ESTRO 2025 - Abstract Book

S2902

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

ESTRO 2025

Conclusion: New a/b have been derived that are consistent with published dose constraints, yielding different gynaecological dose limits. Depending on the boost fractionation scheme, this can bring more or less boost dose delivery margin. Serious discussion within the department is ongoing before clinical implementation of the new a/b values.

Keywords: alfa beta ratio, dose constraints, gynecology

References: Timmerman R.A. Story of Hypofractionation. Int J Radiat Oncol Biol Phys. 2022 Jan 1;112(1):4-21.

3877

Digital Poster A Deep Learning-Based Predictive Model: Can AI Accurately Predict VMAT Dose Distributions in Lung Cancer? Thitaporn Chaipanya 1 , Kampheang Nimjaroen 2 , Sasikarn Chamchod 1,2 , Panatda Intanin 2 , Patiparn Kummanee 1 , Dhammathat Owasirikul 3 , Chirasak Khamfongkhruea 1,2 1 Princess Srisavangavadhana College of Medicine, Chulabhorn Royal Academy, Bangkok, Thailand. 2 Radiation Oncology Department, Chulabhorn Hospital, Chulabhorn Hospital, Bangkok, Thailand. 3 Radiological Technology Department, Kanchanabhishek Institute of Medical and Public Health Technology, Nonthaburi, Thailand Purpose/Objective: Treating lung cancer, especially large tumors, is challenging due to the need for effective tumor control while minimizing radiation to healthy tissue. The process of developing optimal radiotherapy plans is complex and time intensive, involving critical decisions, such as choosing between combined chemoradiotherapy (CCRT) and chemotherapy alone. Artificial intelligence (AI)-based dose prediction has emerged as a promising solution [1,2], offering faster and more efficient optimization of treatment planning. To address these challenges, this study developed a deep learning-based dose prediction model for volumetric modulated arc therapy (VMAT) to streamline planning for locally advanced non-small cell lung cancer (NSCLC). Material/Methods: This study retrospectively analyzed 72 NSCLC cases treated with VMAT from 2015 to 2024, following National Comprehensive Cancer Network guidelines. Patients were grouped into prescription doses of 50 Gy, 54 Gy, and 60 Gy. CT images, contours, and dose distributions were preprocessed into normalized 128 × 128 × 128 matrices. Four 3D U-Net models were trained: Model 1 (50 Gy), Model 2 (54 Gy), Model 3 (60 Gy), and Model 4 (combined 50 Gy and 60 Gy). Training used 70% of data, with 10% for validation and 20% for testing, employing data augmentation and hyperparameter tuning (batch size 4, learning rate 0.0001, 500 epochs). Prediction accuracy was assessed using Mean Absolute Error (MAE), Homogeneity Index (HI), and dose-volume histogram (DVH) analysis. Results: All models accurately predicted VMAT dose distributions (Fig.1), particularly for PTV metrics, with MAE values ranging from 0.48 to 4.01 Gy. Homogeneity Index (HI) predictions aligned closely with ground truth (1.12–1.15), ensuring dose uniformity. DVHs confirmed strong agreement in high-dose regions, but low-dose predictions for OARs like the spinal cord and esophagus showed deviations. Statistically significant differences (p < 0.05) were observed in some OAR metrics, indicating areas for improvement while highlighting the model's clinical potential.