ESTRO 2024 - Abstract Book

S3618

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

Keywords: Hippocampus, radiosurgery, benign

References:

1. Pazzaglia S, Briganti G, Mancuso M, Saran A. Neurocognitive Decline Following Radiotherapy: Mechanisms and Therapeutic Implications. Cancers (Basel). 2020;12(1):146. Published 2020 Jan 8. doi:10.3390/cancers12010146

2. Brown PD, Gondi V, Pugh S, et al. Hippocampal Avoidance During Whole-Brain Radiotherapy Plus Memantine for Patients With Brain Metastases: Phase III Trial NRG Oncology CC001. J Clin Oncol. 2020;38(10):1019-1029. doi:10.1200/JCO.19.02767 3. Brown PD, Ballman KV, Cerhan JH, et al. Postoperative stereotactic radiosurgery compared with whole brain radiotherapy for resected metastatic brain disease (NCCTG N107C/CEC·3): a multicentre, randomised, controlled, phase 3 trial. Lancet Oncol. 2017;18(8):1049-1060. doi:10.1016/S1470-2045(17)30441-2

2008

Digital Poster

A Generative Adversarial Network for Radiotherapy Dose Predictions of Head and Neck Cancers

Jacob S Buatti 1 , Alexandre Cafaro 2 , Sruthi Sivabhaskar 1 , Kristen Duke 1 , Nikos Papanikolaou 1 , Neil Kirby 1 , Nikos Paragios 3 1 University of Texas Health Science Center, Radiation Oncology, San Antonio, USA. 2 Paris-Saclay University, Computer Science, Paris, France. 3 TheraPanacea, AI, Paris, France

Purpose/Objective:

Current dose prediction architectures achieve excellent global dose and point dose agreement (i.e. mean absolute error, mean dose, and max dose). However, some deep learning architectures used for dose predictions lack fine regional details in the predicted dose volumes which are characteristic of linear accelerator gantry motion, continuous modulation of the multi-leaf collimators (MLC), modulated beam fluence, and heterogeneity of tissue. The purpose of this study was to explore a generative adversarial network (GAN) for generating high precision dose predictions for head and neck radiotherapy treatments. Specifically, we elaborated recently developed architectures by incorporating a discriminator during model training with goal of improving coarse prediction details.

Material/Methods:

607 HNC patients who were treated with Intensity Modulated Radiation Therapy (IMRT) from a linear accelerator were selected for training and evaluating model performance. We used a PatchGAN network consisting of a 3 dimensional cascade U-Net motivated by Lie et al1 and a patch-based discriminator motivated by Radford et al.2 The generator input was a 12-channel volume consisting of (1) the planning computed tomography (CT) image, (2) a planning target volume (PTV) mask labeled with each PTV’s prescriptions, and (3-12) 10 organ at risk (OAR) contour masks for the brainstem, spinal cord, esophagus, thyroid, oral cavity, mandible, total parotids, total