ESTRO 2024 - Abstract Book

S4752

Physics - Quality assurance and auditing

ESTRO 2024

324

Digital Poster

External validation of an AI tool for the detection of radiotherapy errors- A multicentric study

Petros Kalendralis 1 , Samuel M.H. Luk 2 , Alan M. Kalet 3 , Tomasz Piotrowski 4,5,6 , Adam Ryczkowski 4,5 , Joanna Kazmierska 4,7 , Inigo Bermejo 1 , Andre Dekker 1 1 Department of Radiation Oncology (Maastro), GROW School for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, Netherlands. 2 The University of Vermont Medical Center, Medical Physics, Burlington, Vermont, USA. 3 University of Washington Medical Center, Department of Radiation Oncology, Seattle, USA. 4 Poznan University of Medical Sciences, Department of Electroradiology, Poznan, Poland. 5 Greater Poland Cancer Centre, Department of Medical Physics, Poznan, Poland. 6 Adam Mickiewicz University, Department of Biomedical Physics, Poznan, Poland. 7 Greater Poland Cancer Centre, Radiotherapy department II, Poznan, Poland

Purpose/Objective:

Bayesian Networks (BNs) represent a graphical modeling approach derived from the field of artificial intelligence (AI). They have demonstrated promise in assisting with the identification of errors in radiotherapy treatment plans. Last year, Kalendralis et al. 1 trained and validated an updated version of the BN created by Luk et al. 2 using data from the University of Washington (UW), University of Vermont (UVM), and MAASTRO clinic in the Netherlands. Nevertheless, it is essential to evaluate the model's performance when applied to data from different institutions. Therefore, in this study, we conducted an external validation using data from the radiotherapy department of the Greater Poland Cancer Centre (GPCC) in Poznan.

Material/Methods:

We used a dataset of 379 Volumetric Modulated Arc Therapy VMAT treatment plans, consisting of 802 treatment arcs, from patients treated between January and May 2022 at GPCC with TrueBeams accelerators (Varian Medical Systems, Palo Alto, USA). We simulated errors in 3% of the plans categorized into three types: patient positioning, Linear Accelerator (gantry and collimator angle errors), and dose prescription, and tested the error-detecting performance of BNs trained at other centers.

Results:

As presented in figure 1 (https://docs.google.com/document/d/1-wWjWeesmMd58T0N0QjeIm33_pmRcUCVwBO vlYgP2c/edit), in terms of the area under the receiver operating characteristic curve (AUC), the BN trained on the three centres (MAASTRO, UVM, UW and GPCC) outperformed (AUC=78%) of the BNs trained in single centres (UW-66% and MAASTRO 75%).

Conclusion: