ESTRO 2025 - Abstract Book

S2708

Physics - Dose calculation algorithms

ESTRO 2025

A protocol was established with the following recommendations: in vivo dosimetry in all breast patients, in the first three fractions, and once a week thereafter; creation of a task in the oncology information system (ARIA TM ) to report the failures analysis and feedback. Further investigation is due to establish corrective future actions and extension to other treatment sites.

Keywords: In Vivo Dosimetry, Breast Radiotherapy

References: - E. Bossuyt, R. Weytjens, D. Nevens, S. De Vos, and D. Verellen, “Evaluation of automated pre-treatment and transit in-vivo dosimetry in radiotherapy using empirically determined parameters,” Phys Imaging Radiat Oncol , vol. 16, pp. 113–129, Oct. 2020, doi: 10.1016/J.PHRO.2020.09.011. - I. Lage, "SUNCHECK™ TRENDING - Longitudinal Registration of Machine and Patient Quality Assurance Tests", 2023 - N. Dogan et al. , “Use of electronic portal imaging devices for pre ‐ treatment and in vivo dosimetry patient ‐ specific IMRT and VMAT QA: Report of AAPM Task Group 307,” Med Phys , Aug. 2023, doi: 10.1002/MP.16536.

4222

Digital Poster Comparing sequential DNN models in convolutional encoder-decoder framework for proton dose calculation Lina S Bucher 1 , Ahmad Neishabouri 2 , Luke Voss 1 , Niklas Wahl 1 1 Division of Medical Physics in Radiation Oncology, German Cancer Research Center –Dkfz, Heidelberg, Germany. 2 Heidelberg Institute for Radiation Oncology – HIRO, Heidelberg Ion-Beam Therapy Center (HIT), Heidelberg, Germany Purpose/Objective: Precise and rapid dose calculation, fundamental for modern radiation therapy, is experiencing an upsurge of new academic approaches applying neural networks. The goal: reaching Monte Carlo (MC) simulation accuracy with millisecond prediction speeds for spot-wise dose calculation. This study focused on comparing different sequence based deep neural networks (DNNs), the LSTM 1 (Long Short-Term Memory), Transformer 2 and RetNet 3 (Retentive Network) model, in an unified convolutional encoder-decoder. Performance as well as prediction and training times were compared. Aiming to further reduce computation times, the effect of complexity reduction of the encoder decoder structure was analyzed. Material/Methods: This study was conducted on 3D water-box phantom and lung patient CTs with corresponding single pencil beam 104.25MeV MC simulated dose distributions, split into 2D image sequences of length 80 and 150 4 . Each model was trained individually over a preset number of epochs using adaptive optimizers and learning rate (LR) schedulers for back-propagation. The mean squared error (MSE) loss was computed on training and unseen validation data for each epoch to track model training. For further performance evaluation gamma-analysis [1%, 1.5mm] was carried out on single-image predictions to include spatial deviations overseen by the MSE loss. The encoder-decoder architecture was inspired by the deep-learning based millisecond speed dose calculation algorithm (DoTA) 5 . In this study a simplified version was developed by reducing the number of convolutional layers and incorporating dimensionality reduction, previously performed by additional pooling layers, directly into the existing convolutions. Results: In the phantom case study, the Transformer model performed best out of the three models, with the lowest MSE (3x10 -6 Gy 2 ), highest mean gamma-pass-rate (99.9%) and lowest inference time (68ms, GPU: RTX A5000) among the three models. Results of the patient case study showed little difference between the models, each performing with