ESTRO 2025 - Abstract Book

S3400

Physics - Machine learning models and clinical applications

ESTRO 2025

Conclusion: The findings suggest that relevant imaging features vary across time-points, and that this information is potentially lost in the current multi-output approaches. Therefore, per time-point deep learning models provide the best approach for full toxicity score trajectory prediction, yet more optimization is needed to draw a definitive conclusion.

Keywords: Dysphagia, NTCP, temporal

2502

Proffered Paper Standardising AI Model Reporting for Radiotherapy: A Domain-Specific Model Card

Ana Maria Barragán Montero 1 , Margerie Huet-Dastarac 1 , Carlos Cárdenas 2 , Marco Fusella 3 , Geoffroy Herbin 4 , Yvonne de Hond 5 , Franziska Knuth 6 , Ciaran Malone 7 , Peter van Ooijen 8 , Charlotte Robert 9 , Benjamin Tengler 10 , Michele Zeverino 11 , Coen Hurkmans 5 , Tomas Janssen 12 , Stine Sofia Korreman 13 , Charlotte L. Brouwer 8 1 MIRO, UCLouvain, Brussels, Belgium. 2 Department of Radiation Oncology, University of Alabama at Birmingham, Birmingham, USA. 3 Radiation Oncology Department, Policlinico Abano Terme, Abano Terme, Italy. 4 Ion Beam Applications, IBA, Louvain-La-Neuve, Belgium. 5 Department of Radiation Oncology, Catharina Hospital Eindhoven, Eindhoven, Netherlands. 6 Erasmus MC Cancer Institute, University Medical Center Rotterdam, Rotterdam, Netherlands. 7 Dept. of Medical Physics, St.Luke’s Radiation Oncology Network, Dublin, Ireland. 8 Department of Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands. 9 Department of Radiation Oncology, Gustave Roussy, Villejuif, France. 10 Department of Radiation Oncology, University of Tübingen, Tübingen, Germany. 11 Institute of Radiation Physics, Lausanne University Hospital, Lausanne, Switzerland. 12 Department of Radiation Oncology, The Netherlands Cancer Institute, Amsterdam, Netherlands. 13 Danish Center for Particle Therapy, Aarhus University Hospital, Aarhus, Denmark Purpose/Objective: Artificial intelligence (AI) is becoming ubiquitous in radiation therapy (RT), with both in-house and commercial models reaching clinical practice. However, there is a lack of transparency when transferring these models from development to end-users, due to insufficient documentation of their specifications and performance. This hinders clinicians' trust, model comparability and selection. Tools like model cards [1] and datasheets [2] aim to overcome this, but despite recent efforts to adopt them in RT [3], there is not yet a comprehensive, practical, and domain specific standard for reporting AI models. We present an RT-specific model card to standardise documentation and increase transparency, operability, and trust in clinically-used AI models. Material/Methods: This initiative was born in the “AI for the fully automated radiotherapy treatment chain” ESTRO physics workshop 2023 and included in the ESTRO AI focus group activities. Our working group consisted of 16 persons from 13 institutions. We performed a literature review on existing tools, guidelines, and examples for AI model reporting (Table 1). Moreover, several participants gathered examples from internal documentation of AI models implemented in their institutions. Based on this, an initial model card was proposed and sent for review to all participants. Over 2 months, three review rounds were performed, where suggested changes were voted in a live Google doc. Unclear fields and conflicting votes were discussed at online meetings, and consensus was reached by majority voting. Three popular RT applications were selected to define task-specific fields: image-to-image translation (synthetic CT), segmentation, and dose prediction.