ESTRO 2021 Abstract Book

S452

ESTRO 2021

exploit tumour type-specific therapeutic opportunities. Against this background, the potential value of combining DNA damage response inhibition (with the ataxia telangiectasia and Rad3-related inhibitor, AZD6738/ceralasertib) will be discussed. The effects of ATR inhibition on the tumour microenvironment in multiple pre-clinical models will be presented, in addition to data that illustrate the potential value of combining RT and ATR inhibition with ICPI. In particular, the presentation will highlight the opportunity for modulating natural killer (NK) cell biology through the addition of anti-PD-1 and other novel ICPI agents. Data from immune competent animal models and a phase I study of ceralasertib as a single-agent and in combination with RT (NCT02223923) will be discussed. These data demonstrate NK-mediated superior tumour control and protection against tumour re-challenge. Future perspectives on possible clinical applications of these findings will be presented. SP-0578 Opportunities, challenges and added-value of multi-institutional knowledge-based models R. Castriconi 1 1 IRCCS San Raffaele Scientific Institute, Medical Physics , Milano, Italy Abstract Text The increasing complexity of radiotherapy and the high dimensionality of plan optimization problems has made it challenging to efficiently generate consistent and high quality plans. In order to reduce the inter- operator variability and to spare the time for planning, automatic planning systems were introduced. Within this scenario, there has been significant progress in the development of data-driven treatment planning approaches utilizing the knowledge from the past to predict the outcome of similar but new tasks. The Knowledge-based (KB) method is a powerful tool to collect planning experiences into models able to classify patients according to their individual anatomical features and possibly to automatically drive optimization on new patients. A major value of this approach is the possibility of accounting for the past clinical experience. On the other hand, the performance of any KB approach is highly dependent on the quality of the available plans used during the modelling phase, making the resulting models representative of the clinical experience of the center. Therefore, a KB-model cannot fully guarantee that the ‘best’ plan is achieved. While plan quality may largely vary between planners of the same institute, much larger variations can be expected among different clinical institutions. A major limitation of KB-approaches is that inter-institute variability is not considered, making the possibility of applying a KB model outside the generating center challenging. Quantifying inter-institute planning variability and extending the KB-method on multi-institute scale could have huge impact and relevant applications. Different initiatives focusing on the multi-institute purpose have been proposed in recent years. Importantly, the commercialization of KB tools (first, the RapidPlan Varian system) may facilitate studies on multi-center scale and help the sharing of the models between centers. It is already well established that the KB-approach could identify systematic differences in planning performances between campuses using the same planning system, guidelines and protocols. Other studies demonstrated that KB-models developed by cross-institutional groups using large-scale data may standardize the plan quality within clinical trials. Recent publications demonstrated the possibility of sharing models between institutes. However, in order to share models, it is necessary to assess the degree of plan transferability between different Institutes, first considering the acceptability of inter-Institute PTV and OARs segmentation variability. One of the first big challenging steps of any KB multi-institutional approach is the pooling of libraries of consistent plans and the implementation of methods to incorporate the inter-institute variability into plan prediction. Of note, there is no well-established workflow or strategy for harmonization referred to KB model creation while these issues critically influence the final result when merging/pooling different institutional KB models. Another relevant aspect concerns the impact of the dimension of the training set on the robustness of KB-model predictions. Not all institutions can collect a sufficient number of high-quality plans to build their own robust KB-models. This is particularly important for more rare diseases as an adequate number of patients to develop reliable models may only be possible across several centers. Multi-institutional initiatives would have the advantage of helping in establishing consistent criteria of modelling and proper cross-validation methods to further score similarities between different populations in terms of contouring/anatomy. National cooperative groups are actively working in the field and provided first evidence of the possibility to develop multi-institutional KB-models. Relevant examples prevalently involve more ‘standardized’ clinical situations, such as prostate and breast cancer. Concerning this last example, first results from the Italian project ‘MIKAPOCo’ (Multi-Institutional Knowledge-based Approach in Planning Optimization for the Community), are available. The project involves seven institutes with KB expertise: preliminary results demonstrate the feasibility of quantifying the inter-institute variability of DVH prediction models in the case of whole-breast irradiation with tangential fields suggesting a sufficiently high inter-institute interchangeability. Although in its infancy, the field of multi-Institutional KB-plan prediction is rapidly growing and is expected to have large impact on the community in the next years. Libraries of interchangeable multi-Institutional KB-plan prediction models have the potential of assisting large-scale implementation of KB- automated planning at national (and international) level as well as to develop tools and services for plan audit/QA, supporting clinical trials and optimizing plan benchmark assessment. Symposium: Next steps in automated treatment planning

SP-0579 Deep learning in planning – Knowledge-based planning or more? E. Sterpin 1 1 KU Leuven, Oncology, Leuven, Belgium

Abstract Text