ESTRO 2025 - Abstract Book

S3411

Physics - Machine learning models and clinical applications

ESTRO 2025

Conclusion: This work introduces a methodology for developing the CTS CTCAE and the CTS EORTC . Aggregating multiple toxicities into a single endpoint, CTS facilitates treatment comparisons and has potential to reduce sample sizes required to detect significant differences in toxicity after radiotherapy in clinical trials. Weighting toxicities based on their association with QOL, CTS engages patients’ experience in clinical trials design. External validation will be essential for broader application and generalizability of CTS.

Keywords: Clinical trials,toxicity endpoint,quality of life

References: 1) Pötter, Richard, et al. 2021. “MRI -guided adaptive brachytherapy in locally advanced cervical cancer (EMBRACE-I): a multicentre prospective cohort study.” The Lancet. Oncology 22 (4). 2) Kirchheiner K, et al. 2016. “Health -Related Quality of Life in Locally Advanced Cervical Cancer Patients After Definitive Chemoradiation Therapy Including Image Guided Adaptive Brachytherapy: An Analysis From the EMBRACE Study.” Int J Radiat Oncol Biol Phys. 1;94(5). 3) Vittrup AS, et al. 2023. “EMBRACE Collaborative Group. Overall Severe Morbidity After Chemo -Radiation Therapy and Magnetic Resonance Imaging-Guided Adaptive Brachytherapy in Locally Advanced Cervical Cancer: Results From the EMBRACE- I Study.” Int J Radiat Oncol Biol Phys. 15;116(4).

2900

Digital Poster A new centre customizable Deep Learning approach for Patient-Specific Quality Assurance outcome prediction Laura Verzellesi, Elisabetta Cagni, Giulia Paolani, Valeria Trojani, Matteo Orlandi, Roberto Sghedoni, Adriana Barani, Daniele Lambertini, Mauro Iori, Andrea Botti Medical Physics, Azienda USL-IRCCS di Reggio Emilia, Reggio Emilia, Italy Purpose/Objective: Patient-specific quality assurance (PSQA) is crucial for ensuring radiotherapy treatment accuracy and safety, yet it is resource-intensive. While existing literature links lower plan complexity with improved delivery accuracy, predicting when PSQA might fail remains challenging. Recent literature indicates that deep-learning models trained to predict PSQA failures in one institution might not perform as well elsewhere. Our research introduces a new deep-learning strategy trained on an innovative image-based representation of VMAT plans that encodes the inherent plan complexity. Encoding the plan information with an image has two advantages: first, it allows avoiding reliance on predefined complexity indices, and second, it enables the use of CNN models and the application of a transfer-learning strategy so that other centers could adapt the model to suit their specific realities and equipment. Material/Methods: We analyzed 1209 consecutive VMAT treatments, from 2023 to 2024, generated with Eclipse treatment planning system and delivered with a TrueBeam STx 2.1 with a High-Definition 120-leaf-MLC. PSQA was performed with Octavius-4D and 1000/1500 chamber-array, and discrepancies were quantified with the gamma analysis. We considered the gamma passing rate (GPR – 3%/3mm with a 50% isodose threshold) and employed the 25th percentile to determine PSQA pass (GPR≥93.80, outcome=1, N=914) or fail (GPR<93.80, outcome=0, N=295). Each DICOM-RT plan was transformed into a six-channel-image, encoding the MLC mechanical movement, dose rate, gantry's angular velocity, and time interval between control points (as shown in Figure 1).