ESTRO 2025 - Abstract Book

S3438

Physics - Machine learning models and clinical applications

ESTRO 2025

Strong IT capabilities were also instrumental to support the integration with imaging systems, EHR and other critical software. The QCA analysis indicated that even in the absence of certain factors like explainable AI features, the combination of configurable AI tools and proactive stakeholder engagement could drive successful adoption. This suggests that customization to local needs and involving end-users in the implementation process can mitigate some of the challenges posed by limited data quality or legislative constraints. Conclusion: The study identifies AI configurability, stakeholder involvement, and high-quality data sets as essential for AI adoption in radiotherapy. Data set characteristics, particularly their quality and standardization, directly impact AI performance. By focusing on these key facilitating factors, radiotherapy centers can overcome challenges such as legislative barriers and varying IT capabilities. This study underscores the multi-faceted nature of AI adoption in healthcare and provides insights into practical deployment strategies, emphasizing the importance of data standardization and stakeholder engagement for policymakers and practitioners.

Keywords: AI, Adoption, Qualitative Comparative Analysis

References: Brouwer et al, 2020 Hurkmans et al, 2024 Kawamura et al, 2024 Landry et al, 2023

3932

Digital Poster Uncertainty quantification of a machine learning model for patient-specific quality assurance with conformal prediction Nicola Lambri 1,2 , Jocelyn Japnanto 3 , Victor Hernandez 4 , Jordi Sáez 5 , Andrew Nisbet 3 , Pietro Mancosu 1,2,3 1 Radiotherapy and Radiosurgery Department, IRCCS Humanitas Research Hospital, Rozzano, Italy. 2 Department of Biomedical Sciences, Humanitas University, Pieve Emanuele, Italy. 3 Department of Medical Physics and Biomedical Engineering, University College London, London, United Kingdom. 4 Department of Medical Physics, Hospital Universitari Sant Joan de Reus, Tarragona, Spain. 5 Department of Radiation Oncology, Hospital Clínic de Barcelona, Barcelona, Spain Purpose/Objective: VMAT plans should undergo measurement-based PSQA, acknowledged as a labour-intensive and time-consuming task. Machine learning (ML) algorithms offer promising solutions to reduce the PSQA workload by correlating the complexity of a VMAT arc to the gamma passing rate (GPR). However, current literature lacks quantification of models’ uncertainty in predicting pass rates, a factor crucial for addressing generalizability limitations [1, 2]. The aim of this study was to develop an ML model to predict GPRs using complexity metrics and quantify its uncertainty through conformal prediction for multicentric applications. Material/Methods: A publicly accessible dataset of 12,473 VMAT arcs delivered from 2018 to 2022 was used (Inst.1) [1]. Nineteen numerical features described plan parameters and complexity, while PSQA measurements were conducted with EPID. For the GPR analyses, a 3%(global)/1mm criterion and a 95% action limit were used. The dataset was split into training, calibration, and test sets. A LightGBM model was trained using 9,975 arcs. Conformity scores were calculated on 623 arcs from the calibration set, and a confidence level of 10% was used to produce prediction intervals (PIs) for the test set, consisting of 2,494 arcs, with guaranteed 90% coverage. The mean absolute error (MAE), mean PI width, and coverage (i.e., fraction of PIs including the measured GPR) were computed on the test set.