ESTRO 2025 - Abstract Book

S3678

Physics - Quality assurance and auditing

ESTRO 2025

Conclusion: Despite advances in thoracic treatment, there remains room for improvement. By applying advanced machine learning techniques, this study not only identifies critical factors but also provides an interpretable framework for understanding model predictions, enabling better-targeted improvements in clinical practice. These tools also allow for the development of personalized QA programs tailored to specific treatment setups, aligning more closely with clinical needs. These findings offer valuable insights into how specific treatment characteristics influence dosimetry outcomes, providing a foundation for refining QA protocols and improving the consistency and accuracy of thoracic radiation therapy delivery across different clinical settings.

Keywords: IROC, machine learning, feature interpretability

3821

Poster Discussion Multivariate control chart for plan complexity metrics Antonio Gañán Mora, Javier de Areba Iglesias, Virginia Álvarez Sánchez, Zulima Aza Villarrubia, Rocío Bermúdez Luna, Borja Blanco García, María Paz Martín Niella, David Flavio Martínez Barrio Servicio de Física Médica, Hospital Clínico San Carlos, Madrid, Spain Purpose/Objective: There are numerous metrics to quantitatively express plan complexity, many of them being correlated. Control charts are a useful statistical tool for establishing control limits in processes, such as complexity metrics (CM). Shewhart individuals control charts are one of the most common types of control chart in radiotherapy being used for just one normally distributed variable. It has been shown that it is not possible to have non-parametric multivariate individuals control charts, hence some assumption about the distribution must be made (or use heuristic methods) [1]. Hotelling’s T 2 control chart assumes a multivariate normal distribution considering the correlation of variables through the covariance matrix. In this work we apply a Hotelling’s T 2 multivariate individuals control chart to establish control limits for plan complexity of prostate treatments, the most common treatment in our institution. Material/Methods: A total of 108 prostate treatments from six different clinicians were retrospectively collected. Eight different physicists planned the treatments using Eclipse versions 15.6 and 16.1. For each plan, four CM were computed: monitor units divided by dose per fraction (MU/cGy), mean MLC gap, tongue-and-groove index (TGI), and modulation complexity score (MCS). For each plan, Hotelling’s T 2 computes the square vector distance of the CMs to the mean CMs weighted by the covariance matrix. Obtaining control charts follows a two-phased process. In Phase I, out-of-control plans were removed iteratively using the “Identify-Eliminate-Recalculate” method until all plans are in control (inside the 99% probability upper control limit). In Phase II, final upper control limits for future prostate plan (prospective analysis) were established using the in-control dataset. Note that data in Phase I follows a beta distribution while future data in Phase II follows an F distribution because of statistical independence of prospective data from retrospective mean vector and covariance matrix. Results: Seven iterations were needed to obtain an in-control dataset. Statistics of this in-control dataset is shown in Table. To check normality, Phase I T 2 quantile function assuming normality –beta quantile function– is always inside the 95% bootstrap confidence intervals for the T 2 empirical cumulative distribution, as shown in Figure. The final Phase II 90% (95%, 99%) confidence upper control limit for the T 2 statistic is 8.3 (10.3, 14.7).