ESTRO 2024 - Abstract Book

S4447

Physics - Machine learning models and clinical applications

ESTRO 2024

intensities. The AI models are more complicated than in the model set approach and take longer time to train and predict.

Results:

Both approaches demonstrated promising plan quality and consistency with assigned priorities. Meanwhile, both approaches can generate multiple AI plans with variable dosimetric tradeoffs, which offer physicians more options for customizing patient treatment. Compared to the model set approach, the modeling approach achieved lower maxima dose and less target dose heterogeneity.

Conclusion:

Both approaches enabled dosimetric tradeoff in AI planning. The model set approach features better integration into clinical applications in the form of protocol selection as well as easy integration of more tradeoff options. The modeling approach features continuous dosimetric tradeoff changes and a greater potential to be generalized to other treatment sites.

Keywords: Machine learning, auto planning, AI

892

Proffered Paper

AI informed discussion on how dose outside the PTV affects distant failure in SBRT NSCLC patients

Denis Dudas 1,2 , Thomas J Dilling 3 , Issam El Naqa 2

1 Czech Technical University in Prague, Faculty of Nuclear Sciences and Physical Engineering, Prague, Czech Republic. 2 H. Lee Moffitt Cancer Center & Research Institute, Department of Machine Learning, Tampa, USA. 3 H. Lee Moffitt Cancer Center & Research Institute, Department of Radiation Oncology, Tampa, USA

Purpose/Objective:

In recent years, several publications have described the impact of lung SBRT dose outside the PTV on the probability of distant metastasis (DM) in early-stage non-small cell lung cancer (NSCLC) patients. Unfortunately, their conclusions were conflicting, leaving multiple important questions unanswered. One study, by Diamant et al. [1,2], reported lower rate of DM for patients with higher dose delivered to a 3cm margin around the PTV. However, a study by Hughes et al. [3] did not confirm Diamant’s results, and a later publication by Lalonde et al. [4] presented higher rate of DM for patients with higher dose delivered to a 3cm margin around the PTV. In this work, we provide an independent analysis of a large institutional patient cohort, using statistical and machine learning outcome modelling with artificial intelligence (AI) explainability techniques (xAI), which allow us to reconcile and explain these earlier conflicts in the context of treatment planning with xAI.