tools, and transparent
information on segmentation and quality control procedures are key prerequisites for
a robust clinical implementation of radiomic applications.
Despite these advances, there are still major challenges that are delaying the widespread
clinical adoption of AI-powered procedures. One important aspect is validation and
clinical evidence, as many algorithms have been trained under laboratory conditions
but have not been systematically tested for their generalizability in real clinical
environments. Inadequate external validation can lead to reduced diagnostic accuracy,
especially when the algorithms are applied to data sets that differ significantly
from the training data [31 ]. In order to ensure the generalizability and comparability of AI models, it is also
important to standardize acquisition and transparently document image quality and
training data. Another issue is data bias, which can derive particularly from unbalanced
training cohorts. Bias refers to data sets in which certain patient groups – for example,
with regard to age, gender, ethnic origin, or disease prevalence – are systematically
over- or underrepresented, potentially limiting the generalizability of the models.
As a result, it can be difficult to transfer AI models trained primarily with data
from a specific population to other patient groups. This can lead to systematic differences
in diagnostic accuracy between patient groups, and it poses one of the biggest challenges
to widespread clinical use.
In addition, there are regulatory hurdles to consider, because AI systems in medicine
are classified as medical devices and therefore fall under the European Medical Device
Regulation (MDR) and the new EU AI Act. These regulations require the algorithms'
decision-making processes to be traceable and transparent, and they also require continuous
performance monitoring to ensure safety and quality standards [32 ]. In the field of radiomics, further challenges arise with regard to standardization
and reproducibility, as the extracted image features are often influenced by technical
factors such as image acquisition and reconstruction parameters [33 ]. For this reason, practitioners need to have robust validation protocols to ensure
the clinical applicability of radiomic models [34 ]. In addition to the obstacles mentioned above, it will be important to educate and
continue to train radiologists in the future to ensure AI models are used properly
and in a meaningful way. Especially “explainable AI” (i.e. AI methods that disclose their decision-making processes to users) are gaining
significance here. It is important that doctors understand how an AI system arrives
at its decision, in order to critically evaluate this decision and understand it in
a clinical context.
Despite rapid developments in some AI methods, major challenges remain: The quality,
standardization and availability of annotated image data is limited; ethical and legal
questions of responsibility are often unclear, and the generalizability of many algorithms
to new patient populations has not yet been sufficiently validated.
Promoting young talents in the digital age
Promoting young talents in the digital age
Many prospective radiologists have expressed that they are uncertain about their future
role in the increasingly digital field of medicine. A multicenter survey of resident
physicians in the US revealed that over 80% of respondents have a strong interest
in AI in medicine, but at the same time over 50% do not feel adequately prepared to
deal with digital technologies [35 ]. Current data from German-speaking countries confirm this trend and indicate that
the perception of the role of AI in radiology work is affecting interest in specialist
training in radiology [36 ].
Particularly in radiology, one of the fields most strongly influenced by digital technology,
there appears to be a discrepancy between the rapid development of new AI applications
and the level of integration of digital skills in the curriculum for medical training.
Initial pilot projects in Germany are integrating AI-related case seminars into medical
internships and structured training offerings into specialist medical training. Digital
learning platforms from the German Radiological Society (DRG) and instructional videos
on AI applications are also being tested and evaluated.
To meet these challenges, it is critical to incorporate digital skills in medical
training as early as possible and in a systematic way [37 ]. In Germany, the National Competency-based Learning Objectives Catalog for Medicine
(NKLM 2.0) serves as the basis for curriculum design at medical schools. In addition
to conventional medical and clinical skills, NKLM 2.0 increasingly includes requirements
for future doctors to develop digital skills. This includes, for example, critically
evaluating and applying AI-powered diagnostic systems; securely working with electronic
patient records; incorporating telemedicine concepts in everyday clinical practice;
understanding the fundamentals of data analysis, and IT security [38 ]. This targeted update of curriculum aims to ensure that medical students are prepared,
as part of their training, for the challenges of an increasingly digital healthcare
system.
Beyond university studies, digital skills are increasingly being promoted in specialist
medical training. The DRG offers special qualification programs to effectively prepare
radiologists to use digital technologies. An outstanding example is the Q1 certificate
from the Information Technology Working Group (AG IT) that provides practical knowledge
about digital applications in radiology. The program includes content on AI and machine
learning, structured report formats, data management, and IT security in radiology.
Another innovative program designed to promote digital skills is AI-RADS, a novel
AI training curriculum for radiologists, offered in both face-to-face and online formats.
An evaluation study showed that participants significantly gained knowledge about
AI applications in radiology through the program and reduced their uncertainties in
dealing with these technologies [39 ]. Programs such as this demonstrate that targeted digital training is essential to
prepare radiologists for future challenges in the increasingly data-driven field of
medicine.
By clearly communicating new requirements and promoting analytical skills, programs
can make radiology more attractive for young medical talent. Incorporating digital
skills in a systematic way in curriculum, as well as adding specialized training offerings,
can enable prospective radiologists to prepare for their role in the increasingly
technological field of medicine; it can also enable them to actively shape the digital
transformation. Nevertheless, it is particularly important to keep a balance between
teaching technical skills and preserving core radiological competencies such as clinical
judgment.
Prospects for the future
Radiology is developing into a vital platform for providing personalized and precision
medicine [40 ]. Advances in radiomics and molecular imaging are making it possible to extract quantifiable
(molecular) features beyond morphological image analysis. Combined with molecular
profiles such as liquid biopsies , this method can lead to more precise diagnoses and more personalized treatment decisions
[40 ]. Key aspects of this development are multimodal imaging and the integration of radiological
data with clinical parameters, genetic information, and laboratory analyses. These
advances in response to the complexity of modern medicine support a personalized diagnostic
approach in which radiology can play a coordinating role.
As radiology becomes more digital, it is also more interdisciplinary. In addition
to close cooperation with the fields of computer science, physics, biology, and engineering,
collaboration with medical technologists for radiology (MTR) is playing a key role.
While automated analysis methods, AI-powered image acquisition, and digital workflows
gain in importance, this professional group will also increasingly need to have advanced
qualifications in the field of data processing and AI-supported image analysis [41 ]. Close collaboration between radiologists and MTRs is crucial for the smooth implementation
and efficient use of digital technologies in everyday clinical practice.
In the coming years, multimodal AI-powered support in decision-making, the integration
of image, genome, and routine data, and the use of digital twins are expected to further
strengthen radiology’s role in interdisciplinary clinical management.
Summary
Current developments in radiology present great opportunities for improving diagnostic
and therapeutic workflows. Innovations in the area of digital infrastructure, artificial
intelligence, and promoting young talents offer tremendous potential that can be actively
shaped by radiology. At the same time, successfully implementing AI-powered digital
technologies will require continually adapting to regulatory requirements, as well
as a critically evaluating the clinical benefits and practicality of these methods.
Through interdisciplinary collaboration, innovative research, and systematically promoting
young talents, radiology will continue to play a key role in the digital healthcare
system and contribute significantly to more personalized patient care in medicine.
