ESTRO 2023 - Abstract Book

S145

Saturday 13 May

ESTRO 2023

oncology : journal of the European Society for Therapeutic Radiology and Oncology, 122(3), 406–410. https://doi.org/10.1016/j.radonc.2016.12.016 [8] Thwaites, D. I et al (1992). A dosimetric intercomparison of megavoltage photon beams in UK radiotherapy centres. Physics in medicine and biology, 37(2), 445–461. https://doi.org/10.1088/0031-9155/37/2/011

Symposium: Preparing for, and handling of, automated brachytherapy treatment planning in clinical practice

SP-0207 The use of data repositories for developments in brachytherapy automatisation L. Tagliaferri 1 , A. Damiani 2

1 Fondazione Policlinico Universitario "A. Gemelli" IRCCS, Dipartimento di Diagnostica per Immagini, Radioterapia Oncologica ed Ematologia, UOC di Radioterapia Oncologica, Rome, Italy, Italy; 2 Fondazione Policlinico Universitario "A. Gemelli" IRCCS, Dipartimento di Diagnostica per Immagini, Radioterapia Oncologica ed Ematologia, UOC di Radioterapia Oncologica, Rome, Italy Abstract Text Modern Interventional Radiotherapy (brachytherapy, BT, IRT) is characterized by a patient-centered approach (PCA). Indeed, it is necessary to consider in clinical practice on the one hand the classic oncologic outcomes but on the other hand also the patient's preferences with a focus on quality of life (Porter 2010, Lievens 2019). In this modern scenario, Artificial Intelligence (AI) can help the clinical decision-making process in each step of the IRT workflow considering the different fields of application (Fionda, 2020). Regarding the role of modern technology in facilitating repetitive tasks and therefore optimizing time, it is important to highlight that an optimized process could impact several steps, in particular: implant, delineation and planning. Artificial Neural Networks (Millar 2001) and Deep Learning machines (Boussion 2021, Han 2021) could be very useful to optimize the pre-treatment analysis, to suggest the optimal sources arrangement or the best applicator selection (Stenhouse, 2021), also considering 3D printing technology (Arenas 2017). However, the greatest impact in saving time is probably in the fields of target and organ-at-risk contouring (Suchanek, 2020, Lancellotta 2022) and in the treatment planning (Jaberi 2017). Finally, it is important also to consider the role of OMICS-analysis and the impact in the clinical practice, so automated process should be encouraged to offer a personalized PCA (Scott 2021). We can find several benefits in implementing automated procedures AI based but, to achieve a significant impact in the clinical data, it is important to have a large amount of data (large database) for analysis, implementing and validating the algorithms. There are several models of data repositories and data sharing. At the hospital level, available data can be organized in an infrastructure called “Data Lake”, in which the original information content, and relations among different data items, are preserved, without any particular provision for any given use case. In this way, at a lower level, a specific data mart can be built for every specific use. For instance, a complex project can have its own data mart, while other simpler projects, maybe on the same pathology, share a common data mart for the sake of simplification and reusability (Damiani 2021). Patient privacy and hospital intellectual property protection, also assured by pseudonymization of data, are combined with continuous data quality control, which can be automated to some extent, for instance when checking that no foreign values are included. A check on data distributions is also possible, and keeping the incidence of missing data under control is of the greatest importance. These are the pillars on which the first segment of an integration framework is built: “traditional” clinical data collected during ordinary clinical practice (Real World Data) are combined with ad-hoc data, specific for the single project, and with -omics data, coming from images (radiomics) and other fields (proteomics, genomics …). The quality level of learned models can be improved by increasing the number of available patients. This often requires the inclusion of more institutions in the project, which in turn poses the problem of data sharing. It is often a blocking issue, in which patient privacy and hospital intellectual property are strictly connected, with the consequence that the complexity of the study is increased and so are the ethical authorizations and legal requirements for the establishment of a research consortium. A solution to this issue can be found in the Federated Learning approach [DAMIANI 2018] in which patient data remain in the institutions that have collected them, while the same model can be learned with an iterative approach, as if all data had been collected in one place. On the same line of privacy preservation, the sandbox is a protected computation environment, managed by hospital personnel, with access control for external users, offering researchers the possibility of learning models without actually seeing, or creating a personal copy of, the data. Regarding the data quality assurance process, we can classify the clinical data in “Not organized, not ‘ontologized’ data”, “Organized, not ‘ontologized’ data”, and “Organized and ‘ontologized’ data”. (Marazzi, 2021). Each data item, based on its properties, can be automatically collected, and analyzed using a dedicated approach. Within some limits, some quality checks can also be automated. In conclusion, the use of AI can be very useful in the analysis and implementation of automated procedures in IRT but it is mandatory to have a suitable data storage system that in one hand should be useful with the aim of the analysis but in the other hand needs to be compliant with national and international regulations and should guarantee the quality assurance in data storage, usage and prospecting validation and improving of the algorithms. Abstract Text In the light of this year’s congress theme, “ From innovation to action ”, let us embrace artificial intelligence (AI) to transform the field of radiation oncology, specifically brachytherapy, and reflect on how this will affect our traditional way of working and thinking. In this talk I will first paint a rough picture of where and how AI can play a role in the brachytherapy workflow. I will furthermore shortly touch upon the fact that there are different types of AI. Next, I will dive a bit deeper into two different AI-based approaches (i.e., protocol-based or example-based) for brachytherapy treatment planning and discuss how each will in a different way affect the traditional way of treatment planning. Specifically, I will address development/tailoring of the approach, updating the approach in case of e.g., protocol changes, how to decide which AI-based brachytherapy treatment plan to use for a new patient, and how to interpret the obtained quality of such a plan. This talk will primarily be based on previous and ongoing efforts by my own research group at the Leiden University Medical Center (LUMC) in SP-0208 How artificial intelligence will change the traditional way of treatment planning T. Alderliesten 1 1 Leiden University Medical Center (LUMC), Radiation Oncology, Leiden, The Netherlands