ESTRO 2025 - Abstract Book

S3808

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

MRSI. Future work could benefit from spatial regularization, as neighboring voxels are likely to have similar relapse probabilities.

Keywords: MRSI, Glioblastoma, Relapse prediction

References: 1. Petrecca, K., et al. "Failure pattern following complete resection plus radiotherapy and temozolomide is at the resection margin in patients with glioblastoma." Journal of Neuro-Oncology 111 (2013): 19-23. 2. Labriji, W., et al. "Bayesian Sparse Model for Complex-Valued Magnetic Resonance Spectroscopy Restoration." In Proceedings ISBI, 2024. 3. Chawla, N. V., et al. "SMOTE: synthetic minority over-sampling technique." Journal of Artificial Intelligence Research 16 (2002): 321-357. 4. Zhu, Z., et al. "Glutathione reductase mediates drug resistance in glioblastoma cells by regulating redox homeostasis." Journal of Neurochemistry 144.1 (2018): 93-104.

3011

Mini-Oral Supervised Learning or Reinforcement Learning for Intensity Modulated Radiation Therapy (IMRT) Treatment Planning? A Head-and-Neck Study Dongrong Yang, Xin Wu, Xinyi Li, Yibo Xie, Qiuwen Wu, Jackie Wu, Yang Sheng Radiation Oncology, Duke University Medical Center, Durham, USA Purpose/Objective: There has been rising interest and development in the last decade about improving planning efficiency or automating the treatment planning workflow. This study aims to evaluate and compare the efficacy of two popular yet distinctive machine learning-based approaches for automating head-and-neck (HN) intensity-modulated radiation therapy (IMRT) treatment planning. Material/Methods: Two autoplanning frameworks were studied in this work. The first, an in-house convolutional neural network-based framework (CNN-Auto), was developed and commissioned at our institution in 2022 1 . Trained in a supervised manner, CNN-Auto takes patient anatomy information and beam configuration as inputs and learns a mapping relation to generate the fluence map intensity as the output. Training and prediction were performed on local workstations, then the output fluence maps were imported into the clinical treatment planning system (TPS) for dose calculation and plan quality evaluation. The second approach involves a deep reinforcement learning based framework (DRL-Auto) trained to optimize objective constraints during the inverse planning process 2 . This framework learns iteratively through trial-and-error constraint adjustments, leveraging feedback from clinical dose criteria to autonomously refine treatment plans. Compared to supervised approaches trained on hundreds of expert-generated plans, this method eliminates the need for human labeling, instead relying on autonomous learning through trial-and-error optimization. Furthermore, the DRL-Auto framework operates within the inverse planning space, eliminating the need for fixed beam geometry as model input. This flexibility enables DRL-Auto to provide a more adaptable and comprehensive solution for automated IMRT plan generation. To evaluate the efficacy of CNN -Auto compared to DRL-Auto, we curated a dataset consisting of 20 HN cancer patients who underwent IMRT at our institution. Both frameworks were tasked with generating treatment plans with 44Gy prescription in 22 fractions. The quality of both plan groups was assessed using several clinically relevant dosimetric endpoints. They were also compared against clinical plans that were manually generated.

Results: The DRL-Auto achieved comparable plan quality as CNN-Auto plans, as well as manually generated clinical plans