ESTRO 2025 - Abstract Book

S3375

Physics - Machine learning models and clinical applications

ESTRO 2025

PTV Breast . VMAT 200 offered the best OAR sparing, lowering mean doses to the heart, contralateral breast, and ipsilateral lung (0.6 Gy vs 0.8 Gy, 0.8 Gy vs 1.1 Gy, and 4.1 Gy vs 4.6 Gy, respectively). Median MU was lowest for VMAT 200 (655 MU) and highest for IMRT 400 (1307 MU). Conclusion: This automated segment reduction technique generated individualized gantry angles that improved dose quality across treatment methods. It combined the use of AI in dose prediction and the dose mimicking algorithm to consistently produce DL plans showing optimal dose distributions and delivery efficiency according to the treatment method. Furthermore, it can be seamlessly integrated into the existing DL pipeline to extract beam angles prior to executing the mimicking process.

Keywords: DL autoplanning, Breast, Beam angle optimization

1231

Mini-Oral Multi-modality learning-based dose prediction and automatic lung dose painting for functional lung avoidance radiotherapy Tianyu Xiong 1 , Guangping Zeng 1 , Zhi Chen 1 , Yu-Hua Huang 1 , Bing Li 2 , Dejun Zhou 1 , Yang Sheng 3 , Ge Ren 1 , Jackie Wu 3 , Hong Ge 2 , Jing Cai 1 1 Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China. 2 Department of Radiation Oncology, The Affiliated Cancer Hospital of Zhengzhou University and Henan Cancer Hospital, Zhengzhou, China. 3 Department of Radiation Oncology, Duke University Medical Center, Durham, NC, USA Purpose/Objective: Functional Lung Avoidance Radiotherapy (FLART) shows promise in reducing radiation-induced pulmonary toxicity. 1 Lung dose painting refers to FLART planning utilizing voxel-wise lung function information. 2 This study aims to develop a multi-modality-guided dose prediction (MMDP)-based automatic lung dose painting algorithm to achieve auto-panning for FLART. Material/Methods: We retrospectively collected data from 196 lung cancer patients across three institutions, comprising 114/28 cases with lung ventilation (V) surrogate maps for training/validation and 21/33 cases with SPECT V/perfusion (Q) images for testing. A meta-optimization (MO)-based auto-planning algorithm 3 was employed to generate conventional radiotherapy (MO-ConvRT) and voxel-based FLART (MO-FLART) plans. High-quality plans were selected to train a novel MMDP model depicted by Figure 1 (a). An innovative instance weighting anatomy-to-function training strategy, illustrated by Figure 1 (b), was tailored to enhance prediction accuracy. Two ablation studies were conducted to validate the effectiveness of the training strategy and multi-modality learning. A function-guided dose mimicking algorithm was developed to convert predicted dose distributions into FLART (DL-FLART) plans, which were compared against MO-FLART, manual conventional radiotherapy (MA-ConvRT) and FLART (MA-FLART) plans.