S2380
Physics - Quality assurance and auditing
ESTRO 2026
criteria [3]. Results:
73–81 (2023). Keywords: TBI, VMAT, Quality Assurance, Field overlap
The results are highly promising. The average gamma passing rate across all PSQA test cases was 97 ± 4%, demonstrating strong agreement between the predicted and calculated PDs. Representative examples of predicted and reference dose distributions are shown in Figures 1 and 2.Figure 1
Digital Poster 3409
Pre-treatment Patient-Specific Quality Assurance (PSQA) Using EPID Devices: An innovative AI-Based Method Carlotta Mozzi 1,2 , Lorenzo Marini 3,4 , Marta Casati 5 , Francesca Lizzi 4 , Michele Avanzo 6 , Emmanul Uwitonze 4 , Icro Meattini 7 , Livia Marrazzo 5,1 , Alessandra Retico 4 , Aafke Kraan 4 , Cinzia Talamonti 1,2 1 Dipartimento di Scienze Biomediche Sperimentali e Cliniche, University of Florence, Florence, Italy. 2 INFN, Istituto nazionale di fisica nuclare, Florence, Italy. 3 Physics, univesity of Pisa, Pisa, Italy. 4 INFN, Istituto nazionale di fisica nuclare, Pisa, Italy. 5 Medical Physics Unit, Careggi Hospital, Florence, Italy. 6 Centro di riferimento oncologico, Ospedale, Aviano, Italy. 7 Radiotherapy unit, Careggi Hospital, Florence, Italy Purpose/Objective: The aim of this study is to develop and validate an innovative AI-based method for performing pre- treatment patient-specific quality assurance (PSQA) using Electronic Portal Imaging Device (EPID) data [1]. This approach aligns with the European EURATOM recommendations for accurate dose delivery. The proposed method employs a validated U-Net convolutional neural network to predict portal dose (PD) distributions at the EPID level [2], based on EPID images acquired during plan delivery in air with a fixed
Figure 2
Conclusion: This study demonstrates the feasibility of an AI-based approach for pre-treatment PSQA using EPID data. The proposed U-Net model accurately predicts portal dose distributions, showing excellent agreement with TPS- calculated doses. This innovative methodology can significantly enhance the efficiency and automation of PSQA procedures in radiotherapy.Future work will focus on training the U-Net exclusively with PSQA cases to further specialize the model for this application. Additionally, the model will be trained on patient treatment data to extend its use to in vivo dosimetry in IMRT configurations. References: [1] DOGAN N. et al. AAPM Task Group Report 307: use of EPIDs for patient - specific IMRT and VMAT QA. Medical Physics 2023, 50(8): e865–e903.[2] RONNEBERGER O., FISCHER P., BROX T. U-net: Convolutional networks for biomedical image segmentation. In: Int. Conf. on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer, 2015, pp. 234–241.[3] MIFTEN M. et al. Tolerance limits and methodologies for IMRT measurement - based verification QA: recommendations of AAPM TG 218. Medical Physics 2018, 45(4): e53–e83.[4] MOZZI C., MARINI L. et al.
gantry angle of 0°. Material/Methods:
Portal dose distributions at the EPID level were simulated by calculating the dose in air on a plane positioned 160 cm from the source, using PSQA plans generated in the Monaco Treatment Planning System (TPS). Since Monaco requires a CT dataset for dose calculations, a dummy CT was used with all voxel densities set to air. A dedicated EPID structure was modeled within this CT dataset. The EPID was represented as a 4.7 cm thick water-equivalent slab (relative electron density = 1). The equivalent water thickness of each EPID layer was optimized to reproduce the detector’s attenuation and scattering properties.The U-Net architecture had been previously validated on a dataset of 250 images obtained by irradiating various phantoms with beam fields of different geometries [4]. For the present PSQA application, the network was retrained by adding 30 PSQA images to the original training dataset. The model’s performance was evaluated using 30 independent PSQA EPID images. Predicted PDs were compared with TPS-calculated PDs (used as ground truth) through gamma analysis with 3%/3 mm local
Made with FlippingBook - Share PDF online