ESTRO 2025 - Abstract Book

S3418

Physics - Machine learning models and clinical applications

ESTRO 2025

3233

Digital Poster Combining different metastatic anatomical locations for the deep learning prediction of the monitor units per control point Vanessa Panettieri 1,2,3 , Lachlan McIntosh 1 , Katrina Woodford 4,5,6 , Jason Li 7 , Susan Harden 8,2 , Sandro Porceddu 8,2 , Nicholas Hardcastle 1,2,9 , Mathieu Gaudreault 1,2 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 Therapy Services, Peter MacCallum Cancer Centre, Melboure, Australia. 5 Sir Peter MacCallum Department of Oncology, Melbourne, Melbourne, Australia. 6 Medical Imaging and Radiation Sciences, Monash University, Melbourne, Australia. 7 Bioinformatics, Peter MacCallum Cancer Centre, Melbourne, Australia. 8 Radiation Oncology, Peter MacCallum Cancer Centre, Melbourne, Australia. 9 Centre for Medical Radiation Physics, University of Wollongong, Wollongong, Australia Purpose/Objective: In hypofractionated radiotherapy of oligometastases, a fine-tuned ablative dose is required to target the tumour while sparing adjacent healthy organs. Generating this personalised dose distribution is a laborious process that often needs iterative input from an experienced workforce extending treatment preparation [1]. Emerging deep learning (DL) automation has the potential to facilitate advanced treatment planning of multiple metastases [2] and single-visit treatment [3]. Among various strategies, as the dose magnitude is controlled by the monitor units (MU) per control point (CP) we can exploit DL for their prediction. However, due to the limited occurrence of radiation therapy treatment for metastases, it might be challenging to collect sufficient datasets for training. This study explores if combining anatomical locations for training can lead to accurate DL predictions of individual anatomical sites. Material/Methods: Consecutive patient plans (01/2019-06/2024) from a single institution treated for liver, bone, or lung metastatic cancer were considered. Predictions of models trained for each anatomical location ( single-site model ) were compared to predictions of a model trained with all anatomical locations ( multi-site model ) on unseen liver, bone, and lung datasets during training (Fig. 1). The predicted MU per CP were converted to meterset weight (MW) per CP and MU per beam to create an artificial intelligence radiotherapy plan (AI-RTPlan), which was imported into the treatment planning system for dose calculation. The predicted dosimetry was compared to the clinical dosimetry with 3%/2mm gamma testing rates and selected dose metrics.