Abstract: Over the last century radiotherapy (RT) has had a remarkable success in cancer treatment. Decades of research has led to novel machinery and procedures – such as image guidance or charged particle treatments – that allow targeting tumors with unprecedented precision while protecting healthy tissues as much as possible. Computational tools and sophisticated mathematical modelling played a crucial role in this success and Artificial Intelligence (AI) methods are increasingly the key drivers behind further improvements, holding the promise for breakthroughs via enabling (real-time) adaptive and ultimately biologically optimized combined modality treatments.

In this presentation I will focus on 3 recent research topics showcasing how computational and AI methods can improve modern RT treatment planning workflow. First, I will present the results of using our Polynomial Chaos methodology to perform a comprehensive robustness analysis of clinical proton treatments, using retrospective patient data from all 3 Dutch proton therapy centers. This multicenter study revealed that the DUPROTON Dutch national robustness evaluation protocol is consistent and safe, but is also overconservative, leaving room for individual improvements via both dose escalation and de-escalation. Second, I will describe a semi-supervised autoencoder framework that can jointly classify and generate realistic breathing signals, allowing the accurate analysis of interplay effects in pencil beam scanned proton therapy and revealing the clear superiority of 4DCT robustly planned treatments over the internal target volume approach. Last, I will present DoTA, our Dose Transformer Algorithm, a Deep Learning (DL) based convolutional transformer network that offers Monte Carlo quality dose calculations in millisecond speed, which can be key to enable online adaptive proton therapy and serve as a fast, reliable quality assurance for daily treatments.

About the speaker: Zoltán Perkó is an expert of computational science and deep learning, and their use in radiation applications. He received his master’s degree in physics from Budapest University of Technology and Economics in 2010 and his PhD with cum laude distinction from Delft University of Technology (TU Delft) in 2015. After 2 years as postdoctoral fellow at Massachusetts General Hospital and Harvard Medical School he established his own research group at TU Delft, pursuing unique research by leveraging state-of-the-art physics simulations with the power of artificial intelligence (AI), focusing on unsolved computational challenges in cancer care. Zoltán and his team are working on improving all aspects of the radiotherapy workflow, enhancing both traditional photon therapy and modern charged particle therapies, working towards clinically feasible online - and ultimately real-time – adaptive treatments offering maximal efficacy with minimal side-effects. Within the new Biological Intervention Optimization AI Lab he is also looking at developing AI tools to biologically tailor and personalize patient treatments via combining irradiations with chemo-, immuno- and other therapies.

Zoltán has successfully raised >1.7MEur funding for his research and his work is being regularly published in top journals. Via close collaborations with Dutch, European and international radiotherapy and proton centers his computational methods are already being used in clinical settings, enabling multi-center studies and having direct societal impact. His algorithms allowed deriving practical robustness recipes, ensuring safe and effective proton treatments, and his group’s latest result - an AI-based millisecond speed proton dose engine with Monte Carlo accuracy - represents the current state-of-the-art in proton transport calculations.