ESTRO 2025 - Abstract Book

S3428

Physics - Machine learning models and clinical applications

ESTRO 2025

Figure 2: Confusion matrix displaying classification accuracy.

Keywords: Treatment planning, Clustering, Classification

References: [1] K. He et al. “Deep residual learning for image recognition,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2016December, 2016. doi: 10.1109/CVPR.2016.90. [2] M. J. Cardoso et al. “MONAI: An open - source framework for deep learning in healthcare,” Nov. 2022. [Online]. Available: http://arxiv.org/abs/2211.02701.

3604

Digital Poster Artificial Intelligence-driven failure prediction on TomoTherapy® systems Kélian Poujade 1,2 , Jérémy Pirard 3 , Louise Travé-Massuyès 4 , Laure Vieillevigne 1,2

1 Medical Physics Department, Oncopole Claudius Regaud, Institut Universitaire du Cancer de Toulouse (IUCT), Toulouse, France. 2 RADOPT, Centre de Recherches en Cancérologie de Toulouse (CRCT), Toulouse, France. 3 Engineering Department, Airbus, Toulouse, France. 4 Laboratoire d'analyse et d'architecture des systèmes (LAAS CNRS), Université de Toulouse, CNRS, Toulouse, France Purpose/Objective: During radiotherapy treatment sessions, equipment failures can occur, leading to variable downtimes that complicate patient care. These failures vary in frequency and are rarely predictable at present. As radiotherapy machines are entirely digital, they generate log files for each irradiation that provide a chronological record of events, machine states and procedure progress. However, this data is still rarely used to anticipate failures in machine components. This study aims to develop a novel Artificial Intelligence (AI)-based method capable of proactively detecting early signs of subsystem failure using TomoTherapy® (Accuray, Madison, WI) log data. Material/Methods: Log messages were grouped into sequences corresponding to machine procedures. Subsystem behavior was assumed normal in the month following a failure and associated intervention, allowing partial labeling of the data as normal. Word2Vec, traditionally used in Natural Language processing (NLP), learnt semantic representations in