ESTRO 2025 - Abstract Book

S2788

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

ESTRO 2025

ED values well below the average in the rectum and small bowel correspond to patients with a large amount of air in these organs. The upper outlier densities in femoral heads correspond to the same patient. As the small bowel is far from the target, a prostate dosimetry (37.25Gy/5 fx) was planned on the CT with the most unfavourable rectum and on the patient with the highest femoral head values. The patient whose rectum had the lowest ED had a D 95% of the prostate of 39.813Gy when the structures had their own mean EDs and 39.834Gy when the rectum ED was overwritten by the ED value obtained in the first part of the study. The D 2cc of bladder and rectum differed by 0.8cGy and 6cGy respectively. For the patient with outlier femoral head EDs, D 95% values differed by 3.3cGy, while for D 2cc for bladder and rectum the differences were 9.3cGy and 1.1cGy respectively. Conclusion: For prostate cancer patients to be treated with MR-LINAC, we can guarantee that using in-house average EDs, creating an MR-only workflow, does not significantly increase the uncertainty in the calculation. Patients with prostheses or foreign bodies,should undergo a CT scan to clearly delineate these objects and related artifacts. Digital Poster Isodose Prediction and Cardiac Dose Assessment Using Neural Networks to Improve Breast Cancer Treatment Martina Mori, Maria Giulia Ubeira Gabellini, Gabriele Palazzo, Alessia Tudda, Cecilia Riani, Paola Mangili, Antonella Del Vecchio, Claudio Fiorino Medical Physics, IRCCS San Raffaele Scientific Institute, Milan, Italy Purpose/Objective: Post-surgery radiotherapy (RT) for breast cancer (BCa) using tangential fields (TF) is an efficient technique, still delivered in most patients/Institutes. However, heart may receive clinically relevant dose in a fraction of patients. This study presents a rapid, CT-based tool to predict heart dose, allowing for proactive selection of alternative techniques (i.e.: breath-hold). Material/Methods: A U-Net model, selected from the MONAI library (v1.3), was previously developed in Python for automatic breast segmentation using data of 861 patients of our Institute; it was successfully trained, with excellent Clinical Target Volume (CTV) contour prediction (Dice=0.90, consistent with measured inter-observer variability). The same model was adapted to predict 20Gy and 36Gy isodose curves, considered as structures shaped to the CTV. In an internal cohort of 264 left-sided BCa patients treated with TF, the isodose curves were converted into structures using a script running on MIM Assistant package. The original U-net was trained on 164 patients and validated on 100. The network was finally tested on a different dataset of 34 patients, comparing predicted vs. planned isodoses using Dice coefficients. Additionally, boolean operations quantified the heart volume receiving 20Gy (H_V20Gy), both as planned (H_V20Gy_plan) and predicted (H_V20Gy_pred). Using the distribution of values of H_V20Gy_plan and mean heart dose (MHD) of each patient of test group, a threshold value (V20Gy_thr) was identified, corresponding to a MHD=3Gy, taken as a safe threshold, according to the Danish Breast Cancer Group (DBCG) Proton Trial [1]. Then, model's ability in identifying patients at risk was assessed by comparing H_V20Gy_pred against H_V20Gy_plan. Keywords: MR-LINAC, electron density, CT 1841