AI shows great promise in improving image quality, personalising dosages (both in diagnosis and theranostics) and helping in image interpretation and subsequent analysis. It opens up ways of fully exploiting the potential of NM, which has recently witnessed technological developments such as total-body PET, where the large amount of data acquired will also benefit from AI. NM has also strived to quantify molecular processes using PET and SPECT, and here again, AI may help with the process. As such, AI has the potential to improve clinical workflows that will increase overall efficiency but also facilitate personalised medicine for the benefit of individual patients. There remains, however, a long road ahead before the potential of AI in NM is realised in a manner acceptable to both healthcare professionals and patients. EANM can, and will, play a key role in this process.

Defining unmet needs

EANM should help to define unmet needs in the operational, physics, and clinical fields. By working together as a community, we can identify the issues that are most pressing along with those that are most likely to benefit from AI. Thereafter, we will need to prioritise and define clear objectives. Oncology is currently top of the list of the most widely published applications of AI in NM, as shown by a systematic review in 2019 [17]. More specifically, 86% of all publications in the AI and radiomics field dealt with oncology [17]. Nonetheless, cardiology, as illustrated by a recent position paper [18], neurology, inflammation, and infection, as well as therapy, should all benefit from those developments. It is here that the different committees of the EANM, together with their counterparts in other societies, will have to take the lead.

Setting standards

Second, we need to define the methods and set the standards against which the AI solutions will be evaluated and “calibrated”. This implies defining methodological details (DL algorithm architectures), statistics, and sample sizes used in the different stages of algorithm development, evaluation and validation, endpoints including ground truth, etc. The current literature is very heterogeneous in most of these aspects. More importantly, basic concepts such as the metrics used for measuring the performance of the models in the different targeted applications need to be clearly outlined [19]. For every solution, the different validation processes need to be clearly determined in advance. Harmonisation, transparency, and generalisability are key for a trustworthy clinical implementation. The development of similar initiatives to the “Image Biomarker Standardization Initiative” in the radiomics field will be a major step forward [20]. The EANM’s EARL Initiative could in time also become an important player, as it is familiar with standardisation aspects.

One of the difficulties with DL algorithms is not only the multiple parameters that need to be optimised but also the different types of networks and their implementation details in terms of, e.g. the number of layers and associated connections, optimisers, loss functions, etc. Each of these parameters can be varied. Consequently, it is crucial to compare the performance of all these various implementations under controlled conditions. Given all the potential variations, it is impossible to reproduce results from the literature by reimplementing the proposed developments. Instead, alternative approaches such as software challenges need to be implemented. In these challenges, a given dataset is made available for developers to evaluate the performance of their algorithms within a controlled environment. Numerous software challenges have been organised throughout the years, mainly within the field of image segmentation but also in other fields of interest and increasingly targeting clinical endpoints (https://grand-challenge.org/challenges/).

Software challenges should be based on a predetermined categorisation of potential applications, using different factors such as impact and potential utility and uptake in clinical practice. Challenges should address one or more of the following points:

1.

Specify the need for data and associated requirements (volume, annotation, QC), defining training and validation datasets.

2.

Clarify the dependency of data volumes necessary for a given task, which should also include standardisation and/or harmonisation for the exploitation of multicentre datasets.

3.

Define algorithm-related aspects such as the number of parameters, their optimisation, robustness, levels of uncertainty, and transferability to datasets from different instruments or body locations.

4.

Deal with model interpretability (white/grey box concept as opposed to black box), and thus with acceptability to professionals (medical physicists, physicians) and the public (patients and families).

5.

Integrate established domain knowledge (e.g. PBPK modelling) in AI algorithms or training procedures to reduce data volume requirements and improve the robustness of the algorithms.

6.

Consider training issues for implementation in clinical practice.

In order to accelerate clinical adoption, validation of the outcomes/results of individual trials/studies is required, and this is where multicentre and multigroup cooperation is of the utmost importance. Such cooperations within the field of NM can and should be triggered by challenges that link methodological and clinical objectives. A European platform such as the EANM with the expertise within its committees is ideally suited for overseeing such challenges, and it also represents a unique opportunity for the EANM to participate in the worldwide efforts of open science by facilitating pan-European collaborations.

Increasing awareness and knowledge

We need to increase the overall level of knowledge and competence in all fields related to AI. This implies revamping the nuclear medicine curricula for both medical physicists and physicians to take account of this evolution. Computational sciences, in particular, must be further integrated into education and training [21]. Although training is the responsibility of each individual European state through their national requirements, societies like the EANM can help increase awareness of the ongoing (r)evolution and promote standardisation within Europe. For example, the next update of the “European Training Requirements” published by the Nuclear Medicine Section of the UEMS should reflect this new dimension and provide guidance as to the minimum content of AI-related matters in the training curriculum, including quality criteria for trainers and training sites. The EANM will play a leading role within the continuing education field, making sure that both its flagship Annual Conference and the education programmes run by ESMIT facilitate the transmission of knowledge for the appropriate and reasoned implementation of AI in NM. These educational programmes can target different AI-related aspects, be they scientific, clinical, or ethical. Such programmes should meet the requirements for high-risk systems set out in the European Artificial Intelligence Act, which state that no AI systems are to be approved without providing users with information on the capabilities and limitations of AI and how to use it.

The time has come to incorporate AI as another partner discipline. Revamping the training of NM physicians implies enhancing existing collaborations between medical physicists, engineers, and specialists in different clinical applications but also the development of closer cooperation with other scientific societies dedicated to medical image computing and analysis. The EANM may also play a key role in helping industrial partners target appropriate developments and associated clinical applications in our field. At the same time, it can contribute to initiatives coordinated by the EU in the context of the deployment of trustworthy AI. Such initiatives include the Artificial Intelligence Regulation, the Data Act, the Digital Governance Act, and the European Health Data Space.

Ethical standards

Last but not least, ethical standards for the implementation of AI in our field and associated clinical applications need to be set out by the EANM. Currie et al. have recently proposed a set of ethical standards to be followed when evaluating and developing AI in NM as listed in Table 1 [22]. The EANM fully embraces these principles, which are applicable to any medical speciality that will employ AI as part of its clinical practice in the future. We believe they constitute a solid and necessary framework within which AI can be incorporated into nuclear medicine.

Table 1 Ethical standards to be followed when evaluating and developing AI in nuclear medicine (adapted from Currie et al.)

Many of those principles appear self-evident, albeit not equally so in all parts of the world. Yet, some are easier said than done. For instance, the “human-in-the-loop” process when making decisions regarding a diagnosis, and the shared accountability of all stakeholders in the implementation of AI solutions, raise questions when projected into the routine setting. Having a human in the loop is indeed all well and good, but what will happen if the AI system proves to be more trustworthy? Past experience with computer-aided diagnosis, which is a separate process from machine learning, showed troubling instances where radiologists largely ignored the correct computer prompts [23]. Again, at this stage, these are theoretical questions, but they ought to be considered part of any clinical implementation of such technology in our practice.

In conclusion, AI is here to stay, and with-it, NM will likely thrive in the foreseeable future. However, AI does not come without a cost. Several preconditions need to be met before AI is able to show its full potential. By its very nature, nuclear medicine is ready to adapt to the necessary changes, and the EANM will fully support the education on AI and the efficient implementation of AI at all levels in NM.