ESTRO 2025 - Abstract Book

S3372

Physics - Machine learning models and clinical applications

ESTRO 2025

[12] Sekuboyina et al., Med. Imag. Anal. 73: 102166 (2021) 1048

Digital Poster Deep learning monitor units per control point prediction for automated VMAT treatment planning in prostate cancer Mathieu Gaudreault 1,2 , Vanessa Panettieri 1,2,3 , Lachlan McIntosh 1 , Katrina Woodford 4 , Jason Li 5 , Susan Harden 2,4 , Sandro Porceddu 2,4 , Nicholas Hardcastle 1,2,6 1 Physical Sciences, Peter MacCallum Cancer Centre, Melbourne, Australia. 2 Sir Peter MacCallum Department of Oncology, the University of Melbourne, Melbourne, Australia. 3 Central Clinical School, Monash University, Melbourne, Australia. 4 Radiation Oncology, Peter MacCallum Cancer Centre, Melbourne, Australia. 5 Bioinformatics, Peter MacCallum Cancer Centre, Melbourne, Australia. 6 Centre for Medical Radiation Physics, University of Wollongong, Wollongong, Australia Purpose/Objective: In volumetric modulated arc therapy (VMAT) treatment, the delivered radiation dose is modulated to maximise the dose to the tumour and minimise the dose to healthy organs [1]. The modulation pattern is determined from a dose optimisation algorithm that requires several iterations to achieve a satisfactory solution [2]. The monitor units (MU) per control point (CP) result from dose optimisation and drive the dose magnitude. Deep learning is a type of artificial intelligence that makes predictions once trained [3]. We investigate the feasibility of deep learning MU per CP prediction for automated VMAT treatment planning in prostate cancer. Material/Methods: Consecutive patients treated for prostate cancer with 60 Gy in 20 fractions in a single institution between 01/2019 and 06/2024 were considered. The VMAT plans had two arcs of 360 o spaced every 2 o . The inputs were either three dimensional (3D) dose distributions per CP or two-dimensional (2D) dose intensity projections averaged along the beam’s -eye view. Four strategies were considered and compared (Fig. 1). In 2D/3D single-model, one model was trained using all 2D/3D samples. In 2D/3D multi-model, one model was trained per CP (360 models in total) using 2D/3D samples of each CP. The outputs were the MU per CP, which were converted to meterset weight per CP and MU per beam to create a radiotherapy plan (AI-RTPlan). The AI-RTPlans were imported into the treatment planning system for dose calculation. The dose obtained with the AI-RTPlan was compared to the clinical dose with 3%/3mm gamma passing rate [γPR(3%,3mm)] and clinical objectives.