25–29 May 2026
La Biodola - Isola d'Elba (Italy)
Europe/Rome timezone
NB: The submission deadline for the Student Paper Awards is Monday, 11 May.

114 Deep Learning–Based Real-Time Error Detection in Radiotherapy with EPID Images

27 May 2026, 10:38
2m
Maria Luisa Room (Hotel Hermitage)

Maria Luisa Room

Hotel Hermitage

Mini Oral Real Time Diagnostics, Digital Twin, Control, Monitoring, Safety and Security Mini Orals

Speaker

Emmanuel Uwitonze (National Institute for Nuclear Physics, Section of Pisa, Pisa, Italy)

Description

Accurate in-vivo dose monitoring is increasingly important in radiotherapy to comply with stringent dose delivery guidelines. Electronic Portal Imaging Devices (EPIDs), routinely used for patient positioning, offer the potential for real-time treatment monitoring. However, EPID images are only indirectly related to patient dose, and their use for error detection is challenging. In this study, an in-house developed DL model was used to transform raw EPID images into water equivalent portal dose images. Particularly, we assessed the feasibility of using these DL-generated portal dose images for application in a real time error detection alert system in phantom experiments, performed at the Careggi university hospital (Firenze, Italy).

Different phantoms were irradiated under reference conditions and with intentionally introduced treatment errors, including deviations in monitoring units (MU) and phantom shifts. EPID images were converted to portal dose images using the DL model, and several comparison metrics were evaluated, including gamma-index analysis and relative mean absolute dose difference (reMADD). The results demonstrate that MU errors can be reliably detected using a combination of metrics. Gamma passing rates decreased with increasing MU errors, while reMADD provided consistent sensitivity across different phantoms and field sizes. Setup errors were more challenging to detect, particularly for narrow fields, with reMADD showing slightly better sensitivity than gamma analysis.

Overall, the proposed DL-based portal dose comparison framework is appropriate for application in a real time alert system for detection of treatment errors. Future work will focus on extending the approach to clinical patient data.

Minioral Yes
IEEE Member No
Are you a student? No

Authors

Emmanuel Uwitonze (National Institute for Nuclear Physics, Section of Pisa, Pisa, Italy) Rossana Lanzillotta (University and National Institute for Nuclear Physics, Pisa, Italy) Carlotta Mozzi (University of Florence and INFN Florence, Italy) Lorenzo Marini (University and INFN, Pisa, Italy) Michele Avanzo (Centro di Riferimento Oncologico di Aviano (CRO) IRCCS, Aviano, Italy) Francesca Lizzi (INFN Section of Pisa, Pisa, Italy) Livia Marrazzo (University of Florence, Florence, Italy) Icro Meattini (University of Florence, Florence, Italy) Stefania Pallotta (University of Florence and INFN, Florence, Italy) Giovanni Pirrone (Centro di Riferimento Oncologico di Aviano (CRO) IRCCS, Aviano, Italy) Alessandra Retico (National Institute for Nuclear Physics, Section of Pisa, Pisa, Italy) Cinzia Talamonti (University of Florence and INFN, Florence, Italy) Aafke Christine Kraan (INFN Pisa)

Presentation materials

There are no materials yet.