ESTRO 2025 - Abstract Book

S3447

Physics - Machine learning models and clinical applications

ESTRO 2025

Conclusion: This work demonstrated the feasibility of training a GAN-based model for vPSQA purposes. Despite current limitations, primarily due to the limited dataset size and the imbalance between "success" and "failure" cases, the proposed approach offers a more flexible alternative to simpler PR prediction models, allowing for a more comprehensive and informative evaluation of PSQA results.

Keywords: virtual-PSQA, GANs, innovation

References: [1] Mehdi Mirza, Simon Osindero (2018). Image-to-image translation with Conditional Generative Adversarial Networks. https://arxiv.org/pdf/1611.07004 [2] Miften, M., Olch, A., Mihailidis, D., Moran, J., Pawlicki, T., Molineu, A., Li, H., Wijesooriya, K., Shi, J., Xia, P., Papanikolaou, N. and Low, D.A. (2018), Tolerance limits and methodologies for IMRT measurement-based verification QA: Recommendations of AAPM Task Group No. 218. Med. Phys., 45: e53-e83. https://doi.org/10.1002/mp.12810

4168

Digital Poster Prospective clinical evaluation of fully automated contouring and treatment planning for prostate radiotherapy David Neugebauer 1 , Jan-Hendrik Bolten 1,2 , Hoi Hin Lau 1 , Stephan Mende 1,2 , Tilmann Rackwitz 1,2 , Thomas Held 1,2 , Christoph Grott 1,2 , Johanna Rademacher 1,2 , Fabian Weykamp 1,2,3 , Philipp Hoegen-Sassmannshausen 1,2,3 , Jürgen Debus 1,2,3 , Sebastian Klüter 1,2 , Jakob Liermann 1,2 1 Department of Radiation Oncology, University Hospital Heidelberg, Heidelberg, Germany. 2 National Center for Tumor Diseases, University Hospital Heidelberg, Heidelberg, Germany. 3 Clinical Cooperation Unit Radiation Oncology, German Cancer Research Center (DKFZ), Heidelberg, Germany Purpose/Objective: First studies investigated approaches combining models for automated segmentation and treatment planning for fully automated workflows in an artificial research setting, showing promising results but seeing a need for studies in clinical settings 1,2 . In this work we established a one-click triggered workflow for prostate treatment and evaluated it in comparison to the standard workflow in a prospective study. Material/Methods: For 22 clinical patients, contouring of OAR and targets as well as treatment planning was performed according to the clinical workflow for prostate only treatment. Subsequently, a fully automated workflow (FAW) was run, utilizing the built-in scripting interface to concatenate deep-learning based segmentation and an in-house adapted machine learning planning model in RayStation 11B (RaySearch Laboratories, Sweden). The time needed was tracked for both workflows and each resulting treatment plan and structure set were reviewed by one experienced radiation oncologist in a blinded fashion, scoring target and OAR structures along with different aspects of the dose distribution as ‘good’, ‘acceptable’ or ‘unacceptable’. The preferred option underwent further steps of the clinical workflow, eventually being used for treatment (see Fig.1).