European Society of Paediatric Radiology Artificial Intelligence taskforce: a new taskforce for the digital age

The role of AI in paediatric imaging is still emerging [10]. Whilst there are several novel uses for AI within radiology applied to adults and an increasing number of commercially available products entering a progressively crowded marketplace, few are tailored to the paediatric population [3]. New rules and regulations concerning safety and legal standards across different nations are important for these medical devices to follow [11, 12], bringing with them ethical considerations not previously addressed in other medical domains [13]. As such, it is likely that the function and goals of this task force may evolve along with the developments in the field. At the time of writing, however, the consensus view of this committee regarding its primary role and function is based upon collaborative and educational efforts, as highlighted from our members [9] and outlined below.

A training programme dedicated to understanding the basic principles, applications and terminologies surrounding development, testing, implementation and safety of AI tools is needed. This will empower radiologists to understand the available literature on this topic, engage with multicentric studies to develop AI solutions and have meaningful discussions with vendors of AI tools when considering whether to integrate these into their own clinical practice [14].

We acknowledge that most radiologists do not have a background in computer science, informatics or coding, and will not see it as their role to acquire this knowledge; nonetheless, it is every doctor’s responsibility to be engaged with new developments within our subspecialty and to put patients’ outcomes and safety first. Our aim will therefore be to ensure all educational material has a clinical relevance that can be demonstrated rather than focusing on seemingly technical, mathematical or engineering-heavy topics.

In order for AI and machine learning advancements to be effectively utilised in clinical care, it is necessary for radiologists to be prepared to handle these changes. A foundational understanding of this incorporated into radiology teaching curriculums with possible involvement of university partners may be the longer-term key goal. Courses teaching the basics of AI exist but are usually offered as a single one-off event [15], institution specific (only for local staff) or aimed at attendees with some coding experience. In the international survey of opinions regarding AI in children’s imaging [9], it was interesting that more than half (68%) of respondents were more likely to learn about it if it was specifically tailored to paediatric radiology. This task force, therefore, believes it is important that dedicated material for paediatric radiologists should be provided and is already making plans for a training strategy in collaboration with representatives from other international imaging informatic society counterparts (the Society of Imaging Informatics in Medicine [SIIM] and the European Society of Medical Imaging Informatics [EuSoMII]) to foster network ties, promote mutual understanding and share areas of good practice.

In the meantime, we would like to direct readers interested in furthering their paediatric radiology AI knowledge to excellent articles previously published within Pediatric Radiology [16], including the recent special issue on this topic (published 2021) [10, 17‐26]. Educational courses on AI and informatics for general radiology are organised by the European Society of Radiology (ESR), such as a dedicated foundation course for senior radiology trainees [27]; for those wanting some recognition of their training, a recently launched AI certificate programme from the Radiological Society of North America (RSNA) is available covering foundational knowledge in the use of AI products for clinical practice [28].

There needs to be a general understanding that AI can play a pivotal role in medical care, and that through collaboration and sharing of multicentric data we can help validate and establish the performance of novel algorithms for paediatric use. Many of the publications describing AI algorithms fail to conduct external validation or create tools based on very biased data sets [29], which could potentially exclude certain populations (i.e. children younger than 2 years or those affected by rare diseases). There is also a lack of data and information regarding post-market surveillance and whether AI performance maintains similarly high accuracy and standards with time. To overcome these issues, we aim to share experiences of streamlined ways of working and sharing large data sets among hospitals, whether it is related to assisting in setting up federated learning databases at local sites or a secure cloud-based method of accessing anonymised information to deploy and evaluate the performance of the algorithms.

There is a clear need for collaboration among interested professional groups to form a strong clinical, radiologic and data/computational scientist support network. Advanced technologies and AI solutions are only as strong as the people who drive and support them. Their development depends on a user-centric design that aids health care outcomes, with proactive involvement of the end users and those who potentially benefit from their input. Partnerships and collaborations are key to delivering innovative solutions that optimise patient care. We will therefore develop and support partnerships amongst paediatric radiologists across multiple centres and subspecialties along with adult radiology colleagues (who may have some domain experience from their own practice to share), expertise from computational and engineering sciences (including software engineers and statisticians), assistance from implementation scientists, Picture Archiving and Communication System (PACS) teams and hospital managers regarding integration into clinical workflow and research partnerships.

We should also bear in mind that many stakeholders may work with or for commercial/industrial AI partners who also provide their own key insights into how to produce and incorporate their AI tools into health care. By combining these skill sets and harnessing the expertise from each domain, we can assist colleagues wishing to understand how to approach the integration and adoption of AI into their clinical department.

Given our above stated objectives, we have put together an ambitious yet achievable set of goals we will be working toward across the next five years that broadly cover three main domains: