ESTRO 2024 - Abstract Book

S4419

Physics - Machine learning models and clinical applications

ESTRO 2024

For treatments with a typical 10 mm PTV margin, a possible gain of 1 mm, 2 mm, 3 mm and up to 4 mm were identified for 6.8%, 57.2%, 10.0% and 3.2% of patients respectively. For 7-8 mm typical PTV margins, 44.4% of the patients could have benefited from a 2 mm margin reduction. Meanwhile, several patients with insufficient margins were identified based on their intra-fraction motion measured with AlignRT.

For the four different cross-validation iterations, the predictivity of the CIT-based motion groups was strong ranging between 97.8% and 98.9%.

Conclusion:

A strong correlation between clinical/technical parameters and SGRT data of intra fraction motion amplitude was identified allowing for individualized PTVs with up to 4 mm margin reduction.

Keywords: PTV personalization, SGRT, Big Data

392

Proffered Paper

Dose predictions in all IMPT robustness scenarios with a single deep learning model

Hazem A.A. Nomer 1 , Franziska Knuth 1 , Joep van Genderingen 1 , Dan Nguyen 2 , Margriet Sattler 1 , Uwe Oelfke 3 , Steve Jiang 2 , Linda Rossi 1 , Ben Heijmen 1 , Sebastiaan Breedveld 1 1 Erasmus MC Cancer Institute, University Medical Center, Department of Radiotherapy, Rotterdam, Netherlands. 2 UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, Dallas, USA. 3 The Institute of Cancer Research and The Royal Marsden NHS Foundation Trust, Joint Department of Physics, London, United Kingdom

Purpose/Objective:

Intensity modulated proton therapy (IMPT) is susceptible to patient setup and proton range uncertainties. IMPT treatment planning uses robust optimisation to obtain adequate Clinical Target Volume (CTV) coverage in a set of predefined scenarios, at the cost of extended planning time. Deep learning (DL) has demonstrated to be capable of predicting realistic dose distributions in the order of seconds, but has so far not been explored for dose prediction for individual robustness scenarios. We investigated three strategies for using DL for dose prediction in all scenarios, for the chosen robustness settings, in head-and-neck (HN) cancer robust IMPT treatment planning, enabling direct assessment of plan robustness.

Material/Methods: