The session titled "Artificial Intelligence in Clinical Practice: Life Past The Hype," held on the opening day of the European Congress of Radiology (ECR) 2025 in Vienna, provided a deep dive into the transformative role of AI in medical imaging. Expert panellists engaged in a rigorous discussion on the implications of AI for radiology, particularly focusing on the evolving regulatory framework established by the EU AI Act and its ramifications for practice. As the adoption of AI technologies escalates, radiologists must adeptly navigate a landscape filled with both potential benefits and formidable challenges to facilitate the seamless integration of these tools into clinical workflows.
Clinical AI Products in Body Imaging: An Overview
Renato Cuocolo from the University of Salerno discussed the transformative impact of artificial intelligence in radiology, with a particular focus on body imaging applications. He noted the rapid evolution of AI, highlighting a significant increase in approved AI-assisted tools in both the United States and Europe; however, he pointed out that adoption rates vary significantly, particularly in Europe.
Cuocolo traced the progression of AI from early cybernetic models to contemporary deep learning frameworks, emphasising a current focus on data modelling and automated processes. A major component of AI's role in radiology is radiomics, which involves extracting quantitative image features to develop predictive models that enhance diagnostic accuracy.
He categorised AI applications into two main types: interpretative, which aid in diagnosis by detecting anomalies, and non-interpretative, aimed at streamlining workflow efficiencies. Specific imaging domains such as breast, thoracic and musculoskeletal radiology are leading the uptake of AI technologies, facilitating tasks like lesion characterisation, nodule detection and preoperative planning.
Surveys indicate that approximately 50% of radiologists in Europe are now utilising AI, though challenges remain due to performance variability. Workflow automation within radiology, including protocol optimisation and triaging of cases, is becoming increasingly integrated into standard clinical procedures. Additionally, advancements in image reconstruction through AI are proving effective in enhancing image quality and reducing noise artefacts.
However, the speaker underscored existing hurdles, particularly concerning regulatory literacy among radiologists, many of whom lack familiarity with medical device regulations. He emphasised the necessity for robust evidence from large-scale clinical trials to substantiate the reliability of AI implementations, which is critical for justifying investment. Public trust in AI's capabilities within the healthcare sector is also waning, with less than half of surveyed patients expressing confidence in AI-assisted healthcare solutions.
Looking forward, the speaker suggested that AI will significantly influence radiology training and the necessary skillsets for the future workforce. He acknowledged ongoing debates surrounding the autonomy of AI in diagnostic processes, noting that current regulations prohibit fully autonomous systems. In conclusion, he urged radiologists to take a proactive role in the integration of AI, ensuring development aligns with clinical requirements rather than external pressures.
EU Artificial Intelligence Act: impact on Medical Imaging
Dr. Hugh Harvey from Hardian Health (UK) delivered an insightful session on the EU AI Act, focusing on its implications for the healthcare sector, particularly in the domain of medical imaging. This comprehensive 400-page legislation is the most extensive the EU has produced to date, categorising AI systems into four distinct risk levels. High-risk AI applications, notably those utilised in healthcare, are subject to rigorous regulatory oversight.
The Act mandates conformity assessments for medical AI systems, which are already being performed via CE marking, but it adds further obligations, particularly concerning post-market surveillance. As of February, certain AI applications classified as unacceptable risk have been banned. Additionally, obligations for general-purpose AI providers will take effect by August 2024, while full enforcement for high-risk AI systems is scheduled for August 2026, complete with penalties for non-compliance.
A significant emphasis was placed on the necessity of AI literacy among clinicians, radiologists and other healthcare professionals. Education initiatives will occur through self-directed learning, hospital-led programmes and the integration of AI concepts into radiology curricula to ensure that professionals are adept at both utilising and monitoring AI technologies effectively.
Risk management is a pivotal component of Article 9 of the Act, extending its focus from developers to healthcare providers. Hospitals and AI practitioners are required to evaluate risks within real-world contexts, ensuring robust cybersecurity, data protection and the timely reporting of adverse events. Article 26 imposes legal accountability on radiologists and healthcare departments regarding the deployment and ongoing supervision of high-risk AI systems.
The session also shed light on the concept of regulatory sandboxes outlined in Article 57, which mandates that each EU member state establish controlled testing environments for AI innovations by August 2026. Furthermore, Article 13 reinforces transparency, obligating AI developers to furnish comprehensive documentation, maintain human oversight protocols and provide performance metrics accessible to end-users.
Dr. Harvey concluded that the responsibility for compliance with the EU AI Act is shared equally between hospitals, AI users and developers, with potential fines reaching up to €7.5 million for disseminating misleading information. As enforcement deadlines draw near, adherence to these regulations is imperative for all stakeholders involved.
Raising The Bar: Clinical Trials for Healthcare AI Products
Ritse Mann from Radboud University Medical Centre (the Netherlands) explored the clinical integration of AI within medical diagnostics, with a particular emphasis on radiology. The discussion centred on the evidence hierarchy necessary for AI applications to evolve from demonstrating technical capabilities to showcasing practical benefits for healthcare systems.
Many AI studies remain at the level of demonstrating technical feasibility, primarily revealing associations from limited datasets while falling short of validation in broader, external settings. This limitation is often due to the reliance on insular institutional data, which complicates external validation efforts. Although several algorithms may show superior performance compared to radiologists in controlled trials, their actual effectiveness in real-world scenarios is still uncertain, influenced by issues such as data drift and regulatory challenges.
The session underscored the necessity for a structured approach to AI model development, advocating for the division of datasets into training, validation and external testing subsets. It also highlighted the imperative of moving beyond mere reader evaluations toward incorporating tangible clinical outcomes to substantiate claims of efficacy. Despite the proliferation of commercial AI solutions, the majority lack robust evidence supporting meaningful improvements in patient outcomes.
The occurrence of randomised controlled trials (RCTs) focusing on AI remains infrequent, with only a handful of studies indicating any significant clinical advantages. While some trials have reported modest efficiencies, such as reduced time in stroke management and enhanced detection rates in mammography, the implications for patient outcomes are still ambiguous. As such, post-marketing surveillance emerges as a critical strategy for assessing AI performance across a variety of clinical contexts, addressing potential data shifts and system updates.
In conclusion, the speaker highlighted the role of AI in diagnostics transcends imaging modalities, necessitating a synthesis with comprehensive clinical datasets. Progressing AI in healthcare requires stringent validation through clinical trials and ongoing monitoring to ensure its reliability and applicability across diverse populations and clinical environments.
Artificial Intelligence is set to transform the field of radiology, yet its effective implementation relies on rigorous clinical validation, adherence to regulatory standards and the preparedness of healthcare professionals. AI-enhanced automation has the potential to significantly improve diagnostic precision and operational efficiency. However, several challenges remain, including the complexities of regulatory frameworks, the necessity of fostering public trust and ensuring the relevance of AI solutions in real-world clinical scenarios. A multidisciplinary approach involving radiologists, software developers and policymakers will be crucial to fully leverage AI's capabilities while upholding patient safety and ensuring clinical effectiveness.
Source & Image Credit: ECR 2025
