ESTRO 2024 - Abstract Book

S4451

Physics - Machine learning models and clinical applications

ESTRO 2024

963

Digital Poster

Two machine learning models based on explainable features to reduce PSQA workload

Gloria Miori 1 , Matteo Chieregato 1 , Rosaria Maio 1 , Francesca Andreoli 2 , Marco Galelli 1

1 Fondazione Poliambulanza, Medical Physics Department, Brescia, Italy. 2 University of Padova, Postgraduate School of Medical Physics, Padova, Italy

Purpose/Objective:

Patient-specific quality assurance (PSQA) is an essential part of the radiotherapy workflow. AAPM recommends measurements-based methods to evaluate PSQA performances. Comparison between calculated and measured plan dose distributions is typically evaluated by gamma analysis. However, measurements of PSQA are time-consuming and cannot be done in case of an online adaptive workflow.

In this study we introduced a new plan complexity metric and created two ML models based on plan delivery metrics to optimize the workload of PSQA measurements.

Material/Methods:

We analysed 87 prostates with pelvic nodes VMAT plans treated from 2017 to 2023. Contrary to previous studies, a single target type was considered. We believe it is important to have a homogenous population and not mixing results from different treatment sites: dose distributions and complexity metrics can vary substantially between different target anatomies. PSQA measurements were delivered on an Elekta LINAC with Agility head, acquired with Octavius (PTW) and compared to Monaco (Elekta) calculated plans. Local gammas 2%/2mm with 50% lower dose threshold were calculated in VeriSoft (PTW) with gamma passing rate = 90%. We derived a complexity metric from the complexity index (CI) previously published in literature (1) . Delivery errors can be related to variations in complexity between consecutive control points: Plan Complexity Deviation (PCD) is calculated as the absolute values sum of CI deviations for each control point i :

For each plan we calculated 8 metrics: MU, segments, MU/segment, CI (1) , Mean Area Metric Estimator (1) , Area Metric Estimator (1) , Aperture Irregularity Metric (2) and PCD.

We used complexity metrics only to create ML models based on explainable features.